Journal of Networks - Academy Publisher

Journal of Networks 

ISSN 1796-2056 

Volume 6, Number 7, July 2011 

Contents 

Special Issue: Selected Best Papers of the International Workshop on Computer Science for 

Environmental Engineering and EcoInformatics (CSEEE 2011) 

Guest Editors: Tianlong Gu and Shenghui Liu 

Guest Editorial 

Tianlong Gu and Shenghui Liu 

SPECIAL ISSUE PAPERS 

Multi-Constrained Routing Algorithm for Multimedia Communications in Wireless Sensor Networks 

Xin Yan, Layuan Li, and F. J. An 

Reputation-aware Service Selection based on QoS Similarity 

Shenghui Zhao, Guoxin Wu, Guilin Chen, and Haibao Chen 

Cost Aggregation Strategy with Bilateral Filter Based on Multi-scale Nonlinear Structure Tensor 

Li Li and Hua Yan 

A Collaborative Nonlocal-Means Super-resolution Algorithm Using Zernike Monments 

Lin Guo and Qinghu Chen 

Mathematical Model and Hybrid Scatter Search for Cost Driven Job-shop Scheduling Problem 

Jie Bai, Kai Sun, and Gen Ke Yang 

Multi-objective Genetic Algorithm for System Identification and Controller Optimization of 

Automated Guided Vehicle 

Xing Wu, Peihuang Lou, and Dunbing Tang 

WebVR—Web Virtual Reality Engine Based on P2P Network 

Zhihan Lv, Tengfei Yin, Yong Han, Yong Chen, and Ge Chen 

An Energy-Efficient Communication Protocol for Wireless Sensor Networks 

Fengjun Shang 

Robust Cross-layer Design of Wireless Multimedia Sensor Networks with Correlation and 

Uncertainty 

Lei You and Chungui Liu 

The E-Commerce Model of Health Websites: An Integration of Web Quality, Perceived Interactivity, 

and Web Outcomes 

Chung-Hung Tsai 

937 

939 

950 

958 

966 

974 

982 

990 

999 

1009 

1017

A New Method of Time-frequency Synthesis of Harmonic Signal Extraction from Chaotic 

Background 

Erfu Wang, Zhifang Wang, Jing Ma, and Qun Ding 

Provable Data Possession of Resource-constrained Mobile Devices in Cloud Computing 

Jian Yang, Haihang Wang, Jian Wang, Chengxiang Tan, and Dingguo Yu 

Image Compression Based on Improved FFT Algorithm 

Juanli Hu, Jiabin Deng, and Juebo Wu 

Correlative Peak Interval Prediction and Analysis of Chaotic Sequences 

Qun Ding, Lu Wang, and Guanrong Chen 

REGULAR PAPERS 

An Energy Efficient Dynamic Clustering Protocol Based on Weight in Wireless Sensor Networks 

Ming Zhang and Suoping Wang 

Performance of UWB Systems with Direct-Sequence Bipolar Pulse Amplitude Modulation and 

RAKE Reception over IEEE 802.15.3a Channel 

Jingjing Wang and Hao Zhang 

Data Accuracy Estimation for Spatially Correlated Data in Wireless Sensor Networks under 

Distributed Clustering 

Jyotirmoy Karjee and H.S Jamadagni 

Networking as a Service: a Cloud-based Network Architecture 

Tao Feng, Jun Bi, Hongyu Hu, and Hui Cao 

1025 

1033 

1041 

1049 

1057 

1065 

1072 

1084

JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 937 

Special Issue on Selected Best Papers of the International Workshop on Computer Science for 

Environmental Engineering and EcoInformatics (CSEEE 2011) 

Guest Editorial 

This special issue comprises of 14 selected papers from the International Workshop on Computer Science for 

Environmental Engineering and EcoInformatics (CSEEE 2011). The conferences received 860 paper submissions from 

15 countries and regions, of which 450 were selected for presentation after a rigorous review process. From these 450 

research papers, through two rounds of reviewing, the guest editors selected 14 as the best papers on the Networking 

Technologies and Information track of the Conference. The candidates of the Special Issue are all the authors, whose 

papers have been accepted and presented at the CSEEE 2011, with the contents not been published elsewhere before. 

2011 International Workshop on Computer Science for Environmental Engineering and EcoInformatics will continue 

the excellent tradition of gathering world-class scientists, engineers and educators engaged in the fields of Computer 

Science and Environmental Biotechnology to meet and present their latest activities. CSEEE 2011 held on July 29-31, 

2011, Kunming, China. This conference is sponsored by International Association for Scientific and High Technology, 

and is in cooperation with Yunnan University, and it is technical co-sponsored by Kunming University of Science and 

Technology. 

“Multi-Constrained Routing Algorithm for Multimedia Communications in Wireless Sensor Networks”, by Xin Yan, 

Layuan Li and F. J. An, proposes a novel routing model that can depict multiple service requirements in multimedia 

sensor networks, and designs a new multi-constrained routing algorithm MCRA for multimedia communications. 

Theoretical analysis and simulation experiments are provided to validate their claims. 

“Reputation-aware Service Selection Based on QoS Similarity”, by Shenghui Zhao, Guoxin Wu, Guilin Chen and 

Haibao Chen, proposes a reputation evaluation method for Web Services which can gradually adjusting the reputations 

based on eliminating the collusive behaviors of consumers step by step. The experimental results show that the model 

can identify the conclusive consumers and improve the exact rate of reputation evaluation and success rate of service 

selection. 

“Cost Aggregation Strategy with Bilateral Filter Based on Multi-scale Nonlinear Structure Tensor”, by Li Li and Hua 

Yan, proposes a novel cost aggregation method for stereo matching with modified bilateral filter. By constructing the 

multi-scale nonlinear structure tensor and adding the new corresponding weight in cost aggregation, more pixels similar 

with central pixel are aggregated in a support window and the final disparity map are more accurate.. 

“A Collaborative Nonlocal-Means Super-resolution Algorithm Using Zernike Monments”, by Lin Guo and Qinghu 

Chen, proposes an efficient improved algorithm by introducing Zernike moments as representation of image invariant 

features into similarity measure. Experimental results indicate the proposed method can handle real video sequences 

with general motion pattern, and performances better than the comparing methods. 

“Mathematical Model and Hybrid Scatter Search for Cost Driven Job-shop Scheduling Problem”, by Jie Bai, Kai Sun 

and Gen Ke Yang, proposes a cost driven model of the job-shop scheduling problem in which the solutions are driven 

by business inputs, such as the cost of the product transitions, revenue loss due to the machine idle time and 

earliness/tardiness penalty. 

“Multi-objective Genetic Algorithm for System Identification and Controller Optimization of Automated Guided 

Vehicle”, by Xing Wu, Peihuang Luo and Dunbing Tang, proposes a multi-objective genetic algorithm (MOGA) with 

Pareto optimality and elitist tactics for the control system design of automated guided vehicle (AGV). 

“WebVR—Web Virtual Reality Engine Based on P2P network”, by Zhihan Lv, Tengfei Yin, Yong Han, Yong Chen 

and Ge Chen, introduces a multi-user online virtual reality engine -- WebVR. The core model innovation of WebVR 

engine is mapping the geographical space and virtual space to the P2P overlay network space, and build quad-tree index 

for the three spaces, and they identify the geocoding based on Hash value, which is used to index the user list, terrain 

data, and the model object data. The model greatly improves the hit rate of 3D geographic data search under P2P 

overlay network. 

“An Energy-Efficient Communication Protocol for Wireless Sensor Networks”, by Fengjun Shang, proposes an 

energy-efficient Single-Hop Active Clustering (SHAC) algorithm for wireless sensor networks. Through both 

theoretical analysis and numerical results, it is shown that SHAC prolongs the network lifetime significantly against the 

other clustering protocols such as LEACH-C and EECS. 

“Robust Cross-layer Design of Wireless Multimedia Sensor Networks with Correlation and Uncertainty”, by Lei You 

and Chungui Liu, uses a cross-layer method to deal with the robust lifetime optimization of wireless multimedia sensor 

network (WMSN) with energy consumption uncertainty and proposes to model the uncertainty as a polyhedral set. 

“The E-Commerce Model of Health Websites: An Integration of Web Quality, Perceived Interactivity, and Web 

Outcomes”, by Chung-Hung Tsai, integrates web quality (system quality, information quality, and service quality), 

perceived interactivity (human-message, human-human), and web outcomes (web usage, web satisfaction, and web 

loyalty) to explore the e-commerce model of health websites. A survey of 1076 users of health websites was conducted 

© 2011 ACADEMY PUBLISHER 

doi:10.4304/jnw.6.7.937-938

938 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 

to validate the proposed model. The findings show that web quality has significantly positive effect on perceived 

interactivity, web usage, and web satisfaction separately, which in turn influence web loyalty. 

“A New Method of Time-frequency Synthesis of Harmonic Signal Extraction from Chaotic Background”, by Erfu 

Wang, Zhifang Wang, Jing Ma and Qun Ding, proposes a new synthesis about wavelet threshold and empirical mode 

decomposition (EMD) complementary of new harmonic signal extraction by experimental simulation. 

“Provable Data Possession of Resource-constrained Mobile Devices in Cloud Computing”, by Jian Yang, Haihang 

Wang, Jian Wang, Chengxiang Tan and Dingguo Yu, proposes a novel PDP scheme, in which a trusted third-party 

agent (TPA) takes over most of the calculations from the mobile end-users. By using bilinear signature and Merkle hash 

tree (MHT), the scheme reduces communication and storage burden, and is fit for mobile devices. 

“Image Compression based on improved FFT Algorithm”, by Juanli Hu, Jiabin Deng and Juebo Wu, adopts Radix-4 

Fast Fourier transform (Radix-4 FFT) to realize the limit distortion for image coding, and to discuss the feasibility and 

the advantage of Fourier transform for image compression. It aims to deal with the existing complex and 

time-consuming of Fourier transform, according to the symmetric conjugate of the image by Fourier transform to 

reduce data storage and computing complexity. 

“Correlative peak interval prediction and analysis of chaotic sequences”, by Qun Ding, Lu Wang and Guanrong Chen, 

proposes a digital circuit design for the logistic-map module used in chaotic stream ciphers, analyzes the factors that 

may affect the output of the sequences, and develops a calculation method for estimating the output sequential 

correlative peak interval. 

We wish to thank the Kunming University of Science and Technology for providing the venue to host the conference. 

We would like to take this opportunity to thank the authors for the efforts they put in the preparation of the manuscripts 

and for their valuable contributions. We wish to express our deepest gratitude to the program committee members for 

their help in selecting papers for this issue and especially the referees of the extended versions of the selected papers for 

their thorough reviews under a tight time schedule. Last, but not least, our thanks go to the Editorial Board of the 

Journal of Networks for the exceptional effort they did throughout this process. 

In closing, we sincerely hope that you will enjoy reading this special issue. 

Guest Editors: 

Tianlong Gu, Guilin University of Electronic Technology, P.R. China 

Shenghui Liu, Harbin University of Science & Technology, P.R. China 


Tianlong Gu was born in Shanxi, China, on 1st October 1964. He received the Bachelor Degree from Taiyuan 

University of Technology in 1984, received Master Degree from Xidian University in 1986 and received his 

Ph.D. degree from Zhejiang University in 1996. From 1998 to 2002, he was a postdoctoral research fellow and 

visiting professor within Murdoch University and Curtin University of Technology, Australia. He has published 

more than 130 papers, and authored 6 books. His main research interests include formal method, knowledge 

engineering and mobile computing.He is a full professor in school of computer science & engineering at Guilin 

University of Electronic Technology, Guilin, China, and Ph.D. supervisor in school of computer science & 

technology at Xidian University, Xian, China. 

Shenghui Liu was born in Heilongjiang, China, on July 24, 1961. Bachelor of Automatic Control(6/1982). 

Master of Computer Science (3/1985). Doctor of Management Engineering(6/2009). Harbin University of 

Science & Technology, Harbin, Heilongjiang, P.R.China. Now he is a professor and the dean of Software School 

in Harbin University of Science & Technology.He has wide research interests, mainly information technology. 

He has published above 50 papers in journals or conference proceedings and some of the papers are indexed by 

SCI, EI. He has won various awards in the past. He served as many workshop chair, advisory committee or 

program committee member of various international conferences.


Multi-Constrained Routing Algorithm for 

Multimedia Communications in Wireless Sensor 

Networks 

Xin Yan and Layuan Li 

Department of Computer Science, Wuhan University of Technology, Wuhan 430063, P.R. China 

Email: yanxin@whut.edu.cn, jwtu@public.wh.hb.cn 

F. J. An 

Faculty of EEMCS, Delft University of Technology, 2600 GA Delft, The Netherlands 

Email: anfengju@hotmail.com 

Abstract — The existing routing protocols designed for 

real-time or multimedia applications in sensor networks 

usually adopt relatively simple routing models where fewer 

service metrics are considered, which is not sufficient for 

real-time or multimedia data transportations. Furthermore, 

for the sake of route discovery or the acquisition of a target 

location, they usually need extra localization equipments or 

beacon exchanges to obtain the geographic location of each 

sensor node or construct a coordinate system for sensor 

nodes, which imports extra costs to routing algorithms. In 

this paper, firstly we propose a novel system model that can 

comprehensively depict the service requirements of 

multimedia applications, and on the basis of this system 

model, we design a new multi-constrained routing algorithm, 

MCRA, for multimedia communications in sensor networks. 

MCRA not only can provide end-to-end delay guarantee 

and packet loss ratio guarantee for multimedia 

communications, but also can balance and improve the 

energy consumption in sensor nodes. Besides, MCRA adopts 

several effective policies to suppress message flooding and 

lessen data redundancy. In MCRA, neither the acquisition 

of target location nor the route discovery process requires 

any extra measurement equipment or coordinate system 

based on location message exchange, however, the target 

location we concern can be easily figured out by a 

localization scheme without message exchange. In addition, 

we may optionally apply MAC multicast and differentiation 

service in MCRA so as to further lower its control message 

overhead and differentiate forwarding priority levels for 

real-time data and best-effort traffic in MAC layer. 

Theoretical analysis and simulation experiments are 

provided to validate our claims. 

Index Terms — routing algorithm, multimedia applications, 

sensor networks, QoS, message suppression, localization 

Ⅰ. INTRODUCTION 

A wireless sensor network (WSN) is comprised of small, 

low powered, self-organizing sensor nodes, densely 

deployed in the area to be monitored. These networks can 

support a wide rang of applications, such as earthquake 

response, health monitoring, battlefield surveillance etc. 

Some of these applications may be augmented due to the 


doi:10.4304/jnw.6.7.939-949 

use of real-time or multimedia data. Real-time or 

multimedia applications have stringent requirements of 

quality of service (QoS), for instance, end-to-end delay, 

packet loss ratio etc, during the data transmissions. 

Several main MAC layer protocols have been 

developed for multimedia communications in sensor 

networks. IEEE 802.11e scheme has provisions for 

service differentiation at MAC layer in sensor networks, 

though it was proposed for ad hoc networks initially [1]. 

In this scheme, the service differentiation is obtained by 

changing the duration of the Inter-Frame Spacing (IFS) 

and the Contention Window (CW) size based on the 

priority of the packet. The scheme can provide QoS 

services for multimedia communications from two 

aspects: timeliness and reliability, thanks to its broadcast 

and multicast functions [2]. In addition, IEEE 

802.15.4/ZigBee specification designed for the low data 

rate, low power consumption, and low cost networks can 

provide a Guaranteed Time Slot (GTS) mechanism to 

allocate a specific duration within a super frame structure 

for real-time transmissions [3,4]. 

Although each layer in sensor network stack may 

provide QoS services for multimedia communications, 

routing protocol in network layer is always playing the 

most important role. Routing protocol can provide not 

only QoS guarantees but also network load-balance and 

congestion management for multimedia data streams. 

As we know, majority of routing protocols in sensor 

networks are oriented to various applications. For 

multimedia applications, routing protocols should aim at 

providing timeliness and reliability services for them, and 

manage to balance the energy consumption in sensor 

nodes [5]. Additionally, routing protocols designed for 

multimedia communications are supposed to have better 

capacities of message suppression and data aggregation 

than common applications, because of more data 

redundancies in multimedia applications [6]. Therefore, 

routing algorithms working in multimedia sensor 

networks should be able to not only provide QoS services 

for multimedia communications but also suffice for other 

communication requirements in WSN (e.g., optimal


energy consumption, localization, message suppression 

and data aggregation etc). Meanwhile, they also ought to 

be resilient and adapt to the dynamics and scalability of 

multimedia sensor networks. 

One the other hand, for most multimedia applications 

in sensor networks, location is more important than a 

specific node ID. Target data without position 

information retrieved from sensor nodes is usually 

unmeaning. It requires sensor nodes in WSN are 

location-aware, which needs to resort to GPS etc 

measurement equipments or some localization algorithms 

to locate these nodes. However, GPS etc measurement 

equipments are impractical even do not work completely 

in some especial environments [7]. To overcome this 

weakness, some virtual or logical coordinate systems are 

proposed [8,9,10], which construct a coordinate space for 

route discovery and target localization by node-location 

message exchange. Whereas, the construction of virtual 

or logical coordinate system needs to consume drastically 

the limited resources in sensor nodes. Worse for them, 

both geographic information based routing algorithms 

and virtual (or logical) coordinate system based routing 

algorithms, the inaccurate node-location information 

sometimes emerging could result in the failure of route 

discovery process. Hence, we should evade the risk of 

routing failure caused by the imprecise location 

information as possible as we can, when designing a 

routing algorithm for multimedia sensor networks. 

Ⅱ. RELATED WORK 

For the sake of QoS provision and adaptation to the 

communication characteristics (energy constrained, 

limited computation capacity, and less memory 

availability) in sensor networks, T. He et al proposed a 

routing protocol named SPEED to provide soft real-time 

guarantees for real-time communications in sensor 

networks [11]. In SPEED, end-to-end soft real-time 

communication is achieved by maintaining a desired 

delivery speed across the sensor network through a 

combination of feedback control and non-deterministic 

geographic forwarding. SPEED is a stateless and 

geographic position based routing protocol without 

end-to-end path set-up before packets forwarding. 

However, SPEED neither takes another important QoS 

metric (packet loss ratio) into account, nor balances well 

the energy consumption in sensor nodes. Moreover, before 

computing data routes, SPEED must assume each sensor 

node is location-aware. 

In [12], E. Felemban et al designed a routing protocol 

called MMSPEED, which presents not only the service 

timeliness by guaranteeing multiple packet delivery speed 

options, but also the service reliability in a probabilistic 

multi-path forwarding manner. The QoS provision is 

realized in a localized way without global network 

information by employing localized geographic packet 

forwarding with dynamic compensation to offset the local 

decision inaccuracies. Although MMSPEED adds 

reliability of route discovery in SPEED, and is more 

suitable for large-scale dynamic sensor networks, it still 

has some disadvantages similar to SPEED, for instance, 


geographic forwarding based and lack of important QoS 

constraints. 

L. Shu et al presented a two phase geographic greedy 

forwarding (TPGF) routing algorithm for multimedia 

sensor networks, which supports multipath transmission 

and hole-bypassing, as well as shortest path transmissions 

[13]. TPGF consists of two phases: geographic 

forwarding and path optimization, wherein geographic 

forwarding is responsible for exploring a delivery 

guaranteed route while bypassing the holes in WSN, path 

optimization is responsible for optimizing the found path 

with the least number of nodes by a method of label 

based optimization. Nevertheless, TPGF and MMSPEED 

fall into the same category essentially. 

In order to differentiate video and audio applications 

which both employ TPGF as their routing protocol, on 

the basis of TPGF, a multi-priority multi-path selection 

scheme (MPMPS) is proposed for transport layer in WSN 

[14]. MPMPS supports multiple transmission priorities 

and chooses the maximum number of paths for 

maximizing throughput of streaming data transmission 

and guaranteeing the end-to-end transmission delay. 

K. Akkaya et al proposed an energy-aware QoS 

routing protocol for sensor networks [15]. The protocol 

finds a least-cost, delay-constrained path for real-time 

data in terms of link cost that captures nodes’ energy 

reserve, transmission energy, error rate and other 

communication parameters. Moreover, the throughput for 

non-real-time data is maximized by adjusting the service 

rate for both real-time and non-real-time data at the 

sensor nodes. The main problem of this protocol is that it 

requires complete knowledge of network topology at each 

node in order to compute a route, so that it is unsuitable 

for the large-scale sensor networks. In addition, like 

SPEED etc, this protocol does not take packet loss ratio 

into account, and must resort to extra localization 

equipments or algorithms to locate sensor nodes before 

route discovery. 

Seen from the mentioned above, like GPSR [16], 

SPEED, MMSPEED, TPGF, MPMPS etc are all 

geographic greedy forwarding based routing algorithms, 

which need the geographic coordinate of each node in 

sensor networks to compute the routing path. They also 

leave out some important QoS constraints in their routing 

models. Besides, all of the routing algorithms are based 

on data-driven delivery mode, which mode is unsuitable 

for the multimedia applications with periodic data. 

On the basis of the discussed above, we manage to 

design a routing algorithm for multimedia 

communications with periodic data in sensor networks, 

which is supposed to satisfy the following design goals: 1) 

end-to-end delay guarantee; 2) end-to-end packet loss 

ratio guarantee; 3) optimal node energy consumption; 4) 

minimum MAC layer support (i.e., it does not need 

special QoS-aware MAC support); 5) optional MAC 

layer services (MAC multicast and differentiation service 

are available); 6) extra position measurement equipment 

or location message exchanges unnecessary; 7) message 

suppression and data aggregation; 8) resilience and 

reliability.


Ⅲ. SYSTEM MODEL 

A WSN can be represented as a weighted, connected 

graph G=(V,E), where V is the set of nodes and E denotes 

the set of wireless communication links connecting the 

nodes. 

A. Network Service Model 

It is believed that in WSN the communications from 

multiple source nodes to one sink node construct a 

reversed multicast tree. Suppose that s(s∈V) is a source 

node of the multicast tree, and M( M ⊆{ V − { d}} 

) is the 

set of source nodes of the multicast tree, where d(d∈V) is 

the sink node. Let us use T(M,d) to denote the multicast 

tree. In order to describe the service requirements from 

multimedia applications in sensor networks (i.e., the 

timeliness and reliability of communications, as well as 

the optimal energy consumption in sensor nodes), we 

define the network service model as follows, which is 

able to comprehensively depict these requirements. 

⎧e2e 

_ delay( p( s, d)) ≤ D 

⎪ priority _ levels( f ( T )) ≥ 2 

⎨ 

⎪ arrival _ probability( p( s, d )) ≥ P 

⎩ 

⎪ energy _ cos t( T ( M , d )) = min[...] 

Where e2e_delay, priority_levels, arrival_probability and 

energy_cost denote the end-to-end delay, the number of 

traffic priority levels, the packet arrival probability, and 

the energy consumption respectively; p(s,d) is the path 

from the source node s( ∀ s∈M) to the sink node d in the 

reversed multicast tree T(M,d); and f(T) refers to the 

traffic type in T(M,d); D and P denote the constraints to 

the end-to-end delay and the packet arrival probability 

respectively. 

B. Routing Model 

Aiming at sufficing for QoS requirements of 

multimedia applications, formula (1) represents the 

network services that should be provided by WSN. The 

network services mainly consist of the service from MAC 

layer and the service from network layer. However, with 

respect to the network layer, the routing service model 

should be defined as follows: 

⎧ residual _ energy( T ( M , d )) ≥ E 

⎪ 

⎪e2e 

_ delay( p( s, d)) ≤ D 

⎨ 

⎪ packet _ loss( p( s, d )) ≤ R 

⎩ 

⎪ hopcount( p( s, d )) = min[...] 

Where residual_energy, e2e_delay, packet_loss and 

hopcount denote the residual energy ratio in sensor node, 

end-to-end delay, end-to-end packet drop ratio and 

hop-count respectively; E, D and R denote the constraints 

to the residual energy ratio, end-to-end delay and packet 

drop ratio respectively. 

In (2), there exist the following relationships: 


(1) 

(2) 

⎧residual 

_ energy( T ( M , d )) = 

⎪ 

⎪ 

min{ residual _ energy( n), n ∈ T ( M , d )} 

⎪ e2e _ delay( p( s, d)) 

= 

⎪ 

⎨ ∑ delay( e) + ∑ delay( n) 

⎪ e∈p( sd , ) n∈p( sd , ) 

⎪ packet _ loss( p( s, d)) 

= 

⎪ 

⎪1 

− ∏ (1− packet _ loss( n)) 

⎩ n∈p( s, d ) 

Where n∈V, e∈E; delay(n) and delay(e) are the delay 

functions of sensor nodes and wireless links respectively. 

Note that hereby we use the hop-count of a path to 

represent the accumulated node energy consumption 

along the path. 

Ⅳ . PROPOSED ALGORITHM 

In this section, we present a routing algorithm named 

MCRA (Multi-Constrained Routing Algorithm), which is 

based on query-flooding and query-driven data delivery 

mode, because this mode has its intrinsic resilience and 

reliability [17,18]. In advance suppose that the delay 

metric of duplex bi-directional wireless links in sensor 

networks has symmetric property. 

A. Routing Procedure 

The message used to query an event occurred within a 

surveillance area is usually called interest. In MCRA, the 

format of interest message is defined as Fig. 1, where each 

item between a pair of parentheses is the comment to the 

corresponding field name. 

interest.type (query event type) 

interest.nodes (visited node set) 

interest.hopcount (trip hop count) 

interest.e2e_delay (trip time record) 

interest.packet_loss (accumulated packet drop ratio) 

interest.D (end-to-end delay constraint) 

interest.R (constraint to packet drop ratio) 

interest.E (constraint to energy consumption) 

interest.neighbors (temporary neighbor table) 

interest.TTL (time to live) 

Figure 1. The format of interest message 

1) Starting from a sink node, interest messages (copies 

of one interest) are flooded to all of neighbors of the sink 

node. When some intermediate node in the network (e.g., 

node k) receives an interest, the node k begins to measure 

its residual energy and the packet drop ratio in it (the 

packet drop ratio is a statistical value during a period of 

time, which has been stored in this node), as well as the 

current system time from the synchronous clock in node k. 

Afterwards, node k uses the detected information to 

calculate and rewrite the two fields interest.e2e_delay and 

interest.packet_loss in this interest, as illustrated in Fig. 1. 

Of course here k ∉ interest. nodes , that is to say, 

interests will never visit the nodes which they have already 

visited. 

(3)


2) The intermediate node k starts to check QoS 

constraints, as illustrated in Fig. 2. If the result of the 

following expression is true: 

residual _ energy( k) ≥ E ∧ 

interest. packet _ loss ≤ R ∧ 

(4) 

interest. e2e_ delay ≤ D 

Then interest.hopcount=interest.hopcount+1, and the ID 

of node k is added into the list interest.nodes in the interest. 

Otherwise, the interest will be discarded by node k. 

Monitored 

Area s 

m 

i 

Data Interest 

Figure 2. The routing process of MCRA 

k 

n 

j 

sink 

3) Delaying for a period of time, if node k receives 

multiple interests that have the same query event type (i.e. 

interest.type) but traveled along different paths, node k 

will select the best interest (i.e., that one with the 

minimum value of interest.hopcount, or 

interest.e2e_delay, or interest.packet_loss) in them, and 

drops the others (details in subsection B.2). Afterwards, 

node k forwards the interest to its neighbors excluding the 

nodes visited by the interest, in terms of a restraining 

forwarding scheme (details in subsection B.1). 

4) The above steps are repeated until the interest arrives 

at a node s that matches the content of field interest.type in 

this interest, i.e., source node, eventually. 

5) Node s performs operations in a similar way to other 

nodes except for not forwarding the interest. It begins to 

read the list value from field interest.nodes in this interest, 

and sends data towards the sink node by using the node list 

value as the travel path of its packets. The node list 

information will guide the data forwarding by means of 

piggybacking, i.e., the path information is carried by the 

packets, as shown in Fig. 2. Note that the logical 

coordinate vector (see subsection D) of node s is also sent 

towards the sink by means of piggybacking. 

6) When the sink node receives the data sent by source 

node s, it also receives the logical coordinate vector of 

node s (i.e., the hop-count information from multiple sink 

nodes to the source node s) carried back by the packets. 

Now the sink node may adopt the coordinate system 

proposed in subsection D to calculate the coordinate 

value of source node s by its logical coordinate vector. 

The differentiation service in MAC layer may be 

applied to MCRA, so that real-time traffic and best-effort 

traffic in networks are classified into different priority 

levels to forward, which provides differentiation service 

for sensor networks [19]. Meanwhile, in order to avoid 

the collisions in routing process as soon as possible, here 

we may optionally apply a MAC protocol supporting 

MAC multicast/broadcast, e.g., CAPWAP protocol (RFC 


5416), though MAC protocols without 

multicast/broadcast are also able to accomplish the 

routing mission in MCRA [20]. 

It should be mentioned here that MCRA does not need 

the support of any special QoS-aware protocol in MAC 

layer. The common MAC protocols also can cooperate 

with the MCRA well. 

B. Message Suppression 

During the routing forwarding process mentioned 

above, we adopt some policies of message suppression in 

order to reduce the message redundancies and the 

re-transmission probability caused by collisions. Message 

suppression focuses on reducing the number of interests 

occurred during the routing process, so that it can help not 

only save the energy consumption in sensor nodes but also 

reduce the time of convergence of route discovery. In 

MCRA, the policies of message suppression consist of two 

aspects: restraining forwarding and deferring forwarding. 

1. Restraining Forwarding 

First of all, similar to other routing algorithms, each of 

nodes in MCRA periodically broadcasts a beacon packet 

(HELLO message) to its neighbors, so that every node can 

keep a neighbor table to store the information passed by 

beacons. Each entry inside the table has the following 

fields: (NeighborID, ExpireTime), where field ExpireTime 

is used to timeout this entry. If a neighbor entry is not 

refreshed after certain timeout, it will be removed from the 

neighbor table. Since geographic position information is 

not necessary in our routing process, no position 

information inside our neighbor tables. 

The main ideal of restraining forwarding policy is how 

to lessen the amount of interest messages by restraining 

some of nodes from forwarding interests. As shown in Fig. 

3, after an interest from node n1 is sent to node n2, node n2 

floods this interest to its neighbors, node n3, n4, n5, and n6 

excluding the node n1, in manner of multicast (because 

interest cannot be sent back to its visited nodes). In 

addition, node n2 stores its neighbor information into field 

interest.neighbors in this message, so did node n1. Note 

that the neighbor information of node n2 replaces the 

existing neighbor information of node n1 in this interest 

field interest.neighbors. 

ID 246 

ID 

2 

3 

5 

... 

n3 

n6 

n4 

n5 

n2 

ID 

1 

3 

4 

5 

6 

n1 

Unrestrained Node 

Restrained Node 

Unknown Node 

ID 

2 

Figure 3. The restraining forwarding policy 

r 

Interest 

When node n3 receives the interest, it begins to 

compare each element in the list interest.neighbors with


each entry in its neighbor table and calculate the value of 

following expression: 

ΦN⊆Φ INTN 

(5) 

Wherein N Φ denotes the set of its neighbors, and Φ INTN 

is the set of the elements in the interest field 

interest.neighbors. If the value of (5) is true, this node is a 

restrained node. That means it discards this interest 

immediately, e.g. node n3 and n6 in Fig. 3. Otherwise, this 

node is an unrestrained node, which forwards interest to its 

neighbors excluding the visited nodes, moreover rewrites 

field interest.neighbors in the interest, e.g. node n4. So the 

amount of interest messages in network can be lessened 

drastically. 

2. Deferring Forwarding 

The main purpose of deferring forwarding policy is to 

lower the amount of interests by dynamically deferring the 

forwarding actions on nodes. Deferring forwarding policy 

is able to make these nodes have enough time to collect 

and merge the interests from their neighbors as many as 

possible. 

Let the forwarding delay on some node be ∆ τ . As 

shown in Fig. 2, during a period of time ∆ τ , node k 

receives multiple interests with the same interest.type 

value from its neighbors, and all of which suffice for QoS 

constraints, i.e., the result of expression (4) is true. Node k 

will list these interests in some order, e.g., in 

interest.e2e_delay, or interest.packet_loss, or 

interest.hopcount order. And then, node k selects the best 

element from the ordered list and forwards it, drops the 

others. After forwarding the best interest, if the node k 

again receives some interests with the same interest.type 

and satisfying the constraint criteria (4), it also discards 

them. 

Considering the factors that influence forwarding delay, 

according to our results of extensive simulation 

experiments, we give an empirical formula that calculates 

the value of ∆ τ . For the forwarding delay on sensor node 

k, the formula is as follows: 

D D−T N 

V D V 

trip k 

∆ τk = ρk 

× × (6) 

Wherein D is the end-to-end delay constraint, |V| is the 

total number of sensor nodes, T denotes the trip time of 

trip 

an interest, N is the number of neighbors of sensor node 

k 

k, and ρ is the instantaneous queue size (in bytes) in 

k 

node k. From (6), we can see that an arbitrary sensor node 

(of course, apart from sink nodes) is able to easily 

calculate its forwarding delay ∆ τ by reading the 

information in interest and its neighbor table. 

C. Data Aggregation 

In multimedia applications, since sensor nodes in a 

monitored area might generate significant redundant data, 

and that duplicate or similar data packets from multiple 

sensors need to be aggregated, so that the amount of 

transmissions would be reduced. 


Suppose that node s and node i in a same monitored area 

send the data they detected respectively to their 

downstream node k, as shown in Fig. 4. If there exists 

redundancy in the data, node k will perform the 

aggregation computation by using some aggregation 

functions, such as suppression (eliminating duplicates), 

min, max and average etc. Some of these functions can be 

performed either partially or fully on each node in WSN. 

Monitored Area 

s 

m 

i 

Data Interest 

k 

Figure 4. The process of data aggregation 

n 

j 

sink 

In Fig. 4, if node i selects node j as its downstream node 

rather than node k, the aggregation computation will be 

performed on the sink node. 

Recognizing that computation consumes much less 

energy than communication, substantial energy savings 

can be obtained through the above process of data 

aggregation. 

D. Localization Scheme 

Borrowed from the idea of logical coordinate system 

[10], we design a new localization approach based on 

hop-count information for MCRA. The key difference is 

that instead of constructing a coordinate space for routing, 

we only use the hop-count information acquired from 

routing process to calculate the location of target. It not 

only avoids the message overhead induced by the 

coordinate space construction, but also eliminates the 

negative impact upon routing process due to the imprecise 

node coordinate information. 

1 

0 

Sink Common Node Source 

m 

s(6,7,4,3) 

Figure 5. The localization example 

Suppose that network nodes are placed on a plane of 

rectangle m units in length and n units in width (or other 

given size shapes, e.g., triangle, ellipse etc), where there 

are 4 perimeter nodes (or called landmarks), as shown in 

Fig. 5. Here we may take these landmarks as sink nodes 

(multiple sinks are used to improve the positioning 

precision of source nodes). We also assume that these sink 

nodes have known their respective hop-counts to some 

2 

n 

3


source node, e.g., the node s(6,7,4,3), where the numbers 

between the pair of parentheses form the logical 

coordinate vector of node s. Each of elements in this 

vector in order represents the respective hop-count from 

each sink to node s. The serial number of each element is 

the ID of its corresponding landmark (sink node). So the 

actual coordinate value of source node relative to these 

sinks can be computed in the following manner: 

∑ 

∑ 

L 

L 

− 

− 

i 

j 

x= m , y= n 

L + L L + L 

∑ − 

i ∑ + − + 

i ∑ j ∑ j 

(7) 

Where L ur denotes the logical coordinate vector of target; 

i − , i + , j − , and j + are the IDs of left, right, down, and 

up perimeter nodes in this coordinate plane, respectively. 

E. Network Dynamics 

Although most of sensor network architectures assume 

that sensor nodes are stationary, it is sometimes deemed 

necessary to support the mobility of sink nodes, and that 

the communications usually fail due to the energy 

exhaustion in sensor nodes or other causes. Besides the 

dynamics of network topology, the dynamics caused by 

the imprecision of network state information also demands 

that routing protocols are able to adapt to these variations 

of network state. 

In MCRA, there are two policies used to implement this 

function. The first policy is notification update, as shown 

in Fig. 6, during the data delivery, once the node k detects 

the communication failure from node s to k, it sends a 

notification message towards the sink node, immediately, 

which informs the sink node restarts a new routing process. 

Meanwhile, we also may use hold-down timer in sink 

node, if necessary. The period of hold-down timer is 

usually set to 3ω , where ω is the average time 

interval of the query occurrences on sink nodes. 

Monitored 

Area s 

Data Interest 

i 

m 

k 

n 

j 

Notification 

Figure 6. The adaptation to the network dynamics 

sink 

The second policy is periodic update, which is that sink 

node restarts periodically a new routing process in certain 

time interval ∆ t . Compared to cable networks, here ∆ t 

should have a larger value. We usually set ∆t≥ 30ω 

, 

where ω has the same meaning as above. 

Ⅴ. DISCUSSION 

A. Correctness Proof 

The correctness and feasibility of MCRA can be 

approved by the following two non-formalized theorems. 


Theorem 1 If a feasible path that suffices for QoS 

constraints exists, MCRA is able to find it. 

Proof: In the routing process of MCRA, interest 

messages are diffused in a restricted flood manner to seek 

all of the feasible paths. Hence, only if the paths that 

satisfy QoS constraints exist, MCRA must be able to find 

them. The theorem holds. 

Theorem 2 The paths found by MCRA form a 

loop-free reversed multicast tree with optimal energy 

consumption, which suffices for the end-to-end delay and 

packet drop ratio requirements. 

Proof: As mentioned before, in MCRA, interest 

message neither visits those paths that do not suffice for 

QoS constraints, nor visits those paths visited by it. In 

addition, each of paths from each source node to the sink 

node is the path with minimal energy consumption (i.e., 

minimal hop-count), because the hop-count of a path 

represents the accumulated energy consumption along the 

path (mentioned in section II). Besides, MCRA balances 

the energy consumption in networks by using the 

constraint to the residual energy ratio in sensor nodes. 

Thus, we may deem that the tree constructed by these 

optimal paths is a loop-free and reversed multicast tree that 

has the optimal energy consumption and satisfies the 

end-to-end delay and packet drop ratio requirements. The 

theorem holds. 

B. Complexity Analysis 

In MCRA, the overhead of message transmissions 

determines its complexity, since not only the total energy 

consumption but also the computation complexity in 

networks are proportional to the number of transmissions. 

Suppose that there exist |Q| queries in G=(V,E). We 

also may assume no collisions or very few collisions 

when messages transmit among nodes if the value of |Q| 

is not large, so each node in G performs 2|V| 

transmissions (|V| HELLO messages plus |V| interest 

messages) at most, where |V| denotes the number of nodes. 

Hence, we can draw that the complexity of message 

overhead is O(|V||Q|) theoretically. 

C. Simplified MCRA 

For the sake of suppression messages, we adopt 

restraining forwarding and deferring forwarding policies 

in MCRA. However, the restraining forwarding policy 

also imports extra control overhead (HELLO messages) 

so as to establish neighbor tables. We design a simplified 

MCRA named MCRA-S, which does not apply the 

restraining forwarding policy during its routing process. 

According to the complexity analysis above, the 

complexity of MCRA-S should be O(|V||Q|) theoretically, 

which is similar to MCRA, maybe even slightly better 

than MCRA. But this case only happens when the 

network has relatively low node density. As the sensor 

node density increases, MCRA would be getting better 

than MCRA-S because of the increasing collisions in 

MCRA-S. Our simulation experiments will prove this 

conclusion. 

In MCRA-S, despite the lack of neighbor table on 

sensor node, we may calculate the N value in (6) by 

k


the following equation used to compute the neighbor 

number of a sensor node in [21,22]. 

V 

N r 

S π 

2 

k = (8) 

Where |V| is the total number of sensor nodes, S denotes 

the total surface area covered by all sensor nodes, and r is 

the communication radius of node k. 

Ⅵ . SIMULATION 

In our evaluation, we compare the performance of four 

different routing algorithms: MCRA, SPEED, Directed 

Diffusion (DD) [23,24] and MCRA-S. Directed Diffusion 

is a typical data-centric algorithm based on query-driven 

data delivery mode, which optimizes single objective (e.g. 

energy savings) by selecting empirical good paths and by 

caching and processing data in network. SPEED is a 

representative algorithm that can guarantee the timeliness 

of multimedia communications by a combination of 

feedback control and non-deterministic geographic 

forwarding. In addition, both Directed Diffusion and 

SPEED are correlative with our algorithm MCRA. 

We simulate MCRA on NS2 (ns-2.33), because this 

version has implemented many MAC layer protocols 

applied in WSN and Directed Diffusion algorithm, as 

well as XCP (explicit congestion control protocol) that is 

similar to SPEED algorithm [25]. Table I describes the 

main setting parameters and scenarios for our 

simulations. 

TABLE I 

THE SIMULATION SCENARIO SETTINGS 

Configuration Options Setting Values 

Routing MCRA, SPEED, DD, MCRA-S 

llType (link layer type) LL (delay_: 0.25ms, bandwidth_: 

not used) 

macType (MAC layer protocol) Mac/802_11 

function) 

(without multicast 

ifqType (interface queue type) Queue/DropTail/PriQueue 

ifqLen (interface queue length) 50 

antType (antenna type) Antenna/OmniAntenna 

propType (propagation model) Propagation/TwoRayGround 

phyType (network interface type) Phy/WirelessPhy 

channel (channel type) Channel/WirelessChannel 

energyModel (energy model) EnergyModel 

Terrain (300m, 300m) 

Node placement Uniform distribution 

Node number and (Variables) 

Communication range 

We present the following metrics to evaluate and 

compare the performance of the four routing algorithms: 

1) end-to-end delay under different node number; 2) 

end-to-end packet loss ratio under different node number; 

3) control message overhead under different node number 

and different communication range (radio radius) 

respectively; 4) packet delivery ratio 

N sec (Nsec and 

N req 

Nreq are the successful connection number and sum of 

request connections respectively) under different node 

number; 5) average energy consumption J 

(J and n are 

n 


the total energy consumption and the node number 

respectively) under different node number. 

In our simulations, sink nodes and source sensor nodes 

are selected randomly in the scenarios, and the flows 

between them (4 flows in each simulation scenario where 

there are 4 queries and 4 sink nodes for DD, MCRA and 

MCRA-S) are CBR traffic pattern with a rate of 50 

packets/second. End-to-end delay constraint D, packet 

drop ratio constraint R, and residual energy ratio constraint 

E are uniformly distributed in [100ms, 250ms], [10%, 

30%], and [5%, 10%], respectively. The communication 

ranges in the scenarios with varying node number are set 

to 30m, on the other hand, the node number in the 

scenarios with varying communication range is 100. Note 

that our simulations do not count in the extra costs 

imported by measurement equipments or location 

message exchanges in SPEED and DD. In our 

experiments, each of the reported values is the average 

result over 100 runs with different random seeds and 

different random node topologies. 

A. End-to-end Delay 

End-to-end delay measures the network delay 

performance of these algorithms. Fig. 7 plots the 

end-to-end delay for the four different routing algorithms. 

At each point, we average the e2e delays of all then 

packets from the 24 flows (100 runs with 4 flows each). 

Delay (ms) 

300 

260 

220 

180 

140 

100 

60 

MCRA 

SPEED 

DD 

MCRA-S 

Node Number 

20 

10 40 70 100 130 160 190 220 

Figure 7. End-to-end delay vs. node number 

Seen from Fig. 7, SPEED has the best e2e delay 

performance as its optimized objective, in particular, in 

the route acquisition phase, because it is a 

non-deterministic geographic routing with less initial 

delay cost. Obviously, Directed Diffusion that only 

optimizes the energy savings by flooding has the worst 

e2e delay performance. Both MCRA and MCRA-S 

perform much better than DD, because the e2e delay is 

considered as an important element in their routing model. 

However, when the network has low density (


in Fig. 8 are the summary of 100 randomized runs. From 

the reported data, we can clearly see MCRA and 

MCRA-S are better than SPEED and DD. The main 

cause is that the e2e packet loss ratio in MCRA is an 

important performance constraint during the route 

discovery. In comparison to MCRA and MCRA-S, when 

the network has fewer sensor nodes (


Energy Consumption 

(mWhr) 

40 

35 

30 

25 

20 

15 

10 

5 

MCRA 

SPEED 

DD 

MCRA-S 

Node Number 

0 

10 40 70 100 130 160 190 220 

Figure 12. Average energy consumption vs. node number 

F. Localization Error 

Unlike SPEED, GPSR, and BVR [9] etc., MCRA is 

not a routing protocol based on location information, i.e., 

the route discovery does not need the location 

information of sensor nodes. So the localization error in 

MCRA only affects the positioning precision of targets. 

Let δ = δδ denote the position error, where x y 

δ and x 

δ are the position errors in horizontal and vertical 

y 

direction respectively. By extensive experiments, we plot 

the position error with respect to different sensor node 

number and different sink (i.e., perimeter node) number, 

as shown in Fig. 13. From it, we can find the relationships 

between the localization error and the node density as 

well as the perimeter node number. 

Position Error (%) 

50 

40 

30 

20 

10 

4 Perimeter Nodes 




Node Number 

0 

10 40 70 100 130 160 190 220 

Figure 13. Position error vs. sensor node and sink number 

Ⅶ . CONCLUSION 

This paper presents a multi-constrained routing 

algorithm MCRA based on query-flooding and 

query-driven data delivery mode for multimedia 

applications with periodic data in sensor networks. 

MCRA can not only provide end-to-end delay guarantee 

and packet loss ratio guarantee for multimedia 

communications, but also improve and balance the 

energy consumption in sensor nodes. Besides, MCRA 

adopts efficient policies to suppress message flooding and 

lessen data redundancy. In MCRA, extra position 

measurement equipment or location message exchanges 

are unnecessary, and that routing computation does not 

require the geographic or logical coordinate information 

of sensor nodes, however, target locations we concern 

still can be figured out on sink nodes by using hop-count 

information. In addition, we may optionally adopt MAC 


multicast in MCRA in order to further lessen its control 

message overhead. Meanwhile, MAC differentiation 

service can be applied to MCRA, so that real-time traffic 

and best-effort traffic in WSN can be classified into 

different forwarding priority levels. Theoretical analysis 

and extensive simulations not only demonstrate the 

correctness of MCRA, but also show that it has a good 

overall performance, thanks to the low end-to-end delay 

and loss ratio of data delivery, the low average energy 

consumption, the high packet delivery ratio, and the 

moderate control message overhead. Our further work is 

to investigate the performance of MCRA when the 

number of queries in WSN increases, and to manage to 

improve the localization precision of sensor nodes. 

APPENDIX A: THE METACODE MCRA 

//1. The definition of interest packet header 

struct interest_header { 

char *type; char **nodes; long hopcount; double 

e2e_delay; double packet_loss; double D; double R; 

double E; char **neighbors; double TTL; } 

//2. The main metacode MCRA 

MCRA (V, M, D, R, E, TTL) { 

Initialization (V); 

while (M≠0) { 

for each sin k∈ M { 

interest.nodes = AddList(null, sink.ID); 

sink.broadcast_interests(interest); 

for each k∈{ V − sin k} 

{ 

if ( k ∉ {int erest. nodes} 

) { 

interest.e2e_delay += k T ∆ ; /* Here k T ∆ is 

the trip time from the upstream of node k to it. */ 

if (interest.e2e_deay >= TTL) { 

k.drop_interest(interest); } 

interest.packet_loss *= packet _ loss ; k 

if ( kresidual . _energy≥ 

E && 

interest. packet _loss 

≤ R && 

interest.2 e e _delay 

≤ D ) { 

interest.hopcount += 1; 

interest.nodes = AddList(interest.nodes, 

k.ID); 

} else { k.drop_interest(interest); } 

/* Check if the node is source node. */ 

if ( k = interest.type) { s = k ; 

s.merge_interests( ∆ τ ); /* Here s ∆ τ is s 

the delay of source node s. */ 

s.coordinate_vector = 

AddList(interest.hopcount); 

s.send_data(interest.nodes, 

s.coordinate_vector, sink); 

} else { 

if (k.RFF(interest) == “Restrained Node”) 

{k.drop_interest(interest); 

} else { k.merge_interests( ∆ τ ); k


N = DeleteList(k.neighbors, 

interset.nodes); /* Here N is the set 

of node k’ neighbors except for the visited nodes. */ 

k.forward_interest(N); } } 

} 

} 

sink.TLS(s.coordinate_vector); /* Calculate the actual 

coordinate values of each monitored target node by Target 

Localization Scheme (TLS) */ 

M=M-sink; 

} 

if (M=0) end;}} 

//3. The meta-code of initialization subroutine 

Initialization (V) { 

for each k∈ V { 

k.exchange_beacons(HELLO); 

k.get_neighbors(NeighborID, ExpireTime); 

} 

} 

//4. The meta-code Restraining Forwarding Function 

(RFF) 

RFF (INT) {/* Here INT denotes interest. */ 

if (this.neighbors ⊆ INT.neigbors) { 

this = “Restrained Node”; 

} else {this = “Unrestrained Node”; 

INT.neighbors = this.neighbors; 

} 

return (“Restrained Node” || “Restrained Node”); 

} 

//5. The meta-code Target Localization Scheme (TLS) 

TLS ( L ur ) {/* Here L ur is the logical coordinate vector of 

target. */ 

if ( L ur >= 4) { 

∑ 

∑ ∑ ); 

return ( x= m 

L − 

i 

L − + i L + 

i 

return ( 

y = n 

∑ L − 

j 

L + L 

); 

∑ ∑ 

− + 

j j 

/* Here i − , i + , j − , and j + are the IDs of left, right, 

down, and up landmarks in this coordinate plane, 

respectively, which are known; m and n are also known 

planar shape parameters. */ 

} 

} 

ACKNOWLEDGMENT 

This work is supported by the Ph.D. Program 

Foundation of Ministry of Education of China under Grant 

No. 200804971030, the Natural Science Foundation of 

Hubei Province of China under Grant No. 2008CDB347, 

and the Fundamental Research Funds for the Central 

Universities of China under Grant No. 2010-Ia-049. 


REFERENCES 

[1] J. R. Gallardo, P. Medina, and W. Zhuang, “QoS 

mechanisms for the MAC protocol of IEEE 802.11 

WLANs,” Wireless Networks, vol. 13, no. 3, pp. 335-349, 

June 2007. 

[2] I. H. Hou and P. R. Kumar, “Admission control and 

scheduling for QoS guarantees for variable-bit-rate 

applications on wireless channels,” in Proc. of the 10th 

ACM International Symposium on Mobile Ad Hoc 

Networking and Computing, New Orleans, LA, USA, pp. 

175-184, May 2009. 

[3] J. K. Song, J. D. Ryoo, and S. C. Kim et al, “A dynamic 

GTS allocation algorithm in IEEE 802.15.4 for QoS 

guaranteed real-time applications,” in Proc. of the IEEE 

International Symposium on Consumer Electronics, pp. 1-6, 

June 2007. 

[4] J. J. Garcia and T. Falck, “Quality of service for IEEE 

802.15.4-based wireless body sensor networks,” in Proc. of 

the 3rd International Conference on Pervasive Computing 

Technologies for Healthcare, pp. 1-6, April 2009. 

[5] I. F. Akyildiz, T. Melodia, and K. R. Chowdhury, “A survey 

on wireless multimedia sensor networks,” Computer 

Networks: the International Journal of Computer and 

Telecommunications Networking, vol. 51, no. 4, pp. 

921-960, 2007. 

[6] N. Dimokas, D. Katsaros, and Y. Manolopoulos, 

“Cooperative caching in wireless multimedia sensor 

networks,” Mobile Networks and Applications, vol. 13, no. 

3-4, pp. 337-356, August 2008. 

[7] F. Ingelrest and S. R. David, “Localized broadcast 

incremental power protocol for wireless ad hoc networks,” 

Wireless Networks, vol. 14, no. 3, pp. 309-319, June 2008. 

[8] A. Jadbabaie, “On geographic routing without location 

information,” in Proc. of the 43rd IEEE Conference on 

Decision and Control, vol. 5, pp. 4764-4769, Dec. 2004. 

[9] R. Fonseca, S. Ratnasamy, and J. Zhao et al, “Beacon vector 

routing: scalable point-to-point routing in wireless 

sensor-nets,” in Proc. of the 2nd conference on Symposium 

on Networked Systems Design & Implementation, vol. 2, 

pp. 329-342, 2005. 

[10] Q. Cao and T. Abdelzaher, “A scalable logical coordinates 

framework for routing in wireless sensor networks,” in 

Proc. of the 25th IEEE International Real-Time Systems, 

pp. 349-358, Dec. 2004. 

[11] T. He, J. Stankovic, and L. Chenyang et al, “SPEED: a 

stateless protocol for real-time communication in sensor 

networks,” in Proc. of the 23rd International Conference 

on Distributed Computing Systems, pp. 46-55, May 2003. 

[12] E. Felemban, C. G. Lee, and E. Ekici et al, “MMSPEED: 

multipath multi-SPEED protocol for QoS guarantee of 

reliability and timeliness in wireless sensor networks,” 

IEEE Transactions on Mobile Computing, vol. 5, no. 6, pp. 

738-754, 2006. 

[13] L. Shu, Y. Zhang, and L. T. Yang et al, “Geographic 

routing in wireless multimedia sensor networks,” in Proc. 

of the Second International Conference on Future 

Generation Communication and Networking, vol. 1, pp. 

68-73, Dec. 2008. 

[14] L. Zhang, M. Hauswirth, and L. Shu et al, “Multi-priority 

multi-path selection for video streaming in wireless 

multimedia sensor networks,” in Proc. of the 5th 

International Conference on Ubiquitous Intelligence and 

Computing, vol. 5061, pp. 439-452, June 2008. 

[15] K. Akkaya and M. Younis, “Energy and QoS aware routing 

in wireless sensor networks,” Springer Cluster Computing 

Journal, vol. 8, no. 2-3, pp. 179-188, 2005.


[16] J. Chen, Y. Guan, and U. Pooch, “Customizing GPSR for 

wireless sensor networks,” in Proc. of the IEEE 

International Conference on Mobile Ad-hoc and Sensor 

Systems, pp. 549-551, Oct. 2004. 

[17] F. Tang, M. Guo, and M. Li et al, “Secure routing for 

wireless mesh sensor networks in pervasive 

environments,” International Journal of Intelligent Control 

and Systems, vol. 12, no. 4, pp. 293-306, 2007. 

[18] M. Busse, T. Haenselmann, and W. Effelsberg, 

“Energy-efficient forwarding in wireless sensor networks,” 

Pervasive and Mobile Computing, vol. 4, no. 1, pp. 

3-32, 2008. 

[19] I. S. Hwang and J. H. Wu, “Performance assessment of 

service differentiation in IEEE 802.11e wireless LANs,” 

International Journal of Ad Hoc and Ubiquitous 

Computing, vol. 3, no.1, pp. 21-32, 2008. 

[20] R. Fantacci, G. Vannuccini, and G. Vestri, “Performance 

analysis of a multiple access protocol for voice and data 

support in multiuser broadband wireless LANs,” Wireless 

Networks, vol. 14, no. 1, pp. 17-28, Jan. 2008. 

[21] R. Nagpal, H. Shrobe, and J. Bachrach, “Organizing a 

global coordinate system from local information on an ad 

hoc sensor network,” in Proc. of the 2nd International 

Workshop on Information Processing in Sensor Networks 

(IPSN '03), Palo Alto, CA, US, pp. 333-348, April 2003. 

[22] J. Bachrach, R. Nagpal, and M. Salib et al, “Experimental 

results and theoretical analysis of a self-organizing global 

coordinate system for ad hoc sensor networks,” 

Telecommunication Systems, vol. 26, no. 2-4, pp. 213-233, 

2004. 

[23] C. Intanagonwiwat, R. Govindan, and D. Estrin, “Directed 

diffusion: a scalable and robust communication paradigm 

for sensor networks,” in Proc. of the 6th Annual 

international Conference on Mobile Computing and 

Networking (MobiCom'00), Boston, Massachusetts, US, pp. 

56-67, Aug. 2000. 

[24] L. Khelladi and N. Badache, “On the performance of 

directed diffusion in dense sensor networks,” in Proc. of 

4th International Conference on Innovations in 

Information Technology, Dubai, pp. 113-117, Nov. 2007. 

[25] I. A. Qazi and T. Znati, “On the design of load factor based 

congestion control protocols for next-generation 

networks,” in Proc. of IEEE INFOCOM 2008 - The 27th 

Conference on Computer Communications, pp. 96-100, 

April 2008. 

Xin Yan: received his M.Sc. degree in 

electrical engineering from the Hubei 

University of Technology, China, in 1997, 

and his Ph.D. degree in computer science 

from the Wuhan University of 

Technology, China, in 2006. 

He is an associate professor at the 

Department of Computer Science, Wuhan 

University of Technology, China. He was a postdoctoral 

researcher at the Network Architectures and Services Group, 

Delft University of Technology, The Netherlands, from Jan. 

2009 to Jan. 2010. His main research interests lie in new 

Internet-like network architectures, and the modeling and 

performance analysis of network behavior and complex 

infrastructures. 

F. J. An: received her B.S. degree in electrical engineering from 

the Delft University of Technology, The Netherlands, in 2010. 


She is currently pursuing the M.Sc. degree in telecommunication 

electrical engineering at the Delft University of Technology, The 

Netherlands. Her research interests lie in mobile computing, 

wireless sensor networks, and ad hoc wireless networks. 

Layuan Li: received his B.S. degree from the Harbin Institute of 

Military Engineering, China, in 1970 and his M.Sc. degree in 

communication and electrical systems from the Huazhong 

University of Science and Technology, China in 1982. 

He is a professor and Ph.D. supervisor at the Department of 

Computer Science, Wuhan University of Technology, China, and 

the Editor-in-Chief of the Journal of Wuhan University of 

Technology. His research interests include high speed computer 

networks and protocol engineering. He received the National 

Special Prize by the Chinese government in 1993.


Reputation-aware Service Selection based on 

QoS Similarity 

Shenghui Zhao 1,2 , Guoxin Wu 2 

1 Department of Computer Science and Technology, Chuzhou University, Chuzhou, China 

2 School of Computer Science & Engineering, Southeast University, Nanjing, China 

Email:zsh@chzu.edu.cn, gwu@seu.edu.cn 

Guilin Chen, Haibao Chen 

Department of Computer Science and Technology, Chuzhou University, Chuzhou, China 

Email:glchen@chzu.edu.cn, chb@chzu.edu.cn 

Abstract— For the up-and-coming computing models like as 

cloud computing, service is the standard package for 

meeting all kinds of consumers' requirements. Web Services 

are the concrete implement of the service. When users 

request and consume Web Services, services' reputations 

will play a vital role in users' selection. A gradually 

adjusting reputation evaluation method of Web Services is 

proposed based on eliminating the collusive behaviors of 

consumers step by step, and a reputation-aware model for 

service selection is designed. In order to adjust reputations, 

QoS similarity is computed firstly according to the 

differences between advertised QoS from service providers 

and delivered QoS from service consumers' evaluation, next, 

current reputation is attained; then the consumers are 

sorted based on reputation using clustering algorithm and 

the potential collusive consumers are mined using 

association rules algorithm; finally, the updated reputation 

is recalculated and saved in the reputation center included 

in the model. The experimental results show that the model 

can identify the malicious consumers and improve the exact 

rate of reputation evaluation and success rate of service 

selection. 

Index Terms—Web Service, quality of service (QoS), 

reputation update, clustering algorithm, collusive 

consumers 

I. INTRODUCTION 

With the widespread of SOA, Web Services has 

become the main computing paradigm across Internet, 

new computing patterns are springing up such as cloud 

computing and CPS (Cyber Physical Systems) etc. A Web 

Service is a self-described and self-contained application 

that uses standard Internet technologies to interact with 

other Web Services, which can be published and accessed 

through the web. At present, many corporations and 

organizations have implemented their core application 

through buying the Web Services on Internet. For example, 

salesfore.com provides ERP service for users. Along with 

the maturation of service market, more and more service 

providers can provide the same or similar service, how to 

rationally select satisfied service has been turned into one 

of the key problems in Web Services research fields. 


doi:10.4304/jnw.6.7.950-957 

When service requestors select required services 

among many services with similar functionality, services' 

non-functional properties is an important considerable 

criterion, such as QoS (Quality of Service), reputation, 

etc. Generally speaking, QoS of Web Service is described 

by response time, reliability, availability, security and 

execution cost and so on. In the early service transactions, 

QoS information was published by service providers, but 

it was not always exact and up-to-date. For the interest of 

ensuring the veracity of QoS properties, it should be a 

direct and valid method to appraise the QoS by requestors 

after invoking the Web Service. These values can be 

acted as the references for subsequent consumers to select 

the service. Many researches on service selection adopt 

this scheme. 

However, in the practical transactions, some feedbacks 

about QoS are falsity information due to the vicious 

estimation aiming at service providers. Thus, relying only 

on feedback estimation of QoS can not provide accurate 

methods for service selection. Reputation based service 

selection methods were proposed later, most of which 

were reputation evaluation on the basis of appreciable 

QoS after invoking a service, and then computed 

predicted reputation integrating multi historical values 

and current value. Above methods can wipe off influence 

of little vicious users at a certain extent and improve the 

success rate of service selection. But, the community 

collusions may be occurred among consumers or among 

consumers and providers, services' reputation may be 

either lower or rose up which leads to distortion of 

reputation. 

This paper discusses service selection based on 

reputation, in which distinguishes and filters out the 

collusive consumers through collusive behavior analysis 

methods. Then these collusive consumers' ratings are 

ignored, and decrease the influence of the malicious 

consumers, which can improve the veracity of reputation 

and the success rate of service selection. The rest of the 

paper is organized as follows. In section 2, we introduce 

related work. Section 3 proposes a method for Web 

Service reputation evaluation. A model for service


selection is set up in Section 4. Next is experimental 

analysis. At last, we conclude the paper. 

II. RELATED WORK 

It is essential to acquire the QoS information when 

service selection is depending on QoS. [1] presented that 

both service selection and composition were QoS-aware; 

the QoS was measured by monitoring system according to 

service operations. An approach for measuring quality of 

Web Services based on the superposition of uncertain 

factors was proposed, and a judging method for 

determining priorities among Web Services, which can 

help users select satisfied service[2]. A dynamic and QoSdriven 

model for service selection was proposed in [3], 

and the dynamic QoS data were computed according to 

users' feedback. In [4], the QoS attributes’ data obtained 

from service providers was revised, and feedback 

similarity came from service consumers was used to 

weight QoS data' trustworthiness, that strengthened the 

accuracy of the service selection. In [5], service-level 

agreements were discussed in order to set the penalties 

over the lack of QoS for web services. It ensured that the 

trustworthiness of a service-oriented environment relies 

on reliable QoS monitoring in certain sense. 

Although QoS based service selection is essential, due 

to the services' marketability, it is hard to avoid the 

dishonest service providers. So, service selection must be 

trust or reputation based [6], that can assure service's 

trustworthiness. In [7], Yao Wang et al. reviewed and 

concluded the service selection' criteria, and presented that 

it was necessary to implement service selection depending 

on trust and reputation. The authors in [8] suggested a 

framework of service selection based on reputation in a 

semantic network. The reputation was computed by 

different service consumers. In [9], Malik et al. had 

proposed a model to compute the reputation of a web 

service in accordance with the personal evaluation of the 

previous users. The characteristic of this method was the 

credibility of the users of evaluating services has been 

taken into account. If the rater tried to provide a fake 

rating, then its credibility would be decreased and the 

rating of this user would become less important in the 

reputation of the web service. 

Obviously, QoS can help consumers select the service 

with high quality, and reputation has been used to make 

consumers select the service providers which honestly 

offer the service with advertised QoS. Making use of 

reputation, consumers can find or select secure, reliable 

and trusted Web Services. So, the service quality's 

reputation is vital important to select the genuine service 

required by the consumers. In [10], Maximilien and Singh 

designed a multi-agent framework based on ontology for 

QoS. The users’ ratings which depended on the different 

qualities satisfied varied consumers' trust requirement 

used for computing the reputation of the web service, and 

it would be the selection criterion. That was dynamic 

selection. 

Ping Wang et al.[11] expressed an idea that aggregating 

previous assessment records (bodies of evidence) via 

consumers' feedbacks and witness of network referrals to 


derive a more objective reputation score on the specific 

service. Then two factors was defined, confidence degree 

and support degree based on evidence theory, to enhance 

the discrimination of the quality of existing evidence to 

help providers avoid malicious assessment. In [12], 

service providers' reputations were figured out through 

applying the current reputation and historical data with 

various weights. According to providers' reputation and 

services' reputation, a method for measuring service 

providers' trust was proposed. By ranking the trust value, 

consumers can select more trusted service. In [13], the 

authors developed a framework aiming to select Web 

Services based on the trust policy expressed by the users. 

The framework allowed the users to select a web service 

matching their needs and expectations. 

In above-mentioned literatures about service selection, 

some only proposed the models of applying reputation, 

and some gave the rating methods on reputation, but they 

lacked of controlling the situations of providing falsity 

reputation information. That is, there are only few 

researches on consumers' collusion and deception. 

Although some researches [14][15] have considered 

collusion among consumers, their analysis objects and 

methods are different from our paper. Moreover, most 

literatures didn't give solutions on integrality of 

reputation data. 

For the sake of solving above problems, in our 

reputation-aware service selection model, the reputation is 

computed based on the similarity of service quality's 

attributes, and a method for collusion behavior analysis is 

proposed. Besides, while constructing a system reputation 

model, we take into account reputation storage and 

security, which can provide a relative secure and trusted 

service selection scheme for consumers. 

III. A METHOD FOR RATING WEB SERVICE 

The evaluation method is based on following 

suppositions: 

1) One service provider offers one service only, and the 

provider's reputation can be apperceived by its' service 

reputation. 

2) The reputation published to UDDI by service 

providers is authentic. 

3) The reputation center is trustable. It can be acted as a 

broker for service consumers and providers' behaviors. 

4) If one service consumer selects a service, it must 

trust in the service provider. 

5) If the consumer is honesty, its evaluation is honesty 

too. 

A. Computation of Web Service's Reputation. 

Supposing that a Web Service s j has m attributes of 

QoS, what is expressed as ( q1, q2,..., q m) 

. For any user 

u i , its invocation to s j 's attributes is represented as 

ij ij ij 

( q1 , q2,..., q m) 

. The advertised QoS of s j is shown as 

j j j 

( Ad_ q , Ad_ q ,..., Ad_ q ) . After s j being invoked by 

1 2 

m


u i , its feedback rate on s j 's QoS is denoted as 

( Eval _ q , Eval _ q ,..., Eval _ q ) . 

ij ij ij 

1 2 

m 

Definition 1 Similarity of service s j 's quality 

j 

Sim q 

After current user i u invoking service j s , u i will give 

a new value on QoS' attributes. The similarity degree 

Sim can be figured out in (1). 

j 

q 

Sim 

= 1− 

m 

∑ 

j k = 1 

q 

ij j 

( Eval _q−Ad_ q ) 

k k 

m is the number of QoS attributes. 

Definition 2 Service s j ' s current reputation 

m 

2 

rep 

j 

cur 

(1) 

s j 's current reputation is the newest reputation after 

j 

this invoking. repcur is computed using (2) which is 

figured as Fig. 1. 

j j 

rep = 1−sinh(1 − Sim ) . (2) 

cur q 

Thereinto,sinh( x) = (exp( x) −exp( − x)) 

/ 2 . 

Figure 1. Result of Function sinh(x) 

It can be seen from Fig. 1, when advertised QoS is 

closer to received QoS, the higher reputation is gained. 

j 

j 

Especially, if Simq is equal to 1, repcur can achieve 

j 

j 

highest value 1. When Sim is less than 0.2, rep q 

cur is 

nearly 0. This is very similar to the actual application 

occasions. So we adopt the function in formula (2). 

But current reputation had something to do with rater's 

credibility degree. If rater is trustful user, its reputation is 

trustworthy and can be contained in the global reputation. 

Otherwise, its reputation is not trustful and will be ignored 

in the global reputation. 

B. Reputation Update and Adjustment 

When j s is invoked at each time, s j 's current 

j 

reputation repcur can be computed using formula (2). 

However, for other users, they not only use the last result, 

but also consult historical data. A reputation center being 

involved in the service selection model is designed in 


Section 4. Reputation center plays the role of computing 

and keeping each service's global reputation. Hence, after 

each time invoking s j , it should update the stored s j 's 

global reputation. 

Assuming parameter δ is the difference between 

received QoS and advertised QoS, that is, 

m 

∑ 

ij j 

δ = ( Eval _q−Ad_ q )/ m .If δ is larger than 

i= 

1 

k k 

0, it can be explained that received QoS is prior to 

advertised QoS. Then the reputation will be increased 

apparently; otherwise it will be decreased. 

j 

Definition 3 Service s j 's global reputation rep glo . 

j 

s j 's global reputation repglo is associated with all 

historical reputations and current reputation. After each 

invoking, 

rep and historical data are updated and 

j 

cur 

j 

rep glo is adjusted using formula (3). 

j j 

rep = rep × (0.9 + 0.1× exp( δ )) (3) 

glo upd 

upd 

n 

∑ i cur 

n 

i / ∑ i , i 

i= 1 i= 

1 

j j di 

In (3), rep = rep _ × γ γ γ = λ 

j 

rep is the ith current reputation of service 

i_cur s j computed by formula (2). γ is the aging factor for the 

i 

ith service reputation, λ ∈ [ 0,1] 

. A smaller λ means only 

recent reputations are included and a larger λ means more 

reputations are included. d i is the time interval of 

between last rating time and current time. For instance, 

when using current reputation, d i almost equals to 0, so 

γ is 1. That is, the current reputation has not aged. 

i 

C. Reputation Storage 

In general, there are three places to store reputation 

information, rater (service consumer), ratee(service 

providers) and a third party(reputation center). This paper 

stores the information in ratee and a third party, 

respectively. The advantage of saving the global 

reputation in ratees is that consumers can find the satisfied 

service's reputation from providers at the time of 

reputation center collapsing. Due to computing global 

reputation in reputation center, each service's rating 

information should be saved. A database is created in 

reputation center to save rating information. The storage 

format includes five items, listed in Table I. 

Rater ID 

(UID) 

TABLE I. STORAGE FORMAT OF REPUTATION 

Service ID 

(SID) 

Reputation 

Value with 

encrypted 

Time 

Stamp 

Transaction 

Number 

(TID) 

Rate0001 Serice001 Repkey TM1 Tran0001 

When a consumer finds the satisfied service providers, 

he will query the providers' reputation in either reputation


center or providers selves. In order to prevent the 

providers to tamper with the saved reputation, the 

reputations will be handled in particular. The method is 

adopting the RcertX[16] idea to save the reputation 

information, where each service's reputation is put in one 

certificate, so all reputations are linked as a body to avoid 

the provider juggling with the reputation. After each 

global reputation is updated, it will be signed digital and 

stored in reputation center and sent to ratee. When users 

take out the reputations from the providers, they will use 

public key published by reputation center to decrypt the 

messages. Of course, public key has been already 

delivered to all consumers by broadcast, and it can be 

transmitted among consumers too. 

D. Algorithm for collusion behavior analysis 

After service transactions, honest users will give 

fairness and objectivity valuation. But the malicious 

consumers will give unpractical estimation, like as rising 

up or playing down the reputation intentionally, as well as 

collusive users. If the proportion of collusive users is 

higher, it will affect the service providers' reputation more 

greatly. Therefore, reputation model should be capable of 

identifying the collusive users and reduce the negative 

influence. 

For the sake of finding the collusive consumers, we 

design an algorithm named as CBA (Collusive Behavior 

Analysis) which has two main operations: classification 

and mining. We will use k-means cluster algorithm [17] to 

create multi clusters and find the correlations among 

consumers adopting the association rule mining algorithm. 

The CBA algorithm steps are as follows: 

Step 1 Take out all records of service s j from the 

database in reputation center, and sort the reputation 

values into three clusters (marked as 0,1,2) using k-means 

cluster algorithm. The 0th cluster denotes the class with 

the lowest reputation, and the 2th cluster represents the 

class with the highest reputation. Average the reputation 

distributed in the 1th cluster, and makes the mean as the 

references to analyze the collusion. 

Step 2 Owing to the reputations in the class of 0th 

clusters are all lower generally, which express that all 

assessment are on the low side to service s j . The reason 

has two sides, on the one hand, the service s j is bad 

service originally; on the other hand, it may be the 

collusive consumers purposely debasing s j ' s reputation. 

For differentiating them, taking out some others service 

from all services except for s j , the extraction proportion 

is 33%. Then, each service records construct their data 

collection which will be analyzed applying the first step, 

respectively. Moreover, take out users included in 0th 

cluster deriving from each service as a list. This user list is 

merged into a new list with service s j ' s 0th cluster user 

list. 

Step 3 Applying the association rules algorithm, on 

given support degree, the largest frequent items can be 


mined, in which there are different users. 

Step 4 The different users in the largest frequent items 

may be looked as collusive community. So, if the user 

exists in the community, it will be marked as dishonest 

user and set hi=0; otherwise, set hi=1. 

j 

After detecting the collusive users, recalculate rep , cur 

j j 

viz. repcur = hi × rep . So, when hi=0, the current 

cur 

reputation is zero, too. 

For example, there are two services, s1 and s2. Their 30 

transaction records are listed as Table II. 

TABLE II. TRANSACTION RECORD 

TID UID SID RepValue TID UID SID RepValue 

1 u1 s1 0.8 16 u6 s1 0.66 

2 u2 s2 0.5 17 u7 s2 0.83 

3 u2 s1 0.6 18 u1 s1 0.89 

4 u2 s1 0.65 19 u8 s2 0.83 

5 u3 s2 0.55 20 u9 s1 0.91 

6 u1 s2 0.8 21 u6 s2 0.95 

7 u2 s2 0.85 22 u5 s1 0.78 

8 u1 s1 0.86 23 u4 s2 0.68 

9 u3 s1 0.2 24 u5 s1 0.88 

0 

1 

u4 s1 0.6 25 u6 s2 0.95 

1 

1 

u5 s1 0.75 26 u6 s2 0.75 

2 

1 

u4 s2 0.7 27 u1 s1 0.9 

3 

1 

u5 s2 0.55 28 u4 s2 0.85 

4 

1 

u5 s2 0.6 29 u3 s1 0.58 

5 

1 

u4 s2 0.7 30 u1 s2 0.79 

Based on the above algorithm and using k-means 

cluster algorithm, we extract all s1 records and s2 records 

and divide them into three clusters, respectively. The 

results are shown in Table III and Table IV. 

TABLE III. DIVIDE S1 RECORDS INTO THREE CLUSTERS. 

---------Cluster0--------- 

00029[0.58,u3,s1] 

00010[0.6,u4,s1] 

00003[0.6,u2,s1] 

00009[0.2,u3,s1] 

---------Cluster1--------- 

00001[0.8,u1,s1] 

00022[0.78,u5,s1] 

00011[0.75,u5,s1] 

00016[0.66,u6,s1] 

00004[0.65,u2,s1] 

-----------Cluster2--------- 

00018[0.89,u1,s1] 

00024[0.88,u5,s1] 

00027[0.9,u1,s1] 

00008[0.86,u1,s1] 

00020[0.91,u9,s1]


TABLE IV. DIVIDE S2 RECORDS INTO THREE CLUSTERS 

---------Cluster0--------- 

00002[0.5,u2,s2] 

00005[0.55,u3,s2] 

00014[0.6,u5,s2] 

00013[0.55,u5,s2] 

-----------Cluster1--------- 

00030[0.79,u1,s2] 

00015[0.7,u4,s2] 

00006[0.8,u1,s2] 

00026[0.75,u6,s2] 

00023[0.68,u4,s2] 

00012[0.7,u4,s2] 

-----------Cluster2--------- 

00007[0.85,u2,s2] 

00019[0.83,u8,s2] 

00025[0.95,u6,s2] 

00028[0.85,u4,s2] 

00017[0.83,u7,s2] 

From Table III and Table IV., we can see, users' list 

(u3,u4,u2) and (u2,u3,u5) derived from cluster0 are made 

up of two lines. If there are more services, they will 

3, 4, 2 

formed as ⎧u u u ⎫ 

⎪ ⎪ in which collusive consumers could 

⎨u2, u3, u5⎬ 

⎪...... ⎪ 

⎩ ⎭ 

be mined using association rule mining algorithm. 

IV. SERVICE SELECTION OF REPUTATION- AWAREE 

Service selection model based on reputation is designed 

and shown as Fig.2. It includes two roles: service 

providers, service requestors, and a data center: UDDI, as 

well as three agents: discovery agent, selection agent and 

rating agent. Discovery agent helps service consumers to 

find services meeting requirements, and selection agent 

selects the service with highest reputation for consumers 

from the satisfied services. The purpose of rating agent is 

to evaluate the newest reputation of service providers and 

update the global reputation. 

Figure 2. Model of Service Selection 

Satisfied 

Services 

The process of service selection is as follows. 

1) If a service provider joins the system, it will publish 

its advertised QoS to UDDI acquiescently. 


2) Service requestors (viz. consumers) send their 

requirements which contain functional and nonfunctional 

descriptions to discovery agent. 

3) Discovery agent queries the UDDI. If UDDI has the 

satisfying services, it returns the results to discovery agent. 

4) Discovery agent submits the results to selection 

agent. 

5) Selection agent will query the reputations of all 

matched services to reputation center. 

6) Selection agent sorts the reputations of all satisfied 

services, and returns service with the highest reputation to 

requestor. 

7) Service provider and requestor carry out their 

transaction. 

8) After using service, consumer evaluates the aware 

quality of service, and sends information back to 

reputation center. 

9) Rating agent sends request to UDDI for promised 

QoS of service provider. Then updates the service' global 

reputation according to formula (2) and (3). 

10) Reputation center makes digital signature for the 

reputation and feed it back to the provider. 

CBA algorithm will be executed by reputation center 

after a period of time. The results can be used in 9). 

V. EXPERIMENTATION 

In order to validate our reputation model, we develop a 

simulation program written in Java. It simulates multi 

service provider and consumers' transaction behavior. The 

transaction results are saved like as the format listed in 

Table I. For the purpose of finding the collusive users, 

CBA algorithm is employed to make clustering and mine 

data for the transaction records at regular intervals. As 

time goes on, collusive consumers will gradually be 

steady. So the diminished reputations also tend towards 

stability. In experiments, the records being used are in a 

fixed time of transactions. 

A. Experimental Environment 

In our simulation environment, the number of service 

providers is 100, and the number of service consumers is 

60. For the convenience of experiment, the kind of 

services offered by providers is 1. Each service's QoS 

includes availability, response time, reliability, security 

and cost. Each service's attributes has four levels: bad, 

ordinary, good, and excellent. We assumed that there are 

10% excellent, 20% bad, 30% good and 40% ordinary 

among the 100 providers. 

However the service level is, the collusive consumers 

will give bad evaluation, and honest consumers will give 

authentic estimation. The appraisal is like as Table V. 

service 

level 

TABLE V. ESTIMATION RANGE FROM CONSUMERS 

proportion 

of all 

providers 

default 

of the 

attributes 

Estimation 

Range from 

Collusive 

Consumers 

Estimation 

Range from 

Honest 

Consumers 

Excellent 10% 1 0.3~0.5 0.9~1 

Good 30% 0.85 0.3~0.5 0.7~0.9 

Ordinary 40% 0.7 0.3~0.5 0.5~0.7 

Bad 20% 0.5 0.3~0.5 0.3~0.5


At the beginning of experiment, assumed all users are 

honest and all services' initial reputation are 0.6.The value 

of execution transaction is 0.6, meaning that if reputation 

is greater than or equal to 0.6, a transaction will be 

executed. At the same time, the successful transaction 

refers to the reputation greater than a threshold. Given 

threshold 0.7, viz. if the estimation value is no less than 

0.7, the transaction is successful. We will take success rate 

of service as a validity measurement of reputation model. 

Success rate of service is defined as the ratio of the 

number of times to the number of total transactions. When 

finding maximum frequent items, we set the support 

degree 0.9. 

B. Experimental Data 

In order to examine the validity and veracity of 

reputation computing in service selection model, we use 

rate of service success and exact rate of reputation to 

measure, respectively. Comparison is implemented in the 

two situations of considering collusion and no considering 

it. 

(1) Comparison of Success Rate 

At first, we study the success rate when the ratio of 

collusion (RoC) varies from 10% to 40% increased by 

10%. The results of considering collusion and no 

considering collusion are shown in Fig.3 and Fig.4. 

Figure 3. Success Rate of Considering Collusion under different RoC 

Figure 4. Success Rate of No Considering Collusion under different 

RoC 

From Fig.3 and Fig.4, the success rate reduces 

evidently as the RoC increasing, it can be explained that 


along with the number of round grew higher, success rate 

goes to steady. When the RoC is 10%, the both success 

rates are approaching to 90%, what can be explained the 

effect of collusion is not distinctness. But, when the RoC 

arrives at 40%, the success rate of considering collusion is 

higher than no considering collusion. Obviously, 

considering collusion has better effect for improving 

success rate. If the ratio of collusive users is bigger, it will 

have a larger effect on the service providers' reputation. 

TABLE VI. PROPORTION OF DIFFERENT SERVICE LEVEL 

service level Proportion of service providers 

Excellent 10% 20% 30% 40% 50% 

Good 20% 20% 20% 20% 20% 

Ordinary 50% 40% 30% 20% 10% 

Bad 20% 20% 20% 20% 20% 

Secondly, we compare the success rate of different ratio 

of excellent services under the circumstances of 20% 

RoC. The ratio of excellent services (RoE) is listed in 

Table VI. When the ratio of excellent service changes 

from 10% to 50%, the rates of success are displayed in 

Fig.5. 

Figure 5. Success Rate of Different Ratio of Excellent Services 

From Fig.5, we can see, no matter how change the RoE 

is, success rate has no great effect in certain collusive 

consumers. The phenomenon is consistent with the fact, 

because of the success rate only be affected on the number 

of collusive consumers, as well as the ratio of collusion is 

fixed. 

(2) Exact Rate of Reputation 

A service s j ' exact rate is represented as: 

j j 

exactRate = 1 − repcur / rep . glo 

The experimental results are shown in Fig.6. Each 

value is the average of exact rates in each round. With the 

increase of rounds, about 500 rounds later, the exact rate 

goes to stable. The exact rates of considering collusion 

and no considering collusion are about 92% and 72%, 

respectively. Above result illustrates our reputation model 

considering collusive behavior can eliminate the influence 

of collusion effectively.


Figure 6. Exact Rate of Reputation of Considering and no Considering 

collusion 

(3) Only considering QoS 

In order to differentiate reputation-aware selection 

from QoS-aware selection, we set the experiment of the 

QoS-aware service selection only relying on the 

evaluated QoS. Table V is hold good. The threshold of 

QoS is set to 0.6, that is, if the evaluated QoS is 0.6, the 

transaction succeeds. The colluding consumers are filtered 

out using our CBA algorithm and total QoS value but not 

global reputation is computed. Total QoS value is the 

average of received QoS. The service selection is based on 

the ranking QoS, obviously, the service with highest QoS 

will be selected. 

Figure 7. Service Selection only based on QoS 

The result is shown as Fig.7. Compared with Fig.3, the 

rate of success is lower than 15% in general. That 

illustrates service selection of reputation-aware is superior 

to QoS-aware'. 

VI. CONCLUSION 

This paper introduces a reputation model considering 

collusive consumers. We get the current reputation by 

utilizing the similarity between advertised QoS from 

service providers and delivered QoS from consumer’s 

evaluation, then update the global reputation and save 

them into reputation center and service providers. At the 

same time, in order to prevent providers tampering with 

reputation, we use the digital signature. Experimental 

results show that the success rate of transaction 


considering collusion is higher than no considering 

collusion. 

In order to find collusive consumers, we make use of 

k-means cluster algorithm to classify the consumers, and 

use association rule algorithm to mine collusive 

consumers, then adjust the service reputation through 

eliminating the collusive consumers gradually. In fact, 

collusion exists not only in consumers, but also exists 

between consumers and providers. Because of profits, 

they have enthusiasm to make collusion. How to reduce 

the second collusion is another challenge. 


This research was partially supported by Anhui 

Educational Department Natural Sciences Project 

(KJ2010A251). 

REFERENCES 

[1] Liangzhao Zeng, Hui Lei, and Henry Chang. "Monitoring 

the QoS for Web Services," Proceedings of the 5th 

international conference on Service-Oriented 

Computing(ICSOC 2007). 2007, LNCS 4749, pp.132-144. 

[2] Liu Yue, Weiyi Liu, Xiaoling Wang,Jin Li. "An Approach 

for Measuring Quality of Web Services Based on the 

Superposition of Uncertain Factors," Journal of Computer 

Research and Develeopment. vol.46, no.5, 2009, pp.841- 

849. 

[3] Liu Y, Ngu A, Zengl Z. "QoS computation and policing in 

dynamic Web service selection," Proceedings of t he 13th 

International World Wide Web Conference. New York , 

USA : ACM Press , 2004, pp.66-73. 

[4] Yan Li,Minghui Zhou, Ruichao Li, Donggang Cao,Hong 

Mei. "Service Selection Approach Considering the 

Trustworthiness of QoS Data," Journal of 

Softwate.Vol.19,no.10, 2008, pp. 2620-2627. 

[5] R. Jurca, B. Faltings, and W. Binder. "Reliable QoS 

monitoring based on client feedback," Proceedings of the 

16th International World Wide Web Conference 

(WWW07), 2007, pp. 1003-1011. 

[6] T. Sobh and K. Elleithy. "Service Selection Should be 

Trust and Reputation-Based," Advances in Systems, 

Computing Sciences and Software Engineering, 2006, 

pp.359-364. 

[7] Yao Wang, Julita Vassileva ."Toward Trust and Reputation 

Based Web Service Selection: A Survey," In International 

Transactions on Systems Science and Applications, 2007, 

pp.118-132. 

[8] Ali Shaikh Ali, Shalil Majithia, Omer F. Rana and et al. 

"Reputation-based semantic service discovery," 

Concurrency and Computation: Practice & Experience - 

First International Workshop on Emerging Technologies 

for Next-generation GRID (ETNGRID 2004) . vol.18, 

issue 8, 2006, pp.817-826. 

[9] Zaki Malik, Athman Bouguettaya. "RATEWeb: Reputation 

Assessment for Trust Establishment among Web Services," 

The VLDB Journal. vol.18,no.4, 2009, pp.885-911,doi: 

10.1007/s00778-009-0138-1. 

[10] E. M. Maximilien and M. P. Singh. "Multiagent System for 

Dynamic Web Services Selection," Proceeding of first 

Workshop on Service-Oriented Computing and Agent- 

Based Engineering (SOCABE at AAMAS), 2005, pp.25- 

29. 

[11] Ping Wang, Kuo-Ming Chao, Chi-Chun Lo, Ray Farmer, 

Pu-Tsun Kuo. "A Reputation-based Service Selection


Scheme," 2009 IEEE International Conference on e- 

Business Engineering. IEEE Computer Society,2009, 

pp.501-506. 

[12] Keting Yin, Bo Zhou, Shuai Zhang, Yixi Chen,et al. "An 

effective approach to select trustable web services," 

WiCOM '08. 4th International Conference on Wireless 

Communications, Networking and Mobile Computing, 

2008, pp.1-6. 

[13] A.S. Ali, S.A. Ludwig, and O.F. Rana. "A cognitive trustbased 

approach for web service discovery and selection," 

Proceeding of the third European Conferenc. on Web 

Services,2005, pp. 38-40. 

[14] Wanita Sherchan, Seng W. Loke and Shonali 

Krishnaswamy. "Explanation-aware service selection: 

rationale and reputation.Service," Oriented Computing and 

Applications, vol.2, no.4, 2008,pp.203–218,doi: 

10.1007/s11761-008-0032-5. 

[15] Babak Khosravifar, Jamal Bentahar, Philippe Thiran, 

Ahmad Moazin, and Adrien Guiot."An Approach to 

Incentive-based Reputation for Communities of Web 

Services," 2009 IEEE International Conference on Web 

Services, 2009, pp.303-310. 

[16] Beng Chin Ooi, Chu Yee Liau and Kian-Lee Tan. 

"Managing trust in peer-to-peer systems using reputationbased 

techniques," LNCS 2003, Volume 2762/2003, pp.2- 

12, doi: 10.1007/978-3-540-45160-0_2. 


[17] MacQueen, J.B. "Some methods for classification and 

analysis of multivariate observations," Proceedings of 5th 

Berkeley Symposium on Mathematical Statistics and 

Probability, University of California Press, Berkeley, 1967, 

pp. 281-297. 

Shenghui Zhao Ph.D candidate of Southeast University, 

associate Professor of Chuzhou University. Her major research 

fields include network security, Web Services and distributed 

computing. 

Guoxin Wu Professor of Southeast University, doctoral 

supervisor. His major research fields include trust network, 

distributed computing and network security. 

Guilin Chen Professor of Chuzhou University. His research 

interests include distributed computing, pervasive computing 

and virtualization. 

Haibao Chen Teacher of Chuzhou University. His major 

research fields include trust and Cloud computing.


Cost Aggregation Strategy with Bilateral Filter 

Based on Multi-scale Nonlinear Structure Tensor 

Li Li 

Shandong Provincial Key Laboratory of Digital Media Technology, Shandong Economic University, Jinan, China 

School of Computer Science and Technology, Shandong University, Jinan, China 

Email :lily_jn @ sina.com 

Hua Yan 

Shandong Provincial Key Laboratory of Digital Media Technology, Shandong Economic University, Jinan, China 

Email: yhzhjg @ sdu.edu.cn 

Abstract—This paper proposed a novel cost aggregation 

method for stereo matching with modified bilateral filter. In 

original bilateral filter, only spatial and range weights are 

used to compute the similarity of two considering pixels and 

a new weight based on structure tensor is added in our 

method. By smoothing each element of the structure tensor 

considering both the spatial and gradient distances of 

neighboring pixels, the nonlinear structure tensor for each 

pixel is constructed. We adopt the Log-Euclidean calculus as 

tensor dissimilarity function to compute the structure tensor 

distance of two considering pixels. Then the multi-scale 

value is computed by summing of the tensor distances in 

each scale. So a new weight based on multi-scale structure 

tensor distance is set up and included in bilateral filter for 

cost aggregation. By constructing the multi-scale nonlinear 

structure tensor and adding the new corresponding weight 

in cost aggregation, more pixels similar with central pixel 

could be aggregated in a support window and the final 

disparity map could be more accurate. Experimental results 

have confirmed the effectiveness of our proposed method. 

Index Terms—stereo matching, cost aggregation, multi-scale 

nonlinear structure tensor, Log-Euclidean tensor distance, 

bilateral filter 


Stereo matching is a key problem in computer vision. 

According to authors [1] stereo matching algorithms 

usually perform four steps: cost computation, cost 

aggregation, disparity computation or optimization and 

disparity refinement. Cost aggregation is mandatory for 

local stereo matching algorithms to improve signal noise 

rate (SNR) and often adopted by global ones. The cost 

aggregation step is to aggregate initial matching costs in a 

support window. An ideal support window should be 

adjusted according to image content to include only the 

pixels with the same disparity. Many cost aggregation 

methods have been presented while this behavior is far 

from ideal. This paper proposed a novel cost aggregation 

strategy with modified bilateral filter based on multi-scale 

nonlinear structure tensors. 

Many adaptive window methods have been presented 

to include more pixels having the same disparity values 


doi:10.4304/jnw.6.7.958-965 

with the central pixel by varying the size, the shape and 

the position of the support window [2-4]. Different from 

the adaptive window methods, the adaptive weight 

method (AW) [5] adopts a fixed window and assigns a 

weight to each pixel in the support window according to 

the spatial and color similarity with the central pixel and 

gains a high performance. The AW method is based on 

bilateral filter which is a non-iterative feature-preserving 

image smoothing technique. Bilateral filter assigns a 

geometric (spatial filter) and a color proximity (range 

filter) constraint independently and a higher weight is 

assigned to the pixel with both smaller spatial and color 

distances to the central pixel. In AW method, the weight 

of a pixel within the support window is obtained by 

applying two independent bilateral filters in the 

neighborhood of potential correspondence. To further 

improve the disparity accuracy and decrease the 

computational time, many modified approaches against 

the AW method have been presented in recent years. The 

segment-based support method (SS) adds the segment 

information [6] in cost aggregation step. By using 

segment information and removing the spatial weight, the 

SS method can further improve the accuracy of disparity 

maps. But the computational time is almost double of that 

of the AW method. To decrease the execution time, a 

simplified asymmetrical strategy was proposed in [7]. 

The bilateral filter is enforced on the reference image 

only and weights are computed by means of a two pass 

approach. These simplifications yield a real-time 

implementation and worse but reasonable results 

compared with the AW method. A fast bilateral stereo 

method (FBS) combines traditional local approach with a 

symmetric adaptive weight strategy based on two 

independent bilateral filters applied on a regular block 

basis [8]. Disparity maps yielded by the FBS method are, 

in general, less noisy compared with the AW method and 

one can trade accuracy for speed and vice versa by 

modifying the block size. 

Danny Barash pointed out that the nature of bilateral 

filter resembles that of anisotropic diffusion [9]. So 

recently many stereo matching methods based on partial 

differential equations (PDE) have also been presented


[10-12]. Usually these methods have solutions of 

disparity maps by constructing partial differential 

equations which are derived from minimizing energy 

functions and include different diffusion equations 

modeling the smoothness assumption. The diffusion 

equations as regularization terms in PDE model are all 

constructed basing on structure tensors of images. The 

structure tensor can be used as a local geometry indicator 

to analyze the geometric structure of image and widely 

used in image segmentation, corner detection and object 

tracking areas [13-15]. However they have not been used 

in cost aggregation method directly until now. 

To fill the void in existing stereo approaches, this 

paper presented a novel cost aggregation method against 

the AW method. Our approach adds a new weight in the 

original bilateral filter based on structure tensor proximity 

of corresponding pixels. By using structure tensor 

information, the final weight can take into account the 

local geometric structure of image and then can more 

accurately detect the similarity of two pixels. Usually a 

Gaussian kernel is used to smooth each element of the 

structure tensor over a local window to remove noise. 

While the Gaussian kernel is isotropic, some weak 

features of the image will be smoothed out. So we 

construct a nonlinear structure tensor based on bilateral 

filter which assigns a weight to a pixel according to both 

spatial and gradient distances with the given one. The 

neighboring pixels that have shorter spatial and gradient 

distances to the central one should have higher weights in 

averaging process. In this way, the nonlinear structure 

tensor, which is adaptive to the image local structures, 

could be constructed and hence similarity of 

corresponding pixels could be better distinguished. Then 

the Log-Euclidean function is selected to estimate 

structure tensor proximity (that is structure tensor 

distance) of corresponding two pixels for easy 

implementation and efficient computation. To deal with 

the scale difference of structure tensor of the image, we 

compute the multi-scale value by summing of structure 

tensor distances in each scale. So a new weight based on 

the multi-scale structure tensor distance could be set up 

and used in cost aggregation function. After cost 

aggregation the final disparity map could be obtained by 

the winner-takes-all (WTA) strategy without any post 

processing step. The flowchart of the algorithm is shown 

in Fig. 1. The experimental results on Middlebury test set 

indicate that the performance of our proposed method is 

competitive with the other state-of-the-art cost 

aggregation strategies. 

The rest of paper is organized as follows. Section II 

briefly introduces the adaptive weight method. Section III 

presents our proposed cost aggregation algorithm in detail. 

Experimental results are given in section IV. Section V 

gives conclusions and an outlook to possible future work. 

II. THE ADAPTIVE WEIGHT METHOD 

In this section, we briefly introduce the adaptive 

weight method which bases on the original bilateral filter 

and will be compared with our method in experimental 

test section. 


Figure 1. The flowchart of the algorithm 

In bilateral filter, given an image I , the value of a pixel 

p in the filtered image I ~ is a weighted average value 

described as follows 

~ 

I ( p) 

= 

∑ 

C 

q∈S 

( p) 

W ( I( 

p) 

− I( 

q)) 

W ( p − q) 

I( 

q) 

∑ 

C 

q∈S 

( p) 

S 

W ( I( 

p) 

− I( 

q)) 

WS 

( p − q) 

. (1) 

where S( p) 

is a support window centered in pixel p, WS 

and W are weighting functions related to spatial distance 

C 

and color distance between p and q respectively. The 

higher weight should be assigned to the pixels with both 

smaller spatial and color distances to the central pixel. 

The adaptive weight method uses the two independent 

bilateral filters to execute cost aggregation. Given a pixel 

p in the reference image and the potential 

l 

I L 

corresponding pixel p in the matching image with 

r 

I R 

disparity d , the aggregated cost 

~ 

C( pl 

, pr 

, d) 

is computed 

as follows 

~ 

C( 

p , p , d) 

= 

l 

r 

∑ 

W ( D ( p , q )) W ( D ( p , q )) W ( D ( p , q )) W ( D ( p , q )) C( 

q , q , d) 

C 

ql∈S 

( pl 

) 

qr 

∈S 

( pr 

) 

∑ 

C 

W ( D ( p , q )) W ( D ( p , q )) W ( D ( p , q )) W ( D ( p , q )) 

C 

qr 

∈S 

( pr 

) 

ql∈S 

( pl 

) 

l 

C 

l 

l 

S 

l 

S 

S 

l 

S 

l 

l 

C 

l 

C 

C 

r 

C 

r 

r 

S 

r 

S 

S 

r 

S 

r 

r 

r 

l 

r 

. (2) 

where the initial matching cost C( q is single pixel 

l , qr 

, d) 

truncated absolute differences (TAD) score between 

corresponding pixels q and assuming the disparity 

l 

r 

value is d , the spatial distance and range distance 

both are Euclidean, the weighting functions and 

q 

DS 

W 

DC S


W both are Gaussian. The adaptive weight method 

C 

provides excellent results in a WTA framework which are 

comparable to some global methods without any complex 

reasoning. To more accurately distinguish the similarity 

of neighboring pixels, in our approach a new weight 

based on structure tensor proximity is added in cost 

aggregation function against the AW method considering 

the high performance of structure tensor to detect the 

local geometric features of the image. 

A. Initial Matching Cost 

III. OUR PROPOSED METHOD 

Given a reference image I and a matching image L 

I , R 

the matching cost indicates similarity of the two images 

with a disparity map. Many stereo matching algorithms 

adopt a criteria based on the color difference between two 

corresponding pixels. To improve the robustness to noise 

or distortions, we adopt a truncated L1 norm as initial 

matching cost function given by 

C ∇ 

2 

2 

( pl 

, pr 

, d) 

= || I L ( pl 

) − I R ( pr 

) || + λM 

|| ∇I 

L ( pl 

) − I R ( pr 

C 0 ( p l , p r , d ) = − log[ δ M + ( 1− 

δ M ) exp( −C( 

p l , p r , d ) / σ M )] 

. (3) 

where p , are corresponding two pixels 

l ( x, 

y) 

pr ( x − d, 

y) 

in the reference image I and the matching image 

L 

I R 

T 

respectively assuming the disparity is d , ∇ = ( ∂x 

∂y) 

is 

the gradient operator, λ , M δ and M σ are predefined 

M 

parameters and C is the initial matching cost 

0 ( pl 

, pr 

, d) 

which will be used in the next cost aggregation step. 

B. Nonlinear Structure Tensor 

In the cost aggregation step, the initial matching cost 

will be aggregated using the bilateral filter process 

expressed by (2). A pixel in the support window will be 

assigned a weight based on the spatial and color distances 

with the central pixel as described before. To more 

accurately detect the similarity of considering two pixels, 

a new weight is added into (2) based on the proximity of 

structure tensors. Firstly the nonlinear structure tensor of 

the reference image is computed. 

The classic differential geometry theory [16] provides 

a method to analyze the local geometric structure of an 

image. Let us consider a multi-valued reference image: 

n 

2 

I ( p ) : Ω → R defined on a domain Ω ∈ R where 

L 

l 

+ 

n ∈ N is the number of the image channel, p is a 

l ( x, 

y) 

pixel in the domain. The local variations of the vector 

norm || dI || can be given by 

L 

|| dI 

L 

2 

|| 

) || 

n 

T 

T ⎛dx⎞ 

= dX GdX G = ∑ ∇I 

i∇I 

i dΧ 

= ⎜ ⎟ 

i= 

1 

⎝dy⎠ 

. (4) 

where G is symmetric as well as semi-positive-definite 

and its coefficients are 

g 

11 

= 


n 

n 

n 

2 

2 

∑ I i g = = ∑ = 

x 12 g 21 I i I x i g y 22 ∑ I iy 

i= 

1 

i= 

1 

i= 

1 

2 

. (5) 

We call G a structure tensor because it indicates the 

local geometry of the image. In fact, the eigenvalues λ , + 

2 

λ of G are the maximum and minimum of 

− 

|| dI L || 

while the orthogonal eigenvectors θ , + θ of G are 

− 

corresponding variation directions. Structure tensor has 

been used in many image applications to present the local 

geometric feature of the image. However, the computing 

derivatives is sensitive to noise, it needs to smooth the 

derivatives for noise reduction. Usually an isotropic 

Gaussian kernel is used to smooth each of the four 

elements in the 2× 2 structure tensor in a local window. 

As we all know, such a smoothing operation will smooth 

out some weak features and the results will not be 

accurate. So we construct a nonlinear structure tensor 

using bilateral filter same as [14]. During the smoothing 

the structure tensor, we consider both the spatial distance 

and gradient distance in the averaging weight assignment. 

Here the gradient distance for a neighboring pixel 

ql ∈ S( 

pl 

) of the central pixel p is given by 

l 

D − 

2 

G ( pl 

, ql 

) = ( I x ( pl 

) − I x ( ql 

)) + ( I y ( pl 

) I y ( ql 

)) 

2 

. (6) 

where I , and , are the first 

x ( pl 

) I y ( pl 

) I x ( ql 

) I y ( ql 

) 

order partial derivatives of the reference image I along 

L 

the horizontal and vertical directions at pixel p and l ql 

respectively. For the multi-valued image, we first 

transformed the image into the intensity image before 

computing the gradient distance. The spatial distance is 

still Euclidean expressed by 

D − 

2 

S ( pl 

, ql 

) = ( x p − x ) ( 

l q + y 

l p y l ql 

) 

2 

. (7) 

Then by considering both the spatial and gradient 

distances, we define a bilateral weighting function of 

smoothing the structure tensor for each pixel p in the 

l 

reference image I as follows 

gˆ 

i, 

j 

( p ) = 

l 

L 

∑ 

G 

ql∈S 

( pl 

) 

W ( D ( p , q )) W ( D ( p , q )) g ( q ) 

∑ 

G 

G 

ql∈S 

( pl 

) 

l 

G 

l 

l 

l 

S 

S 

S 

l 

S 

l 

l 

i, 

j 

W ( D ( p , q )) W ( D ( p , q )) 

where is the element of structure tensor for each 

g i, 

j 

l 

l 

.(8) 

neighboring pixel of the central pixel p in the support 

l 

window, the range of indices is , j = 1, 

2 , ˆ is the 

i g i, 

j 

filtered element value of structure tensor for the central 

pixel p , l WG = exp( −DG 

/ 2σ 

G ) and WS = exp( −DS 

/ 2σ 

S ) 

both are Gaussian functions based on the spatial distance 

D and the gradient distance respectively. Compared 

G 

DS 

with the Gaussian kernel, the weighting function is 

anisotropic and adaptive to the image local structure by 

using bilateral filter. 

C. Multi-scale Nonlinear Structure Tensor Distance 

Multi-scale structure tensor was firstly defined in [17] 

and named as multi-scale fundamental forms. It has been 

widely used in multi-valued images fusion or merging,


noise filtering and segmentation etc. By considering the 

scale differences of the multi-valued image, we construct 

a multi-scale structure tensor by using tensor information 

at each scale. The images at different scales can be 

obtained by smoothing the original image I with a 

L 

series of Gaussian kernels k with different standard 

deviations ξ . By increasing the values of ξ , a fine to 

coarse scale space can be formed. Then a multi-scale 

structure tensor can be expressed by 

G 

m 

= 

n 

∑ 

i= 

1 

ξ 

∇( 

I ∗ Fξ 

) ∇( 

I ∗ Fξ 

) 

L 

m 

i 

L 

T 

m i 

. (9) 

where m ∈ L is the current scale level, L is the total 

number of scales, ∗ is the convolution operator and F 

is the Gaussian function with the deviation ξ . Finally 

m 

the computed multi-scale structure tensor is bilateral 

filtered for different scales as described before. We 

should notice that each scale is nonlinearly filtered 

separately and each scale has three different channels. 

One key factor in the tensor space analysis is a proper 

choice of the tensor distance norm to measure the 

similarity between tensors. For simplicity, we use the 

recently proposed Log-Euclidean calculus [18] as the 

measure function of tensor similarities. The Log- 

Euclidean distance between two tensors and G for 

pixels and q respectively is given by 

pl l 

G pl 

l q 

ξm 

ˆ 

ˆ 2 

DT ( pl 

, ql 

) = tr((log 

m( 

G p ) − log m( 

G )) ) 

l 

ql 

. (10) 

where tr (.) is the trace operator of a matrix, log m(.) 

is 

the logarithm operator of a matrix, the “hat” denotes that 

the structure tensor has been bilateral filtered as described 

before. The logarithm of a matrix is computed by 

decomposing the matrix to its eigenvalues and 

eigenvectors, taking the logarithms of the eigenvalues, 

and constructing a matrix from the eigenvectors and the 

newly computed eigenvalues. Thus the matrix logarithm 

operation is particularly easy to compute when structure 

tensor G is given in terms of its eigenvalues and 

eigenvectors. 

ˆ 

The tensor distance for multi-scale structure tensor can 

be defined as the square root of the sum of the Log- 

Euclidean distances for all scales and can be rewritten as 

∑ − L 1 

~ 

= 

ˆ m 

2 

) − log ( ˆ m 

DT 

( pl 

, ql 

) tr((log 

m( 

G p m Gq 

)) ) 

l 

l 

m= 

0 

. (11) 

~ 

Then the new weight WT = exp( −DT 

/ 2σ 

T ) between 

two considering pixels p and is computed which is 

l q l 

based on the similarity of their tensors. 

D. Cost Aggregation and Disparity Computation 

By the above procedures, the final cost aggregation 

equation can be expressed as follows 


~ 

C( 

p , p , d) 

= 

l 

r 

∑ 

~ 

W ( D ( p , q )) W ( D ( p , q )) W ( D ( p , q )) C ( q , q , d) 

C 

ql∈S 

( pl 

) 

qr 

∈S 

( pr 

) 

C 

∑ 

~ 

W ( D ( p , q )) W ( D ( p , q )) W ( D ( p , q )) 

C 

ql∈S 

( pl 

) 

qr 

∈S 

( pr 

) 

l 

C 

l 

l 

S 

l 

S 

S 

l 

S 

l 

l 

T 

l 

T 

T 

l 

T 

l 

l 

0 

l 

l 

r 

. (12) 

where W and are spatial weight and range weight 

S WC 

respectively same as the expression in (2) used by the 

AW method. Compared with (2), a new weight is 

included which reflects the similarity of two considering 

pixels based on their structure tensors in our approach. A 

neighboring pixel is assigned a high weight to the central 

pixel not only if their spatial and range distances are 

small but also if they have similar local geometric 

structures. The final weight could more accurately 

distinguish the similarity of two relevant pixels. It is 

worth noting that, in the above equation, we simply 

execute the filter on the reference image only different 

from the adaptive weight method to decrease the 

execution time. 

Finally the disparity map is obtained in a WTA 

framework and expressed as below 

~ 

d ( pl 

) = arg min( C( 

pl 

, pr 

, d )) 

d∈D 

. (13) 

where D represents the set of all allowed disparities. 

IV. EXPERIMENTAL RESULTS 

In this section, we aim at assessing the performance of 

our proposed cost aggregation approach based on the 

modified bilateral filter with multi-scale nonlinear 

structure tensor. We used the Middlebury test bed 

provided by authors of [1] to evaluate our approach 

performance compared with the cost aggregation method 

based on the original bilateral filter and the other state-of- 

the-art methods. 

Firstly we compared the results of our method with 

that of the method using the original bilateral filter (OBF). 

Both methods adopt asymmetric strategy which executes 

filter on the reference image only for simplicity and is 

different from that of the adaptive weight method. For 

comparison our method and the OBF method both select 

the L1 norm as initial cost function which is also different 

from the AW method. A constant set of parameters are 

run for all test images. The initial cost function 

7 

parameters are λ M = 2 , 10 and − 

δ M = σ M = 2 , the 

bilateral structure tensor parameters are σ G = 20 , 

σ S = 2. 

5 and the window size = 5× 5 . For multi-scale 

structure tensor computation, the total number of scales 

L is 3, the Gaussian kernelξ 0 = 0. 

6 , 1 1. 

0 = ξ and 2 1. 

4 = ξ . 

The cost aggregation parameters are σ C = 15 , σ S = 10. 

5 , 

σ T = 5 and the window size = 21× 21. 

The corresponding 

disparity maps are plotted in Fig. 2. From the figures, we 

can see that our proposed method has better results than 

the OBF method using the original bilateral filter. This 

manifests that including new weight in cost aggregation 

function can actually improve the disparity accuracy,


Figure 2. The disparity maps of two methods. The top is input reference images of “Tsukuba”, “Venus”, “Teddy” and “Cones” images. The second 

row is the truth disparity maps. The third row is the results of the OBF method. The last row is our results. 

especially in discontinuity area, because of structure 

tensor reflecting the local geometric feature of the image. 

On the other hand, the nonlinear filter and the 

construction multi-scale values of structure tensors can 

further remove the noise and improve the accuracy of the 

discontinuity location fitting for real images which 

usually have different light sources or distortions. To 

manifest this, we have the same comparison experiments 

on other two datasets which are available at Middlebury 

test bed. Each dataset has 9 different images that exhibit 

3 different exposure and 3 different lighting variations. 

Fig. 3 shows the both exposure and lighting variations of 

Figure 3. The left image of the Art dataset with three different 

exposures and under three different light conditions 


left image of the Art dataset. Quantitative comparative 

results are given in Table I. The experimental parameters 

are all the same with those in first the experiment. The 

focus here is evaluation of the raw cost aggregation 

methods which don’t deal explicitly with occlusions and 

TABLE I. 

QUANTITATIVE COMPARATIVE RESULTS OF OUR METHOD WITH THE 

OBF METHOD FOR REAL TEST IMAGES. 

Art 

Dataset 

Book 

Dataset 

OBF method Our method 

Vis. Dis. Vis. Dis. 

Art1-1 15.03 23.59 14.57 21.47 

Art1-2 12.39 19.13 12.40 17.08 

Art1-3 14.26 22.01 14.08 19.50 

Art2-1 13.91 21.74 13.48 19.54 

Art2-2 11.07 18.16 10.74 16.05 

Art2-3 15.88 22.48 15.62 20.15 

Art3-1 13.24 18.06 12.88 15.48 

Art3-2 13.67 16.66 13.12 14.58 

Art3-3 26.80 24.65 26.61 21.40 

Book1-1 12.57 25.39 12.29 24.63 

Book1-2 13.21 25.47 13.13 25.17 

Book1-3 23.02 36.26 23.45 36.11 

Book2-1 13.85 28.78 13.68 28.55 

Book2-2 15.19 27.55 15.35 27.30 

Book2-3 14.46 28.31 14.65 28.33 

Book3-1 12.39 25.26 12.01 24.54 

Book3-2 12.61 25.26 12.61 24.70 

Book3-3 17.81 33.36 17.95 32.75


we only report the percentage of bad pixels (i.e. pixel 

whose absolute disparity error is greater than 1) in nonoccluded 

region (Vis.) and in near depth discontinuity 

( Dis.) described by [1]. From the table we can see that 

our proposed method can outperform the OBF method 

for real images having different exposure and light 

source changes due to including multi-scale nonlinear 

structure tensor information. Our results have been 

plotted in Fig. 4. 

In the third experiment, to evaluate the effectiveness 

of the multi-scale nonlinear structure tensor, our method 

is compared with the method using simple structure 

tensor information for clean and real test images. The 

same parameters are adopted by both methods for 

fairness. The quantitative comparative results are given 

in Table II. From the table, we find that our method 

using the multi-scale nonlinear structure tensor is more 

effective than the method using simple structure tensor 

for real images while not for clean images, mainly 

because real images have more noise or distortions and 

multi-scale nonlinear structure tensor can be used to 

effectively remove noise which is approved by the last 

experiment. We should notice that the computational 

time mainly depends on the bilateral filter process not 

multi-scale nonlinear structure tensor computation, so 

execution time of our approach increases a little when 

compared with the method using simple structure tensor. 

Lastly we compared our proposed method with the 

state-of-the-art cost aggregation strategies. For our 

method, we adopt the optimal parameters minimizing the 

Vis.+Dis. error on the whole test images: window 

size = 31× 31 , σ S = 15. 

5 and the other parameters are 

same as those in the first experiment. The results of the 

other four top methods used here were reported in [7]. It 

is worth noting that these results reported by [7] are 

obtained using the original cost function proposed by the 

authors of each paper and the results for AW and SS 

available on the Middlebury evaluation sites including 

the post processing steps are not used. We have reported 

quantitative comparative results in Table III. From the 

table, we can see that our proposed method have results 

comparable to the best performing cost aggregation 

strategies. The most similar method with our approach is 

the AW method based on the original bilateral filter. The 

AW method outperforming our results is mainly due to 

its symmetric strategy while our approach adopts the 

asymmetric strategy. However our method runs faster 

than the AW method and the AW run is 3226 seconds 

according to [7] while our method takes 210 seconds 

without any accelerating techniques on the Teddy image. 

V. CONCLUSIONS 

In this paper, we proposed a novel cost aggregation 

algorithm based on modified bilateral filter with multiscale 

nonlinear structure tensor for local stereo matching. 

A new weight is included in cost aggregation equation 

based on the structure tensor distance. For reducing noise 

influence, in structure tensor computation, each element 

of the structure tensor is smoothed using a bilateral filter 

and the nonlinear structure tensor is constructed. Then a 

multi-scale tensor is computed at each scale for 

considering the scale difference of the multi-valued 

image. The nonlinear structure tensor can be used in the 

Log-Euclidean measure function as tensor similarities. 

Figure 4. The disparity maps of our method for the Art and Book datasets. The first image of the two datasets is truth disparity maps respectively for 

the Art and Book datasets. 

© 2011 ACADEMY PUBLISHER


TABLE II. 

DISPARITY MAPS OF OUR METHOD AND THE METHOD USING SIMPLE STRUCTURE TENSOR. 

The method using simple 

structure tensor 

Our method 

Vis. Dis. Vis. Dis. 

Tsukuba 3.01 9.76 3.11 10.44 

Venus 6.97 12.49 6.68 13.15 

Teddy 10.86 19.67 10.86 19.70 

Cones 6.45 12.76 6.41 12.77 

Art1-1 14.57 21.74 14.57 21.47 

Art1-2 12.25 17.82 12.40 17.08 

Art1-3 13.77 20.12 14.08 19.50 

Art 

Dataset 

Art2-1 

Art2-2 

Art2-3 

13.42 

10.74 

15.57 

19.98 

16.72 

20.70 

13.48 

10.74 

15.62 

19.54 

16.05 

20.15 

Art3-1 12.92 16.09 12.88 15.48 

Art3-2 13.20 15.08 13.12 14.58 

Art3-3 26.22 22.14 26.61 21.40 

Book1-1 12.48 24.76 12.29 24.63 

Book1-2 13.17 25.08 13.13 25.17 

Book1-3 24.01 36.25 23.45 36.11 

Book 

Dataset 

Book2-1 

Book2-2 

Book2-3 

13.97 

15.52 

15.00 

28.51 

27.56 

28.77 

13.68 

15.35 

14.65 

28.55 

27.30 

28.33 

Book3-1 12.05 24.41 12.01 24.54 

Book3-2 12.80 24.91 12.61 24.70 

Book3-3 18.44 32.86 17.95 32.75 

TABLE III. 

QUANTITATIVE COMPARATIVE RESULTS OF OUR METHOD WITH THE OTHER FOUR TOP ALGORITHMS. 

Tsukuba Venus Teddy Cones 

Vis. Dis. Vis. Dis. Vis. Dis. Vis. Dis. 

SS[4] 2.19 7.22 1.38 6.27 10.50 21.20 5.83 11.80 

AW[3] 3.33 8.87 2.02 9.32 10.52 20.84 3.72 9.37 

FBS[6] 2.95 8.69 1.29 7.62 10.71 20.82 5.23 11.34 

VW[2] 3.12 12.40 2.42 13.30 17.70 25.50 21.20 27.30 

Our method 2.70 10.72 4.93 11.83 10.86 19.70 6.41 12.77 

Finally the multi-scale tensor distance is set up which is 

the square root of sum of tensor distances at each scale. 

So the new weight can be computed and related to the 

multi-scale tensor distance. The proposed new algorithm 

not only considers the spatial and range distances of two 

pixels same as the original bilateral filter, but also the 

local geometric feature distance of them. Our new weight 

can more accurately reflect the similarity of two pixels 

which is important for cost aggregation approach. The 

experimental results confirm the effectiveness of our 

approach compared with the OBF method using clean 

and real test sets and the other state-of-the-art strategies. 

In the future, we plan to devise new cost aggregation 

methods based on the structure tensor to further improve 

the accuracy of disparity maps and decrease the 

computational time. We are also interested in using the 

diffusion equation constructed by structure tensor in the 

variational framework to devise the global stereo 

matching method. 


The authors would like to thank financial supports 

from National Natural Science Foundation of China 


under Grant Nos. 60970048, Natural Science Foundation 

of Shandong Province Grant Nos. 2009ZRB019SF. 

REFERENCES 

[1] D. Scharstein and R. Szeliski, “A taxonomy and 

evaluation of dense two-frame stereo correspondence 

algorithms”, International Journal of Computer Vision, 

pp.7-42, 2002. 

[2] O. Veksler, “Fast variable window for stereo 

correspondence using integral images”, In Proc. Of Conf. 

on CVPR 2003, pp.556-561, 2003. 

[3] H. Hirschmuller, P. Innocent, and J. Garibaldi, “Real-time 

correlation-based stereo vision with reduced border 

errors”, International Journal of Computer Vision, 47:1–3, 

2002. 

[4] J. Kim, K. Lee, B. Choi, and S. Lee, “A dense stereo 

matching using two-pass dynamic programming with 

generalized ground control points”, In Proc .of Conf. on 

Computer Vision and Pattern Recognition 2005, pp. 

1075–1082, 2005. 

[5] K. J. Yoon and I. S. Kweon, “Adaptive support-weight 

approach for correspondence search”, IEEE TPAMI, 28(4), 

pp.650-656, 2005. 

[6] F. Tombari, S. Mattoccia, and L. Di Stefano, 

“Segmentation-based adaptive support for accurate stereo 

correspondence”, In Proc. of Conf. on PSIVT, 2007.


[7] M. Gong, R.G. Yang, W. Liang, and M.W. Gong, “A 

performance study on different cost aggregation 

approaches used in real-time stereo matching”, 

International Journal Computer Vision, 75(2), pp.283-296, 

2007. 

[8] S. Mattoccia, S. Giardino, and A. Gambini, “Accurate and 

efficient cost aggregation strategy for stereo 

correspondence based on approximated joint bilateral 

filtering”, In Proc. of Conf. on ACCV, 2009. 

[9] Danny Barash, “A Fundamental relationship between 

bilateral filtering, adaptive smoothing and the nonlinear 

diffusion equation”, IEEE TPAMI, Vol. (24), No.6, June, 

2002. 

[10] D. Scharstein and R. Szeliski, “Stereo matching with 

nonlinear diffusion”, International Journal of Computer 

Vision, 28(2), pp.155-174, 1998. 

[11] R. Ben-Ari and N. Sochen, “A geometric approach for 

regularization of the data term in stereo-vision”, 

International Journal of Math Imaging, vol.31, pp.17-33, 

2008. 

[12] R.B. Ari and N. Sochen, “Variational stereo vision with 

sharp discontinuities and occlusion handling”, In Proc. of 

Conf. on ICCV, Rio de Janeiro, Brazil, pp.1-7, 2007. 

[13] S. Han, W. Tao, D. Wang, X.Ch. Tai and X.L. Wu, 

“Image segmentation based on grabcut framework 

integrating multi-scale nonlinear structure tensor”, IEEE 

Transactions on Image Processing, 18(10), pp. 289-302, 

June 2009. 

[14] L. Zhang, L. Zhang and D. Zhang, “A multi-scale bilateral 

structure tensor based corner detector”, In Proc. of Conf. 

on ACCV, 2009. 


[15] M. Donoser, S. Kluckner and H. Bischof, “Object tracking 

by structure tensor analysis”, In Proc. of International 

Conference on Pattern Recognition, 2010. 

[16] M. D. Carmo, “Differential geometry of curves and 

surfaces”, Prentice Hall, 1976. 

[17] J. Weickert, B. Romeny and M. A. Viergever, “Efficient 

and reliable schemes for nonlinear diffusion filtering”, 

IEEE Transactions on Image Processing, vol.7, pp.398- 

410, 1998. 

[18] V. Arsigny, P. Fillard, X. Pennec, and N. Ayache, “Log- 

Euclidean metrics for fast and simple calculus on diffusion 

tensors”, Magnetic Resonance in Medicine, vol. 56(2), pp. 

411-421, 2006. 

Li Li works as an instructor in the School of Computer Science 

and Technology at the Shandong Economic University. She 

received the B.E. degree in Motor Engineering from Shandong 

Technology University in 1998 and M.E. degree from 

Shandong University in 2000, and now is currently a Ph.D. 

student in the School of Computer Science and Technology at 

the Shandong University. Her research interests lie in computer 

vision, especially in 3D reconstruction and stereo matching. 

Hua Yan works as an associate professor in the School of 

Computer Science and Technology of Shandong Economic 

University. She received the B.S. degree in Physics in 1997, 

M.E. degree in Communication and Information System in 

2004 and D Sc Tech degree in Communication and Information 

System in 2007 from Shandong University. Dr. Yan’s research 

interests include image and video processing, multimedia data 

retrieval and super resolutions.


A Collaborative Nonlocal-Means Superresolution 

Algorithm Using Zernike Monments 

Lin Guo 1,2 * , Qinghu Chen 1 

1 School of Electronic and Information, Wuhan University, Wuhan, 430074, China 

2 School of Physics and Electronics, Hubei University, Wuhan, 430062, China 

Email: gunglin@gmail.com, cqh@eis.whu.edu.cn 

Abstract—Super-resolution (SR) with probabilistic motion 

estimation is a successful algorithm to circumvent the 

limitation of motion estimation upon conventional superresolution 

methods. However, the algorithm can’t match 

similar patches with rotation or scale. This paper presents 

an efficient improved algorithm by introducing Zernike 

moments as representation of image invariant features into 

similarity measure. A collaborative strategy is proposed 

combining the moment based proximity and the bilateral 

proximity of nonlocal means (NL-means) algorithm for joint 

determination of weights. For the invariant property of 

Zernike moments, structure-similar pixels with rotation or 

scale can also be matched for computation of weights. 

Furthermore, the collaborative mechanism ensures higher 

accuracy of weights for a better estimation of each pixel in 

SR images. Experimental results indicate the proposed 

method is able to handle general video sequences with 

superior performance in SR reconstruction to the compared 

algorithms. 

Index Terms—super resolution, Zernike moments, 

probabilistic motion estimation, nonlocal means, 

collaborative 


Super-Resolution (SR) technique is the fusion of a 

sequence of low-resolution noisy, blurred images to 

produce a higher resolution image or sequence 

overcoming the inherent resolution limitation of LR 

imaging systems. Since Huang and Tsai first proposed the 

concept of SR in 1984, the SR technique has attracted a 

lot of attention in the image processing community due to 

its wide variety of application in image enhancement, 

medical imaging, high definition televisions and 

computer vision. A great deal of literature about SR can 

be found, and the representatives are referred in [1-4]. 

SR reconstruction is an ill-posed inverse problem. A 

wisdly used model for this problem is described as 

follows: 

y � DHM z � n . (1) 

t t , s s 

Where the measurements yt, t = 1, 2, …, T, are results of 

different motion, noise, blur, and decimation parameters 

* Corresponding author: Lin Guo, gunglin@gmail.com 


doi:10.4304/jnw.6.7.966-973 

from an original high resolution reference image zs. The 

matrix Mt,s indicates the geometric warp of yt relative to 

the high resolution image zs. And H is the blur matrix. 

Both of them are assumed for simplicity to be linear 

space and time invariant. Similarly, D denotes the fixed 

spatial resolution decimation. Gaussian random noise n is 

assumed to be added to the measurements. 

To recover zs from yt, Mt,s and H must be known or can 

be reliably estimated from inputs. Most of the existing SR 

methods are roughly based on an estimation of the motion 

between frames followed by the super-resolution fusion 

of inputs according to these motion vectors. As it is well 

know, motion estimation of sub-pixel precise between 

frames is indispensable and commonly regarded very 

critical for successful SR reconstruction. 

However, it is a challenging task to obtain highly 

accurate motion estimation with an affordable 

computation load. In fact, it’s almost impossible to 

handle actual scenes with complex motion patterns or 

very low quality. Inaccurately registration often leads to 

deteriorated reconstruction results even compared to a 

simple interpolated version. So motion estimation has 

become the bottleneck for the conventional SR methods 

to get excellent performance. 

In order to overcome the above problem, several recent 

articles [4-7] attempted to deliver SR methods avoiding 

explicit motion estimation apart from above conventional 

methods. The algorithm in [5] relies on extending their 

previous steerable kernel regression method to multiframe 

super-resolution. The approach in [6] is based on 

the sparse 3D transform-domain collaborative filtering 

and iterative projection on the observation constrained 

subspace. The method in [7] develops the notion of 

probabilistic motion estimation into the classical SR 

formulation, which is regarded as a generalization of the 

very successful nonlocal means (NL-means) denoising 

method [8] to serve the super-resolution task [4]. The 

main idea of the NL-means is that the pixel is estimated 

as a weighted average of similar pixels in its nonlocal 

neighborhood, and the weights are computed according to 

the similarity between two pixels. It shows simple and 

robust to noise. However, the similarity measure only in 

intensity is crude to ensure the accuracy of weights for no 

any information about the underlying image features are 

considered. For example, structure-similar patches with 

rotation or various scales are unable to be matched. As a 

result, unsuitable weights are calculated and assigned to


pixels, and hence lead to the estimation value to deviate 

from the true one. 

Several recent papers have tried to improve the NLmeans 

algorithm in image denoising. Ref. [9] and [10] 

renders a similar approach by employing affine gray scale 

transformations to find patches at the same or different 

scales. Ref. [11] uses cross-scale (i.e., downsampled) 

neighborhoods in the NL-means filter [12]. Ref. [13] 

introduces SIFT as rigid invariant features to compute the 

similarity between different patches. SIFT features as 

local descriptors are suitable for image retrieval, affine 

registration etc., but they are not for denoising and SR. 

Ref. [14] develops Hu moment as rotationally invariant 

feature into similarity measure strategy. Hu moment is 

the simplest moment, but it’s not efficient in many cases. 

Ref. [15] replaces the geometrical Hu moment in [14] 

with Zernike moment, and shows a competitive 

performance in image denoising application. Zernike 

moment is proved to be superior to geometrical moment 

for better capabilities of invariant feature representation 

because of its orthogonal property. This motivates us to 

introduce Zernike moments as invariant descriptors of 

image shape features into the similarity measure to 

improve the super-resolution results. Here, we need to 

point out that the intuitive approach through independent 

interpolation of each frame followed by the Zernike 

moment based denoising processing is unable to provide 

super-resolution results [4]. 

Our super-resolution algorithm using the Zernike 

moments is based on the SR framework with probabilistic 

motion estimation [7]. Similarity for two pixels is 

computed on two small local patches around them in the 

Zernike moment images. For the invariant property of 

Zernike moment, the algorithm is enabled to match more 

similar patches not only with translation but also with 

rotation or scale. However, the images of super-resolution 

generally may contain complex degradation involving 

downsampling, aliasing as well as noise, especially for 

higher order moments that are more sensitive to the noise. 

So Zernike moments of SR images are usually unreliable 

to be the sole basis for computation of weights. To tackle 

the problem, a collaborative algorithm is designed 

combining the Zernike moment based proximity and 

intensity based bilateral proximity of NL-means 

algorithm for joint determination of weights. 

The algorithm proposed in this paper has the following 

major features. Firstly, besides retaining the advantage of 

avoiding explicit motion estimation, our algorithm 

extends the notion of probabilistic motion estimation in 

[7] to include not only intensity-similar patches with 

translation but also structure-similar patches with rotation 

or scale. Secondly, the collaborative similarity measure 

strategy balances the influence of gray-level based 

proximity and invariant moment based proximity. Then 

weights with higher accuracy are computed for better 

estimation of a pixel. 

The remainder of the paper is as follows. Section II 

presents the super-resolution framework with 

probabilistic motion estimation on which our method is 

based. Section III describes the proposed collaborative 


super-resolution algorithm using Zernike moments in 

details. A simplified numerical algorithm in iterative 

form is given at last in this section. Section IV shows 

experimental results on several general video sequences, 

followed by conclusion and discussion in Section V. 

II. THE SR FRAMEWORK WITH PROBABILISTIC MOTION 

ESTIMATION 

According the observation model (1), the Maximum- 

Likelihood (ML) estimation of high resolution image is 

expressed as 

T 1 

2 

zˆ s � arg min � DHMt , szs � y t . (2) 

2 

2 t�1 

The matrix Mt,s in classical SR methods denotes a oneto-one 

mapping between pixels in the s-th and the t-th 

image. And as such, it introduces sensitivity to errors. 

According to the idea of probabilistic motion estimation 

[7], the one-to-one mapping between pixels in classical 

SR methods is substituted to a probabilistic movement 

domain. That means every estimated pixel in the 

reference image with many possible correspondences in 

all the frames of the sequences (including itself). Each 

pixel inside the domain is assigned a value of weight to 

denote the probability of being correct. The movement 

domain is a spatial and temporal neighborhood centered 

at the estimated pixel in the reference image with radius 

R among all the sequences. For given s and t, the 

displacement between the estimated pixel and every pixel 

inside the domain is written as [dx(n), dy(n)], n = 1, … , 

N, N = (2R+1) 2 . The location relationship is described by 

a matrix Mn with size of S1N1S2N2×S1N1S2N2 and value 

of 1 in one position and 0 for others. S1 and S2 are 

sampling factors respectively in horizontal and vertical 

direction. For the pixel whose displacement is indicated 

by the 1 in Mn, the corresponding weight is denoted by 

Wn;t,s , a diagonal matrix with the same size as Mn. Thus, 

we get the following equation: 

N 

M z � �W 

M z . (3) 

t, s s n; 1t 

, s n s 

n�1 

According to (2) and (3), the probabilistic ML 

estimation of the high resolution image is formulated as 

follows: 

1 

zˆ � arg min DHM z � y , (4) 

s PML 2 

N T 

2 

�� n s t L 

Wn 

; t , s 

n�1 t�1 

L 

where W , with size of N1N2 × N1N2, is the 

n; 1t 

, s 

corresponding weight matrix in low resolution space by 

being downsampled from Wn;t,s. 

Since both H and Mn are space-invariant, they can be 

exchanged in position. Thus, defining x = Hz, the task of 

SR is turned to be a two-step process: first estimation of 

the “blurry” high resolution image x according to (5) and 

the subsequent acquirement of z from x by using existing 

deblurring algorithms.


1 

x DM x y . (5) 

N T 

2 

ˆ s PML � arg min �� n s � t Wn 

; t , s 

2 n�1 t�1 

Minimization of (5) leads to a solution represented in 

pixel-wise as 

( k , l) �N ( i, j) t�1 

n; t, s t 

s i j � T 

xˆ 

( , ) 

T 

� � 

� � 

( k , l ) �N ( i, j) t�1 

W ( k, l, i, j) y ( k, l) 

, 

W ( k, l, i, j) 

n; t, s 

where (i, j) is an arbitrary coordinate on high resolution 

grid. And (k, l)�N(i, j) denotes the (k, l)-th pixel within 

the movement domain for pixel (i, j), but is located on 

low resolution grid. That is, (k, l) s.t. (S1k, S2l)� N(i, j), 

which ensures that the center pixel of the patch is on the 

decimation grid. The weight Wn;t,s is computed based on 

the bilateral proximity strategy according to 

W 

� 2 

� � 

� Ri, jDM n xs � Rk , l y � 

t 

2 

n; t, s ( k, l, i, j) 

exp� 

� 

� 

� 

2 � 

� � � 

g 

� � � 

� 

�� 

�� 

2 2 2 

� ( ( )) ( ( )) ( ) � 

(6) 

f dx n � dy n � s � t , (7) 

where the function f may be an arbitrary monotonically 

non-increasing function, such as Gaussian or box form. 

The parameter σg controls the effect of the gray-level 

difference between two patches. Ri,j is an operator that 

extracts a patch of a fixed size centred at the (i, j)-th pixel 

from an image. The square differences of all the pixels of 

two patches are accumulated. Both gray-level proximity 

and geometric proximity are considered for similarity 

measurement, which helps to enhance the effectiveness of 

the algorithm and robustness to noise. However, the 

similar patches in case of rotation and various scales can 

not be matched in this algorithm because the invariance 

property of a patch is not taken into account. 

III. THE PROPOSED METHOD 

A. Zernike Moment Based Image Representation 

Moment based image feature representation has a very 

wide range of applications in the field of image 

processing and pattern recognition. Zernike moments are 

proved to be superior to other moments in noise 

sensitivity, redundancy and expression efficiency for the 

property of orthogonality and invariance. Zernike 

Moments with orthogonal basis functions can be used to 

represent image features by a set of mutually independent 

descriptors, with a near zero value of information 

redundancy [16]. 

The kernel of Zernike Moments is the set of orthogonal 

Zernike polynomials defined over a unit disk in the polar 

coordinate space. The Zernike basis function for order n 

and repetition m is 

V ( x, y) � V ( r, �) � R ( r)exp( jm� 

) , (8) 

nm nm nm 


where n is a positive integer or zero, and m is an integer 

subject to the following constraints: n- |m| = even and 

|m| ≤ n. In addition, θ and r is, respectively, the phase in 

polar coordinate space and the distance from point (x, y) 

to the origin. And j = � 1 . 

The radial polynomial Rnm is defined as 

( n�| m|) 

/ 2 

s 

( �1) ( n � s)! 

Rnm ( r) � � 

r 

s� 

0 � n� | m | � � n� | m | � 

s! � � s �! � � s �! 

� 2 � � 2 � 

n�2 s 

. (9) 

Given that f is a complex-valued function on the unit 

disk, the Zernike moment for f of order n and repetition m 

is 

n �1 

Z � �� f ( x, y) V ( x, y) dxdy , (10) 

nm 

� 2 2 

x � y �1 

* 

nm 

* 

where V nm is the complex conjugate of V nm . 

When f is a digital image, the Zernike moment means 

the projection of image f(x, y) on above orthogonal bases. 

Then (10) becomes 

n �1 

Z � �� f x y V x y . (11) 

* 

nm ( , ) nm ( , ) 

� x y 

To reckon the Zernike moments for f (x, y), the image 

(or a patch) is first mapped to the unit disk of polar 

coordinates, moving the origin of the unit disc to the 

centre of the image. In this paper, a square-to-circular 

mapping transformation [16] is used so that the 

polynomials Rnm(r) need be computed only once for all 

pixels mapped to the same circle. Furthermore, fast 

computation for Rnm(r) in [17] is adopted to speed up the 

calculation. 

The magnitude of Zernike moment is rotation invariant 

as reflected in the mapping to the unit disc. The scale 

invariance can be achieved through normalizing the 

moments by the zero-order geometric moment [18]. 

In terms of (11), Zernike moments of different orders 

are calculated with varying n; accordingly, for given n 

each moment of order n is computed with varying m. And 

moments of different orders correspond to independent 

characteristics of the image, which constitutes a multilevel 

representation for describing various shape features 

of the image. The magnitudes of these moments can be 

presented as images [15]. Fig. 1 shows an image and its 

Zernike moment images up to the third order. Fig. 1 (a) is 

the image “lena”. Fig. (b)-(g) are the moment images of 

Z00, Z 11, Z 20, Z 22, Z 31 and Z 33, respectively. It can be 

seen that the lowest order moment Z 00 displays the main 

content of the image, the same as the average filtering 

result. And the higher moments deliver more detailed 

shape characteristics, but are also more sensitive to the 

noise. So Zernike moments of only up to third order are 

used in this paper. 

B. Collaborative SR Algorithm 

In this section, Zernike moments are first introduced as 

invariant features into the similarity measure strategy.


Figure1. The image with noise and its Zernike moment images of various orders. 

Figure 2. The original image (the left column) and the comparison of the weight distribution decided by three different methods. 

The Zernike moment based similarity measurement for 

SR reconstruction is proposed in this paper as 

w 

Z 

2 

� 

i, j nZ k ( s ) � k, l Zk ( Yt 

) � 

� � R M x R 

2 � 

k 

� exp �� 2 

� , (12) 

� � Z 

� 

� � 

where k ( s ) x Z and k ( t ) Y Z mean the k-th moment image 

for high resolution image xs and Yt. And Yt is the 

interpolation result of the measurement yt. σZ is the 

controlling parameter similar with σg. They can be 

roughly decided by estimation of noise from inputs. So 

when the weights are calculated according to (12), a 

Zernike moment based SR algorithm can be executed 

through (6). 

However, in the practical SR task, images may 

undergo complex motion and degradation, so their 

Zernike moments are usually not accurate enough to be 

the only basis for computation of weights, especially for 

the higher order moments that are more sensitive to the 

noise. Moreover, very inadequate information is rendered 

for the weak textures in the Zernike moment images, 

while the gray-levels of images contain the rich 

underlying details, and they are also more loyal to the 

original images. Thus, a collaborative algorithm is 

developed to combine Zernike moment features and the 

gray-level based proximity in (7) into the computation of 

weights. 


The computation formula of the final weight for a 

searching pixel is designed as 

1 1 

w � wB � wZ � ( wB � wZ 

) , (13) 

2 4 

where B w is an abbreviation for Wn;t,s in (7). B w and Z w 

is calculated, respectively, according to (7) and (12). In 

the collaborative algorithm, the final weight is jointly 

determined by the moment images and bilateral proximity 

of gray-levels, which leads to a more accuracy calculation 

of weights in our method. This can be seen in Fig.2, 

where a comparison of the weight distribution with 

different algorithm is given. The 1st column in Fig. 2 is 

the original image without noise, the 2nd to the 4th 

columns are the weights distribution of the NLM 

algorithm, Zernike moment based algorithm and the 

collaborative algorithm. Differences between the right 

three columns in Fig. 2 show that the NLM algorithm 

strictly matches similar pixels only with translation. The 

moment based algorithm finds more pixels with similarity 

both in translation and rotation, but little difference is 

reflected for pixels with different similarity. The 

collaborative algorithm also can match similar pixels both 

with translation and rotation, but greater weights are 

assigned to pixels that are more similar. So the weights 

are more precise in comparison in the collaborative 

algorithm.


In addition, in order to improve the super-resolution 

algorithm, the following points are considered. Firstly, in 

our method whenever a pixel is SR estimated, the old 

value is replaced for the new one, which is closer to the 

true value than the old value. Hence, more accurate 

information is provided for computation of weights. That 

practically helps not only acquire more precise weight but 

also speed up the SR process. Secondly, in order to 

reinforce the reliability of Zernike moments in presence 

of noise, Yt in (12) is processed by a NLM denoising 

before computation of moments. Finally, the image of 

moment Z00 viewed just as an average version of an 

image. Since gray-level information has been combined 

into the collaborative mechanism, the moment Z00 is not 

necessary any more. Hence, we set Z = {Zk | k = 1, 2, 3, 4, 

5} = {Z11, Z20, Z22, Z31, Z33} in our experiments to 

decrease the computation. 

Summarily, the proposed collaborative SR 

reconstruction for video sequences can be expressed in a 

simplified numerical algorithm. The iterative form of this 

numerical algorithm is represented as: 

T n 

1 

( , ) ( , ) 1 s, t y 

n� 

k l �N i j t� 

t 

s ( , ) � T n 

� ( k , l ) �N ( i, j) �t 

�1 

s, t 

X i j 

� � 

n n 

2 

R 

1 i, j X s � RS1k 

, S2l X 

n� 

t 

2 

B; s, t ( , , , ) exp 

2 

� g 

w k l i j 

w ( i, j, k, l) ( k, l) 

, (14) 

w ( i, j, k, l) 

� � 

� � 

� �� 

� 

� 

� 

� 

| i � S k | � | j � S l | 

1 2 

exp{ � 

} 

max(| i � S k | � | j � S l |) 

1 2 

, (15) 

5 � n n 

2 � 

� � Ri , jZ k (V s ) � RS1k 

, S (V ) 

2lZ 

k t 

2 

n�1 � 

� 

k �1 

� 

Z; s, t ( , , , ) � exp �� 2 

� 

� Z 

w k l i j 

n 

where s, t 

� � 

�� 

,(16) 

w in (14) is computed according to (13). 

n T 

Especially, when n = 0, { X s } s� 

1 in (15) are obtained 

by the bilinear interpolation of { } 1 

T 

n T 

y s s� . And { Vt } t � 1 in 

n T 

(16) are the denoised results of { X t } t � 1 by NLM 

n n 

algorithm [8]. Otherwise, when n > 0, V t = X t . Both 

n 

n 

V t and X t are updated after each iteration. 

IV. EXPERIMENTS 

In this section, the performance of the proposed 

algorithm is validated. The obtained results of processing 

several real video sequences with a general motion 

pattern are presented. The comparison is provided with 

several methods: the bilinear and bicubic interpolation of 

single image as well as the state-of-the-art SR algorithm 

[4]. The results are evaluated from both the visual effects 

and the objective quality measure (PSNR = 


10log10( 

255 

2 

N 

Xˆ � X 

2 

2 

) dB, where N is the number of pixels 

in the true image ˆ X or the constructed image X ). 

All the tests in this section were prepared in the 

following degradation: The input sequences were blurred 

using a 3×3 uniform kernel, down-sampled with a factor 

of 1:3 in each axis, and then added by additive white 

zero-mean Gaussian noise with std = 3. All images were 

in the input range [0,255]. In processing all the sequences, 

all 30 frames took part in the iterative reconstruction of 

each image. 

First, sequences “Miss America”, “Trevor” and 

“Foreman” are tested for evaluation of PSNR. Table 1 

gives the average PSNR for each of the three test 

sequences, where two iterations were run for our method 

and the compared method in [4] with no additional 

deblurring followed. Fig. 3 illustrates the PSNR values 

frame by frame for every test sequence. 

In the experiments, the parameter σg and σZ was set 

manually to 2.5 and 2.4 for all the test sequences. The 

size of the patch used for calculating the weights wB and 

wZ was equally selected as 7 × 7 pixels (high resolution 

grid) for all sequences. The searching range for 

movement domain is 7 × 7 pixels (high resolution grid) 

for sequence “Miss America” and “Trevor”, and 19 ×19 

pixels (high resolution grid) for sequence “Foreman”, 

which has greater displacements between frames. 

The table shows that both the method in [4] and our 

method can handle sequences of arbitrary motion patterns 

and achieve effects of the state-of-the-art compared to the 

single image interpolation. And the proposed algorithm 

yields superior performance to all compared methods in 

PSNR. 

Then, an example on sequence “Mobile” with 30 

frames of 330 × 264 pixels is to reveal the visual effects 

for the proposed method and the compared methods. Fig. 

4 represents the selected two frames from the 

reconstructed results for bilinear interpolation, the 

method in [4] and the proposed method. The details are 

unfolded by their enlarged parts in Fig. 5. It can be seen 

that some numbers in the images, such as “15”, “16”, 

“18”, “19”, are obviously clearer in our results than 

others, due to that the invariant moments of Zernike are 

introduced, which improves the NLM algorithm. 

V. CONCLUSION AND DISCUSSION 

This paper proposes a SR algorithm using Zernike 

moments. The algorithm is based on the framework of 

probabilistic motion estimation and needs no explicit 

motion estimation. A collaborative similarity measure 

strategy is developed in our algorithm to combine the 

Zernike moment based proximity and the bilateral 

proximity of NL-means algorithm. As representation of 

image local invariant features, Zernike moments enable 

the algorithm to match more similar pixels not only with 

translation but also with rotation or scale. The 

collaborative mechanism ensures more suitable weights 

assigned to similar patches for better estimation



TABLE I. MEAN-PSNR RESULTS FOR THREE TEST SEQUENCES WITH DIFFERENT METHODS. 

Sequence Bilinear Bicubic 

Protter et al. 

[4] 

Our method 

(1st Iteration) 

Our method 

(2nd Iteration) 

Miss America 33.91 34.31 35.31 35.65 35.87 

Trevor 29.42 29.79 30.39 30.58 30.76 

Foreman 28.38 28.89 29.57 29.65 29.84 

(a)Miss America 

(b)Trevor (c) Foreman 

Figure 3. The PSNR values of each frame reconstructed by different methods 

(a) (b) (c)


(d) (e) (f) 

Figure 4. Results for the 9th (the top row) and 14th (the bottom row) frames from “Mobile” sequence. From left to right: bilinear interpolation; the 

method in [4]; the proposed method. 

of images. Experimental results demonstrate that the 

proposed algorithm is able to process real video 

sequences with general motion patterns with 

improvements both in PSNR and the visual effects 

compared to the stare-of-the-art algorithm. 

Several aspects of the proposed method may be further 

studied to improve the algorithm. Firstly, parameters σB 

and σZ reflecting the size of the noise are constant during 

the iterations. Since noise becomes smaller in the later 

reconstruction, it should be reasonable to decrease 

parameters σB and σZ appropriately with the increased 

iteration. Secondly, Zernike moment used in our method 

may be replaced by the Pseudo-Zernike moment, which is 

proved able to represent image details better with lower 

orders. And fast computation has been proposed to 

directly calculate arbitrary Pseudo-Zernike moment with 

high order without lower order moments first computed. 

Thus, to use Pseudo-Zernike moment instead may help to 

improve the performance without computation increased. 

Finally, several important parameters in our algorithm are 


Figure 5. Enlarged parts of images in Fig. 4. 

manually selected in the experiments, such as σB, σZ, and 

the searching size for the motion domain. An adaptive 

selection strategy may be developed to improve the 

algorithm in the future work. 

REFERENCES 

[1] S. Park, M. Park, M. G. Kang, “Super-resolution image 

reconstruction: a technical review,” IEEE Signal 

Processing Magazine, vol. 20, pp. 21–36, May 2003. 

[2] S. Farsiu, D. Robinson, M. Elad, and P. Milanfar, 

“Advances and challenges in superresolution,” Int. J. Imag. 

Syst. Technol., vol. 14, pp. 47–57, August 2004. 

[3] W. Shao, Z. Hui, “Edge-and-corner preserving 

regularization for image interpolation and reconstruction,” 

Image and Vision Computing, vol. 26, pp. 1591–1606, 

2008. 

[4] M. Protter, M. Elad, H. Takeda, et al., “Generalizing the 

nonlocal-means to super-resolution reconstruction,” IEEE 

Trans. Image Process., vol. 18, pp. 36–51, January 2009. 

[5] H. Takeda, P. Milanfar, M. Protter, et al., “Superresolution 

without explicit subpixel motion estimation,”


IEEE Trans. Image Process., vol. 18, pp. 1958-1975, 

September 2009. 

[6] A. Danielyan, A. Foi, V. Katkovnik, et al., “Image and 

video super-resolution via spatially adaptive blockmatching 

filtering,” in Proc. Int. Workshop Local and Non- 

Local Approx. Image Process., Switzerland, August 2008. 

[7] M. Protter, M. Elad, “Super resolution with probabilistic 

motion estimation,” IEEE Tran. Image Process., vol. 18, 

pp. 1899–1904, August 2009. 

[8] A. Buades, M. Morel, “A review of image denoising 

algorithms with a new one,” Mutiscale Model Simul. vol. 4, 

pp. 490–530, February 2005. 

[9] S. Alexander, E. Vrscay, and S. Tsurumi, “A simple, 

general model for the affine selfsimilarity of images,” in 

Proc. Int. Conf. on Image Analysis and Recognition, 

Lecture Notes in Comput. Sci. 5112, Springer-Verlag, 

Berlin, 2008, pp. 192–203. 

[10] G. Peyre, “Sparse modeling of textures,” J. Math. Imaging 

Vision, vol. 34, pp. 17–31, 2009. 

[11] M. Ebrahimi and E. Vrscay, “Examining the role of scale 

in the context of the non-local-means filter,” in Image 

Analysis and Recognition, Lecture Notes in Comput. Sci. 

4633, Springer-Verlag, Berlin, 2008, pp. 170–181. 

[12] A. Buades, B. Coll, J. Morel, “Image denoising methods a 

new nonlocal principle,” SIAM Review, vol. 52, pp. 113– 

147, Jan. 2010. 

[13] Y. Lou, P. Favaro, S. Soatto, “Nonlocal similarity image 

filtering,” in Reports CAM (8–26), 2008. 

[14] S. Zimmer, S. Didas and J. Weickert, “A rotationally 

invariant block matching strategy improving image 

denoising with non-local means,” in Proc. of the Int. 


Workshop on Local and Non-local Approximation in 

Image Processing, pp. 135–142, 2008. 

[15] Z. Ji, Q. Chen, Q. Sun and D. Xia, “A moment-based 

nonlocal-means algorithm for image denoising,” 

Information Processing Letters, vol. 109, pp. 1238–1244. 

September 2009. 

[16] R. Mukundan and K. Ramakrishnan, “Fast computation of 

legendre and zernike moments,” Pattern Recognition, vol. 

28, pp. 1433–1442, September 1995. 

[17] C. Chong, P. Raveendran and R. Mukundan, “A 

comparative analysis of algorithm for fast computation of 

Zernike moments,” Pattern Recognition, vol. 36, pp. 731– 

742, 2003. 

[18] B. Ye and J. Peng, “Invariance analysis of improved 

Zernike moments,” Journal of Optics A: Pure and Applied 

Optics. vol. 4, pp. 606–614, 2002. 

Lin Guo was born in Hubei, China, in 1978. She is currently 

pursuing the Ph.D. degree in Wuhan University, Wuhan, China. 

Her research interests include super-resolution reconstruction, 

image processing and computer vision. 

Qinghu Chen was born in Hubei, China, in 1957. He 

currently is a professor in School of Electronic Information at 

Wuhan University, Wuhan, China. His main research interests 

include image processing and intelligent recognition.


Mathematical Model and Hybrid Scatter Search 

for Cost Driven Job-shop Scheduling Problem 

Bai Jie 

Department of Automation, Shanghai Jiao Tong University, Shanghai, China 

Email: baijie@sjtu.edu.cn 

Sun Kai* 

Shandong Provincial Key Laboratory of AM&MC Technology for Light Industry Equipment, 

Shandong Institute of Light Industry, Shandong, China 

Email: sunkai79@gmail.com 

Yang Gen Ke 

Department of Automation, Shanghai Jiao Tong University, Shanghai, China 

Email: gkyang@sjtu.edu.cn 

Abstract— Job-shop scheduling problem (JSP) is one of the 

most well-known machine scheduling problems and one of 

the strongly NP-hard combinatorial optimization problems. 

Cost optimization is an attractive and critical research and 

development area for both academic and industrial societies. 

This paper presents a cost driven model of the job-shop 

scheduling problem in which the solutions are driven by 

business inputs, such as the cost of the product transitions, 

revenue loss due to the machine idle time and 

earliness/tardiness penalty. And then, a new hybrid scatter 

search algorithm is proposed to solve the cost driven jobshop 

scheduling problem by introducing the simulated 

annealing (SA) into the improvement method of scatter 

search (SS). In order to illustrate the effectiveness of the 

hybrid method, some test problems are generated, and the 

performance of the proposed method is compared with 

other evolutionary algorithms such as genetic algorithm and 

simulated annealing. The experimental simulation tests 

show that the hybrid method is quite effective at solving the 

cost driven job-shop scheduling problem. 

Index Terms—cost optimization, job-shop scheduling 

problem, scatter search, simulated annealing 


The job-shop scheduling problem (JSP) is one of the 

most well-known machine scheduling problems and one 

of the strongly NP-hard combinatorial optimization 

problems [1]. Historically, JSP was primarily solved by 

the branch-and-bound method and some heuristic 

procedures based on priority rules [2]. During the past 

decade, researches on meta-heuristic methods to solve the 

JSP have been widely studied, such as genetic algorithm 

[3],[4], simulated annealing [5], tabu search [6] and 

Manuscript received Sep 5, 2010; revised Nov 6, 2010; accepted Dec 

15, 2010. 

The project is supported by Shandong Provincial Natural Science 

Foundation, China (No.ZR2010FQ009) and the National Nature 

Science Foundation of China (No. 61074150) 

*Corresponding author 


doi:10.4304/jnw.6.7.974-981 

particle swarm optimization [7]. The majority of studies 

on JSP, however, are driven by production criteria, such 

as total flowtime, maximum complete time (makespan), 

maximum tardiness and number of tardy jobs, etc. 

In nowdays, the critical challenge to manufacturing 

enterprise is to become more flexible and profitable [8]. 

The “goal” of a manufacturing enterprise is not only to 

apply advanced technology, but consistently to make 

money (i.e., profits), as discussed in the book The Goal 

[9]. The development and application of an appropriate 

scheduling solution with seamless integration of business 

and manufacturing play a critical role in any modern 

manufacturer achieving this goal [10]. 

Jiang et al. [11] present a cost driven objective 

function for job-shop scheduling problem and solve it by 

using genetic algorithm, and the experimental results 

demonstrate the effectiveness of the algorithm. However, 

the study didn’t provide precise mathematical model for 

the cost driven job-shop scheduling problem. The major 

contributions of this paper are summarized as follows: 

(1) Based on the characteristics of general JSP, a 

mathematical model of cost driven JSP is presented. The 

cost of operational transitions between products, the 

revenue loss due to machine idle time during the phase of 

product transitions, and the penalty due to missing the 

required on-time delivery date, and so on are included in 

the model. 

(2) A new hybrid evolutionary algorithm, which 

combines the strong global search ability of scatter search 

(SS) with the strong local search ability of simulated 

annealing (SA), is developed to solve the cost driven JSP. 

The organization of remain contents is as follows. In 

section 2, the formulation and model of the cost driven 

JSP is presented. Section 3 presents the conceptual 

introduction to SS and SA and proposes the hybrid scatter 

search (HSS) algorithm to solve the cost driven JSP. 

Section 4 provides experimental results and performance 

analyses. Section 5 offers concluding remarks.


II. PROBLEM STATEMENT 

The JSP may be formulated as follows. There are n 

jobs that plan to process on m machines. Each job 

involves a set of operations, which are performed on the 

machines in a pre-specified order. Each machine can 

process only one job at a time, and it cannot be 

interrupted. Furthermore, the processing time is fixed and 

known. The objective widely used is to find a sequence of 

jobs to minimize makespan or total weighted complete 

time[1-7]. In the study, we focus on developing a cost 

driven model for JSP in order to make manufacturing 

enterprises more profitable. 

The multiple job sequences are defined as decision (or 

manipulate) variables. The illustration of the costs 

generated from a scheduling scenario is schematically 

shown in Figure 1. 

The key ideas behind the developed scheduling system 

are constrained cost driven optimization solution, namely, 

the cost is defined as an objective function subject to the 

relevant constraints. In the paper, the cost of the product 

transitions, revenue loss due to the machine idle time and 

earliness/tardiness penalty are included in the cost model 

for optimizing production scheduling solutions. 

Figure 1. Cost description of job-shop production line 

In order to formulate the cost driven job-shop 

scheduling problem, we introduce a null job (job 0) 

whose processing time and transition cost with other job 

are zero and all sequences on each machine are started 

from job 0 and ended at job 0. And then we define the 

parameter and decision variables as follows. 

Paremeters: 

i, j : index of jobs, and i , j � 0 ~ n 

k : index of machines, and k � 1 ~ m 

t ijk : sequence-dependent transition cost from job i 

to job j on machine k 

Ci : flowtime of job i 

p : processing time of job i on machine k 

ik 

C ik : complete time of job i on machine k 

W k : machine waiting cost of machine k per unit 

time during job transfer 

ei: earliness due date of job i 

di: tardiness due date time of job i 

αi: unit time earliness penalties of job i 

βi: unit time tardiness penalties of job i 

T (S) 

: total transition cost of a schedule scenario S 

R (S ) : total revenue loss of a schedule scenario S 


E (S ) : total earliness/tardiness penalty of a schedule 

scenario S 

Decision variable: 

�1 

if job j is preceded by job i on machine k 

X ijk � � 

(1) 

�0 

else 

The precedence coefficient Y ijk is defined as: 

�1 

if machine k process job i right after machine h 

Yihk � � 

(2) 

�0 

else 

A. Transition cost 

Transition includes work to prepare the machine, 

process, or bench for product parts or the cycle. This 

includes obtaining tools, positioning work in process 

materials, return tooling, cleanup, setting the required jigs 

and fixtures, adjusting tools, and inspecting material [12]. 

Sule and Huang [13] described the activities typically 

associated with sequence-dependent and sequenceindependent 

operations in machine shop environments. 

Transition cost is comprised of labor cost for setup 

operation and profit loss during machine idle time. 

Let T (S) 

denote the total transition cost of a schedule 

scenario S, the total transition costs of the solution S can 

be calculated by the following formula: 

B. Revenue loss 

m 

n 

n 

�� 

� 

ijk � ijk X t 

S T( 

) 

(3) 

k�1 

i�1 

j�1 

During the work process of JSP, each of the jobs is to 

be sequentially processed on machine 1~m. If a machine 

has finished processing a job, while the next job is being 

processed on the previous machine, the machine will be 

idle and waiting for the next job. The idle time of the 

machine will cause revenue loss of manpower and fixed 

assets for the manufacturer to maintain the ready state of 

the machine. Less machine idle time means higher 

utilization of manpower and machines. The revenue loss 

can be calculated by the following formula: 

m 

n 

n 

�� 

R( 

S) 

� ( C jk � Cik 

� pik 

) �Wk 

� X ijk (4) 

k �1 i�1 

j�1 

C. Earliness/tardiness penalty 

In the proposed cost driven JSP, every job has a duedate 

window. Any job completed prior to ei will incur an 

earliness penalty, on the other hand, any job completed 

after di will incur a tardiness penalty. No penalty, 

however, will be incurred if any job can be completed 

within the time window [ e i , i d ]. Let αi and βi, which can 

be determined by the inventory carrying cost and 

tardiness compensation, be the unit time earliness and 

tardiness penalties for job i. The earliness/tardiness 

penalty E can be calculated as:


n 

� 

i �1 

E( 

S) 

� [ � � max( 0, 

e � C ) � � � max( 0, 

C � d )] (5) 

i 

D. Mathematical model of cost driven JSP 

i 

Based on the above discussion of the problem, the 

mathematical model of cost driven JSP can be formulated 

as: 

m n n 

m n n 

Min{ 

�� tijk 

� Xijk 

�� 

( C jk � Cik 

� pik 

) �Wk 

� Xijk 

k �1 i�1 

j�1 

k �1 i�1 

j�1 

Subject to: 

jk 

n 

��[ 

�i 

� max( 0, 

ei 

� Ci) 

� �i 

� max( 0, 

Ci 

� di 

) ]} 

i�1 

n 

� X ijk 

j�0 

n 

� X ijk 

i�0 

jk 

i 

i 

i 

i 

(6) 

� 1 i � j , i � 1,... 

n , k � 1,..., 

m (7) 

� 1 j � i , j � 1,... 

n , k � 1,..., 

m (8) 

C � p � ( 1� 

X ) M � C n 

Cik � pik 

� ( 1� 

Yihk 

) M � Cih 

i 1,... 

n 

ijk 

ik i , j � 1,... 

, k 1,..., 

m 

� (9) 

� , h , k � 1,..., 

m (10) 

X ijk �{ 

0, 

1} 

i , j � 0, 

1,..., 

n , i � j , k � 1,..., 

m (11) 

Constraint (7) defines that a job should be right before 

another job on each machine. Constraint (8) denotes that 

a job should be right after another job on each machine. 

Constraint (9), where M is a very large positive number, 

shows that each machine can process at most one job at 

any time. Constraint (10) shows that each job can be 

processed on at most one machine at any time. Constraint 

(11) ensures that the variable only takes the integer 0 or 1. 

Obviously, the proposed JSP is a hard, constrained, 

m 

combinatorial optimization problem with ( n ! ) possible 

solutions for n-job m-machine problems. Even for a 

simple case with ten jobs and ten machines, the 

computation time for complete enumeration of all 

possible solutions is quite large [1]. If we have a large 

number of jobs and machines, e.g., thirty jobs and fifty 

machines, the complete enumeration of all possible 

solutions is computationally prohibitive, i.e., no exact 

algorithm is capable of solving the optimization problem 

in a reasonable computation time. Frequently, 

evolutionary algorithms as promising approximate 

techniques, such as genetic algorithm [3],[4] and 

simulated annealing [5], are employed to solve the 

scheduling problem of finding a desirable, although not 

necessarily, optimal solution. 

III. HYBRID SCATTER SEARCH FOR JSP 


Scatter search (SS) is an evolutionary method that has 

been successfully applied to hard optimization problems 

[14-17]. Unlike genetic algorithm, a scatter search 

operates on a small set of solutions (called reference set) 

and makes only limited use of randomization as a proxy 

for diversification when searching for a globally optimal 

solution. Glover [14] proposed a template to serve as the 

guideline of implementing SS to solve combinational 

optimization problems. The template consists of five 

components which are diversification generation method, 

improvement method, reference set update method, 

subset generation method and solution combination 

method. A template of the standard SS can be shown as 

follows: 

Step 1: Initialization 

1.1 Use the diversification generation method to 

generate diverse trial solutions. 

1.2 Apply the improvement method to enhanced 

trial solutions. 

1.3 Apply the reference update method to build 

the initial reference set (RefSet) from the 

enhanced trial solutions. 

Step 2: Computation 

Do 

2.1 Generate subsets of RefSet with the subset 

generation method. 

2.2 Use solution combination method to combines 

these subsets and obtain new solutions. 

2.3 Use the improvement method to enhance these 

new trial solutions. 

2.4 Apply the reference update method to update 

the RefSet. 

While (the maximum generation is not meet) 

Step 3: Output the optimization results. 

In the template, five key components of scatter search 

are labeled in bold font. The computation flow chart of 

scatter search can be shown in Figure 2. 

Figure 2. Computation flow chart of scatter search


The paper develop a new hybrid scatter search, which 

introduces the simulated annealing into the improvement 

method to enhance the local search ability of scatter 

search, to solve the cost driven job-shop scheduling 

problem. The detailed discussion of applying hybrid 

scatter search is shown as follows. 

A. Encoding scheme and fitness function 

One of the key issues in applying SS successfully to 

cost driven JSP is how to encode a schedule of the 

problem to a search solution. We utilize an operationbased 

representation [3] that uses an unpartitioned 

permutation with m -repetitions of job numbers. A job is 

a set of operations that has to be scheduled on m 

machines. In this formulation, each job number occurs m 

times in the permutation, i.e. as often as there are 

operations associated with this job. By scanning the 

permutation from left to right, the kth occurrence of a 

job number refers to the kth operation in the 

technological sequence of this job. A permutation with 

repetition job numbers merely expressed the order in 

which the operations of jobs are scheduled. 

For example, suppose a solution is given as {3 2 3 4 2 

4 2 1 1 3 2 3 1 4 4 1} in 4 jobs and 4 machines problem. 

Each job consists of three operations, and is thereby 

repeated four times. Third number of solution in this 

example is 3. Here, 3 implies second operation of job 3 

because number 3 has been repeated twice. 

Fitness function is used to evaluate the performance of 

solutions. In the paper, the objective function of solution 

S is defined as: 

m n n 

m n n 

Fit( 

S) 

�� 

tijk 

� X ijk �� 

( C jk � Cik 

� pik 

) �Wk 

� X ijk 

k�1 

i�1 

j�1 

k�1 

i�1 

j�1 

n 

�� 

[ � i � max( 0, 

ei 

� Ci 

) � �i 

� max( 0, 

Ci 

� di 

) ] � M � feaS 

i�1 

�0 

where feaS � � 

�1 

positive number. 

B. Diversification generation method 

(12) 

if 

S is feasible 

and M is a very large 

otherwise 

The diversification generation method is used to 

generate a collection of diverse trial solutions, using an 

arbitrary trial solution (or seed solution) as an input. This 

element of the SS approach is particularly important, 

given the goal of developing a method that balances 

diversification and intensification in the search. This 

method was suggested by Glover [14], which generates 

diversified permutations in a systematic way without 

reference to the objective function. 

Assume that there are a n� m JSP, a given trial solution 

S used as a seed is representing by indexing its elements, 

so that they appears in consecutive order to yield 

S � {[ 1], 

[ 2],..., 

[ l]} 

, where l � m � n . Define the 

subsequence S ( h : t) 

, where t is a positive integer 

between 1 and h , to be given by: 

S( h : t) 

� ([ t], 

[ t � h], 

[ t � 2h],...., 

[ t � rh]) 

, 


where r is the largest nonnegative integer such that 

t � rh � l . Then define the S (h) 

for h � n , to be: 

S( h) 

� { S( 

h : h), 

S( 

h, 

h �1),..., 

S( 

h : 1)} 

. 

To illustrate the strategy, suppose S is given by: 

S={[1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], 

[12], [13], [14], [15],[16]}. 

If we choose h=4, then 

S(4:4)={[4], [8], [12],[16]}, 

S(4:3)={[3], [7], [11], [15]}, 

S(4:2)={[2], [6], [10], [14]}, 

S(4:1)={[1], [5], [9], [13]}, 

to give: 

S(4)={(4:4), (4:3), (4:2), (4:1)} 

={[4], [8], [12],[16] ,[3], [7], [11], [15], [2], [6], [10], 

[14], [1], [5], [9], [13]}. 

In this illustration, we have allowed h to take the value 

closest the square root of l. The value is interesting based 

on the fact that, when h equals the square root of l, the 

minimum relative separation of each element from each 

other in the new permutation is maximum, compared to 

the relative separation of exactly 1 in the trial solution S 

[14]. In general, for the goal of generating a diverse set of 

trial solutions, preferable values for h range from 1 to l 2 . 

C. Improvement method 

Each of the new trial solutions which are obtained 

from the diversification generation method or solution 

combination method is subjected to the improvement 

method. This method aims to enhance the quality of these 

solutions. In the paper, we take two versions of local 

search meta-heuristics to improve trial solutions. A longterm 

SA-based improvement method is only applied to 

the best new trial solution, and a short-term swap-based 

local search is taken to enhance other new trial solutions. 

With the hybridization of these two local methods, we 

can get a compromise between solution quality and 

computational effort. 

C.1 Swap -based local search method 

A Swap-based local search method is taken to improve 

methods for HSS. In the method, swap operator is 

adopted to obtain neighbors. 

Swap operator: Let [i ] and [ j ] be two randomly 

selected positions whose job numbers are different in the 

trial solution S. A neighborhood of S is obtained by 

interchanging the job in position [i ] and [ j ] . 

For each new trial solution, the local search method 

takes it as the initial solution, and then searches in its 

neighborhood until there is no improvement. If the local 

search yields a better value than the one from the original 

solution, the new solution will replace the original 

solution. If no improvement has been found after the 

local search, no replacement will be made. 

C.2 SA-based improvement method 

Ever since it was introduced by Kirkpatrick [18], the 

SA algorithm has been applied to many combinatorial 

optimization problems. The SA approach can be 

interpreted as an enhanced version of local search or 

iterative improvement, which can avoid being trapped in


local minima by probabilistic jumping. In the paper, a 

SA-based local search method is developed as a longterm 

improvement method for HSS. 

In the SA-based local search method, swap operator 

described above is adopted to obtain neighbors. The new 

solution is accepted if the objective function is improved. 

Otherwise, the new solution is accepted with probability 

exp( � � / T ) , where � is the change of the objective 

function value and T is a control parameter. 

SA process can he controlled by the cooling schedule. 

The selection of initial temperature, cooling rate, 

termination temperature and temperature length 

influences the quality of the solutions. In SA optimization 

process, the temperature is gradually reduced. It is well 

known that specifies temperature with the equation 

Tk=λ·Tk-1 is often a good choice and it can provide a 

tradeoff between computation time and good solutions. 

Tend is chosen to terminate the SA process, when the 

current temperature T


Similarly, filling in the holes of string 2 with 

unselected operations of original trial solution 1, we can 

get: 

New trial solution 2: 3 2 3 1 4 2 2 1 3 4 4 4 1 3 2 1 

IV. COMPUTATIONAL EXPERIMENTS 

A. Instance data generation 

To illustrate the effectiveness of the algorithm 

described in this paper, we consider several instances 

originated from two classes of standard JSP test problems: 

Fisher and Thompson instances (FT06, FT10, FT20) [19], 

and Lawrence instances (LA01, LA02,…,LA31) [20]. 

Originally, each instance only consists of the machine 

number and processing time ( p ik ) for each step of the 

job. Furthermore, to apply these instances to the cost 

driven JSP presented in this paper, some extra instance 

data should be generated. 

In this paper, a method with lower and upper bound on 

including transition cost ( t ijk ), machine waiting cost 

per unit time( k W ), earliness/tardiness penalties( i � , � i ) 

and random selection is used to generate extra data of 

problem instances. Table I shows the bounds used for 

problem data. 

TABLE I. 

LOWER AND UPPER BOUND FOR PROBLEM DATA 

Problem data k W t ijk i � i � 

Lower bound 1 1 1 1 

Upper bound 10 10 5 10 

The due date [ e i , d i ] of problem can be generated by 

the following formula: 

� 

�ei 

� 0. 

5� 

C 

� 

� 

� 

� 

�di 

� 0. 

5� 

C 

�� 

max 

max 

� � � 

� � � 

m 

� 

k �1 

m 

� 

k �1 

p 

ik 

p 

ik 

(14) 

where C max is best makespan known so far, δ and σ 

are tightness factors of due dates, and we can obtain 

proper δ and σ for each instance by several runs. 

B. Experimental setup 

The algorithm for cost driven JSP mentioned above was 

programmed in Borland C++ and the experiments were 

executed on a Intel Pentium 2.8G with 512M RAM. The 

parameters of HSS use the following configuration: 

Parameters for SS: 

The number of trial solutions in Refset1 ( b 1 ) equals the 

number of jobs in the problem, and so is that in RefSet2. 

And after n� m generations for n-job m-machine problem, 

the algorithm is terminated. 

Parameters for SA-subprogram: 


Initial temperature T 0 in this paper is set by T0 � �fmax 

, 

where �fmax is the maximal difference in fitness value 

between any two neighboring solutions. It should be 

adjusted experimentally. Epoch length L is set to the 

number of ( n � 1) 

� m . Decreasing rate � is set as values 

0.98. Termination temperature T end is equal to 0.1. SA 

sub-program is terminated whenever there is no 

improvement in 20 successive generations, which enables 

a reduction in running time. 

C. Simulation results and comparison 

In order to illustrate the effectiveness of the hybrid 

method, we compare the proposed method with other 

evolutionary algorithms such as GA [11], SA and SS. The 

GA parameters: population size is set equal to HSS, 

crossover probability is 0.95, mutation probability is 0.05. 

The termination conditions of these algorithms are set 

equal to the HSS’s CPU time. The statistical performance 

of 20 independent runs of these algorithms are listed in 

table 2, including the optimum value known so far(BKS), 

the best objective value ( C * ) , the percentage value of 

average objective value over BKS (%) and the average 

CPU time ( t ) of the HSS. 

Table II shows that the results obtained by HSS are 

much better than those obtained by GA, SA and SS. The 

superiority of the best optimization quality demonstrates 

the effectiveness and the global search property of the 

hybrid search, and the superiority of the average 

performance over 20 random runs shows that the hybrid 

probabilistic search is more robust than these algorithms. 

It can be seen that the HSS algorithm can get desirable 

solutions in a reasonable computation time even for 

problems with 30 jobs and 10 machines. For more largescale 

problems, we can trade off between computation 

time and solution quality by adjusting the number of 

generations or the parameters of SA-subprogram. 

V. CONCLUSION 

Cost optimization is an attractive and critical research 

and development area for both academic and industrial 

societies. This is also a multi-disciplinary subject with 

optimization, control, business intelligence, computer 

science and operation research, and so on. In the cost 

driven JSP model we proposed, the solutions are driven 

by business inputs, such as market demand and the costs 

of inventory and machine idle time during the product 

transition phase. And then, a hybrid optimization 

algorithm that combines SS with SA is proposed to solve 

the problem. This hybrid method combines the 

advantages of these two algorithms and mitigates the 

disadvantages of them. The obtained results indicate that 

this hybrid method is superior to GA, SA and SS, and is 

an effective approach for the cost driven JSP.


Inst. Size BKS 

TABLE II. 

EXPERIMENT RESULTS. 

GA[11] SA SS HSS 

C * % C * % C * % C * % t 

FT06 6� 6 426 426 0 426 0 426 0 426 0 1.85 

FT10 10� 10 2732 2741 4.05 2848 6.53 2732 2.85 2732 1.61 45.21 

FT20 20� 5 2442 2495 2.32 2504 1.74 2491 1.45 2442 0.92 42.92 

LA01 10� 5 935 935 1.26 935 0 935 0 935 0 5.26 

LA02 10� 5 1028 1028 1.41 1028 0 1028 0 1028 0 5.90 

LA06 15� 5 1670 1670 1.22 1670 0.15 1670 0 1670 0 10.83 

LA07 15� 5 1789 1789 1.25 1789 0.25 1789 0 1789 0 11.57 

LA11 20� 5 1899 1899 2.85 1899 1.92 1899 1.27 1899 0.56 58.63 

LA12 20� 5 2013 2013 3.22 2013 2.01 2013 1.32 2013 0.56 42.44 

LA16 10� 10 3061 3176 3.16 3103 1.95 3061 1.57 3061 1.28 41.23 

LA17 10� 10 2912 3021 2.93 2952 1.78 2912 1.46 2912 1.44 36.45 

LA21 15� 10 3978 4122 3.83 413 5 3.05 4065 2.57 3978 1.61 132.62 

LA22 20� 10 3463 3521 4.92 3539 3.54 3463 3.15 3463 2.37 367.17 

LA26 20� 10 3781 3997 4.50 3844 3.54 3805 3.32 3781 2.61 331.54 

LA27 20� 10 4433 4538 5.89 4507 5.76 4482 4.37 4433 3.29 430.17 

LA31 30� 10 5704 5983 5.75 5875 4.63 5841 3.72 5704 2.86 725.12 

LA32 30� 10 5443 5709 5.41 5607 4.43 5574 3.36 5443 2.74 692.20 

LA36 15� 15 3523 3748 6.36 3675 4.31 3627 3.92 3523 2.35 542.94 

LA37 15� 15 3791 3922 6.83 3955 4.63 3903 4.29 3791 2.68 476.67 


This research is supported by the Shandong Provincial 

Natural Science Foundation, China (No.ZR2010FQ009) 

and the National Nature Science Foundation of China 

(No. 61074150). 

REFERENCES 

[1] M. R. Garey, D. S. Johnson, and R. Sethi, “The 

Complexity of Flowshop and Jobshop Scheduling”, 

Mathematics of Operations Research, Vol. 1, No. 2, 

pp.117-129, 1976. 

[2] J. Adams, E. Balas, and D. Zawack, “The shifting 

bottleneck procedure for job shop scheduling”, 

Management Science, Vol. 34, pp.391–401, 1988. 

[3] L. W. Cai, Q. H. Wu, and Z. Z. Yong, “A genetic 

algorithm with local search for solving the job problems”, 

Lecture Notes in Computer Science, Vol. 1803, pp.363- 

365, 2000. 

[4] Y. Li, Y. Chen, “A genetic algorithm for job-shop 

scheduling”, Journal of Software, Vol 5, pp.269-274, 2010. 

[5] M. Kolonko, “Some new results on simulated annealing 

applied to the job shop scheduling problem”, European 

Journal of Operational Research, Vol. 133, No. 1, pp. 

123-13,6, 1999. 

[6] V. P. Eswaramurthy, and A. Tamilarasi, “Tabu search 

strategies for solving job shop scheduling problems”, 

Journal of Advanced Manufacturing Systems, Vol. 6, No.1, 

pp.59-75, 2007. 


[7] D. Y. Sha, and C. Y. Hsu, “A hybrid particle swarm 

optimization for job shop scheduling problem”, 

Computers & Industrial Engineering, Vol. 51, pp.791– 

808, 2006. 

[8] M. Lansiti, and R. Levien, “Strategy as Ecology’, Harvard 

Business Review, Vol. 3, pp.68-78, 2004. 

[9] E. M. Goldratt, and J. Cox, “The Goal - A Process of 

Ongoing Improvement”, North River Press, Croton-on- 

Hudson, New York, 1984. 

[10] Y. Z. Lu, “Profit driven manufacturing enterprise 

optimization: Problem and Solution”, Plenary Presentation, 

Proceedings of 23rd Chinese Control Conference, August, 

2004. 

[11] L. W. Jiang, Y. Z. Lu, Y. W. Chen, “Cost Driven 

Solutions for Job-shop Scheduling with GA.”, Control 

Engineering of China. Vol. 44, pp. 72—74, 2007. 

[12] A. Allahverdi, J. N. D. Gupta, and T. Aldowaisan, “A 

Review of Scheduling Research Involving Set-up 

Considerations”, Omega, Vol. 27, No.2, pp.219-239, 1999. 

[13] D. R. Sule, and K. Y. Huang, “Sequency on two and three 

machines with set-up, processing and removal times 

separated”, International Journal of Production Research, 

Vol. 21, pp.723-732, 1983. 

[14] F. Glover, “Scatter search and path relinking”, In: Corne, 

D., Dorigo, M., Glover, F. (Eds.), New Ideas in 

Optimization. McGraw-Hill, pp. 297–316, 1999. 

[15] R. A. Russel, and W. C. Chiang, “Scatter search for the 

vehicle routing problem with time windows”, European 

Journal of Operational Research, Vol. 169, pp.606-622, 

2006. 

[16] A. Haq, M. Saravanan, A. Vivekraj, and T. Prasad, “A


scatter search approach for general flowshop scheduling 

problem”, International Journal of Advanced 

Manufacturing Technology, Vol.31, No.7, pp.731-736, 

2007. 

[17] N. Mansour, C. Kehyayan, H. Khachfe, “Scatter search 

algorithm for protein structure prediction”, International 

Journal of Bioinformatics Research and Applications. 

Vol.5, No.5, 501-15, 2009. 

[18] S. Kirkpauick, C. D. Geian, and M. P. Vecchi, 

“Optimization by simulated annealing”, Science, Vol. 220, 

pp.671-680, 1983. 

[19] H. Fisher. and G. L. Thompson, “Industrial scheduling”, 

Englewood Cliffs, NJ: Prentice-Hall, 1963. 

[20] S. Lawrence, “Resource constrained project scheduling: 

An experimental investigation of heuristic scheduling 

techniques”, Graduate School of Industrial Administration, 

Carnegie Mellon University, Pittsburgh, PA, 1984. 

Bai Jie, was born in Shanghai, China in 

April 8, 1981. Received his bachelor 

and master degrees in control theory 

and engineering from Department of 

Automation at Shanghai Jiaotong 

University in 2002 and 2005. 

Now he is a PH.D candidate in 

control theory and engineering of 

Shanghai Jiaotong University. His 

research interest covers industrial production scheduling, 

computer integrated manufacturing, and intelligent 

optimization and application. 


Sun Kai, was born in Yutai city, 

Shandong province, China, in May 30, 

1979. Received his PH.D degree in 

control theory and engineering from 

Department of Automation at Shanghai 

Jiaotong University in 2009. 

Now he is a Lecturer at School of 

Electronic Information and Control 

Engineering, Shandong Institute of 

Light Industry, Shandong, China. His research interests are in 

optimization and scheduling. 

management. 

Yang Gen Ke, was born in Taiyuan city, 

Shanxi province, China, in June 5, 1963. 

Received his PH.D degree in system 

engineering from Department of 

Automation at Xi’an Jiaotong University 

in 1998. 

Now he is a professor at Department 

of Automation of Shanghai Jiaotong 

University. His research interests are in 

operation research and supply chain


Multi-objective Genetic Algorithm for System 

Identification and Controller Optimization of 

Automated Guided Vehicle 

WU Xing, LOU Peihuang and TANG Dunbing 

Nanjing University of Aeronautics and Astronautics, Nanjing, China 

Email: {Wustar5353, meephlou, d.tang}@nuaa.edu.cn 

Abstract— This paper presents a multi-objective genetic 

algorithm (MOGA) with Pareto optimality and elitist tactics 

for the control system design of automated guided vehicle 

(AGV). The MOGA is used to identify AGV driving system 

model and optimize its servo control system sequentially. In 

system identification, the model identified by least square 

method is adopted as an evolution tutor who selects the 

individuals having balanced performances in all objectives 

as elitists. In controller optimization, the velocity regulating 

capability required by AGV path tracking is employed as 

decision-making preferences which select Pareto optimal 

solutions as elitists. According to different objectives and 

elitist tactics, several sub-populations are constructed and 

they evolve concurrently by using independent reproduction, 

neighborhood mutation and heuristic crossover. The lossless 

finite precision method and the multi-objective normalized 

increment distance are proposed to keep the population 

diversity with a low computational complexity. Experiment 

results show that the cascaded MOGA have the capability to 

make the system model consistent with AGV driving system 

both in amplitude and phase, and to make its servo control 

system satisfy the requirements on dynamic performance 

and steady-state accuracy in AGV path tracking. 

Index Terms— multi-objective optimization, genetic 

algorithm, system identification, controller optimization. 


Automated guided vehicle (AGV) is a wheeled mobile 

robot with automatic guidance and driving systems. It can 

move along the designated routes and transport materials 

in flexible manufacturing systems [1]. To correct position 

and attitude error promptly in its movement, AGV servo 

control system should regulate the velocities of driving 

wheels at a frequency and accuracy required by its path 

tracking [2]. In the hierarchical control architecture, many 

sophisticated control laws are used for path tracking at 

the upper layer, but it is usual to adopt a PID control law 

for servo control at the bottom layer. 

How to construct a sufficiently accurate plant model is 

the first step for using most non-empirical control system 

design methods. Classical identification techniques such 

as least square method still have many limitations. If 

model construction is considered as an optimization of 

identification accuracy instead of a mapping from plant to 

model, genetic algorithm (GA) can be used for it [3,4], 


doi:10.4304/jnw.6.7.982-989 

e.g. a time-delay system model is identified by GA from 

step responses [3]. 

Moreover, GA can also be used to optimize parameters 

of PID controller [5-10], e.g. a self-organization GA with 

cyclic mutation [6] and a real-coded adaptive GA with a 

variable crossover and mutation probability [8]. In many 

control systems, it is usual to adopt different controller 

parameters based on a tradeoff in multiple performance 

objectives. A multi-objective GA (MOGA) is proposed to 

find an appropriate setting of PID controller by analyzing 

Pareto optimal surfaces [9]. A modified GA with elitist 

model and niching method is developed to guarantee a set 

of PID parameters with different tradeoffs regarding 

multiple requirements [10]. 

This paper presents a MOGA with Pareto optimality 

and elitist tactics for system identification and controller 

optimization of AGV. The remaining parts are organized 

as follows. Section II introduces the existing GAs used 

for multi-objective optimization. Section III presents the 

MOGA with Pareto optimality and elitist tactics in detail. 

Section IV describes AGV prototype and its test system. 

Section V applies the MOGA to experiments of system 

identification and controller optimization. Finally, section 

VI gives a brief conclusion. 

II. MOGA FOR SYSTEM IDENTIFICATION AND 

CONTROLLER TUNING 

System identification and controller tuning can both be 

converted into multi-objective optimization problems if 

the former is viewed as model parameter optimization by 

minimizing the error between model response output and 

plant response output, and the latter is considered as 

controller parameter optimization by minimizing the error 

between the actual output and the desired output. In this 

sense, they can be handled by one optimization method, 

such as a cascaded GA [11,12]. 

A. Problem Description for Multi-Objective Optimization 

The selection of objective function has a significant 

influence on optimization results. In control engineering, 

it is usual to use rising time t , overshoot r 

� , stead-state 

error e and other time-domain or frequency-domain 

s 

criteria as control objectives. The purpose of multiobjective 

optimization is to find a vector X containing a


set of variables x that can simultaneously minimize a 

i 

function vector F(X) containing a set of objective 

functions f ( xi 

) , which is formulated as 

min� 

f1( x1) 

f 2 ( x ) � f m ( xm 

) � 

�a b � 

min F(X) � 

2 

. 

s.t. x � , 

(1) 

i 

i 

i 

Unfortunately, these objectives may conflict with each 

other, and these functions may not be minimized at the 

same time. For example, a solution with small rising time 

may be the one with higher overshoot. Pareto superiority 

and Pareto optimality are defined to compare different 

solutions in multi-objective optimization problems. 

Definition 1. Pareto Superiority 

Let X and Y denote two vectors in multi-objective 

optimization problems. If the function vector has the 

following relationship: fi ( yi 

) � f i ( xi 

) for all objectives 

(i=1,2,…,m), and fi ( yi 

) � fi 

( xi 

) for at least one 

objective, X is Pareto superior to Y. 

Definition 2. Pareto Optimality 

Let X ∈ [a,b] denote a vector in multi-objective 

optimization problems. If there is no other vector Y∈[a,b] 

Pareto superior to X, X is Pareto optimal. 

Multi-objective optimization problems usually involve 

the minimization of several conflicting criteria that can 

not be achieved simultaneously. Therefore a satisfactory 

tradeoff must be found and a set of optimal solutions 

(instead of a single solution) must be provided. In this set, 

there is no solution superior to others when all objectives 

are taken into account. These solutions (also called nondominated 

solutions) comprise Pareto optimal set. The 

graphical expression of their function values is called 

Pareto front. Take the minimization of two objective 

functions f 1( x) 

and f 2 ( x) 

for example. If it assumes the 

area surrounded by a solid line and a dotted line in Fig.1 

is the value range of objective functions, then the solid 

line is Pareto front of this minimization problem, and 

point X is Pareto superior to point Y. 

Figure 1. Graphical expression of Pareto Superiority and Optimality. 

B. Improved GA for Multi-Objective Optimization 

Conventional GA is only suitable for single-objective 

optimization problems because its fitness function only 

contains one criterion. Fitness function need be modified 

to make it compatible with multi-objective optimization 

problems. The possible improved approaches of GA can 

be classified in three groups. 

(1) Aggregating approaches in which all objectives are 

combined into a single function, such as weighted sum 

approach [6,8]. It is not necessary to modify GA structure 

itself, and multi-objective optimization problems can be 

solved as the same as single-objective ones. However, it 


is difficult to select weights for different objectives and 

an improper selection may lead to optimization failures. 

(2) Non-aggregating approaches that are not Pareto 

based, e.g. some techniques based on population policies 

and special handling of the objectives are used to search a 

solution set. Vector evaluated GA (VEGA) [13] is a wellknown 

example of this group. 

(3) Pareto based approaches in which the amount of 

individuals that are superior to the individual A is used as 

the rank of A [14]. Non-dominated Sorting GA (NSGA) 

divides the entire population into several groups with 

different ranks, and the individuals with the same rank 

have the same reproduction probability [15]. NSGA-II is 

an improved one that preserves the optimal individuals by 

using elitist tactics and replaces fitness sharing parameter 

with crowding distance [16]. 

III. MOGA WITH PARETO OPTIMALITY AND ELITIST 

TACTICS 

In this paper, Pareto superiority or optimality is used to 

construct Pareto sub-population. Elitist selection tactics 

are used to preserve excellent individuals and guide the 

entire population evolution direction. The lossless finite 

precision method and the normalized increment distance 

are proposed to keep the population diversity with a low 

computational complexity. Multi-population evolution 

mechanism is presented to promote the development of 

multiple sub-populations. 

A. Elitist Selection Tactics 

Because of probabilistic behavior existing in evolution, 

the best individuals may be lost in the next generation. 

Elitist tactics are used widely to guarantee the survival of 

the best individuals in many GAs [9,10]. On another hand, 

current researches on MOGA mainly focus on how to get 

non-dominated solutions distributed uniformly in Pareto 

front, but almost neglect the influence of decision-making 

preferences on solution selection. Apart from preserving 

the best individuals, elitist selection tactics in this paper is 

used to inject decision-making preferences into MOGA, 

which limits the scope of non-dominated solutions to the 

area interested by decision-makers (shaded area in Fig.1). 

Elitist selection tactics are implemented by two ways. 

One way is to designate an individual having balanced 

performances in all objectives as the evolution tutor, and 

select these individuals that are Pareto superior to this 

tutor as elitists. If X is designated as the tutor in Fig.1, all 

individuals in the shaded area are potential elitists. They 

will evolve forward to Pareto front under the guidance of 

evolution tutor while keeping balanced performances. 

The other way is to use decision-making preferences 

directly. If a Pareto solution is meaningless to a practical 

problem, eliminate it from Pareto sub-population (PSP). 

If it satisfies decision-making preferences, select it into 

elitist sub-population (ESP). In addition, an aggregated 

function is used to describe the overall performance of 

Pareto solutions for multi-objective optimization. 

Decision-making preferences get into ESP construction 

via elitist selection tactics, pass to the next generation via 

ESP reproduction and mutation, and spread in the entire


population via heuristic crossover of elitists and other 

individuals. They can form an elitist guidance mechanism 

that leads the entire population to Pareto front interested 

by decision-makers. 

B. Diversity Keeping Techniques 

Diversity keeping techniques are used in GA to avoid 

the population premature problem [14,16-18]. Niching 

methods [14] penalize the crowded individuals by a cost 

function. The crowded distance between individuals is 

used in non-dominated sorting [16]. The vector norm 

function is used for multi-objective fitness [17]. These 

methods avoid the closeness and similarity of solutions 

but need much computing time. Finite precision method 

deletes the similar individuals by reducing the computing 

precision of objective functions intentionally [18]. Its cost 

is the reduction of computing precision. 

In this paper, the lossless finite precision method 

(LFPM) and the multi-objective normalized increment 

distance (MNID) are used as diversity keeping techniques 

at two layers, in order to decrease the distribution density 

of solutions and keep a high computing efficiency and 

precision. The former is used to eliminate the individual 

with a serious congestion and a low fitness and the latter 

is used to evaluate the elitist fitness that determines the 

corresponding parameters in genetic operations. 

Let all objective functions in optimization problems be 

{ i 

F , i=1,2,…,m}. Let the entire population in the current 

generation be { D , r=1,2,…,n}. Let all function values of 

r 

individual r D be { 1 2 

f , r f ,…, r 

m 

f }. The steps of LFPM 

r 

are described below. 

(1) All individuals in the current generation are sorted 

in a descending order respectively according to each 

i 

objective function, and the resulting sequences are F 

={ i 

j 

f , j=1, 2,…,n}. The minimum result variation from 

large to small is specified as a fixed step A for each 

i 

objective function. Initialize the individual number to be 

compared in the first function as k=1. 

1 

(2) Consider the sorting sequence F for the first 

objective. Increase the individual number: k=k+1. If k>n, 

then terminate the algorithm, otherwise check whether 

f and that 

the difference between the function value of 1 

k 

of its former 1 

f is larger than k �1 

i 

A . 

function value of individual D (1≤k≤n). 

k 

1 (3) If f - k 

1 

f ≥ k�1 

i 

to step (2). 

(4) If 1 

k 

f - 1 

f < k�1 

i 

function to be compared as s=2. 

(5) Search the function value 

f is the first 

A , then preserve individual D . Jump 

k 

A , then set the number of objective 

1 

k 

s 

f of individual h 

D in k 

s 

the sequence F of the s-th objective. If sorting rank 

s 

h =1, then preserve individual k 

D , and jump to step (2). 

k 

Otherwise, check whether the difference between the 

s 

s 

function value of f and that of its former 

h 

f is larger 

h- 

1 

than A . s 


s 

fh�1 ≥ s 

s 

(6) If f - h A , then preserve individual D . k 

Jump to step (2). 

s s 

(7) If f - h fh� 1 < A , then increase the number of 

s 

objective function to be compared as s=s+1. If s>m, then 

eliminate individual D due to its serious congestion and 

k 

jump to step (2). Otherwise, jump to step (5). 

It is seen that LFPM deletes the individual when the 

difference between its function result and its former’s 

result in each objective sorting is smaller than fixed steps, 

and it is more suitable for keeping population diversity in 

Pareto-based multi-objective optimization. 

MNID adopts the conception of crowded distance and 

vector norm. Elitists are sorted as different sequences 

according to each objective respectively. Each objective 

function value is converted to one component of a vector, 

shown as follow 

� 

� 1 

� � k k 

�( 

oi 

� oi�1) 

/ oi 

k 

Di k 

�1 

i � 1 

i � 1 

k 

k 

Where o and i� 

1 o are the function values of elitist 

i 

E and i�1 

E in the sequence sorted by objective i 

k . All 

objective function components are computed according to 

(2), and MNID is 

(2) 

1 2 2 2 

m 2 

�Di � ( �Di 

) � ( �Di 

) �� 

� ( �Di 

) (3) 

It is seen that relative increments of objective functions 

are used to compute the difference between the former 

elitist and the latter one on each objective, and it is more 

suitable for evaluating crowded degree in Pareto-based 

multi-objective optimization. 

C. Multi-population Evolution Mechanism 

Multi-population evolution mechanism (MEM) is used 

here to promote the development of different individuals. 

The entire current-generation population is divided into 

multiple sub-populations according to different objectives 

and elitist tactics. Single-objective sub-populations (SSP) 

are constructed by selecting the individuals based on each 

objective. Pareto optimality is used to organize Pareto sub 

-population (PSP) in which at least one objective function 

value of each individual is superior to that of others. 

Evolution tutor or decision-making preferences are used 

as elitist selection tactics for elitist sub-population (ESP). 

In the evolution process, sorting rank of individual is 

the base on which reproduction probability is calculated. 

Nonlinear normalized geometric sorting is used here to 

relieve the population premature. For the individual with 

rank r, the reproduction probability is 

r �1 

� p( 

r) 

� q0( 

1� 

q) 

� 

n 

�q0 

� q /[ 1� 

( 1� 

q) 

] 

(4) 

Where q is the probability parameter changing from 0 

to 1, q is the reproduction probability of the individual 

0 

with rank 1, n is the total number of individuals in the 

sorting sequence, p (r) 

is the reproduction probability of


the individual with rank r , and the sum of p(r) of all 

individuals is 1. 

The larger q is, the larger reproduction probability the 

individual with high rank has, the more influence elitists 

have on ordinary individuals, and the heavier selection 

pressure the entire population is under. So the selection 

pressure can be kept in an appropriate range by changing 

the probability parameter q . The function values do not 

determine the reproduction probabilities directly, which 

can decrease the possibility of population premature. 

After reproduction, the neighborhood mutation with 

variable amplitude is used to produce new elitists. The 

mutation amplitude is associated with the reproduction 

number of individual. Let the reproduction number of the 

1 

r 

elitist with rank 1 be N . For the elitist having 

e 

N e 

copies, the amplitude of the i-th neighborhood mutation is 

B 

� 0 

� 

�B1 

� ( i �1) 

B 

i � 1 

r ( i) 

� 1 

r 

2 / N e 1 � i � N e 

Where 1 B and B are the initial value and incremental 

2 

value of mutation amplitude. When an elitist has more 

copies, the variable range of mutation amplitude is larger, 

and this operation can search better individuals around 

this elitist more carefully. If the elitist only has one copy, 

the mutation amplitude is 0, having the same effect as the 

elitist preservation method. 

Produce randomly a mutation factor � in the range 

e 

from 0 to 1 for the (i+1)-th copy of elitist E , and the 

r 

new individual after the i-th neighborhood mutation is 

(5) 

� E ( 1� 

( 0. 

5 �� 

) B ( i)) 

(6) 

X i r 

e r 

The heuristic crossover is carried out between elitists 

and individuals. For the individual X with rank r in the 

r 

sorting sequence, the crossover amplitude is 

C � C rC / N 

(7) 

r 

1 � 

Where 1 C and C are the initial value and incremental 

2 

value of crossover amplitude, and N is the total number 

x 

of the sorting sequence. 

Produce randomly a crossover factor � in the range 

c 

from 0 to 1 for the elitist E and the individual i 

X , and r 

the new individual after the heuristic crossover is 

r 

c 

r 

i 

2 

Y � ( 1�� 

C ) E �� 

C X (8) 

It is seen from (7) and (8) that the crossover amplitude 

is larger if the sorting rank of the individual X is lower, 

r 

and it is influenced by the elitist E to a larger extent in 

i 

the heuristic crossover. Decision-making preferences can 

spread to SSP from ESP by this operation. 

If some components of new individuals are beyond the 

parameter boundary after the neighborhood mutation and 

heuristic crossover, replace them by the boundary values. 


x 

c 

r 

r 

D. Algorithm Description 

In this paper, Pareto optimality is used to guarantee 

solutions with different tradeoffs regarding multiple 

objectives. Elitist selection, neighborhood mutation and 

heuristic crossover are combined to expand the influence 

of decision-making preferences and make a directional 

search in Pareto front. LFPM, MNID and MEM are used 

to enhance the population fitness and diversity. The steps 

of the proposed MOGA with Pareto optimality and elitist 

tactics are detailed in Fig. 2. 

Figure 2. MOGA with Pareto Optimality and elitist tactics. 

IV. AGV PROTOTYPE AND ITS TEST SYSTEM 

This section describes a vision-based AGV prototype 

and its test system, as shown in Fig.3. A CCD camera is 

set in the vehicle center. Two driving wheels are placed 

on each side of its body symmetrically, and their velocity 

and direction are controlled by two sets of driving devices 

(drivers, motors, reducers, etc) respectively. Castors are 

distributed around the vehicle to support its weight. 

AGV movements at the desired linear and angular 

speed are achieved by changing the rotation velocities of 

driving-wheel motors. Path errors are perceived by AGV 

vision navigation, and the speed difference between two 

driving wheels is calculated by path tracking to eliminate 

these errors. Desired driving-wheel velocities are got by 

synthesizing the speed difference and AGV moving speed. 

Actual velocities are detected by processing the encoder 

signals and the errors between them and desired values 

are the inputs of servo controller. PID controller regulates 

driving-wheel velocities by changing motor voltages. Its


performance is associated with its PID parameters, and 

parameter tuning needs the model of AGV driving system. 

In order to get the real-time data in the experiments of 

system identification and controller optimization, remote 

control software is developed on the host computer. The 

vehicular controller transmits the real-time data of speed 

difference, desired velocities and actual velocities to the 

host computer by using wireless communication devices. 

Then the host computer saves the experiment data to the 

database, on which different algorithms can be analyzed 

and compared effectively based. 

V. EXPERIMENTS AND ANALYSIS 

This section uses the MOGA to identify system model 

and optimize servo controller. The actual velocity of 

driving wheel is recorded in the step response experiment, 

and the plant response curve is plotted according to them. 

Different GAs are used here and their optimization results 

are compared. 

A. System Identification by MOGA 

Give a step voltage to AGV motor driver, and record 

the real-time data of actual velocity in the start-up process. 

The plant response curve is plotted as the solid curve in 

Fig. 4. The second-order model identified by least square 

method (LSM) has the response curve as the dashed 

curve in Fig. 4, which is largely different from the plant 

response curve. GA is used to optimize model parameters 

in the following part. Let the perfect second-order model 

of driving system is 

G( 

z) 

b z � b 

* 1 2 � (9) 

2 

z � a1z 

� a2 

Where the object optimized by GA is the parameter 

vector X � [ a1 

a2 

b1 

b2 

] . The vector identified by 

LSM is 

LSM 

X � [�1. 

514 0. 

6704 0. 

005 0. 

1504] 

. 

According to the above result, these parameters are 

limited to the following range. 

� 2 � a 1 � �1, 

0 2 1 � � a , 1 . 0 0 1 � � b , 3 . 0 0 2 � � b . 

Firstly, conventional single-objective genetic algorithm 

(SOGA) is used here. ITSE (integral of the time and the 


Figure 3. Vision-based AGV prototype. 

squared error) objective function is adopted regarding the 

fast rising response curve in the start-up process, shown 

F 

A 

� 

n 

� 

i�1 

* 2 

T ( y � y ) 

(10) 

i 

* 

Where y and i y are the model response output and 

i 

the plant response output at sampling time T . i 

Set the following parameters for SOGA in terms of 

model order and data number. Population scale is N=60. 

The maximum number of population generation is G=100. 

Crossover probability is P c � 0. 

6 . Mutation probability is 

P m � 0. 

2 . SOGA runs randomly 5 times, and the model 

parameters identified by SOGA are listed in table I. 

TABLE I. 

MODEL PARAMETERS IDENTIFIED BY SOGA 

Number Fitness Model parameters 

1 2.9908 -1.6670 0.7969 0.0470 0.0830 

2 2.4510 -1.5437 0.7050 0.0170 0.1433 

3 2.2540 -1.5069 0.6712 0.0265 0.1369 

4 2.7409 -1.6122 0.7535 0.0365 0.1048 

5 2.0128 -1.3975 0.5887 0.0204 0.1707 

In table I, the first model has the highest fitness, and its 

step response curve is shown as the solid-dotted curve in 

Fig.4. It is obvious that the solid-dotted curve approaches 

to the solid curve more closely than the dashed curve 

both at amplitude and at phase in the first wave top and 

bottom, but it has an increasingly larger phase error than 

the dashed curve from the second wave top. 

Figure 4. Step response curve of model optimized by SOGA. 

This phenomenon does not imply that GA is inferior in 

phase optimization to LSM. Careful analysis reveals the 

cause. The objective function (10) defines the amplitude 

error between the model response output and the plant 

response output. Minimizing the amplitude error of the 

perfect response curve can also achieve a minimum of the 

phase error. However, it is almost impossible to minimize 

two errors of the practical response curve simultaneously 

because of many distortions. Only using the objective 

function of amplitude error in GA unavoidably results in 

the lack of phase precision in this optimization process. 

So two objective functions including amplitude error and 

phase error need to be used, and SOGA needs to be 

replaced with the MOGA as well. The phase error is 

defined as 

i 

i


F 

P 

� 

m 

� 

i�1 

* 2 

w ( O � O ) (11) 

* 

Where O and i O are the sampling number of the i-th 

i 

wave top or bottom for the model response output and the 

plant response output. w is the weight related to i. 

i 

The multi-objective aggregated function is 

FC 1FA 

2 

i 

i 

i 

� � �� 

F 

(12) 

Where 1 � and 2 � are the weights. Set 1 � = 2 

� =0.5 for 

a balanced optimization on two objectives. 

Since the model identified by LSM has a balanced 

precision both at amplitude and at phase, the first elitist 

selection tactics are used by employing the parameter 

LSM 

vector X as the evolution tutor, and leading the entire 

population to evolve towards the direction that is Pareto 

LSM 

superior to X . The selection tactics can decrease the 

negative influence of fixed weights to the aggregated 

function (12), which limits all objective items to a finite 

variable range. 

The MOGA proposed uses the same population scale 

and maximum number of population generation as SOGA. 

The probability parameter is q =0.1, the initial value and 

1 

incremental value of mutation amplitude is B =0.5 and 

e 

2 

B =0.5, and the initial value and incremental value of 

e 

crossover amplitude is 1 C =0.5 and C =0.5. MOGA runs 

2 

randomly 5 times, and the model parameters identified by 

MOGA are listed in table II. 

TABLE II. 

MODEL PARAMETERS IDENTIFIED BY MOGA 

Number Amplitude 

error 

Phase 

error 

Model parameters 

1 28.2715 47 -1.6781 0.8396 0.0004 0.1599 

2 28.4033 47 -1.6762 0.8375 0.0022 0.1580 

3 28.3686 47 -1.6750 0.8367 0.0005 0.1605 

4 28.6065 47 -1.6706 0.8320 0.0035 0.1568 

5 28.7471 47 -1.6630 0.8252 0.0026 0.1588 

In table II, model parameters are optimized directly by 

minimizing two objective functions of amplitude error 

and phase error. Five groups of models are similar to each 

other, which reflect a better convergence of MOGA than 

that of SOGA. The step response curve of the first model 

identified by MOGA is shown as the solid-dotted curve in 

Fig.5. This curve approaches to the solid curve more 

closely than the dashed curve both at amplitude and at 

phase in the first three wave tops and bottoms, which 

shows that the model identified by MOGA has a balanced 

high precision both at amplitude and at phase. 

Driving system model is identified by MOGA as 

0. 

0004z 

� 0. 

1599 

G ( z) 

� 

(13) 

2 

z �1. 

6781z 

� 0. 

8396 

B. Controller Optimization by MOGA 

Ziegler-Nichols method is used to tune PID parameters 

for the second-order model (13). 


P 

k =0.6, P 

I k = 5.5245, k =0.0163. 

D 

The object optimized by GA is the parameter vector 

K � [ k P k D k I ] , and their ranges are 

0 � k P � 1, 

0 � k D � 0. 

2, 

0 � k I � 30 . 

Firstly, weighted sum genetic algorithm (WSGA) is 

used here. There objectives combine with each other to 

form a weighted sum function, shown as 

� 

F 3 

2 

� ( w1 

| e( 

t) 

| t � w2u 

( t) 

� w4 

| ey( 

t) 

|) dt � w t (14) r 

Where e (t) 

is the error between the response output 

and the desired output. u (t) 

is the control input. ey (t) 

is 

the overshoot error when the response output overshoots. 

The weights are set as following to avoid overshoot [6,8]. 

w =0.999, 1 

2 w =0.001, w =2, 3 w =200 4 

WSGA uses the same population scale and maximum 

number of population generation as SOGA. It runs 5 

times to optimize the controller for driving system model 

(13), and PID parameters are listed in table III. These 

parameters are very similar, overshoot is almost zero, and 

rising time is equal to setting time. 

k 

Figure 5. Step response curve of model optimized by MOGA. 

P 

TABLE III. 

CONTROLLER PARAMETERS OPTIMIZED BY WSGA 

k 

D 

k F � /% 

I 

t /s t /s s 

0 0.0174 8.3978 8.0267 0.11 0.24 0.24 

0 0.0175 8.4764 8.0594 0.13 0.24 0.24 

0.0014 0.0178 8.6152 8.1217 0.18 0.22 0.22 

0 0.0174 8.4215 8.0379 0.12 0.24 0.24 

0.0018 0.0177 8.5431 8.0176 0.21 0.24 0.24 

Servo controller is designed by using the first group of 

parameters with the smallest overshoot, the third group 

with the highest fitness, and the fifth group with the 

lowest fitness. The step response curves of driving system 

model (13) are shown as the solid-dotted curve, the solid 

curve and the dashed curve in Fig.6. Three response 

curves superpose with each other, and the driving system 

has a step response without any oscillation. Weight 4 w 

related to overshoot error is much larger than other 

weights in (14), and the severe punishment to overshoot 

decreases system response speed unavoidably. 

It is seen that weight selection has a great influence on 

the optimization results of WSGA, and it is difficult to 

get a compromise between speed and stability of system 

r


response. The MOGA with Pareto optimality and elitist 

tactics is proposed to avoid the negative influence. Rising 

t and integral absolute error (IAE) 

time t , settling time r 

s 

| e | are adopted as multi-objective functions, shown as 

F � y 

3 

F � t 

1 

err 

� 

r 

F � t 

2 

s 

n 

� 

i�1 

| e | 

Figure 6. Step response curve of controller optimized by WSGA. 

i 

(15) 

AGV servo control system is required to achieve a fast 

rising and settling response output for velocity regulating 

in path tracking [2], and decision-making preferences for 

Pareto optimal solutions are defined as following. 

(1) If t >1.5s, settling time of driving system is too 

s 

long to satisfy velocity regulating in path tracking. Delete 

the Pareto optimal solutions. 

Number 

1 

2 

3 

4 

Figure 7. Step response curve of controller optimized by MOGA. 


5 

k 

TABLE IV. 

CONTROLLER PARAMETERS OPTIMIZED BY MOGA 

k 

P D 

I 

0.0108 0.0398 18.9429 

0.0107 0.0380 17.6868 

0.0020 0.0383 17.8309 

0.0016 0.0401 19.1172 

0.0004 0.0402 19.2231 

0.0003 0.0387 17.9431 

0.0050 0.0382 17.7343 

0.0046 0.0402 19.1693 

0.0047 0.0383 17.7930 

0.0046 0.0399 19.0549 

k | e | 

t /s 

r 

3.2082 0.10 

3.2653 0.10 

3.2372 0.10 

3.1575 0.10 

3.1596 0.10 

3.2382 0.10 

3.2536 0.10 

3.1760 0.10 

3.2434 0.10 

3.1736 0.10 

(2) If t ≤0.1s and r t ≤0.2s, driving system achieves a 

s 

fast and smooth response for velocity regulating. Select 

the Pareto optimal solutions as elitists. 

(3) If no elitist exists, compare the multi-objective 

optimization performance of Pareto optimal solutions by 

the aggregated function 

F � t � t 

(16) 

C 

MOGA adopts the same parameters as the above subsection 

except using the maximum generation number 

when all elitists remain without any change continuously 

is G =10. MOGA runs 5 times to optimize the controller 

k 

for driving system model (13), and PID parameters are 

listed in table IV. It shows that MOGA can find multiple 

Pareto optimal solutions rather than only one according to 

decision-making preferences, which is different from 

WSGA essentially. The generation number of MOGA is 

only half to that of WSGA, and PID parameters have the 

similar components and performance, which shows that 

MOGA converges to Pareto front interested by decisionmaker 

without falling into the local minimum trap. Servo 

controller is designed by using two groups of parameters 

in the second test, and the step response curves of driving 

system model (13) are shown as the solid curve and the 

dashed curve in Fig.7. Two curves have a shorter rising 

time and settling time than those in Fig.6. Although their 

overshoots are a little larger than that of WSGA, the 

value of less than 10% can still ensure a smooth step 

response output and no more than one oscillation. 

t /s 

s � /s 

0.16 8.68 

0.10 4.98 

0.10 4.95 

0.16 8.73 

0.16 9.03 

0.10 4.97 

0.10 4.74 

0.16 9.08 

0.10 4.97 

0.16 8.71 

Regarding the difference between the second-order 

model (13) and the actual driving system, the first group 

of PID parameters is used to design servo controller on 

ARM LPC2220 and RTOS μC/OS-II. Plot the actual 

response curve of AGV driving system as the solid-dotted 

curve in Fig.7. Although rising time, settling time and 

overshoot of this actual curve have some increase, this 

influence is not so significant to decrease the controller 

performance obviously. In AGV movement control test, 

our servo controller still has the satisfactory performance 

of velocity regulating for path tracking. 

2 

r 

2 

s 

Generation numbers 

when convergence 

35 

50 

68 

43 

32 

VI. CONCLUSION


A cascaded MOGA is used to identify AGV driving 

system model and optimize its servo control system in 

this paper. Pareto optimality is used in genetic algorithm 

to guarantee solutions with different tradeoffs for multiobjective 

optimization. Elitist selection, neighborhood 

mutation and heuristic crossover are combined to expand 

the influence of decision-making preferences and make a 

directional search in Pareto front. LFPM, MNID and 

MEM are combined to enhance the fitness and diversity 

of the entire population. Experiment results show that the 

cascaded MOGA have the capability to make the system 

model consistent with AGV driving system well, and to 

make its servo control system satisfy the requirements on 

dynamic performance and steady-state accuracy in AGV 

path tracking 


This work was supported in part by a grant from 

NUAA Research Funding (Grant No.NJ2010025) and 

Research Start-up Funding (Grant No.S1026-053). 

REFERENCES 

[1] Kelly A., Nagy B., Stager D., et al, “An infrastructure-free 

automated guided vehicle based on computer vision,” 

IEEE Robotics and Automation Magazine, Vol.14, No.3 

(September 2007), pp.24-34 

[2] Wu Xing, Lou Peihuang, “Optimal path tracking control 

based on motion prediction,” Control and Decision, Vol. 

24, No.4 (April 2009), pp. 565-569. 

[3] Shin G. W., Song Y. J., Lee T. B, “Genetic algorithm for 

identification of time delay systems from step responses,” 

International Journal of Control, Automation and Systems, 

Vol.5, No.1 (February 2007), pp.79-85. 

[4] Tan Xin, Yang Huaqian, “The optimization of nonlinear 

systems identification based on genetic algorithms,” 

Proceedings of International Conference on 

Computational Intelligence and Security, Guangzhou, 

China, October 2006, pp. 266-269. 

[5] Meng X. Z., Song B. Y., “Fast genetic algorithms used for 

PID parameter optimization,” Proceedings of IEEE 

International Conference on Automation and Logistic, 

Jinan, China, August 2007, pp.2144-2148. 

[6] Zhang J. H., Zhuang J., Du H. F., et al, “PID controller 

optimization based on the self-organization genetic 

algorithm with cyclic mutation,” Proceedings of the 6th 

International Conference on Artificial Intelligence, 

Aguascalientes, Mexico, November 2008, pp. 277-284. 

[7] Tan G. Z., Jiang B., Yang L. M., “A novel immune genetic 

algorithm-based PID controller design and its application 

to CIP-I intelligent leg,” Proceedings of the 3rd 

International Conference on Natural Computation, Haikou, 

China, August 2007, pp. 282-286. 

[8] Ding Y. M., Wang X. Y., “Real-coded adaptive genetic 

algorithm applied to PID parameter optimization on a 6R 

manipulator,” Proceedings of the 4th International 

Conference on Natural Computation, Jinan, China, 

October 2008, pp. 635-639. 

[9] Arruda L.V. R., Swiech M. C. S., Delgado M. R. B., et al, 

“PID control of MIMO process based on rank niching 


genetic algorithm,” Applied Intelligence, Vol.29, No.3 

(December 2008), pp. 290-305. 

[10] Wang Guoliang, Yan Weiwu, Shao Huihe, “Multiobjective 

optimization based on genetic algorithm for PID 

controller tuning,” Journal of Harbin Institute of 

Technology, Vol.16, No.1 (February 2009), pp. 71- 74. 

[11] Dionisio S. P., Joao O. P. P., “Genetic algorithm based 

system identification and PID tuning for optimum adaptive 

control,” Proceedings of IEEE/ASME International 

Conference on Advanced Intelligent Mechatronics, 

Monterey, United states, July 2005, pp.801-806. 

[12] Valarmathi K., Devaraj D., Radhakrishnan T. K., “Realcoded 

genetic algorithm for system identification and 

controller tuning,” Applied Mathematical Modelling, 

Vol.33, No.8 (August 2009), pp.3392-3401. 

[13] Schaffer J. D., “Multiple objective optimization with 

vector evaluated genetic algorithms,” Proceedings of the 

1st International Conference on Genetic Algorithms, 

Hillsdale, Canada, 1985, pp.93-100. 

[14] Fonseca C. M., Fleming P. J., “Genetic algorithms for 

multiobjective optimization: formulation, discussion and 

generalization,” Australian Electronics Engineering, 

Vol.27, No.2 (February 1994), pp.416-423. 

[15] Srinivas N., Deb K., “Multiobjective optimization using 

nondominated sorting in genetic algorithm,” Evolutionary 

Computation, Vol.2, No.3 (1994), pp.221-248. 

[16] Deb K., Pratap A., Agarwal S., “A fast and elitist 

multiobjective genetic algorithm: NSGA-II,” IEEE 

Transaction on Evolutionary Computation, Vol.6, No.2 

(April 2002), pp.182-197. 

[17] Zhao Liang, Ju Gang, Lu Jianhong, “An improved genetic 

algorithm in multi-objective optimization and its 

application,” Proceedings of Chinese Society for Electrical 

Engineering, Vol.28, No.2 (January 2008), pp. 96-102. 

[18] Qi Rongbin, Qian Feng, Du Wenli, et al, “Multiobjective 

genetic algorithm based on elitist selection and individual 

migration,” Control and Decision, Vol.22, No.2 (February 

2007), pp.164-168. 

WU Xing was born in China in 1982. He received his Doctor 

Diploma on Mechanical Engineering from Nanjing University 

of Aeronautics and Astronautics in 2010. He holds a lecturer 

position in NUAA now, and acts as the key member of some 

important projects such as National Natural Science Foundation 

of China. His research interests include mobile robot, motion 

control and embedded system control. 

LOU Peihuang was born in China in 1962. He is the dean of 

Jincheng College, NUAA now and holds a professor position 

since 2001. He acts as the chief leader of some important 

projects such as Special Project of Jiangsu Province for the 

Transformation of scientific and technological achievements. 

His research interests include manufacturing system control and 

fault diagnosis. He has won the first prize for scientific and 

technological progress of Jiangsu province 1 time, the second 

prize 2 times and the third prize 5 times. 

TANG Dunbing was born in China in 1972. He is the 

deputy dean of Mechanical and Electrical Engineering 

Department, NUAA now and holds a professor position since 

2005. He acts as the chief leader of some important projects 

such as National Natural Science Foundation of China. His 

research interests include creative design and complex system 

modeling. He has won the first and second prize for scientific 

and technological progress of Jiangsu province in 2008.


WebVR——Web Virtual Reality Engine Based 

on P2P network 

Zhihan Lv 

College of Information Science and Engineering, Ocean University of China, QingDao, China 

CNRS UPR9080/ IBPC, 13 rue Pierre et Marie Curie, F-75005, Paris, France 

Email: lvzhihan@gmail.com, lu@ibpc.fr 

Tengfei Yin, Yong Han, Yong Chen, Ge Chen* 

College of Information Science and Engineering, Ocean University of China, QingDao, China 

� Abstract: WebVR, a multi-user online virtual reality engine, 

is introduced. The main contributions are mapping the 

geographical space and virtual space to the P2P overlay 

network space, and dividing the three spaces by quad-tree 

method. The geocoding is identified with Hash value, which 

is used to index the user list, terrain data, and the model 

object data. Sharing of data through improved Kademlia 

network model is designed and implemented. In this model, 

XOR algorithm is used to calculate the distance of the 

virtual space. The model greatly improves the hit rate of 3D 

geographic data search under P2P overlay network. Some 

data preprocessing methods have been adopted to accelerate 

the data transfer. 3D Global data is used for testing the 

engine. The test result indicates that, without considering 

the client bandwidth limit, the more users, the faster 

loading. 

Keyword: Virtual Reality; P2P; WebVR; Web3D; GIS; 

Geocoding; Kademlia 


Geographic Information Science is developed on the 

basis of Geography, Cartography, Surveying and 

Computer Science Disciplines, the software entity 

implied based on which is Geographic Information 

System (GIS). Virtual Reality (VR) technology is 

reflection of the real world to simulate and generate a 

three-dimensional virtual space with computer. The 

combination of Geographic Information System and 

Virtual Reality technology generates VRGIS, which not 

only possesses GIS function such as spatial data storage, 

process, query and analysis, but significant improves 

friendly interface and intuitive interaction combined with 

VR technology. 

With the developing of Internet era this century, 

theories and practices combined with Web demonstrates 

its unprecedented vigor and vitality, and WebVR-GIS 

becomes the inevitable outcome of this trend, which 

implies VR-GIS on Internet, providing data sharing, 

collaborative roaming and GIS analysis function, while 

shows the whole “global virtual environment” in a scene. 

One case of WebVR-GIS extension to multi-user is Web 

Virtual Environment, where each user has a virtual role. 

Virtual scene inflects with real environment and efficient 

information sharing makes each virtual role interactive 

entertainment and work together free from limits of time 

� *Corresponding author, email: webvr@vip.qq.com 


doi:10.4304/jnw.6.7.990-998 

and spatial. It is necessary that virtual districts turn to 

virtual cities and virtual earth in virtual environment for 

the rapid increase of users and people explore desire for 

new things. Virtual earth is the most macro-scale 

implement of virtual environment. People are inspired by 

its real reflection of real world, and eagerly waiting its 

characteristics such as massive user collaborative 

interaction, massive virtual environment data sharing, 

global real geographic location reflection, etc. 

However, a series of new problems will arise while the 

level of virtual environment extends to earth. For 

example: (1) Global geographic location is complex, so 

the space partition methods of traditional GIS based on 

topological can’t meet its demand; (2) Global virtual 

scene data is enormous, which can’t be deposited from 

inside and external memory at one-time; (3) Global-scale 

nodes was enormous, frequent changes, unpredictable 

behavior, so the controllability faces tough challenges. 

The normal solutions for the three problems above are 

as follows: Common methods known as latitude and 

longitude region division model, map projection division 

model and Voronoi diagram region division used in 

geographic information system are used to divide GIS 

region (geographic area dividing) efficiently. The 

problem that loading enormous global virtual scene data 

is attributed to “global spatial data partition model”. 

Earth model division method in three-dimensional space 

include traditional “grid (cell) partition”, “G2PS model”, 

etc. The reasonable division and organization of earth 

model will reduce virtual scene data effectively. The third 

problem can be attributed as “distribute network model”, 

which is much related with hardware. The better solution 

is server cluster [1] technology. For example, the terrain 

and image database of Google Earth containing 70T in 

2007 is support with a huge “cloud storage” server cluster 

in Google Inc. “Second Life” use each computer to 

simulate 90 square meters of virtual scenes, which has 

5000 servers running now in Linden Inc. 

In practice, firstly, the effectiveness and feasibility of 

“geographic area division” and “global special partition 

model” is eagerly to be improved. Secondly, global net 

supporting similar “cloud computing storage” hardware 

requirements coupled with higher maintenance costs will 

discourage many institutions. At the same time, the 

existing distributed network algorithms are with virtual 

network (P2P) on the basis of DHT algorithm based, the 

user ID of which is mostly based on logic distance, rather


than related to geographic location, and even network 

address such as IP. It can reduce index time for user 

based on logic distance, but bandwidth waste and traffic 

congestion while file transferring cross-regional and 

cross-country has become a problem that can’t be 

ignored. 

Taking all the above into consideration, we propose a 

new space division method to encoding map global 

geographic environment model based on Hash, which 

makes it apply to divide real world, virtual scene and 

network, and be the spatial division method of overlap 

world in network virtual environment. The method 

belongs to the reasonable combination of geographic 

information science, virtual reality and virtual network, 

with higher innovation and better scalability. The 

universal of the method will take “earth virtual 

environment” to everyone, providing global users a 

platform in which interacting with 3D avatar in virtual 

reality environment. All these will not only shorten 

distance between people, but also provide a better data 

share and cooperation means for science research work, 

while provide data base and theoretical basis for 

interaction between people and environment. 

II. RESEARCH 

Through the problems on the P2P network of virtual 

geographic environment, the following points are worthy 

of studying: 

(1) A supporting mass, précising model of the real 

world environment, classification and indexing of space, 

mapping with the real location, enhancing fidelity, meet 

the deep planning needs on three-dimensional data. 

Cavagna R., C. Bouville, J. Royan developed a P2P 

model based on theory of space division of Voronoi 

diagram [2] , for the transfer of cites` three-dimensional 

scene, and planed to use for online games in the future . 

However, its support of streaming three-dimensional 

scene index structure PBTree [3] for the 2.5-dimensional 

data, with a high degree of property from 

two-dimensional vector data compression out of the 

three-dimensional model of the true model does not 

support the fine. M. Varvello, C. Diot, and EW Biersack 

combined the KAD network model with the 

mathematical model of the virtual environment [4] , and 

tested in Second life as a framework [5] . He still used the 

original data partitioning Second life and organization as 

the core model, without importing space division and the 

advanced theory of real-world mapping. 

(2) It can improve stability and data retrieval 

efficiency by using distributed Hash structure, which is 

based on space partition to virtual index structure in peer 

networks. In distributed network, frequently joining in or 

quitting of nodes will cause a large number of networks 

Churn, which on the overall robustness of the distributed 

network architecture puts forward higher requirements. 

Hu SY, TH Huang, SC Chang, et al. have developed a set 

of P2P model Flod [6] based on the Voronoi diagram 

theory of the space partition. The model has solved the 

problem of user neighbors distributed storage. Dynamic 

classification method applies to rapidly changing data in 


3D game scene. However, the P2P model based on the 

theory of division space is constantly in a dynamic, it will 

trigger the whole user list traversal each time the user 

moves, which costs a lot of computing performance. 

Under the conditions of the existing hardware, it can not 

be extended to the global field. The global scope of the 

real geo-spatial environment is relatively stationary; there 

is no need of dynamic division. Therefore, Approach 

with pre-partition manner can completely make the land 

division, without changing zoning process with the 

mobile node, greatly improving the system running in 

real-time. Using the distributed Hash structure, based on 

space partition to store, can realize retrieve distributed 

resources, achieve the load balancing requirements, and 

also improve the hit rate and data retrieval efficiency. 

(3) Creating a set of the space partition model applied 

to the real world, the global virtual environment and the 

global network structure can organically couple 

geographic addresses, network addresses and user 

identity. In the regular research, geo-coding, spatial data 

partition model and the virtual peer networks are as an 

independent branch of study. Use the virtual network 

architecture on the virtual environment scene to share 

data, while associate the information of virtual 

geographical world and real users, can effectively solve 

the model for low precision, a large amount of data, 

multi-user interaction delay, waste of network bandwidth 

optimization and other issues. 

III. SYSTEM OVERVIEW 

A. Data flow 

The design principle is trying to support more load 

quantity of online users, to reduce the data transmission 

quantity, to improve data compression ratio, and to 

increase the data download source. Different types of 

data are preprocessed, and the result is stored to data 

servers. 

While the client is browsing scene, according to the 

neighborhood search strategy, it choose the source of data 

transfer, loading from server if the source can`t be find. 

Data 

Data 

Server 

Client 

Client 

Figure 1 Data Flow 

Client 

�. THE KEY ISSUES


A. Graph-based scene topology 

The block index based on Hash 

Terrain and 

scene block 

index 

Spherical Space 

Geocode 

d b 

The client cache 

WebVR rendering 

The scene graph is managed by hierarchical bounding 

box, using bounding sphere and bounding box to achieve 

the scene bounding volume hierarchy. The information is 

stored by a directed acyclic graph structure. A scene 

graph includes a root node, multi-level interior of the side 

nodes, and multiple terminal leaf nodes. The root and 

side nodes take charge of the construction of the level of 


Model and 

other kinds 

of index 

Judge AOI 

Make the data in AOI into 

Download Request list 

Select nearby nodes 

Through the XOR 

Send Download Request List to selected nodes 

On-demand loading 

the nodes, and the completion of certain functions; the 

leaf node is saved to one or more object information can 

Geocode Based On Hash 

Node roaming 

Node 

Localization 

based on Hash 

Geocode 

Figure 2 Engine Architecture 

Node List 

Search node 

Observe User interaction 

and data transmission 

Summarize the 

mathematical model and 

test key data 

Join 

Withdraw 

be drawn. Each node maintains its own bounding volume, 

and so on, constitute a distinct level. This level bounding 

box diagram can speed up the correct information in the 

expression of the composition of the scene graph, and 

also expedite the reduction of scene objects, intersection 

tests, collision detection and a series of operations. This 

structure allows each node to have multiple parent


nodes. When the same geometric object needs to be 

repeatedly referenced by more than one parent node 

pointing to the same child node, with each parent node 

pointing to a new child node of the tree than the total 

number of nodes, memorizing utilization and scene 

traversal steps reduced, rendering the final results remain 

unchanged. 

B. Data file partition 

The nodes in scene graph include terrain, objects and 

other types, based on different data types, different 

partition methods are adopted. 

Quadtree-based multi-scale geographic data block 

Geographic data includes terrain model using Tin 

Triangulation, the real image in the terrain covering the 

surface, and vector data. These geographic data are the 

surface data and little overlap in the vertical direction, so 

as to evenly split the region based on quadtree 

classification and index structure, each quadtree level 

represents a level of precision. An example of mature 

application is worldwind [7] . Quadtree construction 

processes the whole terrain as root, starting from the root 

node, checking whether the root partition to satisfy 

certain conditions. If it is not satisfied, not partition, it 

will be used as a leaf node preserved. Otherwise, 

recursively to the root node continuously divided into 

four equal sub-regional nodes, until it not split up any 

longer. The last step is drawing and rendering all the leaf 

nodes. The greater depth of division, the resolution will 

be higher. That is, each raising separate layer of depth, 

sampling density doubled. For the earth's surface, it is the 

need for separate ways after a projection from the plane 

to the latitude and longitude of the projection 

transformation. 

Figure 3 Quadtree block 

Object node as unit data block 

The scene contains a variety of objects nodes, 

including the following ones. (1) Construction 

information extruded from attributes information with a 

high degree of vector data; (2)3D model information 

import from 3DsMax; (3) geometry data. In the traversal 

of scene, each node of the outermost layer under the root 

is considered as a unit. 

C. Multi-scale data preprocessing 

The topographic data according to the different 

quadtree levels, divided the LOD data and stored it to 

external memory. For object nodes, each node object to 

the crude unit are generated from the refined precision of 

the data L1 4 to L4, in which L2 and L3 as L1 generated 


by collapse of law on the simplified model, L4 Impostor 

generated by image cache node. The texture object node 

based on cell aspect ratios of 2:1 generated three 

simplified texture memory to external memory, L2-L4 in 

four levels corresponding to the simplified model. The 

texture data is compressed as DTX3 by GPGPU in the 

way of calling the CUDA library. External memory 

models in different scales in turn are called as needed, 

more efficient than single MIPMAP file to be transferred. 

D. Data request depends on culling result 

For large-scale scenes, when a large number of models 

are read into memory, the computer system will 

inevitably result in a huge burden and could lead to 

insufficient memory. At this point, we need a dynamic 

scheduling mechanism. Time is unidirectional and cannot 

accurately predict the behavior of users in the future, and 

therefore the data loading to pre-deployment of 

space-related scheduling. Dynamic scheduling can 

produce the node when he was on the scene of some child 

nodes, while drawing long term without any participation. 

Child nodes can be automatically uninstalled, free 

memory space; On the contrary, he cannot load certain 

child nodes in memory, the dynamic scheduling of its 

control of the scene sub-tree. 

E. Memory release in time 

Using the time-related scheduling on the data 

withdraw after called. LOD node of a level of detail 

rendering scenes, if not involved in long-term, it will 

uninstall it, otherwise it is loaded. Design of smart 

pointers in the realization of the base class for all nodes, 

effectively prevent the memory leak caused by 

incomplete release. 

F. P2P-based data sharing 

Advanced Kademlia-based protocol scheduling discipline 

Node ID and files are geo-coded with HASH for an 

index purpose. 

1. Building geo-coding database 

Multiple data formats are involved in this research, 

such as vector data, DEM data, image Data, and 

three-dimensional model mesh and texture data. It is 

necessary to build up an index for data after creating the 

earth three-dimensional model. First, vector data, 

DEM data and image Data are partitioned in a multi-scale 

way according to space information. Each level is labeled 

by a 16-bit binary HASH which called “prefix”. After 

eight-times partition, each type of data can be expressed 

as a 128-bit binary HASH, which is made up of eight 

prefixes in the order left to right, the most significant byte 

is the first-scale data (big-endian byte order). The 128-bit 

binary HASH will be used as its geo-coding, which also 

mapping the sign of different scales. Second, as for the 

three-dimensional model mesh and texture data. The first 

7 partition are combined in the same way as vector data 

and DEM data. Accordingly, the result is a 112-bit binary 

HASH. Meanwhile, the three-dimensional model itself is 

also labeled by a 16-bit binary HASH, adding the 

preceding result then the three-dimensional model mesh


and texture data are expressed as a 128-bit binary HASH 

as well. Similarly, its 128-bit binary HASH will be used 

as its geo-coding, which mapping the sign of different 

scales. Next, the geo-coding and corresponding mappings 

of various data formats are put into different fields of a 

database and established an index. In addition, filename 

of data are renamed to relevant geo-coding. 

In a Kademlia network, node’s ID is created randomly 

and data’s ID is generated according to the contents of 

documents 

[8] , therefore, there is time correlativity 

between data and their forms of expression. By contrast, 

the HASH ID of our study is calculated by using the 

space division method, which has space correlativity. 

128bit Hash 

128bit Hash 

160bit Hash 

Index of terrain 

2 . Constructing Peer-to-Peer network based on 

geocoding 

For each user, its final ID used for positioning is made 

up of two parts. The first part is the geo-coding that is the 

first eight geographical prefix, which is relevant to its 

current location; the second one is its 32-bit binary user 

ID. 

On user neighbor lists, every node keeps a list of 

neighbor’s information( IP address, UDP port, Node ID). 

Those lists are stored in a quadtree. The quadtree is 

divided into 7 layers and each layer is marked with a 

prefix. Every neighbor’s 32-bit binary user ID is included 

in the leaf nodes. 

Logical distance. Given two 160-bit identifiers, A and 

B, we define the distance between them as their bitwise 

exclusive or (XOR) interpreted as an integer, d (A, B) = 

A�B . T he distance is related to their 

geographical position, during to the fact that the 

generation of node’s ID which depends on geographical 

division. Consequently, smaller distance means they are 

geographically closer. 

User neighbor lists generating. The locale-sensitive 

area of nodes is a circle with radius-r. 

The adjacent area is formed by the intersection of this 

circle and the spatial-quadtree leaf nodes. If the 

adjacent area has no nodes at all, the radius-r would 

become bigger constantly until there is node being 

contained in. At the same time, all the nodes being 


Index of nodes object 

Index of users 

contained in the adjacent area are added into the user 

neighbor lists. 

3. Joining and quitting network 

Every time when the node logs on to the network, it is 

expected to finish three tasks. They are generating its ID, 

generating its neighbor lists, and downloading required 

item of data according to its new location. 

When node quit the network, it is impossible for other 

nodes to receive this node’s response. As a result, this 

node will be regarded as off-line. Thus, it is predicted 

that the stability of network won’t be affected. 

Figure 4 Index Structure 

4. Nodes shifting 

There are mainly four changes when a node is moving 

from one place to another. First, its ID will be 

regenerated quickly. Second, starting to search required 

data on the basis of the new position. Third, its neighbor 

lists will be regenerated as well. Fourth, the node has to 

inform its neighbors of leaving, send request to its new 

neighbors and ask for establishing friendship, which 

aiming to insert itself into other nodes’ neighbor lists as 

necessary. 

Figure 5 Node AOI change 

5. Neighborhood Selection 

Some software for file sharing, such as emule would 

be in operative condition except at zero searching hit 

rates. We need to guarantee the occurrence rate of the 

geographical neighbors, or we may lose them. We can 

reach the conclusion that the node searching way of DHT 

cannot be used in virtual environment.


After trigger event of neighborhood selection 

occurring, the peer send its information, I for short, to 

any neighbor in its neighbors list. I, will go through the 

following process when any peer which receives the 

information. P for short, calculating its geocode ID by the 

first 128 bit hashID of the sending peer, then judge 

whether the receiving peer owns the same geocode by 

XOR algorithm. If the answer is yes, it will send I to its 

neighbors list, then any neighbor receives I executes P, 

which is a recursive process. It occurs under the 

following conditions. If the hashID of peer receiving 

information is different, from me at the geocode bits 

between 64 and 128, then P stops. The condition of all 

the searching process stops, the distance between peers 

executing P and starting peer is beyond the value of 

(scene side/24). It is different from the method of limiting 

recursion frequency by depth in Kadmelia. In fact the 

whole process is distributed traversal. 

6. Data Searching 

When trigger event of data searching occurs, the peer 

will recalculate the logical distance between any peer in 

its neighbors list and itself. That is XOR value between 

its ID and other IDs. Then it sends data index value to the 

nearest n peers, after any peer online receives the 

information. It doesn’t only execute local research, but 

also sends geo-code info to its peers list, and so on. 

According to the small world theory, we could get the 

data we need when the recursion frequency is 6. We 

make the recursion frequency is 8 by default, which is 

used by the radius of data searching. When we get the 

data, we add the peer with the data into our local 

neighbors list. Then they begin to execute multi-source 

transmition. 

Server based on IOCP for large numbers of peers 

Because users’ list is stored among lots of clients, 

some clients may not add it into the network. One of the 

extreme situations is that, when a peer logins in, all of its 

users on the list are offline, so it can’t take part in the 

network. We call this situation “Information isolated 

island”. In order to prevent this situation, we designed the 

Server based on IOCP. It stores the recent login user info, 

including the user’s name, IP address, and geocode info. 

After each user login in the network, it will send its info 

to server to store. Then it will totally break the link with 

server. After it login in the server, it will download the 

latest 20 users without any active user on its users’ list, 

and add them into its neighbors list. The model of IOCP 

and multi-thread can be the best way to use resources, 

which supports lots of users. 

Transmission Strategy 

After data searching, we begin to transmit data 

according to resource list by searching. 

1. Cache Mechanism 

The process is manifested in two aspects. First, on the 

sending side, when file block is being sent, the 

mechanism will control the data bytes sent by a single 


running thread all the time, which is a limiting process. 

Second, on the receiving side, it will create a buffer with 

the same size with source file before file transmission. 

Transmission is pigeon-holing. When the software quits, 

it will remember every position of downloaded data in 

the file. When the software restarts, it will execute 

Resume from break point. 

During the transmission process, model data is pressed 

by zip-standard while texture data is compressed by 

Jpeg2000-standard. 

2. File Distribution 

In order to improve the number of data resources, we 

don’t only store the whole data in neighbor peers, but 

also peers nearby after file data is cut into blocks. On the 

data distribution side, cutting and calculating every file 

into a lot of blocks with the same size by hash algorithm, 

each block can generate unique ID by file geocode and 

block content. Distribute and store data resource by the 

principle of file HashID being closed to geocode on 

clients. The process is prepared for data searching. On the 

receiving side, searching process is based on parameter of 

file name. Ask some peers whose hashID is closed to the 

file name ID, if the answer is yes, then we can find some 

peers which owns the file, and ask for downloading. 

Figure 6 P2P file share structure 

Network adaptability 

Realize penetrating NAT by the way of holing, and 

penetrating firewall based on UPNP. 

�. IMPLEMENTATION AND TEST 

A. Implementation 

The component is encapsulated by Microsoft ATL 

library. Pack security certificate and the component into 

the style of CAB. Interact by the way of JavaScript 

calling components interfaces.


B. Test 

We use 1 ° × 1 ° global elevation and image data 

provided by Noaa web site to do data partitioning and 

geocoding index generatation for free. Among them, 

Qingdao City’s elevation and image data is high accuracy 

ASTER G-DEM data. Model data is part of building 

Data 

type 

Globe 

DEM 

Globe 

DOM 

Globe 

9 level 

LOD 

models of Qingdao. We use Sqlserver Database and do 

the geocoding of number and connectivity of parts of the 

database using C#. Rendering and the network part use 

C++. Some pages use script called component functions 

of JavaScript, and achieve interface effects based on exits 

library. 

Usually the rendering engine and networking engines 

are separately tested, such as people usually test the 

amount of data supported by rendering engine and its 

Qingdao 

DEM 

rendering efficiency, while testing network engine on the 

transmission efficiency. For P2P networks, Consistency, 

Persistency, Scalability and other properties are usually 

tested. Some researches on P2P networks of the virtual 

environment also test the changes in performance caused 

by AOI region changes. 

Because WebVR engine has compact architecture and 

Qingdao 

DOM 

Qingdao 

5 level 

LOD 

Building 

Model 

4 level 

LOD 

size 911M 683M 990M 25M 479M 15M 5G 

Time 87min 2min 900Min 

Figure 7 From Globe to City 


Table 1 Data preprocess time 

20 

15 

10 

it is tightly integrated with rendering, Data Division, 

Dispatch and network part, we just test the overall 

performance. 

The users number / User information server load 

Test the server bandwidth changes in the second 

period when users were respectively (1, 5, 10, 50, 100, 

and 200). They login in the user information server at the 

same time. The figure shows that the server only 

provided the neighbor list download in the initial phase. 

Then it disconnected with the node, no longer has traffic. 

The users number / Data server load 

Test the data server bandwidth changes in 10 seconds 

when the number of users were respectively (1, 5, 10, 50, 

100, 200), and they come into a new scene. The figure 

shows that when the node could not get the scene from 

other nodes, the data server provides data for download. 

5 

0 

0.1s 0.2s 0.3s 0.4s 0.5s 0.6s 0.7s 0.8s 0.9s 1s 

-5 

Figure 8 User information server load 

100 

80 

60 

40 

20 

0 

1s 

-20 

2s 3s 4s 5s 6s 7s 8s 9s 10s 

Figure 9 Data server load 

1 

5 

10 

50 

100 

200 

1 

5 

10 

50 

100 

200


The users number / Data loading speed 

Test the loading times changes when users were 

respectively (1, 5, 10, 50, 100, 200), and a node comes 

into a new scene. It can be observed that with the 

increasing of the number of users, data loading speed 

increased. 

800 

700 

600 

500 

400 

300 

200 

100 

0 

1 5 10 50 100 200 

Data size / Loading time and Data size / Rendering frame 

Test the data load time and rendering frame rate 

changes of a node when data volume is respectively (50, 

100, 200, 500, 1000, 5000) MB. When the amount of 

data increases by the various engine optimizations, it still 

maintains a smooth roaming rate. 

5 

4 

3 

2 

1 

Figure 10 Data loading speed 

0 

50 100 200 500 1000 5000 

Figure 11 Loading time 

80 

70 

60 

50 

40 

30 

20 

10 

0 

50 100 200 500 1000 5000 

�. CONCLUTION 

This paper introduced a multi-user online virtual 

reality engine. The main contributions are mapping the 

geographical space and virtual space to the P2P overlay 

network, and dividing the three spatial by 

business-oriented quad-tree method. The geographical 

code is identified with Hash value, which is used to index 

the user list, terrain data, and the model object data. 

Achieve sharing of data through improved Kadmelia 

model. In this model, XOR algorithm is used to calculate 

the distance of the virtual space. 


Figure 12 Rendering frame 


This research was supported by the Open Research 

Project of State Key Laboratory of Coal Resources and 

Safe Mining under Project SKLCRSM09KFB02 and 

Scientists and Engineers Serve Enterprises Program of 

Ministry of Science and Technology under Project 

2009GJA00047 

REFERENCES 

[1] Anderson T. E., D. E. Culler, D. A. Patterson, et al. 

A case for NOW (Networks of Workstations), IEEE 

Micro, 15(1):54--64, February 1995. 

[2] Cavagna R., C. Bouville, J. Royan, P2P Network 

for very large virtual environment, Proceedings of the 

ACM symposium on Virtual reality software and 

technology, 269-276, 2006. 

[3] Royan J., C. Bouville, P. Gioia, PBTree - A new 

progressive and hierarchical representation for 

network-based navigation in densely built urban 

environments, Annales des Télécommunications, 60, 

1394-1421, 2005. 

[4]M.Varvello, E. Biersack, and C. Diot. A networked 

virtual environment over KAD. In Proc. ACM CoNEXT 

conference (CoNEXT), pages 1-2, New York, NY, USA, 

December 2007. 

[5] M.Varvello, C.Diot, and E. W. Biersack. P2P 

Second Life: experimental validation using Kad. In 

Infocom 2009, 28th IEEE Conference on Computer 

Communications, pages 19-25, Rio de Janeiro, Brazil, 

April 2009. 

[6] Hu S. Y., T. H. Huang, S. C. Chang, et al., FLoD: 

A Framework for Peer-to-Peer 3D Streaming, In The 

27th Conference on Computer Communications (IEEE 

INFOCOM ‘08), 2008. 

[7] David G. Bell, Frank Kuehnel, Chris Maxwell, 

Randy Kim, Kushyar Kasraie, Tom Gaskins, Patrick 

Hogan, Joe Coughlan, NASA World Wind: Opensource 

GIS for Mission Operations. New York, 2007. 

[8] Maymounkov P., D. Mazires, Kademlia: A 

peer-to-peer information systems based on the XOR 

metric. In: Proceedings of IPTPS, Cambridge, USA, 

pp.53-65, Mar.2002, 

Zhihan Lv is a Ph. D candidate of 

Marine Information Technology 

laboratory, Ocean University of China, 

China. His research interests include 

virtual reality, 3D Visualization, 

computer network and software 

architecture. 

He has been an intership at the 

Immersion technology Co., LTD for four 

years and at the Key Lab of Marine 

Resource and Environmental Geology, First Institute of 

Oceanography, SOA for four months. From 2010 Sep, he had 

been a visiting Ph.D student at French National Center for 

Scientific Research (CNRS) in France for one year, with the 

support of China Scholarship Council.


Tengfei Yin is a master student of 

Marine Information Technology 

laboratory, Ocean University of China, 

China. His research interests include 

Web based virtual reality and P2P 

network. 

. 

Yong Han is a professor of virtual 

reality in the keylaboratory of Ocean 

Remote Sensing, Ministry of Education, 

Ocean University of China, China. His 

researchinterests include virtual reality, 

computer animation, and GIS. 

Yong Chen is a lecturer in college of 

information science and engineering at 

Ocean University of China, China. His 

research interests include computer 

technique and computer graphics. 

He has gone to New York University 

for his postdoctoral fellow at 2009. 

Ge Chen is a professor of physical 

oceanography at Ocean University of 

China, China, the dean of the School 

of Information Science, Ocean 

University of China. His main research 

interests include marine remote 

sensing, virtual reality, and GIS, PhD 

supervisor. 

He has been to IFREMER in France 

for his postdoctoral fellow for two 

years. 

Prof. Chen is an International Member of New York Academy 

of Sciences, and an international member of AAAS. He have 

served as president of the international conference session six 

times. 



An Energy-Efficient Communication Protocol for 

Wireless Sensor Networks 

Fengjun Shang 

College of Computer Science and Technology 

Chongqing University of Posts and Telecommunications, Chongqing 400065, China 

E-mail: shangfj@cqupt.edu.cn 

Abstract—WSNs (Wireless Sensor Networks) can collect 

reliable and accurate information in distant and 

hazardous environments, and can be used in National 

Defence, Military Affairs, Industrial Control, 

Environmental Monitor, Traffic Management, Medical 

Care, Smart Home, etc. The sensor whose resources are 

limited is cheap, and depends on battery to supply 

electricity, so it’s important for routing to efficiently 

utilize its power. In this paper, an energy-efficient Single- 

Hop Active Clustering (SHAC) algorithm is proposed for 

wireless sensor networks. The core of SHAC has three 

parts. Firstly, a timer mechanism is introduced to select 

tentative cluster-heads. By analyzing relation between 

time of timer and residual energy, it is known that time of 

timer is inversely proportional to residual energy of nodes 

so a timer mechanism can balance the residual energy of 

the whole network nodes which improves the network 

energy efficiency. Secondly, a cost function is proposed to 

balance energy-efficient of each node. Finally, an active 

clustering algorithm is proposed for single-hop 

homogeneous networks. Through both theoretical analysis 

and numerical results, it is shown that SHAC prolongs the 

network lifetime significantly against the other clustering 

protocols such as LEACH-C and EECS. Under general 

instance, SHAC may prolong the lifetime by up to 50% 

against EECS. 

Index Terms—wireless sensor network, active cluster, 

cost function, homogeneous, timer 


The rapid developments and technological advances in 

MEMS(Micro Electromechanical System) and wireless 

communication, has made possible the development 

and deployment of large scale wireless sensor networks. 

Wireless sensor network consists of hundreds to several 

thousands of small sensor nodes scattered throughout 

an area of interest. The potential applications of sensor 

networks are highly varied, such as environmental 

monitoring, target tracking, and battlefield surveillance. 

Sensors in such a network are equipped with sensing, 

data processing and radio transmission units. 

Distinguished from traditional wireless networks, 

sensor networks are characterized by severe power, 

computation, and memory constraints. Due to the strict 

energy constraints, energy efficiency for extending 

network lifetime is one of the most important topics. 

Sensor nodes are likely to be battery powered, and it is 


doi:10.4304/jnw.6.7.999-1008 

often very difficult to change or recharge batteries for 

these nodes. Prolonging network lifetime for these 

nodes is a critical issue. Therefore, all aspects of the 

node, from hardware to the protocols, must be designed 

to be extremely energy efficient. 

Wireless sensor networking is a broad research area, 

and many researchers have done research in the area of 

power efficiency to extend network lifetime. In order to 

achieve high energy efficiency and increase the 

network scalability, sensor nodes can be organized into 

clusters. The high density of the network may lead to 

multiple adjacent sensors generating redundant sensed 

data, thus data aggregation can be used to eliminate the 

data redundancy and reduce the communication load 

[1]. Hierarchical protocols aim at clustering the nodes 

so that cluster heads can do some aggregation and 

reduction of data in order to save energy. 

In this paper we assume that the sink is not energy 

limited (at least in comparison with the energy of other 

sensor nodes) and that the coordinates of the sink and 

the dimensions of the field are known. We also assume 

that the nodes are uniformly distributed over the field 

and they are not mobile. Under this model, we propose 

a new energy-efficient adaptive clustering algorithm. 

Our contributions have four parts. Firstly, a timer 

mechanism is introduced to produce tentative clusterheads 

so that our algorithm may prolong network 

lifetime. Secondly, an estimated average energy method 

is introduced to avoid additional communication 

between BS and cluster-head. Thirdly, a cost function is 

proposed to balance energy-efficient of each node. Last 

but not least, an active clustering algorithm is proposed 

in single-hop homogeneous network. Through both 

theoretical analysis and numerical results, it is shown 

that SHAC prolongs the network lifetime significantly 

against the other clustering protocols such as LEACH- 

C and EECS. 

Owning to constraining the resource of sensor node, 

clustering algorithm aiming at ad hoc networking can 

not be used directly, especially, the energy of WSN is 

limited, so new clustering algorithm must be researched. 

LEACH (low-energy adaptive clustering hierarchy) [1] 

is first proposed as clustering routing protocol in WSN. 

Its clustering idea is used in many clustering routing 

protocol, for example, TEEN (threshold sensitive 

energy efficient sensor network protocol) [2], HEED 

(hybrid energy-efficient distributed clustering) [3] etc.


At the same time, there are some independent designing 

clustering routing protocol, for example, ACE 

(Algorithm for Cluster Establishment) [4], LSCP 

(Lightweight Sensing and Communication Protocols) [5] 

etc. 

The paper is organized as follows. In Section Ⅱ, 

related work is discussed. Section Ⅲ, describes our 

proposed clustering routing algorithm. In section Ⅳ, 

simulation results are presented while Section Ⅴ 

concludes the paper. 

II. RELATED WORKS 

Generally, clustering algorithms for WSNs can be 

categorized into two groups: Single hop and Multi-hop 

clustering. This section describes a number of existing 

clustering algorithms within each of the following 

categories. 

A. Single-hop Clustering Algorithm 

LEACH is one of the most popular hierarchical routing 

algorithms for sensor networks. The idea is to form 

clusters based on the received signal strength and use 

local cluster heads as routers to the sink. This is shown 

to save energy since the transmissions will only be done 

by such cluster heads rather than all sensor nodes. All 

the data processing such as data fusion and aggregation 

are local to the cluster. Cluster heads change randomly 

over time in order to balance the energy dissipation of 

nodes. This decision is made by the node choosing a 

random number between 0 and 1. 

In recent years, a number of modifications have been 

proposed for the LEACH algorithm, for example, 

EECS [6], LEACH-B [7] etc. In EECS, in order to 

cluster, nodes in selecting cluster-head do not only the 

closest cluster-head but also the closest distant from 

cluster-head to BS to balance the load of network. But 

the EECS only balance energy in the area of clusterhead 

and it can not balance the energy in whole 

network. 

In Ref. [8], it is proposed to select cluster-head 

according to the residual energy of node. The main 

disadvantage of this algorithm requires the energy 

information of all nodes of the network not to be 

distributed implementing. SEP [9] mainly aims at twolevel 

heterogeneous network, that is, its initial energy 

has two kinds of level in this network. DEEC algorithm 

[10] aim at the general multi-level heterogeneous 

networks, But it can also adapted to operate in 

homogeneous sensor network. In Ref. [11], DCHS 

algorithm is proposed. In this algorithm, in order to 

En 

_ current 

extend the lifetime, a parameter 

is 

En 

_ max 

introduced. Furthermore, introducing a factor r s is a 

further modification of the threshold equation so that 

this may improve the performance of algorithm. 


HEED (hybrid energy-efficient distributed clustering) 

[3] periodically selects cluster heads according to a 

hybrid of the node residual energy and a secondary 

parameter, such as node proximity to its neighbors or 

node degree. HEED terminates in O(1) iterations, 

incurs low message overhead, and achieves fairly 

uniform cluster head distribution across the network. In 

order to balance the consuming energy, the above 

protocol periodically select cluster heads. 

B. Multi-hop Clustering Algorithm 

In PEGASIS [12], further improvement on energyconservation 

is suggested by connecting the sensors 

into a chain. Its shortcoming is that the algorithm must 

know the topology of network. In Ref. [13], the 

network is grouped into a number of clusters according 

to a randomly selected clustering scheme. TEEN [2] is 

well suited for time critical applications and is also 

quite efficient in terms of energy consumption and 

response time. It also allows the user to control the 

energy consumption and accuracy to suit the 

application. The main drawback of this scheme is that, 

if the thresholds are not reached, the nodes will never 

communicate and the user will not receive any data 

from the network at all and will not be ware of the 

overall operation or availability of the network. Thus, 

this scheme is not well suited for applications where the 

user needs to get data on a regular basis. Another 

possible problem with this scheme is that a practical 

implementation would have to ensure that there are no 

collisions in the cluster. TDMA scheduling of the nodes 

can be used to avoid this problem. This will however 

introduce a delay in the reporting of the time-critical 

data. APTEEN [14] combines the best features of both 

proactive and reactive networks and to provide periodic 

data collection as well as near real-time warnings about 

critical events. It is suitable for a network with evenly 

distributed nodes. But it is difficult to design APTEEN 

protocol so that it can not be applied. ECMR (energyconscious 

message routing) [15] is multi-hop routing 

protocol and calls for network clustering and assigns a 

less-energy-constrained gateway node that acts as a 

centralized network manager. Based on energy usage at 

every sensor node and changes in the mission and the 

environment, the gateway sets routes for sensor data, 

monitors latency throughout the cluster, and arbitrates 

medium access among sensors. But it does not support 

mobile node. 

In Ref. [16], an Energy-Efficient Unequal Clustering 

is proposed, which considers the hot spots problem in 

multi-hop wireless sensor networks. It partitions the 

nodes into clusters of unequal size, and clusters closer 

to the base-station have smaller sizes than those farther 

away from the base-station. Thus cluster heads closer to 

the base-station can preserve some energy for the intercluster 

data forwarding. Simulation results show that 

the unequal clustering mechanism balances the energy 

consumption well among all sensor nodes and achieves 

an obvious improvement on the network lifetime.


M-LEACH [17] aims at multi-hop sensor network and 

the complexity of algorithm is great. In Ref. [18], 

author proposes a novel energy efficient clustering 

scheme for single-hop wireless sensor networks, which 

better suits the periodical data gathering applications. It 

elects cluster heads with more residual energy in an 

autonomous manner through local radio communication 

with no iteration while achieving good cluster head 

distribution; furthermore, it introduces a novel distancebased 

method to balance the load among the cluster 

heads. But the algorithm does not consider the energy 

distribution. 

The clustering algorithm for single-hop networks have 

little delay and been well suited for time critical 

applications. However, its energy consumption is much 

higher between BS and node. The clustering algorithm 

for multi-hop network is complex and difficult to 

implement. In this paper, we propose a single-hop 

clustering algorithm which prolongs the lifetime of 

network. 

In order to conserve node energy and prolong lifetime 

of the network, the previous research have been mainly 

focused on balancing energy consumption among 

cluster members and they do not consider energy 

consumption among cluster-heads. In this paper, we 

propose the SHAC algorithm for homogeneous and 

single-hop sensor network. According to the residual 

energy of node, SHAC algorithm selects tentative 

cluster-heads in order to improve the clustering idea of 

LEACH. At the same time, SHAC algorithm keeps the 

distributed characteristic of algorithm and it does not 

require location information of all nodes of the network. 

Ⅲ. SHAC ROUTING ALGORITHM 

In this paper, a novel clustering idea is proposed 

called active clustering. Generally speaking, clusterheads 

are first selected based on the corresponding rule 

and then nodes are passive adding to that cluster-head, 

for example, LEACH selects them according to 

threshold, etc.. In our idea, nodes select actively 

cluster-heads according to cost function so that it can 

balance energy well. Our idea includes several parts as 

follows. 

In selecting tentative cluster-heads phase, a timer 

mechanism is introduced. Its aim is rational selecting 

cluster-heads according to residual energy. The high 

residual energy is high probability selected cluster-head 

so that this may balance the whole network energy. 

According to above idea, a clustering algorithm is 

proposed based on the average energy of whole 

network. This algorithm is similar to LEACH-C [8], but 

it avoids transmitting residual energy from nodes to BS. 

An estimation algorithm must be introduced so that this 

algorithm may avoid above energy consumption 

problem. After tentative cluster-heads are selected, 

according to the cost function, the tentative clusterheads 

select final cluster-heads according to prior, that 

is, the number of nodes adding to that cluster-head and 

then each tentative cluster-head knows final clusterhead. 

Lastly, the final cluster-head broadcasts 


information around nodes, because selecting clusterheads 

is minimal cost so that it may prolong the 

lifetime of network. 

A. Network Model 

Let us consider a sensor network consisting of N 

sensor nodes uniformly deployed over a vast field to 

continuously monitor the environment. We denote the 

th 

i sensor by s i and the corresponding sensor node 

set Node = { n1, 

n2,..., 

nN 

} , where Node = N . We make 

some assumptions about the sensor nodes and the 

underlying network model: 

1) There is a base-station (i.e., data sink) located far 

away from the square sensing field. Sensors and the 

base-station are all stationary after deployment. 

2) All nodes are homogeneous and have the same 

capabilities. Each node is assigned a unique identifier 

(ID). 

3) Nodes have no location information. 

4) All nodes are able to reach BS in one hop. 

5) Nodes can use power control to vary the amount of 

transmission power which depends on the distance to 

the receiver. 

6) Links are symmetric. A node can compute the 

approximate distance to another node based on the 

received signal strength, if the transmitting power is 

given [16]. 

We use a simplified model shown in figure 1 for the 

radio hardware energy dissipation. Both the free space 

2 

4 

( d power loss) and the multi-path fading ( d power 

loss) channel models are used in the model, depending 

on the distance between the transmitter and receiver. 

Transmission ( E Tx ) and receiving costs ( E Rx ) are 

calculated as follows 8 : 

⎪⎧ 

2 

lEelec 

+ lε 

fsd 

, d < do 

ETx 

( l, 

d) 

= ⎨ 

(1) 

4 

⎪⎩ 

lEelec 

+ lε 

mpd 

, d > do 

Where d is the distance between the transmitter and the 

receiver. 

L bit 

packet 

Transmit 

Electronics 

L 

E elec 

E Tx 

L bit 

packet 

( L, 

d) 

Tx Amplifier 

2 

εLd 

ERxL Receive 

Electronics 

L 

E elec 

Figure 1 Radio Energy Dissipation Model 

To receive this message, the energy used by the radio 

can be expressed following: 

ERx ( l) 

= Eelecl 

(2) 

with l as the length of the message in bits, d as the 

distance between transmitter and receiver node. A 

d


sensor node also consumes E DA (nJ/bit/signal) amount 

of energy for data aggregation. It is also assumed that 

the sensed information is highly correlated, thus the 

cluster-head can always aggregate the data gathered 

from its members into a single fixed length packet. 

B. SHAC Algorithm 

In the network deployment stage, the base-station 

broadcasts a “hello” message to all nodes at a certain 

power level. Using this approach, each node can 

compute the approximate distance to the base-station 

based on the received signal strength. It not only helps 

nodes to select the proper power level to communicate 

with the base-station, but also helps us to produce 

clusters of unequal size. Figure 2 gives an overview of 

the SHAC mechanism, where the anomalistic polygon 

of unequal size represent our clusters of unequal size 

and the traffic among cluster heads illustrates our 

single-hop forwarding method. 

3 

4 

Node i 

200m×200m 

5 

Clusterhead 

1 

6 

2 

BS 

Figure 2 An overview of the SHAC mechanism 

From figure 2, The SHAC algorithm makes nodes cost 

maximum to have a lower chance of becoming a 

cluster-head than nodes cost minimum in order to 

reduce energy consumption, at the same time, in each 

area of cluster-head overlay, the highest residual energy 

is selected as final cluster-head. The process of SHAC 

algorithm is as follows. It firstly starts by selecting a 

tentative cluster-head. This decision is made by the 

timer of nodes. If the timer expires, then the sensor 

declares itself to be a tentative cluster-head. Thus each 

tentative cluster-head receives information from node 

adding its cluster-head and calculates the number of 

node. Lastly, in selecting final cluster-heads phase, 

each tentative cluster-head selects the final cluster-head 

according to the prior so as to acquire final clusterheads. 

Once final cluster-head is selected, it broadcasts 

the information to the neighboring nodes. In forming 

cluster phase, each node adds the selected cluster-head 

according to cost function and then each node returns 

the information to selected cluster-head. In data 

transmitting phase, cluster member transmits data to 


cluster-head according to TDMA slot and then clusterhead 

converges the data and transmits to BS. Once the 

above process is completed, the algorithm begins to 

prepare next round work. 

C. Selecting Tentative Cluster-head 

Selecting tentative cluster-head is the basis for 

creating clusters. After deployment, each sensor sets a 

random waiting timer [20]. If the timer expires, then the 

sensor declares itself to be a cluster-head, a focal point 

of a new cluster. However, events may intervene that 

cause a sensor to shorten or cancel its timer. For 

example, whenever the sensor detects a new neighbor, 

it shortens the timer. On the other hand, if a neighbor 

declares itself to be a cluster-head, the sensor cancels 

its own timer and joins the neighbor’s new cluster. 

Because LEACH does not consider residual energy and 

distance of nodes, an average energy factor for SHAC 

and make the node with the highest level of energy to 

be first tentative cluster-head. The key parameter is as 

follows. 

E( 

i) 

residual 

Energyi 

= (3) 

E( 

r) 

where E ( i) 

residual is the residual energy of the i-th node, 

E (r) 

is the average energy of the node and r is the 

current round number. Energy factor is used to balance 

the whole network energy. 

Definition The number of rounds from first round to 

round which first node dies is called lifetime. 

Every node i maintains a variable x i , which is 

assigned a random value from 0 to 1, namely, x i 

=random(0,1). Obviously, x i is a random variable with 

uniform distribution on the interval [0, 1]. Each node i 

waits for a initiator timer according to an exponential 

random distribution i.e. 

−λit 

i xi 

= e 

(4) 

where λ i = Energyi 

. 

Formula (4) explains that t i is inversely proportional to 

E ( i) 

residual . Formula (4) may be written by 

ln( xi 

) 

ti 

= − 

(5) 

λ 

Substituting (3) into formula (5) and E (r) 

is invariant 

in that round, so Formula (5) is written by 

ln( xi 

) 

ti 

= − 

E( 

i) 

residual 

(6) 

We can find the relation between t i and E ( i) 

residual by 

setting the derivative t i with respect to E ( i) 

residual . 

dti 

dE( 

i) 

residual 

ln( xi 

) 

= 

2 

E( 

i) 

residual 

(7) 

Q 1 ≤ xi 

≤ 1 

i


∴ 

dt 

dE( 

i) 

i 

residual 

< 0 

, that is, t i is inversely 

proportional to E ( i) 

residual . 

From above analysis, we select short time as clusterhead, 

that is, we select high energy as cluster-head, so it 

is beneficial to prolong lifetime of network. After 

sensors are deployed, each sensor sets a random waiting 

timer. If the timer expires, then the sensor declares 

itself to be a cluster-head, a focal point of a new cluster. 

D. Estimating Average Energy 

It is important to estimate the average energy, but the 

disadvantage of this approach is that each node has to 

estimate the aggregate remaining energy in the network 

since this requires additional communication with the 

base-station and other nodes. In order to improve the 

approach, the average energy must be estimated [10]. 

In SHAC, we assume that there are N nodes 

distributed uniformly in a M×M region. If there are k 

N 

clusters, there are on average nodes per cluster (one 

k 

N 

cluster head and −1 

non-cluster head nodes). Each 

k 

cluster head dissipates energy receiving signals from 

the nodes, aggregating the signals, and transmitting the 

aggregate signal to the BS. Since the BS is far from the 

nodes, we can assume that the energy dissipation 

4 

follows the multi-path model ( d power loss). Each 

non-cluster head node only needs to transmit its data to 

the cluster head once during a round. We can also 

assume that the distance to the cluster head is small, so 

the energy dissipation follows the free-space model 

2 

( d power loss). Hence, the total energy consumed 

during a single round can be estimated as: 

Eround = ECH 

+ Enon−CH 

(8) 

where E CH is tentative cluster-head consumption 

energy, Enon− CH is cluster member consumption 

energy. For tentative cluster-head E CH , the single 

round energy consumption is as follows. 

4 N N 

E CH = LEelec 

+ Lε 

mpd 

toBS + × LE DF + ( −1) 

× LEelec 

k 

k 

(9) 

For cluster member Enon− CH , the single round 

consumption energy is as follows. 

2 

Enon−CH 

= LEelec + Lε 

fsd 

toCH (10) 

where l is the number of bits in each data message, 

d toBS is the distance from the cluster head node to the 

BS, d toCH is the distance from the node to the cluster 

head, and we have assumed perfect data aggregation. 

If we know E round , we may estimate the average energy 

Etotal 

− rEround 

Er 

= (11) 

N 


N 

where Etotal 

= ∑ Ei 

is the initial energy of all the 

i= 

1 

nodes, E i is i th node energy, E round is single round 

energy consumed. Furthermore, let single round energy 

consumed to be uniform. On above condition, we may 

estimate E r as follows. 

In Ref. [8], the two parameters k and d toCH are given 

by: 

N ε fs M 

k = (12) 

2 

2π ε mp dtoBS 

M 

dtoCH = (13) 

2πk 

From Ref. [19], we may find the parameters d toBS to be 

equal to: 

M 

dtoBS = 0. 

765 

(14) 

2 

Substituting (12), (13), and (14), into (11) allows for 

estimation of Eround. Furthermore, we may estimate that 

Etotal 

− rEround 

Er 

= 

N 

and avoid additional communication between clusterhead 

and BS. 

E. Active Selecting Cluster-heads 

Clustering a wireless sensor network means 

partitioning its nodes into groups, each one with a 

cluster head and some ordinary nodes as its members. 

The task of being a cluster head is rotated among 

sensors in each data gathering round to distribute the 

energy consumption across the network. SHAC is a 

distributed cluster heads competitive algorithm, where 

cluster head is selected primarily based on the residual 

energy of each node. 

Firstly, several tentative cluster-heads are selected 

using the timer mechanism to compete for final cluster 

heads. Secondly, each tentative cluster-head broadcasts 

a COMPETE_HEAD_MSG, which includes residual 

energy(RE), distance from BS(DBS), broadcast radius R 

and ID of that tentative cluster-head and each node adds 

the cluster-head according to the cost function f(i,j). 

Thirdly, each tentative cluster-head broadcasts 

RECEIVE_NODE_MSG, which includes the number of 

nodes adding this cluster-head and each tentative 

cluster-head receives the information and selects the 

final cluster-head according to prior knowledge, that is, 

selects the cluster-head according to number of nodes 

adding to that cluster-head, we select them about six 

cluster-heads. Lastly, the final cluster-heads broadcast 

FINAL_HEAD_MSG. 

F. Balancing Cluster Member Energy 

After cluster-heads are selected, the key problem is 

assigned each node to particular cluster-head. It is 

important to balance energy consumption in area 

around the cluster-head, for example, node i shown in


figure 2 can use cluster-head 5 or 6, which must de 

decided by the cost function. Intuitively, node i add 

cluster-head 5, because it is close to cluster-head, but, it 

is not beneficial to balance the network energy 

consumption. The proposed cost function f ( i, 

j) 

is as 

follows. 

c( 

ni 

, CH j ) g( 

CH j , BS) 

f ( i, 

j) 

= w× 

+ ( 1− 

w) 

× 

(15) 

E 

E 

i 

The condition is cost function f ( i, 

j) 

is minimum, if 

node i uses the CH j , where E i is current energy of 

the i-th node, ECH denotes current energy of the j-th 

j 

cluster-head and 

2 

d ( ni 

, CH j ) 

c ( ni 

, CH j ) = 

2 

d 

d 

g ( n , CH ) = 

i 

j 

n _ CH 

4 

( CH j , 

4 

dCH 

_ max 

BS) 

where d( CH j , BS) 

denotes the distance from j-th 

cluster-head to BS, E denotes residual energy of j-th 

CH j 

d _ max max{( CH j 

cluster-head, CH = , BS)} 

, 

d ( ni 

, CH j ) denotes the distance from i-th node to j-th 

cluster-head, d n _ CH = max{( ni 

, CH j )} . 

The cost function f includes both distance and 

energy factors. In intuition, it can balance the current 

energy consumption of area around the cluster-head. 

The explanation is as follows: 

The idea is to making f ( i, 

j) 

minimum in selecting 

cluster member, that is 

min{ f ( i, 

j)} 

= } 

) , ( 

c( 

ni 

, CH j ) g CH j BS 

min{ w × + ( 1− 

w) 

× 

E 

E 

2 

i 

d ( ni 

, CH j ) 

d ( CH j , BS) 

= min{ w× 

+ ( 1− 

w) 

× 

} 

4 

d × E 

d × E 

2 

n _ CH 

i 

CH 

4 

CH _ max 

j 

CH 

j 

CH j 

In order to be convenient to analysis, constant is 

introduced, that is 

2 

4 

d ( ni 

, CH j ) d ( CH j , BS) 

min{ a × 

+ b × 

} 

E 

E 

i 

CH j 

1 

1 

where a = , b = . 

2 

4 

dn 

_ CH dCH 

_ max 

Since at each round, there it is communication between 

cluster member and between cluster-head and BS, the 

above formula become: 

2 

4 

d ( ni 

, CH j ) d ( CH j , BS) 

min{ a × ∑ 

+ b × ∑ 

} (16) 

E 

E 

i 

i 

j 

CH j 

In above formula, factors are introduced, that is 


min{ a × 

∑ 

i 

2 

ε fsd 

( ni 

, CH j ) 

+ b × 

ε E 

fs 

i 

2 

∑ 

j 

4 

ε mpd 

( CH j , BS) 

} 

ε E 

mp 

CH j 

In above formula ε d ( n , CH ) is the key for cluster 

fs 

i 

member consumption energy and ε mpd 

( CH j , BS) 

is 

the key for cluster member consumption energy. 

Generally, only cost function f ( i, 

j) 

is minimum, so 

cluster member and cluster-head consumption energy is 

minimum so that the network consumption energy is 

minimum. If so, this may prolong the lifetime of 

network. Or else, if cluster member and cluster-head 

consumption energy is minimum so that cost function 

f ( i, 

j) 

must be minimum. 

G. Analysis Algorithm 

SHAC is a distributed cluster heads competitive 

algorithm, where cluster head selection is primarily 

based on the residual energy of each node. The pseudocode 

for an arbitrary node s i is given as follows. Our 

goal is selecting the final cluster-heads according to 

cost function. Therefore the distribution of cluster 

heads can be controlled over the network. 

Each tentative cluster head maintains a set ACH of its 

“adjacent” tentative cluster heads. Tentative head si is 

an “adjacent” node of sj if si is in sj’s competition range. 

Whether a tentative cluster head sj will become a final 

cluster head depends on the cost function of nodes only, 

i.e., the algorithm is distributed. 

Algorithm1 SHAC Pseudo-code 

BS broadcast average energy message AE_MSG; 

%Selecting ClusterHead 

For i=1:1:NodeNumber 

If the timer fires 

Broadcast ADV_Tentative_Head(R); 

End 

On receiving ADV_Tentative_Head(R) form si 

sj broadcast a JOIN_MSG(ID); 

End 


If si is Tentative Clusterhead 

Broadcast COMPETE_HEAD_MSG(ID, R, RE, DBS); 

End 

End 

On receiving COMPETE_HEAD_MSG form sj 

If ∀ ni∉beTentativeHead with min(f(i,j)) 

Add ni to sj; 

sj.Num=sj.Num+1; 

End 

For k=1:1: TentativeClusterNumber 

Broadcast RECEIVE_NODE_MSG(ID, R, RE, DBS); 

Other ∀ ni∉ beTentativeHead Receive RECEIVE_NODE_MSG; 

End 

For j=1:1: TentativeClusterNumber 

Selecting about six clusterheads with max(sj.Num); 

Broadcast FINAL_HEAD_MSG; 

On receiving a FINAL_HEAD_MSG from sj 

If si∈sj.ACH 

j 

4


Broadcast QUIT_MSG; 

End 

On receiving a QUIT_MSG from si 

If si∈sj.ACH 

Remove si from sj.ACH; 

End 

End 

%Forming Cluster 


If ni∈sj.ACH 

Broadcast HEAD_MSG(ID, Rmax, RE); 

Wait for JOIN_ClusterHead_MSG; 

Else 

Receiving all HEAD_MSG; 

Calculate the f(i,j); 

Select the ClusterHead with min{f(i, j)}; 

Add ni to CHj and Broadcast the JOIN_ClusterHead_MSG; 

End 

End 

In this algorithm, each node broadcast information 

using same power, in order to save energy, this 

broadcasting radius is R. Firstly, according to waiting 

timer, if the timer expires, the tentative cluster-heads 

are produced. Secondly, the tentative cluster-heads 

broadcast the number of receiving node. Thirdly, it 

produces the final cluster-head according to sj.Num. 

Furthermore, each cluster-head broadcasts HEAD_MSG 

information, which has ID of node, Rmax and residual 

energy RE. According to receiving information, nodes 

calculate cost function f ( i, 

j) 

and select minimum 

f ( i, 

j) 

cluster-head. Lastly, nodes send 

JOIN_ClusterHead_MSG information and tell that 

cluster-head. 

After cluster-head is selected, according to cost 

function, nodes add themselves to the cluster-head. 

Furthermore, during the data transmission phase, the 

cluster-head structures a TDMA-based schedule, which 

determines when each cluster member can 

communicate with the cluster-head. 

According to Algorithm 1, the cluster head selection 

process is message driven, thus we must discuss its 

message complexity. 

Theorem The message complexity of the cluster 

formation algorithm is O(N). 

Proof: At the beginning of the cluster head selection 

Area 

phase, BS broadcasts an AE_MSG, (Area is 

2 

πR 

selected that area of scenario) tentative cluster-heads 

broadcast ADV_Tentative_Head and cluster members 

Area 

broadcast N- JOIN_MSG. Next step, tentative 

πR 

2 

cluster heads are produced and each of them broadcasts 

a COMPETE_HEAD_MSG. Then each of them makes 

a decision by broadcasting a RECEIVE_NODE_MSG to 

calculate the total number of node adding that clusterhead. 

Furthermore each of them makes a decision by 

broadcasting a FINAL_HEAD_MSG to act as a final 

cluster head, or a QUIT_MSG to act as an ordinary 

node. Suppose k cluster heads are selected, they send 


out k HEAD_MSG, and then (N-k) ordinary nodes 

transmit (N-k) JOIN_ClsterHead_MSG. Thus the 

Area 

Area 

messages add up to N × +N × (1 - ) + 3N 

2 

2 

πR 

πR 

Area 

Area 

× + k + N -k +1= (3× + 2) × N + 1 in the 

2 

2 

πR 

πR 

cluster formation stage per round, i.e., O(N). 

The above theorem shows the message overhead of 

SHAC is small. In HEED, the upper-bound of message 

complexity is Niter×N where Niter is the number of 

iterations. Because we have avoided message iteration 

in the cluster-head selection algorithm, the control 

message overhead in SHAC is much lower than that in 

HEED. EECS algorithm is also O(N), but when it select 

cluster-head, it does not consider the cost of each node 

and cluster-head. 

Ⅳ. SIMULATIONS AND ANALYSIS 

We select a scenario to simulate our algorithm using 

MATLAB and the parameter set is shown in Table I. 

TABLE Ⅰ. SIMULATION PARAMETERS 

Parameter Value 

Network coverage (0,0)~(200,200) 

Base station location (100,350) 

N 1000 

Initial energy 1J 

50nJ/bit 

Eelec 

εfs 

εmp 

do 

EDA 

10pJ/bit/m 2 

0.0013pJ/bit/m 4 

87m 

5nJ/bit/signal 

Data packet size 4000 bits 

The key problem is how to select parameter w in SHAC 

algorithm, because w is weight value using balancing 

between cluster-head and cluster member. We select w 

from 0 to 1 to test it and the test results are shown in 

figure 3. From figure 3, it is known w=0.6 is the best, 

so we select w =0.6. 

Number of Rounds 

1000 

900 

800 

700 

600 

500 

400 

300 

200 

100 

0 

0 0.2 0.4 0.6 0.8 1 1.2 

Weight w 

Figure 3 Relation between w and number of rounds 

We compare between LEACH-C, EECS and SHAC. 

Figure 4 shows lifetime of network over the simulation 

time, where SHAC is the longest. SHAC prolongs the 

network lifetime significantly against the other 

clustering protocols such as LEACH-C and EECS. 

Under general instance, SHAC may prolong the 

lifetime at least 30%. Additionally, it may prolong the 

lifetime by up to 50% against EECS.


Number of Alive Nodes 

1000 

900 

800 

700 

600 

500 

SHAC 

EECS 

LEACH-C 

Comparing Number of Alive Node 

400 

600 700 800 900 1000 1100 1200 1300 1400 


Figure 4 Lifetime of network over simulation time 

By observing the number of dead nodes from figure 4, 

it can be seen that there are no dead nodes in 1,000 

rounds of SHAC. In 1,000 rounds of EECS, there are 

30 nodes at least, which is 3% of number of total 

nodes, there are 100 nodes at least, which is 10% of 

total number of node in 1,000 rounds in LEACH-C. 

The number of dead node shows balance network 

energy consumption. The less there are dead nodes, 

the better we can do balance network energy. SHAC 

both prolongs lifetime of network and reduces the 

number of dead nodes. Hence, SHAC more efficiency 

balances the energy consumption of network 

compared to the other two strategies. EECS introduces 

a cost function, but its performance is not better than 

SHAC, the reasons have three points. Firstly, it does 

not consider energy consumption of node in selecting 

tentative cluster-head. Secondly, it only considers the 

distance factor and omits residual energy of node and 

cluster-head. Lastly, it does not consider the cost of 

each node and cluster-head. It has been considered the 

residual energy of node in LEACH-C, but overhead 

about transmitting information is bigger, so it has the 

worst performance. 


500 

450 

400 

350 

300 

250 

200 

150 

100 

50 

0 

Comparing Number of Clusterhead 

EECS 

LEACH-C 

SHAC 

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 

Number of Clusterheads 

Figure 5 Number of cluster-head versus number of rounds 

The number of cluster-head is shown figure 5. It is 

known that SHAC is stable. The fluctuation about 

number of cluster-head, EECS is similar to SHAC, 

LEACH-C is the biggest, because LEACH-C only 

uses random number and threshold scheme to select 

cluster-head so that the fluctuation is significant. The 


number of cluster-head in SHAC and EECS is 

converged to the optimized number of cluster-head, 

because two algorithms use competing scheme so that 

they control number of cluster-head, but SHAC 

algorithm introduces active scheme to select clusterheads 

so that energy-efficiency is the highest among 

above three algorithms. 

In this part, we investigate the energy efficiency of 

SHAC, we compare the amount of cumulative residual 

energy by all nodes in three algorithms. 1400 rounds of 

simulations are sampled and the amount of total 

cumulative residual energy by all nodes is shown in 

Figure 6. The residual energy by all nodes per round in 

SHAC is the highest compared to EECS, LEACH-C. 

And because the distribution of selected cluster-heads is 

uncontrollable in EECS and LEACH-C, there is a 

dramatic variation in the energy consumption of the 

cluster-heads. Due to the stability of cluster-heads 

topology in SHAC, the amount of energy spent by all 

nodes is almost the same and the lowest in each round. 

In Figure 6, SHAC achieves the most residual energy in 

the network and this further illustrates why SHAC 

achieves the longest lifetime compared to the other 

strategies. 

Total Cumulative Residual Energy in Network(J) 

Comparing Total Cumulative Residual Energy ofNetwork over Number of Rounds 

750 

SHAC 

EECS 

700 

LEACH-C 

650 

600 

550 

500 

450 

400 

500 600 700 800 900 1000 1100 1200 1300 1400 


Figure 6 Total cumulative residual energy of network versus number 

of rounds. 

It compares the number of information from all cluster 

member nodes to all cluster-heads in figure 7. This 

figure shows that, SHAC forwards the most number of 

packets, when compared to the other strategies, from 

nodes to cluster-heads. Furthermore, from this figure it 

can be seen that LEACH-C and EECS experience a 

significant drop in performance after 850 th round, 

where SHAC continues to forward more packets. 

After first node is dead, the number of information 

about SHAC is bigger than LEACH-C and EECS, 

because the lifetime is longer than LEACH-C and 

EECS. After 850 th rounds, the number of information 

is significantly bigger than LEACH-C EECS, because 

dead nodes add quickly. The number of information 

from node to cluster explains details of real scenario. 

The more is the formation, the more are details, so 

SHAC is better in real application.


Number of Message Received 

x Number 106 of Total Message Received in Clusterhead Over Number of Rounds 

1.4 

1.3 

1.2 

1.1 

1 

0.9 

0.8 

0.7 

SHAC 

EECS 

LEACH-C 

0.6 

700 800 900 1000 1100 1200 1300 1400 


Figure 7 Number of information from node to cluster-head 

Figure 8 shows the total number of dead node, when 

the number of dead nodes is less than 5% of each 

round. From figure 8, we know that the mostly dead 

nodes are in center of scenario and it may explain the 

algorithm can balance energy consumption of network 

and has better energy efficiency. 

400 

350 

300 

250 

200 

150 

100 

50 

Base Station 

0 

0 50 100 150 200 

Figure 8 Distribution of dead nodes, where x axis is the length of 

scenario and y axis is width of scenario 

Ⅴ . CONCLUSIONS AND FUTURE WORKS 

In this paper, a Single-Hop Active Clustering (SHAC) 

algorithm is proposed about wireless sensor networks 

by research current routing algorithms. The core of 

SHAC has three parts. Firstly, a timer mechanism is 

introduced to select tentative cluster-heads. By 

analyzing relation between time of timer and residual 

energy, it is known that time of timer is inversely 

proportional to residual energy of nodes so a timer 

mechanism can balance the residual energy of the 

whole network nodes which improves the network 

energy efficiency. Secondly, a cost function is proposed 

to balance energy-efficient of each node. Last but not 

least, an active clustering algorithm is proposed in 

single-hop homogeneous network. Through both 

theoretical analysis and numerical results, it is shown 

that SHAC prolongs the network lifetime significantly 

against the other clustering protocols such as LEACH- 

C and EECS. Under general instance, SHAC may 

prolong the lifetime 30% at least, especially, it may 

prolong the lifetime up to 50% against EECS. 

In future research, we will consider NS2 simulation 

platform using event-driven mechanism to simulate 


performance of the SHAC algorithm. In LEACH-C, 

EECS and SHAC, we assume that data are transmitted 

at any moment, but for event-driven network, in no 

events, nodes do not consume energy and keep 

sleeping status. Once there is a event, the node is 

waked to collect data and communicate, so this can 

improve energy-efficient of sensor network so that this 

make SHAC is better to apply in real condition. 

ACKNOWLEDGEMENTS 

The author would like to thank the Chongqing Natural 

Science Foundation under Grant No. 2009BB2081 and 

the Science and Technology Research Project of 

Chongqing Municipal Education Commission. The 

Project Sponsored by the Scientific Research 

Foundation for the Returned Overseas Chinese Scholars, 

State Education Ministry. The author would also like to 

thank to MATLAB software. 

REFERENCES 

[1] W. Heinzelman, A. Chandrakasan, and H. Balakrishnan. 

Energy-efficient routing protocols for wireless microsensor 

networks. Proceeding 33rd Hawaii International Conference 

on System Sciences,Volume 8, pp.8020-8030, 2000. 

[2] Manjeshwar A, Grawal DP. TEEN: A protocol for enhanced 

efficiency in wireless sensor networks. Proceeding of the 

15th Parallel and Distributed Processing Symp. San 

Francisco: IEEE Computer Society, 2001, pp. 2009-2015. 

[3] Younis O, Fahmy S. Heed: A hybrid, energy-efficient, 

distributed clustering approach for ad-hoc sensor networks. 

IEEE Trans. on Mobile Computing, 2004, 3(4):660-669. 

[4] Chan H, Perrig A. ACE: An emergent algorithm for highly 

uniform cluster formation. Proceeding of the 1st European 

Workshop on Wireless Sensor Networks. LNCS 2920, Berlin: 

Springer-Verlag, 2004, pp. 154-171. 

[5] Fang Q, Zhao F, Guibas LJ. Lightweight sensing and 

communication protocols for target enumeration and 

aggregation. Proceeding of the 4th ACM Int’l Symp. on 

Mobile Ad Hoc Networking & Computing. ACM Press, 2003, 

pp. 165-176. 

[6] Ye M, Li C, Chen G, Wu J. EECS: An energy efficient 

cluster scheme in wireless sensor networks. Proceeding of 

the IEEE IPCCC 2005. New York: IEEE Press, 2005, pp. 

535-540. 

[7] Depedri A, Zanella A, Verdone R. An energy efficient 

protocol for wireless sensor networks. Proceeding of the 

AINS 2003. Menlo Park, 2003, pp. 1-6. 

[8] Heinzelman WR, Chandrakasan AP, Balakrishnan H. An 

application-specific protocol architecture for wireless 

microsensor networks. IEEE Trans. on Wireless 

Communications, 2002, 1(4):660-670. 

[9] Smaragdakis G, Matta I, Bestavros A. SEP: A stable election 

protocol for clustered heterogeneous wireless sensor 

networks. Proceeding of the Int’l Workshop on SANPA, 

Boston.,no. 4, pp. 660-670, 2004. 

[10] Qing li, Zhou Qing-Xin, Wang Ming-Wen. A Distributed 

Energy-Efficient Clustering Algorithm for Heterogeneous 

Wireless Sensor Networks .Chinese Journal of Software, 

2006, 17(3): 481-489.


[11] Handy MJ, Haase M, Timmermann D. Low energy adaptive 

clustering hierarchy with deterministic cluster-head selection. 

Proceeding of the 4th IEEE Conf. on Mobile and Wireless 

Communications Networks. Stockholm: IEEE 

Communications Society, 2002, pp. 368-372. 

[12] Lindsey S, Raghavenda CS. PEGASIS: Power efficient 

gathering in sensor information systems. In: Williamson DA, 

ed. Proceeding of the IEEE Aerospace Conf. Vol 3, New 

York: IEEE Press, 2002, pp.1125-1130. 

[13] Mhatre V, Rosenberg C. Design guidelines for wireless 

sensor networks: communication, clustering and aggregation. 

Ad Hoc Network Journal, 2004, 2(1):45-63. 

[14] Manjeshwar A, Agrawal DP. APTEEN: A hybrid protocol 

for efficient routing and comprehensive information retrieval 

in wireless sensor networks. Proceeding of the 2nd Int’l 

Workshop on Parallel and Distributed Computing Issues in 

Wireless Networks and Mobile Computing. Florida: IEEE 

Computer Society, 2002, pp.195-202. 

[15] Younis M, Youssef M, Arisha K. Energy-Aware routing in 

cluster-based sensor networks. Proceeding of the 10th IEEE 

Int’l Symp. on Modeling, Analysis and Simulation of 

Computer and Telecommunications Systems. Fort Worth: 

IEEE Computer Society, 2002, pp.129-136. 

[16] Li Cheng-Fa, Chen Gui-Hai,Ye mao, etc..An uneven 

cluster-based routing protocol for wireless sensor network. 

Chinese journal of computers, 2007, 30(1): 27-36. 

[17] V. Mhatre and C. Rosenberg. Homogeneous vs. 

Heterogeneous Clustered Sensor Networks: A Comparative 

Study. Proceedings of 2004 IEEE International Conference 

on Communications, Paris, France, June 2004, Volume 6, pp. 

3646-3651. 

[18] Hsiao-Lan Hsu, Qilian Liang. An energy-efficient protocol 

for wireless sensor networks. In Vehicular Technology 

Conference, 2005, Volume: 4, pp. 2321-2325. 

[19] Bandyopadhyay S, Coyle EJ. An energy efficient 

hierarchical clustering algorithm for wireless sensor 

networks. In: Mitchell K, ed. Proceeding of the INFOCOM 

2003. Vol 3, New York: IEEE Press, 2003, pp.1713-1723. 

[20] Yongtao Cao,Chen He. A Distributed Clustering Algorithm 

with an Adaptive Backoff Strategy for Wireless Sensor 

Networks. IEICE TRANS. COMMUN. , 2006, Vol. E89-B(2): 

609-613. 

Fengjun Shang (1972- ), received 

the Diploma degree in Intelligent 

Instrument at Chengdu University 

of Technology, China, in 2001. He 

finished his Ph.D. in Instrument 

Science and Technology at the 

College of Opto-electronic 

Engineering, Chongqing University, 

China, in 2005. Since then he 

works as an associate professor 

with the Institute of Computer Network Engineer in 

Chongqing University of Posts and Telecommunications, 

China. His research interests include sensor network, traffic 

engineer, network optimization and WiMAX. 

Email: shangfj@cqupt.edu.cn. 



Robust Cross-layer Design of Wireless 

Multimedia Sensor Networks with Correlation 

and Uncertainty 

Lei You Chungui Liu 

School of electronic and information engineering, Tianjin University, Tianjin, China 

Email: youlei@tju.edu.cn, cgliumail@163.com 

Abstract—Due to content-enriched sensing information and 

flexibility, Wireless Multimedia Sensor Network (WMSN) 

has a lot of potential applications. Low channel capacity, 

limited resource, correlation between sensing sources and 

uncertain factors make the design and optimization of 

WMSN challenging. In a densely deployed WMSN, there 

generally exist correlation and redundancy in the 

multimedia information collected by sensors with 

overlapped sensing area. In this paper, we adopt a crosslayer 

method to deal with the robust lifetime optimization of 

WMSN with correlated sources, which also has energy 

consumption uncertainty in the transmission of wireless 

links. To reduce the number of constraints of the source rate 

region, a pairwise Distributed Source Coding (DSC) scheme 

is proposed by matching sensor nodes based their 

correlation. A Distributed Pairwise Matching (DPM) 

algorithm is thus proposed. With a polyhedral set modeling 

the uncertain energy consumption, a robust cross-layer 

problem, which maximizes the lifetime of the WMSN 

through allocating source rate and flow rate on the link (i.e., 

routing) simultaneously under all the possible uncertainty, 

is formulated. The counterpart of the problem is showed to 

be a convex problem with linear constraints, which can be 

divided into several separate subproblems in different layers 

by dual decomposition. Based on the subgradient method 

and DPM, a partially distributed optimization algorithm is 

proposed. The algorithm can be implemented in small scale 

WMSNs with sink node only responsible for the calculation 

and distribution of the value of the lifetime. Simulation 

results verify the performance of the proposed algorithm 

and show its robustness under uncertainty. 

Index Terms—robust cross-layer optimization, wireless 

multimedia sensor networks, correlation, uncertainty, dual 

decomposition, subgradient. 


Wireless multimedia sensor network (WMSN) is 

emerging with the availability of inexpensive visual and 

audio hardware such as CMOS cameras and microphones 

[1-3]. A sensor node in WMSN is typically comprised of 

multimedia sensor module, microprocessor unit and radio 

module. WMSN can sense and transmit video and audio 

streams, still images, as well as scalar sensor data through 

multihop wireless links between sensors. WMSN is able 

to enhance the traditional sensor networks with contentenriched 

sensing information and has many new 


doi:10.4304/jnw.6.7.1009-1016 

applications such as ambient Intelligence [4]. An example 

of WSMN is illustrated in Figure 1. The multimedia 

sensor nodes can be connected wirelessly with neighbors 

and the sensed multimedia information (visual or audio) 

is transmitted through one or multiple hops to the sink 

node. 

Monitored 

objectives 

or events 

Servers 

Wreless multimedia sensor nodes 

Sink node 

Figure 1. An example of a wireless multimedia sensor networks 

Low capacity and limited resource make it challenging 

to deliver multimedia data in WMSNs. Due to densely 

deployed, multimedia information collected by sensors 

with overlapped sensing area is spatially correlated [3]. 

Thus, eliminating redundancy by exploring the 

correlation between nodes can increase the efficiency of 

WMSNs. It was shown that the distributed source coding 

(DSC) (Slepian-Wolf coding theorem) [5] can be used to 

eliminate the redundancy without explicit communication 

between correlated nodes. In addition, the encoder of 

DSC is much simpler than that of the state-of-the-art 

predictive encoding algorithm [6]. Thus, DSC is quite 

suited for WMSN with low capacity and limited resource. 

Optimal DSC schemes are proposed in [14]. However, 

the number of constraints to determine the Slepian-Wolf 

coding rate region grows exponentially in the number of 

correlated sources involved in the DSC. 

Cross-layer optimization of wireless communication 

networks has been an active research area recent years. 

Backpressure-based approaches [15] that determine 

routing and scheduling using queue backlog information 

are widely used for cross-layer design due to its 

optimality. Capacity (stability) region for a time-varying 

wireless network is formulated in [17], an optimal routing, 

scheduling and power control policy is proposed based on 

the backpressure approach in [15]. Joint optimization of 

congestion control, routing and scheduling in wireless 

multi-hop networks is studied in [16] [18] using dual


decomposition and sub-gradient method. A lot of papers 

consider the optimization of wireless (multimedia) sensor 

networks without considering uncertain parameters [7-11] 

using the similar methods. 

However, ignoring the uncertainty in the design of 

WMSN may results in inefficient energy consumption 

and reduced QoS. In the case of variable and uncertain 

system parameters, a solution that is efficient with all the 

possible values of the parameters is preferable (although 

it may not be optimal for given values). The authors in 

[12] considered the uncertainty in the distance between 

sensor nodes in wireless sensor networks (WSN), 

modeled the uncertainty with polyhedral and ellipsoidal 

sets, formulated the robust optimization problems of 

WSN and derived the robust counterpart problems that 

are proved to be convex. The problem in [12] is solved 

using a centralized algorithm in the programming solver, 

which is suitable for the analysis and initialization of 

WSNs but not for the practical operation. In addition, 

there are often strict requirement for the suorce rates from 

sensor nodes to the sink in WMSNs, while it only 

required some percentage of information reaches the sink 

in [12]. 

Taking a cross-layer approach and optimizing 

transmission and source rate allocation simultaneously, 

especially in the present of correlation and uncertainty, 

are important for extending the lifetime of WMSN. In 

this paper, we consider the robust optimization of the 

lifetime of WMSNs with correlated sources and under 

energy consumption uncertainty using cross-layer design. 

Our aim is to develop an algorithm that is able to be 

implemented in the practical operation of WMSN. For 

this, firstly, a Distributed Pairwise Matching (DPM) 

algorithm is proposed to group the sensor nodes in pairs 

in order to implement the DSC with reduced complexity. 

Then, the transmission energy consumption uncertainty in 

each sensor nodes was modeled as a polyhedral set, with 

which we showed that the robust cross-layer optimization 

problem can be transformed into a convex counterpart 

problem with linear constraints. The lifetime 

maximization problem of such a WMSN is solved by 

dual decomposition and subgradient method [13]. The 

algorithm we developed not only is (partially) distributed, 

but also keeps the layering and modularity architecture of 

the communication network. The algorithm can be 

implemented in small scale WMSNs with sink node only 

responsible for the calculation and distribution of the 

value of the lifetime. 

II. SYSTEM MODEL AND PROBLEM FORMULATION 

We consider a static wireless multimedia sensor 

network, the sensor nodes of which are observing the 

same event using certain multimedia sensors (e.g., visual 

or audio sensor). We assume homogeneous sensor nodes, 

i.e., nodes are deployed with the same sensor. The 

WMSN is modelled as a directed graph G � ( N, L) 

with a 

node set N and a link set L. There are n +1 nodes in N , 

including n sensor nodes (the set of which is denoted by 

N ) and a sink node. Link between nodes i and j is 

s 


denoted by (, i j ) . The maximum transmission rate of link 

(, i j) is c ij . The set of nodes directly connected to node i 

is denoted by V i . 

A. Pairwise Distributed Source Coding (Pairwise DSC) 

Due to densely deployed, multimedia information 

collected by nearby sensor nodes are often correlated and 

thus redundant. Since transmission of redundant 

multimedia data is energy- and bandwidth-consuming, we 

employ Slepian-Wolf distributed source coding (DSC) [5] 

to eliminate the redundancy. Denote Si as the source 

coding rate of sensor node i. To construct the sensed 

multimedia event in the sink without distortion, the 

source coding rates of all the sensor nodes must satisfy 

the Slepian-Wolf (S-W) source rate region [5] as follow, 

c 

� Si � H(Z Z ), Z � Ns 

(1) 

where 

i�Z 

Z C is the complement of Z . c 

H (Z Z ) is 

conditional entropy 

However, the S-W source rate region determined by (1) 

is difficult to be used in practical optimization of 

WMSNs due to: 

1) The number of constraints in (1) grows 

exponentially in the number of sensor nodes; 

2) Global correlation information (joint entropy and 

conditional entropy of all the possible partitions of sensor 

nodes) is required. 

Here, we propose to implement the DSC in pairs, i.e., 

group two nodes in pair, which will perform the DSC 

jointly based on their correlation (joint entropy Hij). (, ) 

We thus propose the following distributed algorithm to 

match the nodes into pairs. 

Distributed Pairwise Matching (DPM) algorithm of 

correlated sensor nodes 

Initialization: Setting the weight of node i and one of its 

neighbor j, Wij (, ) � Hij (, ) ; initializing the temporary 

set of neighboring nodes of sensor node i, �� i Ns, 

as 

tmp 

Bi � Vi 

. 

Algorithm: 

(1) Each node i sends a ‘matching’ massage to the 

neighbor j � arg min H( i, j) 

. 

tmp 

j�Bi (2) If node i receives a ‘matching’ massage from its 


(to which it has sent a 

tmp 

j�Bi ‘matching’ massage), it sends a ‘matched’ massage 

to all its other neighbors. Thus, node i and node 

j are matched, i.e., ( i, j) are the pairs to perform 

DSC. 

(3) If node i receives a ‘matched’ massage from its 


(to which it has sent a 

tmp 

j�Bi tmp tmp 

‘matching’ massage), it sets B � B / j and then 

i i


sends a ‘matching’ massage to the 


. 

tmp 

j�Bi (4) (1)-(3) repeat until all the node are matched or 

B � null 

tmp 

i 

In the DPM algorithm, sensor nodes are grouped 

through local negotiation with their neighbors. The sensor 

node tries to find a node which has the minimal joint 

entropy (maximum correlation) in his potential neighbors 

that have not be matched. 

With the pairs determined by DPM algorithm, pairwise 

DSC is performed between sensor nodes in each pairs. 

Note that there may be some nodes that cannot find any 

neighboring nodes as his partner. We assume the set of 

such nodes is Q . Let P the set of all the pairs determined 

by the DPM algorithm. S-W region (1) can be now 

written as 

Si � H( i j), Sj � H( j i), Si �Sj � H( i, j), �( i, j) �P 

Sm� H( m), �m�Q (2) 

B. Uncertainty Model of Energy Consumption 

For the designing of WMSNs, optimization with 

respect to energy is of most important work. However, 

Energy consumption may be uncertain due to several 

reasons: the inaccurate distance measurement, system 

noise or error, changes of ambient environment and 

deviation of the electric circuits. Here, we consider the 

uncertainty in energy consumption of transmitting 

multimedia data on wireless links. 

Denote eij the energy consumed by transmitting one bit 

on the link ( i, j ) . et � ( eij) (, i j) �Lis 

the energy consumption 

vector on all links. We argue that there is little correlation 

in the energy consumption uncertainty among different 

sensor nodes in practice. Thus we can use the following 

polyhedral set � to model the uncertainty, 

�� 

0 

� 

��e eij �eij �hij , � hij �� Ri, �i�Ns, h�0� 

(3) 

�� j�Vi �� 

which specifies a set of energy consumption vectors that 

are within a certain distance ( i R � ) from an nominal 

0 0 

vector, e � ( eij ) (, i j) �L 

, which can be seen as the estimated 

energy assumption. The parameter � �[0,1] controls the 

level of the uncertainty. 

In practical WMSN scenarios, we may know neither 

which energy consumption vectors in � should be use, 

nor the statistic characteristics of vectors in � . In this 

situation, we must assume that all the instances may 

happen. We should guarantee the robustness of our 

optimal solution under all the possible energy 

consumption vectors. Next, we will formulate such robust 

cross-layer optimization problem, and solve it using dual 

decomposition and subgradient method. 


C. Robust Optimization Problem Formation and its 

counterpart 

Our goal is to maximize the lifetime of a WMSN with 

correlated sources and uncertain energy consumptions. 

Assume that each node i has an initial energy of Ei � 0 . 

Let fij be the flow rate of link between node i and node 

j . The lifetime of a node i with flow rate vector 

f � ( fij ) (, i j) �L 

is 

Ti 

� 

Ei 

e f 

. 

� 

j�Vi ij ij 

The lifetime of an WMSN, T , is defined as the time 

when the first sensor node runs out of its energy, i.e., 

T � minT . i 

i�N Robust cross-layer optimization problem of a WMSN 

with correlated source is to determine all the source 

coding rates Si, i�Ns and the flow rate vector f on the 

links, which satisfy the constraints (2) and (4)-(6) under 

all the possible energy consumption vectors in � while 

maximizes lifetime of the WMSN. Specifically, we have 

the following robust optimization problem, 

max T 

st .. Si � H( i j), Sj � H( ji), 

Si �Sj � H(, i j), �(, i j) �P 

Sm� H( m), �m�Q f � f �S , �i�N (4) 

� � 

ij ji i s 

j�Vi j�Vi � 

T( f e ) � E , �i�N , e�� 

(5) 

j�Vi ij ij i s 

0 � fij �cij , �( i, j) �L 

(6) 

where 

� the constraints in (4) is the flow conservation law at 

each sensor nodes; 

� constraints in (5) is the energy conservation law at 

each node under all the possible energy 

� 

consumption vectors; 

constraints in (6) represent that the flow rate on a 

link should be less than the capacity of that link� 

The problem above is not a linear problem due to 

products in the constraints in (5). Note that the initial 

energy Ei � 0 implies T � 0 , and thus we can introduce a 

new variable q � 1/ T to obtain a equivalent linear 

problem which is 

max q 

st . . (2), (4), and (6) 

f e �qE , �i�NE�� (7) 

� 

j�Vi ij ij i s 

(P1) 

For convenience, we denote the above problem as P1. 

For robust optimization of the lifetime of a WMSN, 

constraints in (7) should be satisfied under all the possible 

energy consumption vectors, which is equivalent to


� fe � max � fh � qE, �i�N(8) j�Vi ij 

0 

ij 

� hij ��Ri, �i, h�0 

j�Vi j�Vi ij ij i 

The dual of the linear programming term (i.e., 

max � fh ij ij with variables h ij ) in (8) is 

� hij � Ri, i, h 0 

j�Vi � � � 

j�Vi min 

� Ry 

i i 

st .. y � f , �j�V, y �0 

i ij i i 

where yi is a dual variable. The constraints (8) is thus 

equivalent to 

� 

j�Vi fe ��Ry� qE 

ij 

0 

ij i i i 

y � f , y �0, �i�N, �j�V i ij i i 

Finally, we obtain the robust counterpart problem of 

P1 as 

min 

q 

st . . (2), (4), and (6) 

0 

� fe ij ij ��Ry i i 

j�Vi � qEi, �i�N(9) (P2) 

y � f , �j�V , y �0, 

i ij i i 

III. SOLUTION WITH DUAL DECOMPOSITON 

In this section, we derive the distributed algorithm for 

the robust cross-layer optimization of a WMSN by 

solving P2 using dual decomposition and subgradient 

algorithm. 

A. Dual Problem 

The problem P2 is a linear (and thus convex) problem. 

We can solve it in its dual domain since strong duality 

holds for convex problem under mild condition (which is 

satisfied for this problem). For a detailed discussion of 

this, refer [13]. However, the objective function of P2 

(with linear constraints) is not strictly convex and thus the 

dual function is not differentiable. We use a similar 

approach as [8] to change the objective of P2 as follow: 1) 

2 

Change the objective q to q since objective that 

minimizes q is equivalent to the one that minimizes 

, 

2 

q . 2) 

Add a small quadratic regularization term for each source 

rate and flow rate of each link to the objective. Then, the 

objectives of P3 becomes 

2 2 2 

q �� f �� 

S 

�� 

ij i 

i�N j�Vi i�N By choosing � and � small enough, the solution of the 

regularized problem can be arbitrarily close to that of the 

original one. 

By introducing Lagrange multipliers �i for the flow 

conservation constraints in (4) and �i for the robust 

energy conservation constraints in (9) at each sensor 

nodes, the Lagrangian of P2 can be given by 


Lq ( , f, y; 

�� , ) 

� q �� f �� 

S 

�� 

2 2 2 

ij i 

i�N j�Vi i�N � � � 

� � ( f � f �S 

) 

i ij ji i 

i�N j�Vi j�Vi 0 

��i( � fe ij ij ��Ry i i �qEi) 

i�N j�Vi 2 

��i i � �i i � 

2 

i 

i�N i�N � 

2 

� ( � fij j�Vi 0 

� fij ( �ieij ��i��j) i�N � �Ri�iyi �( q �q E ) � ( S � S ) 

� 

� 

� 

� 

�� 

� 

� 

�� 

Then, the objective function of dual problem can be 

written as 

D( �� , ) � min L( q, f, y; 

�� , ) 

and the dual problem is 

(2) and (6), yi 

�0, 

yi� fij, q�0 

max D( 

�� , ) 

st .. � �0, � � 0 

(DP1) 

B. Subgradient algorithm 

Since dual objective function is not strictly convex, 

subgradient method is used to solve DP1. To obtain the 

subgradient, we should solve the minimization problem 

in D( �� , ) given ��, , which can be decomposed into 

three independent subproblems, SP1, SP2, SP3 and SP4 

as follows. 

SP1: 

2 

min q � q � E 

SP2: 

SP3: 

st . q� 

0 

� 

� 

i�N i i 

� f � f �e �� R� 

y 

min ( 

2 

ij ij ( i 

0 

ij i j ) i i i 

j�Vi st .. 0 � f �c, y � f , y �0 

ij ij i ij i 

�S ��S � � S ��S 

2 2 

min ( i i i ) ( j j j) 

st .. S � H( i j), S � H( j i), 

i j 

Si �Sj � H(, i j), �(, i j) �P 

SP4: 

2 

min ( �mSm ��Sm) 

st .. Sm � H( m), �m�Q The subproblems actually correspond to the problems 

should be handled by different layers: 

� SP1 is to optimize the lifetime of the WMSN, and is 

an application layer problem. 

� SP2 is to determine the flow rate of each link, which 

can be seen as a routing problem of network layer. 

Due to the uncertain energy consumption, SP2 is 

actually a robust routing problem.


� SP3 is to allocating the source rates between 

correlated sources of each pair, and is the rate 

control problem of transport layer. SP4 is also the 

source rate problem, whose solution can be readily 

obtained as Sm� H( m) 

. 

� � � � With the obtained value q , Si , fij and yi 

by solving 

SP1, SP2, SP3 and SP4, the subgradients of �D( �� , ) 

the dual variables �i, �i can be given, respectively, by 

at 

f � 

� 

f � 

� � 

f �S 

, 

� � 

i ij ji i 

j�Vi j�Vi i � 

� 

i �� j�Vi � 

ij 

0 

ij ��i 

� 

i 

g q E f e R y 

k k 

k �1 

(10) 

For the k-th steps of the subgradient method, the dual 

variables are updated as 

k�1k �i �( �i ��k 

fi) 

� , 

k�1k �i �( �i ��kgi) 

� 

(11) 

where ( x) � � x if x � 0 , and ( ) 0 x � � , otherwise. �k is a 

positive scalar step-size. 

According to [13], subgradient method is guaranteed to 

converge to the optimum if step-sizes �k appropriately as follows. 

are designed 

Theorem 1: Dual variables � and � converge to the 

� � optimal dual solutions ( � , � ) if the positive scalar stepsizes 

�k are chosen such that 

� 

lim� �0, � �� 

k �� 

� 

Remark1: Since strong duality holds, the corresponding 

� � � 

primal variables ( q , r , y ) are globally optimal variables 

of primal problem P2 (and also P1) for optimal dual 

� � variables ( � , � ). 

Remark2: We can see that through subgradient update 

algorithm, four subprolems are coordinated by Lagrange 

multipliers � , � 

and cooperatively work with each other to achieve the 

whole goal of maximizing the lifetime. 

� � (i.e., dual variables in the dual problem) 

IV. IMPLEMENTATION OF THE PROPOSED SOLUTION 

In this section, we would discuss how to implement the 

solution in the optimization in a practical wireless 

multimedia sensor network. We will show that the 

solution obtained in the last section implies a partially 

distributed algorithm to calculate the routing (i.e. 

determine the flow rate on each link) and the source rate 

in the present of correlated sources and energy 

consumption uncertainty to maximize the lifetime of the 

WMSN. 

Assume that a small scale wireless multimedia sensor 

network as illustrated in Figure.1. The sink node is 

partially responsible for the organization and 

management of the whole WMSN. The WMSN suffer the 

uncertainty of energy consumption which is modeled in 

(3) in section II. Assume that the time is slotted. Every 

time the correlation of the WMSN changes, for example 


the locations or orientations of the cameras in wireless 

visual sensor networks changes, the whole network 

would calculate the routing and source rates with the 

following partially distributed algorithm (Algorithm 1) 

that is derived from the solution in the last section. 

A. Distributed algorirthm 

With the pairwise DSC and the subgradient method, 

we propose the following distributed algorithm to the 

robust cross-layer optimization of a WMSN with 

correlated sources and uncertainty, which works as 

follow: 

Algorithm 1: Robust cross-layer optimization of 

WMSN with correlated sources and uncertainty 

1) Pairwise matching of source sensor nodes using the 

DPM algorithm. 

0 0 

2) Initializing dual variables with �i , �i 

At the k-th step, nodes in the WMSN implement the 

k k 

following operations (dual variables are �i , �i in k-th 

step): 

3) The sink node calculate the optimal qk ( ) with the all 

k the �i , �� i N , which is sent to it from all the sensor 

nodes at the end of the last step, 

k 

qk ( ) � � �i 

Ei 

/2 

i�N and then sends the resulting qk ( ) 

nodes. 

to all the sensor 

4) Each pair of sensor nodes �(, i j) �P solve the 

constrained quadratic programming SP3 to obtain 

optimal source rate for their pairwise DSC; 

5) Each sensor node m 

Sm� H( m) 

. 

�m�Q uses source rate 

6) Sensor node i, �� i N solve the quadratic 

programming SP2 to obtain optimal fij ( k) and yi( k) 

7) Each sensor node i, �� i N calculate the subgradient 

value i f and gi as (10), and updates the dual 

variables �i , �i as (11). 

k 1 

8) Every sensor sends the value of �i � to the sink node 

k 1 

and its neighboring nodes, and the value of �i � neighboring nodes. 

9) Go to step 3) and repeat until convergence. 

to its 

Remarks 1: To facilitate the distributed implementation 

of subgradient method in the above algorithm and reduce 

the massage pass in each step, we use a small constant 

step size � that is same for all the nodes and in all the 

steps. The subgradient method can converge to a small 

neighborhood of the optimal solution of DP1 with 

constant step size that is small enough [13]. 

Remarks 2: The calculation of the source coding rate in 

step 4) can be implemented in any sensor node of the 

pairs, and the results are reported to the other node by 

massage passing. We assume that nodes in pairs know 

their joint entropy and conditional entropy.


B. Sumary 

We show the whole picture of the robust cross-layer 

design of WMSN with correlated sources and uncertainty 

in Figure 2. The original problem is a lifetime 

maximization problem with constraints of link rate, 

source rate, energy conservation, and robustness 

guarantee. Though we can transform the problem into a 

linear programming, it still involves a lot of variables. 

Centralized algorithm like simplex or interior-point 

methods can be used to solve it. However, they are not 

efficient approaches to be implemented in the practical 

operation of the WMSNs. Sink node needs to gather all 

the information (topology, link, uncertainty, correlation 

and so on), calculates the solution centrally, and 

distributes the results to every node. This will not only 

shorten the lifetime of the whole network due to energy 

consumption in information collection and distribution, 

but also increase the delay in the delivery of multimedia 

data. It may offset the benefits from cross-layer 

optimization. 

Robust lifetime optimization problem with 

correlated sources and uncertainty 

Application layer: 

Objective of the network 

Transport layer: 

Source rate control 

Network layer: 

Routing (flow rates on 

links) 

Goblal variables 

dual variables 

local variables 


local variables 


Cross-layer 

coordination 

by dual 

variables 

update 

Figure 2. Cross-layer design architecture of robust optimization of 

WMSN with correlated sources and uncertainty 

In this work, we deal with the problem in its dual 

domain. The main advantage of doing this is that dual 

problem can be further divided into some separate 

subproblems, which correspond to the functions in 

different protocol layers. As shown in Figure 2, the 

subproblems are, respectively, 

� the application layer problem (system objective, i.e., 

lifetime maximization), 

� transport layer problem (source rate control) and 

� network layer problem (routing). 

Dual problem can be solved using subgradient method. 

Local variables (except lifetime variable in Algorithm 1) 

are needed to update the subgradient, and thus it can be 


used to develop a distributed algorithm. Except 

application layer problem, the other two problems can be 

implemented distributedly in Algorithm 1. 

Dual variables are “bridges” connecting and 

coordinating the separate subproblems in different layers. 

As shown in Figure 2, through dual variables, our crosslayer 

design not only achieves the optimal objective, but 

also keeps the layering architecture developed for the 

computer networks, which has the advantages of 

modularity. 

V. PERFORMANCE EVALUATION 

In this section, we conduct simulations to investigate 

the performance of the proposed algorithms. The 

simulations are implemented in MATLAB software. 

In our scenario setup, we assume the following radio 

model for the nominal energy consumption in the 

0 t 2 

transmission on a wireless link: ei, j� e di, 

j . The unit of 

t 

2 

e is J bytes m and di, j is the distance between 

transmitter i and receiver j . There are 30 sensor nodes 

and one sink, which are randomly deployed in an area of 

2 

200� 200 m . The initial energy of every sensor node is 

10,000 t 

e .Set 100 t 

Ri�e and Ci, j�18, 

�( i, j) � L. 

For simplicity, we assume that the correlation between 

source nodes only depends on the distance between them. 

We use the following correlation model: Hij (, ) � Hd 0 i, j 

and Hij ( ) � H( ji) � Hd 1 i, j . In practical scenarios with 

camera sensors, more complex models, which may 

depend on not only distance but also orientation and focal 

length, can be used. We set H0 � 0.02 and H1 � 0.005 . 

A. Performance of DPM algorithm 

We compare the performance of DPM algorithm with a 

random matching (RM) algorithm. For random matching 

algorithm, each sensor node finds a partner in its 

neighbors randomly by massage passing. We replace the 

DPM in Algorithm 1 with the random matching 

algorithm for Algorithm 1 with RM, and show the 

resulting lifetime of Algorithm 1 with DPM and RM in 

Figure 3. 

The value of lifetime 

2.8 

2.6 

2.4 

2.2 

2 

1.8 

1.6 

Algorithm 1 with DPM 

Algorithm 1 with RM 

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 

Uncertianty parameter � 

Figure 3. Comparison of DPM and RM


We can see from Figure 3 that the lifetime obtained 

by Algorithm 1 with DPM is larger than that with RM 

(more than 23% for this scenario). In addition, the 

lifetime decreases with the increase of the uncertainty 

parameter. This is because the robust algorithm would 

sacrifice some performance to confront the uncertainty. 

This is desirable, especially when the uncertainty is 

relative large (as illustrated in Figure 4). 

B. Performance of Robustness 

We compare Algorithm 1 with a deterministic solution 

which does not consider the uncertainty in energy 

consumption. The deterministic solution optimizes 

0 

WMSN with the nominal energy consumption (i.e., e i, j). 

To show the benefit of the robust optimization, we 

calculate the following two ratios: 

L �L L �L 

R � R � 

L L 

det robust robust det 

0 un un wc 

e e e e 

1 , 2 

det det 

0 

e 

wc 

e 

det where L 0 is the optimal lifetime of deterministic 

e 

robust 

solution; L un is the lifetime obtained by robust 

e 

optimization considering the uncertainty; det 

L wc is the 

e 

lifetime of deterministic solution in the worst energy 

consumption. 

The ratio R1 is the relative decrease of optimal lifetime 

of robust solution in the nominal case, while ratio 

R2 reflects the relative increase of the lifetime of robust 

solution over deterministic solution in the waste case. The 

results are shown in Figure 4. 

We can observe from Figure 4 that R2 is larger than 

R1 and the difference increases with the increase of � . 

This means that robust solution is more desirable when 

uncertainty becomes large. It guarantees the optimality 

under all the instances of the uncertain energy 

consumption vectors with small performance loss 

(compared with the deterministic energy consumption). 

The value of ratios: R1 and R2 

0.12 

0.1 

0.08 

0.06 

0.04 

0.02 

0 

R1 

R2 

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 

Uncertianty parameter � 

Figure 4. Comparison of the ratio R1 and R2 


VI. CONCLUSION AND FUTURE WORK 

Robust cross-layer design of wireless multimedia 

sensor network with correlated sources and uncertain 

energy consumption was investigated. To reduce the 

redundant information and complexity, distributed source 

coding was implemented in pairs which were determined 

by a distributed pairwise matching algorithm. A 

polyhedral set was used to model the uncertainty in 

energy consumption. Robust lifetime maximization 

problem was formulated and its counterpart was showed 

to be a convex problem with linear constraints. Dual 

decomposition and subgradient method were used to 

solve the problem, which leads to a cross-layer design 

approach and a partially distributed solution. Dual 

variables were used to coordinate the separate 

subproblems in each layer, and thus layering architecture 

and modularity were maintained in the cross-layer design. 

The obtained distributed algorithm can be implemented in 

small scale WMSNs with sink node responsible for the 

calculation and distribution of the value of the lifetime. 

Simulation results verified the performance of the 

proposed algorithms, especially the benefits of robust 

optimization under large uncertainty case. 


This work is partially supported by Innovation 

Foundation of Tianjin University (Grant No.60302014), 

Tianjin Science and Technology Support Program (Grant 

No.09ZCKFGX0170), NSFC of China (Grant 

No.60972054), 863 Program of China (Grant 

No.2009AA011507, SQ2009AA01XK1485134), and 

Major Programs of Ministry of Science and Technology 

of China (Grant No.2009ZX03004-006). 

REFERENCES 

[1] I. F. Akyildiz, T. Melodia, K. R. Chowdhury, “A survey on 

wireless multimedia sensor networks,” Comput. Netw. 

(Elsevier), vol. 51, no. 4, Mar. 2007, pp. 921–960. 

[2] Akyildiz I. F., Melodia T., Chowdury K. R., "Wireless 

Multimedia Sensor Networks: Applications and Testbeds," 

Proceedings of the IEEE (invited paper), vol. 96, no. 10, 

October 2008, pp. 1588-1605. 

[3] Akyildiz, I. F., Melodia, T., Chowdury, K. R., "Wireless 

Multimedia Sensor Networks: A Survey," IEEE Wireless 

Communications Magazine, vol. 14, no. 6, December 2007, 

pp.32-39. 

[4] Nakashima Hideyuki, Aghajan Hamid, Augusto Juan 

Carlos, Handbook of Ambient Intelligence and Smart 

Environments. Springer Press. 2010. 

[5] D. Slepian, J. K.Wolf, “Noiseless coding of correlated 

information sources,” IEEE Trans. Inform. Theory, vol. IT- 

19, Jul. 1973, pp. 471–480. 

[6] X. Guo, Y. Lu, F. Wu, W. Gao, S. Li, “Distributed 

multiviewvideo coding,” Proc. Visual Communications 

and Image Processing (VCIP), San Jose, CA, USA, 

January 2006. 

[7] A. Sankar, Z. Liu, “Maximum lifetime routing in wireless 

ad-hoc networks,” in Proc. IEEE INFOCOM 2004, pp. 

1089–1097. IEEE press. Hong Kong (2004). 

[8] R. Madan, S. Lall, “Distributed algorithms for maximum 

lifetime routing in wireless sensor networks,” in Proc.


IEEE GLOBECOM 2004, pp.2185–2193. IEEE press. 

Dallas (2004). 

[9] Ilias Politis, Michail Tsagkaropoulos, Tasos Dagiuklas, 

Stavros Kotsopoulos, “Power Efficient Video Multipath 

Transmission over Wireless Multimedia Sensor Networks”. 

Mobile Netw Appl, Vol.13, 2008, pp.274–284 

[10] F. Ordonez, B. Krishnamachari, “Optimal information 

extraction in energy-limited wireless sensor networks,” 

IEEE J. Select. Areas Commun., vol. 22, no. 6, 2004, pp. 

1121–1129. 

[11] Jamal N. Al-Karaki, Raza Ul-Mustafa, Ahmed E. Kamal, 

“Data aggregation and routing in Wireless Sensor 

Networks: Optimal and heuristic algorithms”. Computer 

Networks, vol. 53, no. 7, 2009, pp. 945-960. 

[12] Wei Ye, Fernando Ord´o˜nez. “Robust Optimization 

Models for Energy-Limited Wireless Sensor Networks 

under Distance Uncertainty”. IEEE transactions on 

wireless communications, vol. 7, no. 6, Jun. 2008, pp. 

2161-2169. 

[13] S. Boyd, L. Xiao, and A. Mutapcic, “Subgradient 

methods,” in lecturenotes of EE392o, Stanford University, 

Autumn Quarter 2003-2004. 

[14] R. Cristescu, B. Beferull-Lozano, and M. Vetterli, 

“Networked Slepian-Wolf: Theory, algorithms, and scaling 

laws,” IEEE Trans. Inform. Theory, vol. 51, no. 12, pp. 

4057–4073, Dec. 2005. 

[15] L. Tassiulas and A. F. Ephremides. , “Stability properties 

of constrained queueing systems and scheduling policies 

for maximum throughput in multihop radio networks, ” 

IEEE Transactions on Automatic Control, vol. 37, no. 12, 

Dec.1992, pp. 1936–1948. 

[16] X. Lin and N. Shroff, “The impact of imperfect scheduling 

on cross-layer congestion control in wireless networks,” 

IEEE/ACM Trans. Networking, vol. 14, no. 2, April 2006, 

pp. 302–315. 

[17] M. Neely, E. Modiano, and C. Rohrs, “Dynamic power 

allocation and routing for time-varying wireless networks,” 

IEEE J. Select. Areas Commun., vol. 23, no. 1, Jan. 2005, 

pp. 89–103. 

[18] L. Chen, S. H. Low, M. Chiang, and J. C. Doyle, “Crosslayer 

congestion control, routing and scheduling design in 

ad hoc wireless networks,” in Proc. of IEEE Infocom, Apr. 

2006, pp.1-13. 

Lei You received the Ph.D degree from Beijing University of 

Posts and Telecommunications, Beijing, China, in 2010. During 

2008.9-2009.8, He was a visiting PhD student with Uppsala 

University, Uppsala, Sweden. He is currently a lecture of the 

school of Electronic Information Engineering at Tianjin 

University. His research interests include wireless multi-hop 

networks and cross-layer optimization. 

Chungui Liu received the Ph.D degree in Computer 

Applications from Tianjin University, Tianjin, China, in 

2009. He is currently working as a Post-Doctor with 

school of Electronic Information Engineering at Tianjin 

University. His research interests include wireless mesh 

networks and Internet of Things (IoT). 



The E-Commerce Model of Health Websites: An 

Integration of Web Quality, Perceived 

Interactivity, and Web Outcomes 


Department of Health Administration, Tzu Chi College of Technology, R.O.C. 

Email: tsairob@tccn.edu.tw 

Abstract—The study integrates web quality (system 

quality, information quality, and service quality), 

perceived interactivity (human-message, humanhuman), 

and web outcomes (web usage, web 

satisfaction, and web loyalty) to explore the ecommerce 

model of health websites. A survey of 1076 

users of health websites was conducted to validate the 

proposed model. The findings show that web quality 

has significantly positive effect on perceived 

interactivity, web usage, and web satisfaction 

separately, which in turn influence web loyalty. This 

study also confirms that perceived interactivity is an 

important mediator between web quality and web 

outcomes. This study emphasizes the importance of 

both web quality and perceived interactivity in the 

progress towards success health websites. The findings 

may be used as theoretical base for future research 

and can also offer empirical foresight to executives 

and managers of hospitals when they initially 

introduce and upgrade the health websites into their 

organizations. 

Index Terms—web quality, perceived interactivity, web 

usage, web satisfaction, web loyalty 


Ever since the Internet emerged in the 1990s, a great 

number of situations and modes of commercial 

competitions have had tremendous changes. Lots of 

customer-oriented service industries have started to set up 

the platforms and portals on the Internet to serve the 

customers so that the customers can be connected with 

the services the organizations offer at any time no matter 

how far they are or where they are. Moreover, many 

traditional commercial behaviors can also be conducted 

with the long-distance virtual transactions through the 

Internet. The Internet has changed not only the modern 

people’s living styles but also the transaction modes 

between contemporary businesses and customers. 

Therefore, the representative organizations, hospitals, 

which offer the service of medical treatments, have also 

gradually valued this tendency and trend. 

Hospitals can take advantage of the websites to 

provide the patients and their families with the health 

information so that they can learn the latest knowledge of 


doi:10.4304/jnw.6.7.1017-1024 

medicine. The mechanism of the online registration can 

also be utilized to create another channel to seek medical 

advice. Alternatively, the conferences or seminars 

regarding the health education can be regularly held in 

the hospitals and post the contents onto the websites for 

the general public to browse the films and download the 

briefings and other materials. Lots of healthcare courses 

can even be offered online. For instance, the prenatal 

healthcare course for the couples, the smoking cessation 

course, bodybuilding course and so on. These functions 

are no longer just the static introduction to each clinic and 

department. Instead, they have become the virtual 

clearinghouse for the health information [1]. At present, 

numerous websites of hospitals and medical institutions 

have been equipped with these features. 

Health website is not only a health communication 

channel but also a full representation of a service 

department to the customer. With the establishment of the 

reliable and popular health websites, the customers can be 

provided with the general healthcare information, and 

hospital-customer relationships can be further reinforced. 

In addition, studies have showed that the customers 

satisfied with a website would have higher level of 

customer loyalty [2]. Thus, it’s critical to investigate how 

the websites quality to affect health web outcomes (e.g. 

web usage, web satisfaction, web loyalty). 

In addition, Internet has transformed the traditional 

physician-patient relationship because those who use the 

Internet frequently ask their physicians more specific 

questions and suggest specific illnesses and treatments [3]. 

Lustria [4] also showed that the use of interactive 

technology could enhance learning and persuasion of 

content on the basis of well health web quality. Therefore, 

it’s very crucial to explore the issue of interactivity of 

health websites. 

Accordingly, the purpose of the study is to examine 

how web quality (system, information, and service 

quality) and perceived interactivity (human-message and 

human-human) affect the websites’ loyalty through users’ 

attitude toward websites (website usage and satisfaction). 

The research model will be empirically tested using the 

structural equation modeling (SEM). Through the 

statistical analysis, we can investigate the interaction 

between technological and social factors, and furthermore 

find out the important antecedents of websites’ loyalty. In 

this way, we hope to provide the managers of hospitals


and the administrators of the IS department with the 

insight and reference regarding the management of 

hospital websites. 

II. LITERATURE REVIEW 

2.1 Web Quality 

Aladwani and Palvia [5] define website quality as 

users’ evaluation of a website’s features meeting users’ 

needs and reflecting overall excellence of the web site. 

Hwang and Kim [6] on the other hand define website 

quality as the user’s perception on the customer service 

and privacy based on the website interface and functions. 

Liu and Arnett [7] conducted a survey on the top 

thousand businesses as listed in Fortune magazine, and 

found four factors that relevantly affected the success of 

the website: (1) information and service quality, (2) 

system use, (3) playfulness, and (4) system design quality. 

In DeLone and McLean’s research [8] on e-commerce 

systems and the measurement of quality, it was found that 

other than system and information quality, the importance 

of customers support in e-commerce system was essential, 

thus emphasizing the importance of the quality of service. 

Ahn, Ryu, and Han [9] believe that the technologyfocused 

approach sees a website as an information system 

and focuses on system and information quality, while a 

service-focused approach see a website as a service 

provider and includes service quality. According to the 

arguments of the above literature, even though 

measurements for website quality may be changed 

according to research purpose or field, the main 

classification still uses the categories of system, 

information, and service quality. 

2.2 Perceived Interactivity 

Most scholars agree that interactivity is a strong and 

important marketing characteristic of the World Wide 

Web as compared to other traditional media (e.g. 

television, newspaper, magazines, etc.). After review of 

29 articles on interactivity, McMillan and Hwang [10] 

classified the various scholarly definitions of interactivity 

in four categories: (1) process, (2) feature, (3) perception, 

and (4) a combination of the three. McMillan [11] found 

that website interactivity was based on two-way 

communication, levels of control, user activity, sense of 

place, and time sensitivity. McMillan [12] further pointed 

out that there are three dimensions of perceived 

interactivity: two-way communication, controls of 

navigation (or choices), and time to load/time to find. Wu 

[13] found that perceived interactivity was the 

psychological state experienced by the website visitor 

during the interaction process. It consists of three 

dimensions: perceived controls, perceived responsiveness, 

and perceived personalization. Ko, Roberts, and Cho [14] 

define perceived interactivity as “the degree to which 

people engage in a communication process by actively 

interacting with mediated messages and other people.” 

They found that the two most frequently occurring and all 

encompassing dimensions were human-message 

interactivity and human-human interactivity. 


From a healthy communication perspective on the 

World Wide Web, Cassell, Jackson, and Cheuvront [1] 

argued that the World Wide Web was suitable for 

persuasive communication. This type of communication 

is a form of social influence that can effectively affect 

internalization of specific attitudes and in turn affect 

behaviors. Also, Lustria [4] found that the users of web’s 

hypertext are able to freely browse through the system, 

and randomly access material, processing information 

according to individual mental models, and redefine 

learning structure and content. This research also showed 

that high levels of perceived interactivity promoted high 

levels of comprehension of the content of website. Thus, 

high levels of perceived interactivity of website would 

lead to strengthen the effect of online learning and 

persuasion so that the user feels the convenience, 

usefulness, and enjoyment of visiting the websites. Due 

to the aforementioned reasons, this study uses “perceived 

interactivity” to measure website interactivity. Besides, 

perceived interactivity includes two dimensions: humanmessage 

interactivity and human- human interactivity. 

2.3 The relationship between website quality and 

perceived interactivity 

Wu [13] suggested a conceptual structure for the 

antecedent and consequential variables in interactivity. 

That is, website factors (actual interactivity, vividness, 

and design), site-visitor factors (personality traits, product 

knowledge, and web skills), and situational factors (visit 

motivation, access speed, and visit location) are three 

types of factors that influence perceived interactivity. 

Song and Zinkhan [15] manipulated 16 different 

versions of a website in an experiment directed at college 

students to prove that speed and message format 

(personalization of messages) would positively influence 

perceived interactivity. Because the response time of the 

website is a characteristic of system quality, and website 

messages are one kind of information quality, it is 

predicted that website quality may affect perceived 

interactivity. Therefore, this research proposes the 

following hypothesis: 

H1a: Website quality has a positive effect on 

perceived interactivity. 

According to information systems success model 

proposed by DeLone and McLean [16], system quality, 

information quality, and service quality are related to 

usage and user satisfaction of an information system. 

DeLone and McLean [8] showed that D&M IS Success 

Model also could be adapted to the measurement of the ecommerce 

systems. 

Hwang and Kim [6] also showed that website quality 

would positively influence affective reaction of users, 

which is a subjective perception or judgment about 

whether such interaction will change their core affect or 

emotion toward the website. Therefore, this research 

proposes the following hypothesis: 

H1b: Website quality has a positive effect on website 

usage. 

Hic: Website quality has a positive effect on website 

satisfaction.


2.4 The relationship between website quality and 

perceived interactivity 

Jee and Lee [17] found that perceived interactivity has 

positive impact on the user’s attitudes toward the website 

and intent of purchase. McMillan, Hwang, and Lee [18] 

found that “perceived interactivity” was a better predictor 

of user attitudes toward the website than “actual 

interactivity”. Similarly, Song and Zinkhan [15] also 

found that perceived interactivity was positively 

correlated with website loyalty and attitude. Therefore, 

this research proposes the following hypotheses: 

H2a: Perceived interactivity has a positive effect on 

website usage. 

H2b: Perceived interactivity has a positive effect on 

website satisfaction. 

2.5 The relationship between website use, website 

satisfaction, and website loyalty 

DeLone and McLean [8,16] argued that system usage 

and user satisfaction both affect the user’s net benefit. 

Besides, Dick and Basu [19] found that sustainable 

loyalty could only be achieved when the customer enjoys 

a high level of positive attitudes (satisfaction) toward the 

product as well as a high level of repetitive patronage. 

III. RESEARCH METHOD 

In order to provide the public with correct and current 

online health information, Department of Health (DOH) 

hold the activities of excellent awards of health 

information websites to assess the websites of all 

hospitals in Taiwan since 2002. We mail the invitation 

letters to the executives of hospitals who have obtained 

the excellent awards to express our need for the research 


Figure 1. The Proposed Research Model 

In the research of virtual community websites, Kuo 

[20] found that there was a relationship between 

continuous usage and overall satisfaction, and continuous 

usage, satisfaction, and loyalty were also related. Otim 

and Grover [21] proved that product satisfaction 

influences customers repeat purchase intention (loyalty) 

in the context of website service. Kassim and Abdullah 

[22] also proved that customer satisfaction has positive 

direct effect on loyalty in e-commercial environments. 

Casalo, Flavian, and Guinaliu [2] also found that 

customers satisfy with previous interactions with the bank 

websites has positive effect on customer loyalty, and 

website usability has also positive effect on customer 

loyalty. Therefore, this research proposes the following 

hypotheses: 

H3a: Website usage has a positive effect on website 

satisfaction. 

H3b: Website usage has a positive effect on website 

loyalty. 

H4: Website satisfaction has a positive effect on 

website loyalty. 

Based on the review of the literature, figure 1 presents 

the conceptual framework from which the proposed 

research model is formed. 

purpose. Of these hospitals contacted, five teaching 

hospitals (located in northern, central, southern Taiwan) 

were willing to participate in the survey. Before we can 

conduct the survey, it must be approved by IRB 

(Institutional Review Board) of hospitals. Distribution 

and collection of survey questionnaires was coordinated 

with the help of the executives, and information systems 

managers of the hospitals. 

We used a self-report questionnaire to empirically 

validate the proposed research model. The questionnaire


was pilot tested using 30 hospitals’ patients who had 

prior experience in online websites. These items were 

revised according to the feedback. After the revision, the 

survey was conducted to a convenient sample of 1200 

patients for four months. Of the 1200 samples, the 

samples with incomplete responses and missing data 

were deleted. Finally, the eligible samples of 1076 

patients were yielded, and the total response rate is 

89.67%. 

IV. RESULT 

The data analysis proceeded according to a two-step 

approach [23]. First, we assessed the measurement model, 

which consists of the six latent factors, including the 

assessment of reliability, discriminant validity, and 

convergent validity of the scales. Second, we validated 

the structural model, which represents the series of path 

relationships linking the six constructs. 

4.1 Sample Characteristics 

Of these respondents, 653 respondents are women 

(60.7%), 37.6% are age 30 and below. The education 

levels of mostly respondents are university (40.8%). The 

majority of respondents’ career belongs to service 

industry (24.1%). Mostly respondents lived in northern 

Taiwan (49.8%). The times using the internet is mostly 

10 times and above per week (32.1%), while 1~3 times 

and above per day (46.6%). Table I presents descriptive 

statistics for the seven constructs in the study. The mean 

scores for seven constructs are all almost on the middle 

point of 5-point Likert-type scales, and show a 

reasonable dispersion in their distributions across the 

ranges. 

Table I Sample demographics 

Construct Mean 

Standard 

Deviation 

Minimum Maximum 

Web Quality 3.96 0.50 2.33 5.00 

System Quality 4.03 0.61 2.00 5.00 

Information Quality 3.97 0.51 2.00 5.00 

Service Quality 3.90 0.61 2.00 5.00 

Perceived Interactivity 4.07 0.61 2.50 5.00 

Human-Message 4.12 0.66 1.00 5.00 

Human-Human 4.01 0.69 2.00 5.00 

Web Usage 3.83 0.75 1.67 5.00 

Web Satisfaction 3.98 0.63 2.00 5.00 

Web Loyalty 3.93 0.74 1.50 5.00 

4.2 Measurement Model Results 

To validate the measurement model, three types of 

validity were assessed: content validity, convergent 

validity, and discriminant validity. Content validity was 

done by interviewing senior system users and pilottesting 

the instrument. And the convergent validity was 

validated by examining Cronbach’s α, composite 

reliability and average variance extracted from the 

measures [24]. As shown in Table II, the Cronbach’s α of 

every subscales range from 0.81 to 0.92, which are above 

the acceptability value 0.7 [25]. Besides, the composite 

reliability values range from 0.81 to 0.90, and the 

average variances extracted by our measures range from 

0.52 to 0.76, are all within the commonly accepted range 

greater than 0.5 [24]. In addition, all measures are 

significant on their path loadings at the level of 0.001. 

Therefore, the convergent validities of all seven 

constructs are confirmed. 

Discriminant validity of the sub-dimensions of 

2 

constructs was validated by comparing the � values of 

the CFA with original sub-dimensions of every construct 


against other CFAs which every possible combination of 

two dimensions (the correlation coefficient of two 

dimensions assigned to be 1) was examined. As shown in 

2 

Table III, the � values of the CFA with original subdimensions 

of web quality (system quality, information 

quality, and service quality) and perceived interactivity 

(human-message and human-human) were significantly 

better than any possible union of any two dimensions. 

Therefore, the discriminant validities of the subdimensions 

of the two constructs are confirmed. 

Besides, according to Fornell and Larcker [26], 

discriminant validity can also be tested among all 

constructs by comparing the average variance extracted 

(AVE) of each construct with the squared correlation of 

that construct and all the other constructs. As shown in 

Table IV, all squared correlations between two 

constructs are less than the average variance extracted of 

both constructs. Therefore, the results confirm that the 

discriminant validity of constructs in the study is 

satisfactory.


Web Quality 

Table II Construct Reliability and Convergent Validity 

Construct Cronbach’s α Composite Reliability 

Average Variance 

Extracted 

Web Quality 0.92 0.90 0.76 

System Quality 0.85 0.85 0.60 

Information Quality 0.82 0.81 0.52 

Service Quality 0.87 0.87 0.63 

Perceived Interactivity 0.88 0.90 0.76 

Human-Message 0.86 0.87 0.68 

Human-Human 0.81 0.82 0.60 

Web Usage 0.86 0.86 0.68 

Web Satisfaction 0.81 0.81 0.60 

Web Loyalty 0.90 0.90 0.67 

Table III Discriminant Validity of Sub-Dimensions of Web Quality and Perceived Interactivity 

Model 

2 

� d.f. Δ 2 

� 

1.Not Restricted 440.878 51 - 

2.System Quality and Information Quality Assigned to 1 908.829 52 467.951*** 

3. System Quality and Service Quality Assigned to 1 778.147 52 337.269*** 

4. Information Quality and Service Quality Assigned to 1 845.428 52 404.550*** 

Perceived Interactivity 

1.Not Restricted 56.569 13 - 

2. Human-Message and Human-Human Assigned to 1 403.170 27 346.601*** 

*** p


Table V Fit Indices for the Structural Model 

Structural Model Statistic Fit Indexes 

Recommended 

Threshold 

Figure 2 illustrate the results of the structural model 

with the estimated standardized path coefficients and 

path significance among constructs (non-significant 

paths as dotted lines). As predicted, all proposed 

hypotheses are supported. Table VI illustrates the 

squared multiple correlations (R 2 ) of all endogenous 

variables in the model. The estimated standardized path 

coefficients indicate the strengths of the relationships 

between the dependent and independent variable. 

Meanwhile the R 2 value represents the proportion of 

variance that is explained by the predictors of the 

variable in the model. 

As expected, web quality (β=0.940) has significant 

effects on perceived interactivity, accounting for 88.4% 

of the variance in the construct. Web quality (β=0.292) 

and perceived interactivity ( β =0.546) have all 


2 

� 1458.789 - 

2 

� / d.f. 4.355 < 5 

GFI 0.90 > 0.9 

RMSEA 0.056 < 0.08 

AGFI 0.88 > 0.8 

NFI 0.93 > 0.9 

RFI 0.92 > 0.9 

IFI 0.95 > 0.9 

TLI 0.94 > 0.9 

CFI 0.95 > 0.9 

Figure 2 Final Proposed Model 

significant effects on web usage, accounting for 68.2% of 

the variance in the construct. Besides, web quality (β 

=0.239), perceived interactivity ( β =0.454), and web 

usage ( β =0.287) have significant effects on web 

satisfaction, accounting for 87.5% of the variance in the 

construct. Web usage (β=0.292), and web satisfaction 

(β=0.627) are both significant predictors of web loyalty, 

accounting for 78.9% of the variance in the construct. 

The results of the structural model show that web 

quality (system quality, information quality, and service 

quality), perceived interactivity (human-message and 

human-human) are two key aspects affecting web 

outcomes of hospitals’ websites (web usage, web 

satisfaction, and web loyalty). The results also 

demonstrate that web quality has significant impact on 

web outcomes mediated by perceived interactivity.


V. DISCUSSION 

This study proposed a research model to better 

understand the e-commerce model of health websites. 

The model considered the relationships among web 

quality (system quality, information quality, and service 

quality), perceived interactivity (human-message and 

human-human), and web outcomes (web usage, web 

satisfaction, and web loyalty). Moreover, the model aims 

to interpret that perceived interactivity is an important 

mediator between web quality and web outcomes. The 

results of this study are discussed below. 

The results of this study suggest that web quality 

consists of three dimensions: (1) system quality, (2) 

information quality, and (3) service quality. Furthermore, 

positive perceptions of health web quality predict 

customers’ perceived interactivity, web usage, and web 

satisfaction. Previous studies having found similar results. 

Hwang and Kim [6] proposed a conceptual framework to 

interpret how web quality influences affective reaction. 

Also, Ha and Stoel [27] proposed the extended 

technology acceptance model of online shopping 

techniques to show high quality e-shopping sites should 

result in the perception that one’s experience is enjoyable 

and trust in e-shopping. It implies that online customers 

are more likely to feel positive affective reaction (usage 

and satisfaction) when they feel the health website is 

well-designed, knowledgeable, and responsive. 

This study also confirms that perceived interactivity is 

an important mediator between web quality and web 

outcomes. Perceived interactivity has significantly 

positive effect on health web usage and web satisfaction. 

The findings also support previous empirically research 

(e.g. [15]). The finding proved the mediating role of 

perceived interactivity in affecting the effect of web 

quality on online customers’ perception of web usage 

and web satisfaction. It is consistent with the results of 

previous research (e.g. [28,29]). Interestingly, the 

empirical evidence supports again the mediating role of 

perceived interactivity. 

The integrative viewpoint implies that an online health 

website is not only an information system but also but 

also a service provider/department to the customer. 

Accordingly, in the developing and maintaining phase of 

online health websites, system engineers and managers 

should value simultaneously system functions, health 

contents, and follow up service in order to pursue better 

web quality. On other hand, perceived interactivity plays 

important roles in web outcomes. It implies that online 

customers are more likely to continue to use a health 

website when they feel the health web is playful and 


Table VI The Squared Multiple Correlations (R 2 ) 

Construct R 2 

Perceived Interactivity 0.884 

Web Usage 0.682 

Web Satisfaction 0.875 

Web Loyalty 0.789 

interactive. Some strategies that managers of hospitals 

could use to increase the level of perceived interactivity, 

such as quick navigation, personalized web page, 

transmission of relevant messages, value-added search 

mechanism, bulletin boards, multi-media contents, etc. 

REFERENCES 

[1] M. M. Cassell, C. Jackson, and B. Cheuvront, 

“Health Communication on the Internet: An 

effective channel for health behavior change?,” 

Journal of Health Communication, Vol. 3, pp. 71-79, 

1998. 

[2] L. V. Casalo, C. Flavian, and M. Guinaliu, “The role 

of satisfaction and website usability in developing 

customer loyalty and positive word-of-mouth in the 

e-banking services,” International Journal of Bank 

Marketing, Vol. 26, No. 6, pp. 399-417, 2008. 

[3] K. Altinkemer, P. De, and Z. D. Ozdemir, “Toward 

a Consumer-to-Healthcare provider (C2H) electronic 

marketplace,” Communications of the Association 

for Information Systems, Vol. 18, pp. 413-430, 2006. 

[4] M. L. A. Lustria, “Can interactivity make a 

difference? effects of interactivity on the 

comprehension of and attitudes toward online health 

content,” Journal of the American Society for 

Information Science and Technology, Vol. 58, No. 6, 

pp. 766-776, 2007. 

[5] A. M. Aladwani, and P. C. Palvia, “Developing and 

validating an instrument for measuring userperceived 

web quality,” Information and 

Management, Vol. 39, pp. 467-476, 2002. 

[6] Y. Hwang, and D. J. Kim, “Customer self-service 

systems: the effects of perceived web quality with 

service contents on enjoyment, anxiety, and e-trust,” 

Decision Support Systems, Vol. 43, pp. 746-760, 

2007. 

[7] C. Liu, and K. P. Arnett, “Exploring the factors 

associated with web site success in the context of 

electronic commerce,” Information and 

Management, Vol. 38, pp. 23-33, 2000. 

[8] W. H. DeLone, and E. R. McLean, “Measuring ecommerce 

success: applying the DeLone and 

McLean Information Systems Success Model,” 

International Journal of Electronic Commerce, Vol. 

9, No. 1, pp. 31-47, 2004. 

[9] T. Ahn, S. Ryu, and I. Han, “The impact of web 

quality and playfulness on user acceptance of online 

retailing,” Information and Management, Vol. 44, 

pp. 263-275, 2007.


[10] S. J. McMillan, and J. Hwang, “Measures of 

perceived interactivity: an exploration of the role of 

direction of communication, user control, and time 

in shaping perceptions of interactivity,” Journal of 

Advertising, Vol. 31, pp. 29-42, 2002. 

[11] S. J. McMillan, “Interactivity is in the eye of the 

beholder: function, perception, involvement, and 

attitude toward the web site,” in Proceedings of the 

2000 Conference of the American Acadenn/ of 

Advertising. Mary A. Shaver, ed. East Lansing, MI: 

Michigan State University, pp. 71-78, 2000. 

[12] S. J. McMillan, “A four-part model of cyberinteractivity: 

some cyber-spaces are more interactive 

than others,” New Media and Society, Vol. 4 (June), 

pp. 271-291, 2002. 

[13] G. Wu, “Conceptualizing and measuring the 

perceived interactivity of websites,” Journal of 

Current Issues and Research in Advertising, Vol. 28, 

No. 1, pp. 87-104, 2006. 

[14] H. K. Ko, M. S. Roberts, and C. Cho, “Crosscultural 

differences in motivations and perceived 

interactivity: a comparative study of American and 

Korean internet users,” Journal of Current Issues 

and Research in Advertising, Vol. 28, No. 2, pp. 93- 

104, 2006. 

[15] J. H. Song, and G. M. Zinkhan, “Determinants of 

perceived web site interactivity,” Journal of 

Marketing, Vol. 72, pp. 99-113, 2008. 

[16] W. H. DeLone, and E. R. McLean, “The DeLone 

and McLean Model of Information Systems Success: 

a ten-year update,” Journal of Management 

Information Systems, Vol. 19, No. 4, pp. 9-30, 2003. 

[17] J. Jee, and W. N. Lee, “Antecedents and 

consequences of perceived interactivity: an 

exploratory study,” Journal of Interactive 

Advertising, Vol. 3, No. 1, [available at http:// 

www.jiad.org], 2002. 

[18] S. J. McMillan, J. Hwang, and G. Lee, “Effects of 

structural and perceptual factors on attitudes toward 

the website,” Journal of Advertising Research, Vol. 

43, pp. 400-409, 2003. 

[19] A. S. Dick, and K. Basu, “Customer loyalty: toward 

an integrated conceptual framework,” Journal of the 

Academy of Marketing Science, Vol. 22, No. 2, pp. 

99-113, 1994. 

[20] Y. F. Kuo, ”A study on service quality of virtual 

community websites,” Total Quality Management 

and Business Excellence, Vol. 14, No. 4, pp. 461- 

473, 2003. 

[21] S. Otim, and V. Grover, “An empirical study on 

web-based services and customer loyalty, European 

Journal of Information Systems, Vol. 15, pp. 527- 

541, 2006. 

[22] N. M. Kassim, and N. A. Abdullah, “Customer 

loyalty in e-commerce settings: an empirical study,” 

Electronic Markets, Vol. 18, No. 3, pp. 275-290, 

2008. 


[23] J. C. Anderson, and D. W. Gerbing, “Structural 

Equation Modeling in practice: a review and 

recommended two-step approach,” Psychological 

Bulletin, Vol. 103, No. 3, pp. 411-423, 1988. 

[24] J. F. Hair, R. E. Anderson, R. L. Tatham, and W. C. 

Black, Multivariate Data Analysis, NJ: Prentice Hall, 

1998. 

[25] J. C. Nunnally, Psychometric Theory, 2 nd Ed., New 

York, NY: McGraw-Hill, 1978. 

[26] C. Fornell, and D. F. Larcker, “Evaluating Structural 

Equation Models with unobservable variables and 

measurement error,” Journal of Marketing Research, 

Vol. 18, pp. 39-50, 1981. 

[27] S. Ha, and L. Stoel, “Consumer e-shopping 

acceptance: antecedents in a technology acceptance 

model”, Journal of Business Research, Vol. 62, No. 

5, pp. 565-571, 2009. 

[28] A. Basso, D. Goldberg, S. Greenspan, and D. 

Weimer, “First impressions: emotional and cognitive 

factors underlying judgments of trust e-commerce,” 

In Proceedings of the 3rd ACM Conference on 

Electronic Commerce, 137-143, Tampa, FL, USA, 

2001. 

[29] G. Wu, “The mediating role of perceived 

interactivity in the effect of actual interactivity on 

attitude toward the website,” Journal of Interactive 

Advertising, Vol. 5, No. 2, [available at 

http://www.jiad.org], 2005. 

Chung-Hung Tsai is assistant professor and director of 

Department of Health Administration at Tzu Chi College of 

Technology. He received his Ph. D. degree from National 

Dong-Hwa University. He is currently one member of the 

editorial board and reviewer of Journal of Healthcare 

Management. 

His current research areas are knowledge management system, 

health information system, e-commerce, and 

telemedicine/telecare/telehealth system management. His 

academic papers have been published in Technological 

Forecasting and Social Change (SSCI), International Journal of 

Information Technology and Management (EI), Key 

Engineering Material (EI), International Journal for Quality 

Research (SCIndeks), Journal of e-Business (TSSCI), Journal 

of Technology Management, MIS Review, Journal of 

American Academy of Business (ABI), Electronic Commerce 

Studies, Journal of Business Administration, Journal Customer 

Satisfaction, and Journal of Health Management.


A New Method of Time-frequency Synthesis of 

Harmonic Signal Extraction from Chaotic 

Background 

Erfu Wang 

Key Laboratory of Electronics Engineering, College of Heilongjiang Province 

School of Electronic Engineering, Heilongjiang University, Harbin, China 

Email: efwang_612@163.com 

Zhifang Wang, Jing Ma and Qun Ding 

Key Laboratory of Electronics Engineering, College of Heilongjiang Province 

School of Electronic Engineering, Heilongjiang University, Harbin, China 

Email: {zhifang.w@gmail.com, majing20041499@126.com, qunding@yahoo.cn} 

Abstract—The separation of chaos and signal is an 

important problem of chaos signal processing. In recent 

years, the time-frequency analysis method is more and more 

mature. It can extract the time-domain character and 

frequency-domain character at the meantime. Timefrequency 

method can mainly carry out the problem of 

extraction from continuous chaos system background; 

achieve separation between chaos and signal according to 

different time-frequency character of chaos signal, noise 

signal and harmonic signal. So it can get useful signal from 

chaotic background. This paper first introduced the basic 

theory of time-frequency methods. Use the wavelet method 

and empirical mode decomposition method to analyze the 

extraction performance of harmonic signal from chaos 

background according to the different noise situation. After 

compare the wavelet method and empirical mode 

decomposition method, we summarize a new 

complementary synthesis method of harmonic signal 

extraction combine the wavelet threshold and empirical 

mode decomposition according to the experiments and 

simulation. Computer simulation verified that the methods 

have high availability. 

Index Terms—Harmonic Signal, Time-frequency Analysis, 

Extraction, Chaos, wavelet 


Chaos is widespread in the various domains, such as 

chaos secure communication, ECM and heart computer 

signal processing[1,2]. Chaos theory attracts a lot of 

attention by the scholars in the last decade. 

Many researchers introduced several methods of 

detecting, separating and extracting signals according to 

the chaos’ different characters. Leung uses smallest phase 

space capacity method to estimate polynomial parameter 

insert chaos[3]. Native Fuping Wang and some other ones 

use chaos attractor geometric properties, realize the 

separation between chaos interference and weak signal by 

the concept of differential manifold tangent space[4,5]. 

Haykin research the extraction of small objectives signal 


doi:10.4304/jnw.6.7.1025-1032 

from the ocean noise which is proved as chaos noise by 

the way of artificial nerve network[6~9]. Short research 

the extraction from chaos communication system in the 

way of chaos forecast method according to the 

characteristic of short-time forecast. These methods 

carved out a new field of chaos signal processing, but 

lack of systematicness. Some methods are rigorous and 

weak applicability, and demand target signal smaller than 

chaos background signal[10~13]. Time-frequency 

analysis theory is more popular in recent years. Timefrequency 

method can mainly carry out the problem of 

extraction from continuity chaos system background, 

achieve separation between chaos and signal according to 

different time-frequency character of chaos signal, noise 

signal and harmonic signal[14~17]. So it can get useful 

signal from chaos background. When the signal-to-noise 

ratio (SNR) is not so weak the extraction effect will be 

perfect. 

Comparing wavelet method with empirical mode 

decomposition(EMD) method, according to the 

performance analysis of harmonic signal extraction from 

chaos background in different noise level and signal-tonoise 

ratio(SNR), finally give the extraction produce 

about complementary scheme. After theory analysis and a 

lot of simulation, the paper will give the corresponding 

results and analysis. 

II. BASIC THEORY OF TIME-FREQUENCY METHODS 

Time-frequency methods are good at deal with nonstationary 

signal. It can extract the time-domain character 

and frequency-domain character at the meantime. Which 

is the most classical wavelet transform method, EMD as a 

new method of time-frequency signal processing in some 

application can get better than the wavelet transform. 

A. Chaotic System 

We choose the chaotic system Lorenz system to 

simulate and experiment. Lorenz system is a three-


(a) Wavelet method 

(b) EMD method 

Figure 1. The recovery harmonic signal by wavelet method and EMD 

method when NL=10% 

dimensional continuous dynamic system, its nonlinear 

state equations are defined as followed: 

� dx 

� 

� �a( 

x � y) 

dt 

� 

� dy 

� � �xz 

� bx � y 

(1) 

� dt 

� dz 

� xy � cz 

�� 

dt 

In the following part of the simulations and 

experimentations we choose a=10, b=28 and c=8/3 as the 

parameters. We get x0=y0=z0=0.1 and the step length is 

0.01.We iterate 4000 points and use the 1900 to 4000 

from x axis to be the chaotic background sequence. 

B. Wavelet Transform 

Wavelet transform started to develop as a timefrequency 

Analysis method from the anaphase 20th 

century. To the present signal x(t) ∈ L2(R), the 

continuous wavelet transform(CWT) of signal x(t) is 

defined as: 


� 

� � 

� � � 

� 1 

� t b 

WT x ( a, 

b) 

x( 

t) 

� [ ] dt 

�� 

a 

a 

(2) 

x( 

t), 

� a, 

b ( t) 

In the Eq.2: a>0 is the scale factor and b is the shifted 

factor. 1 t � b 

� a, 

b( 

t) 

� � [ ] The equation is called wavelet 

a a 

primary function is the shift and scale dilation of the 

generating wavelet. Wavelet transform is a kind of 

correlated calculation between the originality signals and 

the group of wavelet functions after dilation essentially. 

The analysis process based on wavelet transform is the 

decomposition and reconfiguration process virtually. The 

primary wavelet Daubechies is continuous, orthogonal 

and easy to implement. So the wavelet analysis part use 

the Daubechies. The db6 is chosen due to its good 

localized character and orthogonal character. 

C. Wavelet threshold de-noising theory 

Because of wavelet transform is linear, wavelet 

transformation coefficient is additive. When elect wavelet 

matching with signal to conduct wavelet 

transformation ,signal energy mainly focus on wavelet 

coefficient of a few sparse and amplitude relatively large 

on some frequency band, and wavelet transform of white 

noise is still white noise, which is widely distributed in 

each dimension of time axis and amplitude is not big. So 

we can set a threshold, using the threshold to adjust 

wavelet coefficients according to certain rules. After 

adjustment, various wavelet coefficients reconstruct 

signal according to the inversion algorithm for getting 

target signal. This is the wavelet threshold de-noising 

theory basis. This paper choose heursure threshold. 

D. EMD Method 

EMD (Empirical Mode Decomposition) is a 

trenchancy instrument to analyze non-linear and nonstationary 

signal and is instituted by N. E. Huang etc 

originally. EMD method is based on the concept of 

instantaneous frequency on the basis of in-depth study, 

and the corresponding Hilbert transformation is close 

related to the method. Decompose non-linear and nonstationary 

signals can obtain a series IMF (Intrinsic Mode 

Function) which express the signal character and time 

scale. IMF is narrowband stationary signal. IMF must 

satisfied the following two conditions: 

(1)The discrepancy of zero-crossing and the extreme 

point is zero or one; 

(2)At every point, the mean of local maximum 

envelope and local minimum envelope is zero; 

Only decompose the signal into some IMF, analyze 

every instantaneous frequency can reveal true physics 

sense of original signal. 

III.UNDER DIFFERENT CIRCUMSTANCES OF TIME- 

FREQUENCY METHOD 

A. The Influence of noise level(NL) on the Extraction 

Effect 

The Lorenz system, Gauss noise and s(n)=Asin(2πfn) 

mixed the composite signal, and fix A=5,f=5Hz. The






TABLE I. EXTRACTION EFFECT COMPARISON IN DIFFERENT NL 

NL 10% 30% 50% 70% 100% 

R-wavelet 0.9608 0.8370 0.6850 0.5271 0.4273 

R-EMD 0.9451 0.8285 0.8404 0.6237 0.5545 

noise level (NL) means that the standard ratio of Gauss 

noise to chaotic interference. Observe wavelet transform 

method and EMD method on harmonic signal extraction 

effect after changing NL. 

Two methods of extraction effect when NL=10%, 

NL=30% and NL=70%. 

From Fig. 1, Fig. 2and Fig. 3 we can see that two 

methods are perfect when there is definite noise in the 

chaotic background and NL is small. The waveform 

almost distorted by wavelet method when the noise is 

much bigger, then this method can not extract harmonic 

signal. EMD method is still useful and the effect is 

perfect. The correlation coefficient is 0.5271 when use 

wavelet method, and the correlation coefficient is 0.7286 

when use EMD method. From Fig. 3 we can know that 

when the NL=70%, the recovery level has a large 

improvement and the extraction performance is more 

perfect. 






When NL are 10%,30%,50%,70%,100%, signal is 

extracted from chaotic background with noise by use the 

wavelet method and EMD method. TABLE I gives 

quantitative comparison directly. R-wavelet means the 

correlation coefficient between recovery harmonic signal 

and original harmonic signal when use wavelet method. 

R-EMD means the correlation coefficient between 

recovery harmonic signal and original harmonic signal 

when use EMD method. 

We can get a conclude from TABLE I by quantitative 

results compared of wavelet method and EMD 

method .When NL50%, EMD method is better than 

wavelet method. The extraction effect is perfect. 

B. The Influence of SNR on the Extraction effect 

The Lorenz system, Gauss noise and s(n)=A*sin(2πfn) 

mixed the composite signal, and fix A=5,f=5Hz. 

According to the influence of NL on the extraction 

performance, and fix NL=30%. SNR means the energy 

ratio of harmonic signal to chaotic interference and Gauss 

noise. Compare the extraction performance of wavelet 

method and EMD method after changing Harmonic





method when SNR=-5 

TABLE II. EXTRACTION EFFECT COMPARISON IN DIFFERENT SNR 

SNR -1 -5 -10 -20 -30 

R-wavelet 0. 8949 0.7945 0.5907 0.2381 0.0271 

R-EMD 0.8582 0.8398 0.7438 0.2589 0.1789 

Signal Amplitude A and SNR. 

Two methods of extraction effect when SNR=-5 and 

SNR=-20. 

From Fig. 4 and Fig. 5 we can see that two methods are 

perfect when there is a little noise in the chaotic 

background. When SNR=-5 the wavelet method is more 

perfect, the waveform almost distorted by wavelet 

method when the SNR is much lower (SNR=-20), then 

this method can not extract harmonic signal. EMD 

method is more perfect. 

When SNR are -1,-5,-10,-20,-30, we use of wavelet 

method and EMD method for extraction method of 

harmonic signal from chaotic background with noise. 

TABLE II directly gives the quantitative of correlation 

coefficient comparison. 

We can get conclude from TABLE II by quantitative 

results compared of wavelet method and EMD method. 





method when SNR=-20 

When NL=30% in noise level and SNR>-5, wavelet 

method is better than EMD method. This wavelet 

decomposition stability plays a leading role. When 

SNR


(a)NL=10% 

(b)NL=30% 

Figure 6. The extraction performance of Wavelet threshold and EMD method 

TABLE III. DE-NOISING PERFORMANCE OF WT(WAVELET 

TRANSFORM) AND EMD 

SNR Wavelet EMD 

Parameter set first give wavelet primary need not 

function and series 

first set 

Convergence 

rate 

slow 

fast 

Stability stable unstable 

SNR higher lower 

can infer that WT and EMD are complementary in denoising 

performance from the simulation result 

mentioned above. 

Due to complementary characteristics of the Wavelet 

method and EMD in all aspects, it was switched between 

wavelet method and EMD method for choosing better 

extracted performances of harmonic signal in the chaos of 

background according to different noise level and SNR 

condition. 

IV. A NEW METHOD OF HARMONIC SIGNAL EXTRACTED 

W e w i l l d iscuss a comp r e h e n s i v e method 

complementary advantages which combines wavelet 

transform, threshold de-noising and EMD method 

according to complementary characteristics of the 


(c)NL=70%. 

(d)NL=100%. 

TABLE IV. THE CORRELATION COEFFICIENT OF MIXED 

METHOD AFTER EVERY STEP SEPARATION IN THE DIFFERENT NL 

NL 10% 30% 50% 70% 100% 

r-xbyz 0.9809 0.8500 0.6764 0.5665 0.4385 

R-EMD 0.9772 0.7554 0.8241 0.7614 0.5692 

Wavelet method and EMD in the noise level and SNR 

aspect and two methods also can be used in de-noising 

speed and stability. 

A. Simulation experiment and analysis 

Experimental one: The Lorenz system, Gauss noise 

and s(n)=A*sin(2πfn) mixed the composite signal, and fix 

A=5,f=5Hz. Using wavelet method for mixed signal, then 

using EMD method , and observe the separation 

performance. 

We can give the separation performance when 

NL=10%, NL=30%, NL=70% and NL=100%. “original 

signal” is harmonic signal .“recovery-xbyz” means that 

extraction of signal after the wavelet threshold de-noising. 

“recovery-EMD” is recovery harmonic signal after the 

EMD method. 

TABLE IV is that the correlation coefficient of mixed 

method after every step separation in the different NL.


(a)NL=10% 

(c)NL=70%. 

(b)NL=30% 

(d)NL=100% 

Figure 7. The extraction effect of Wavelet threshold and EMD method 

TABLE V. THE CORRELATION COEFFICIENT OF MIXED 

METHOD AFTER EVERY STEP SEPARATION IN THE DIFFERENT NL. 

NL 10% 30% 50% 70% 100% 

r-xbyz 0.9627 0.8722 0.8091 0.6856 0.5684 

R-EMD 0.9579 0.8555 0.7868 0.6682 0.5435 

From Fig. 6, we known that we can use EMD method 

after reconstruction of wavelet threshold de-noising. 

When noise level(NL


Time-frequency complementary advantages scheme 

based EMD and Wavelet. 

Design of Time-frequency complementary advantages 

scheme based EMD and Wavelet. 

(1)Estimate noise level and SNR level for original 

mixed signals ,choose appropriate parameter and series in 

noise . 

(2)Do EMD de-noising for mixed signals to obtain all 

IMF components ,extract harmonic signal components 

first step ,observe correlation coefficient of signal 

recovered this moment. Using EMD could work in lower 

SNR and considering advantage of quicker speed 

convergence . 

(3)Do the second de-noising using wavelet transform 

perfect stability and high-precision, then reduce noise by 

heursure again. 

(4)Make the signal that is removed noise twice as 

harmonic signal extracted. 


This paper first introduce the basic theory of timefrequency 

methods. Comparing wavelet method with 

empirical mode decomposition(EMD) method, according 

to the performance analysis of harmonic signal extraction 

from chaos background in different noise level and 

signal-to-noise ratio(SNR). We summarize a new 

synthesis about wavelet threshold and empirical mode 

decomposition(EMD) complementary of new harmonic 

signal extraction by experimental simulation. Computer 

simulation verified that the methods are high availability. 

Finally give the extraction procedure about 

complementary scheme. 


This work is supported by the National Science 

Foundation of China (no.60672011), Open Fund of Key 

Laboratory of Electronics Engineering College of 

Heilongjiang University (No. D22D20100027), the 

technology research project of Education Department of 

Heilongjiang Province (No. 11511381) and Dr. Start 

funds of Heilongjiang University. 

REFERENCES 

[1] Richer M, Schreiber T and Kaplan D T, “Fetal EGG 

extraction with nonlinear state-space projection,” IEEE 

Trans. Biom. Eng. vol. 45(1), pp. 133~137, 1998. 

[2] Leung Henry and Huang Xing-ping, “Sinusoidal frequency 

estimation in chaotic noise”, ICASSP, vol. 2, 

pp.1344~1347, 1995. 

[3] Wang Fuping, Guo Jingbo, Wang Zanji and Xiao Dachuan , 

“Harmonic Signal Extraction from Strong Chaotic 

interference,” Acta Physica Sinica. vol. 50(6), pp. 

1019~1023, 2001. 

[4] Haykin S and Li X B, “Detection of signal in chaos,” 

Proceeding of IEEE, vol. 83(1), pp. 94~122, 1995. 

[5] Short K M and Parker A T, “ Unmasking a hyperchaoti 

communication scheme,” Physical Review E, vol. 58, pp. 

1159~1162, 1998. 


[6] Donald B.Percival and Andrew T.Walden, “Wavelet 

Methods for Time Series Analysis,” Cambridge University 

Press,2000. 

[7] Huang N E,Shen Z and Long S R, “The empirical mode 

decomposition and the Hilbert spectrum for nonlinear and 

non-station time series analysis,”Proceeding of the Royal 

Society of London A, vol. 454, pp. 903~995, 1998. 

[8] Newland D E, “Wavelet analysis of 

vibration,”part1,2.Journal of Vibrationand Acoustics, vol. 

116, pp. 409~425, 1994. 

[9] Sun yankui, “Wavelet analysis and its application,”China 

Machine Press. pp. 219~243, Oct. 2005. 

[10] Li Hong Guang and Meng Guang, “Harmoic signal 

extraction from chaotic interference based on empirical 

mode decomposition”, vol.53, July, 2004. 

[11] Wang guoguang and Wang shuxun, “Research on Methods 

of Extracting signals from Chaos,” Jilin University, 2007. 

[12] Zhang Defeng, “Matlab and Wavelet analysis,” China 

Machine Press, Jan. pp.65-92, 2010. 

[13] Ying Tan, Jun Wang and J. M. Zurada. Nonlinear Blind 

Source Separation Using a Radial Basis Function Network. 

IEEE Transactions on Neural Networks. vol. 12, pp. 

124~134, 2001. 

[14] S. A. chard, D. T. Pham and C. Jutten. Criteria Based on 

Mutual Information Minimization for Blind Source 

Separation in Post-nonlinear Mmixtures. Elsevier signal 

processing. vol. 85, pp.965~974, 2005. 

[15] W. L. Woo, S. S. Dlay. Nonlinear Blind Source Separation 

Using a Mixture RBF-FMLP Network. IEE Proceedings, 

Vision, Image and Signal Processing. vol. 152, pp.173~183, 

2005. 

[16] A. Ziehe, M. Kawanabe and S. Harmeling. Separation of 

Post-nonlinear Mixtures Using ACE and Temporal 

Decorrelation. Proceeding of Independent Component 

Analysis and Blind Signal Separation (ICA’2001). 

pp.433~438, 2001. 

[17] Kun Zhang, Laiwan Chan. Extended Gaussianization 

Method for Blind Separation of Post-nonlinear Mixtures. 

Neural Computation. pp.425~452, 2005 

Erfu Wang Heilongjiang province, 

China, 1980. Received PhD degree and 

M.S. degree in Harbin Institute of 

Technology, China, in 2009 and 2005 

respectively, and B.S. degree in Jilin 

University, China. The interest fields are 

blind source separation, array signal 

processing, wideband wireless 

communication,etc. 

Zhifang Wang Henan province, China, 

1979. Received PhD degree and M.S. 

degree in Harbin Institute of Technology, 

China, in 2009 and 2005 respectively, and 

B.S. degree in Henan University, China. 

The interest fields are biometric, 

development of media systems, image 

analysis, cluster analysis,etc.



Jing Ma Heilongjiang province, China, 

1985. Received B.S. degree in 

Heilongjiang University, China, in 2008. 

The interest fields are chaotic theory, 

encryption system, signal separation ,etc. 

Qun Ding Heilongjiang province, 

China, 1957. Received PhD degree 

and M.S. degree in Harbin Institute of 

Technology, China, in 2007 and 1997 

respectively. The interest fields are 

secure communication, information 

security, Pattern recognition, etc. 

Currently, she is a professor of 

Heilong Jiang University, China.


Provable Data Possession of 

Resource-constrained Mobile Devices in Cloud 

Computing 

Jian Yang 1,2 

1College of Electronic and Information Engineering, Tongii University, Shanghai 201804, China 

2College of Mathematics and Computer Science, Dali University, Dali, Yunnan 671003, China 

Email: sbjc1215@126.com 

Haihang Wang 1 , Jian Wang 1,3 , Chengxiang Tan 1 and Dingguo Yu 1 

1College of Electronic and Information Engineering, Tongii University, Shanghai 201804, China 

3College of Electronics & Information Engineering, Henan University of Science & Technology, Luoyang, 

China 

Email: wanghh@sh163.net, wangjian_migi@sina.com, cxtan@trimps.ac.cn, zjydg@163.com 

Abstract—Benefited from cloud storage services, users can 

save their cost of buying expensive storage and application 

servers, as well as deploying and maintaining applications. 

Meanwhile they lost the physical control of their data. So 

effective methods are needed to verify the correctness of the 

data stored at cloud servers, which are the research issues 

the Provable Data Possession (PDP) faced. The most 

important features in PDP are: 1) supporting for public, 

unlimited numbers of times of verification; 2) supporting 

for dynamic data update; 3) efficiency of storage space and 

computing. In mobile cloud computing, mobile end-users 

also need the PDP service. However, the computing 

workloads and storage burden of client in existing PDP 

schemes are too heavy to be directly used by the 

resource-constrained mobile devices. To solve this problem, 

with the integration of the trusted computing technology, 

this paper proposes a novel public PDP scheme, in which the 

trusted third-party agent (TPA) takes over most of the 

calculations from the mobile end-users. By using bilinear 

signature and Merkle hash tree (MHT), the scheme 

aggregates the verification tokens of the data file into one 

small signature to reduce communication and storage 

burden. MHT is also helpful to support dynamic data 

update. In our framework, the mobile terminal devices only 

need to generate some secret keys and random numbers 

with the help of trusted platform model (TPM) chips, and 

the needed computing workload and storage space is fit for 

mobile devices. Our scheme realizes provable secure storage 

service for resource-constrained mobile devices in mobile 

cloud computing. 

Index Terms—bilinear signature, merkle hash tree, provable 

data possession, mobile computing, cloud computing, 

trusted computing 


doi:10.4304/jnw.6.7.1033-1040 


In cloud computing, multi-tenants share the external 

resources of computing and storage, which allows 

enterprises and individuals get on-demand computing or 

storage services from cloud service providers(CSP), such 

as Amazon’s S3 and Google’s App Engine, and no 

longer maintain their local physical machines. However, 

In this new model, the users put their data on the cloud 

storage servers maintained by service providers, which 

deprives the users of their control of the physical 

possession of data, even though they are the owners of 

the data. In this case, some new security needs and 

problems have arisen. At the same time, when one’s data 

are outsourced, he wants to know whether the data is 

truly stored at the correct servers and be intact as stated in 

the Service Level Agreement (SLA). In addition, in 

multi-layer cloud services framework, higher layer cloud 

applications need lower layer services (such as storage 

service or virtual machine image service).In this 

circumstance, higher layer CSP needs effective methods 

to verify the basic storage services provided by lower 

CSP. It is the problem that the important direction of the 

current research field in cloud computing - provable data 

possession (PDP) – wants to solve. On the basis of PDP, 

if servers ensure the clients can retrieve correct data files 

with the help of erasure codes, such as Reed-Solomon 

codes, the services are named as Proofs of Retrievability 

(POR). 

The first solutions to this issue are proposed by 

Deswarte et al. [1] and Filho et al.[2], which both use 

RSA-based functions to hash the whole data file for every 

verification challenge. Obviously, both of them are


inefficient for big data files, which need more time to 

compute and transfer their hash values. Ateniese et al. [3] 

proposed a formal definition and related operations of the 

PDP model. They use some homomorphic tokens in the 

encoded file to help verify whether the file was tampered 

without any legal authentication. Juels et al. [4] formally 

put forward the protocol structure and security framework 

of the POR model and proposed a method to detect 

unauthorized changes of data by adding some “sentinels” 

randomly in the original files. But the above schemes 

only have limited times of verification operations and do 

not support public verification. On the basis of the 

security model in [4], Shacham et al. [5] designed an 

improved scheme to realize public data possession 

verification by using bilinear signature. But in their 

framework, the number of the authentication tokens 

stored on the server is proportional to the number of data 

blocks, and it only is fit for static file storage. Similar to 

[5], Wang et al. [6] uses Merkle hash tree (MHT) to build 

verification tags stored at the servers and support 

dynamic data update. They treat the leaf nodes of MHT 

as the left-to-right sequence so that the locations of error 

data can be detected. 

From another perspective, with the prevalence of the 

3G and 4G wireless communication networks, the mobile 

devices, such as mobile phones, PDA, also want to share 

the benefits introduced by the cloud on-demand storage 

and computing service. But the traditional mobile 

terminals are resource-constrained devices (with low 

CPU frequency and small memory) and can not use the 

existing PDP schemes directly, which require clients 

encode files with erasure codes, divide encoded files into 

blocks and sign on every data blocks. Those operations 

on large files are intolerable in our mobile computing 

clients. How to solve this problem is the main goal of our 

work. 

A. Our Contribution 

Combined with bilinear map and the well-studied 

authentication structure-MHT, this paper addresses to 

design a storage service model with public provable data 

possession in mobile computing environment. To achieve 

public verifiability in this environment, we need a trusted 

third-party auditor (TPA) to do most of the calculation 

works done by the clients in other PDP schemes. 

Specifically, our contribution in this paper can be 

summarized as the following two aspects: 

� We use trusted computing technology for the 

mutual authentication between the end-users and 

the TPA so that the most computations of the 

end-users can be done by TPA. The end users 

only need to generate some passwords and a 

small amount of random numbers, which can be 

done by the TPM chips embedded in their 

machines. 

� We improve the existing PDP schemes 

(especially the work proposed in [6]) with 

bilinear mapping signature to make them fit for 

the mobile computing environment. To the best 

of our knowledge, our scheme is the first storage 

framework in mobile cloud computing to support 


the stateless verification, public provable data 

possession, dynamic data update, and is provably 

secure in random oracle model. 

B. Paper Structure 

The rest of the paper is organized as follows. Section II 

introduces the related works. Then we provide the 

description of our scheme in Section III, including model 

structure, notation and preliminaries, important functions 

and detailed implement of our scheme. Section IV gives 

the security analysis and performance evaluation, 

followed by Section V which gives the concluding 

remark of the whole paper and overviews the related 

work finally. 

II. RELATED WORKS 

Ateniese et al. [3] proposed the first formal definition 

of the PDP model and related operations, functions. In 

their system some homomorphic verifiable tags are used 

to verify data file. Juels et al. [4] proposed a formal 

definition of POR and its security model. After being 

encrypted and divided into small data blocks, which are 

encoded with Reed-Solomon codes, the data file is added 

into some "sentinels” to detect whether it was intact. 

However, the both schemes don not support dynamic data 

update and can only verify limited times because that the 

two schemes only have finite number of the “sentinels” in 

a file. When the finite “sentinels” are exhausted, the file 

must be send back to the owner to re-compute new 

“sentinels”. In their improved work, Ateniese et al. [7] 

proposed a new scheme with homomorphic linear 

authenticators (HLA), of which communication 

complexity is independent of the file length. Though the 

scheme supports infinite times of verification, it can not 

verify publicly. 

Chang et al. [8] proposed a remote identity check 

scheme. By using redactable signature (a kind of 

homomorphic signature scheme proposed in [9]), a 

redactor can calculate a effective signature on a redacted 

message x’ without knowing the private key. This idea 

can be used in a third-party verification scheme and 

verifier does not need to know the private key. Shacham 

et al. [5] proposed two POR schemes: the first uses 

bilinear signature, and is provable secure and efficient in 

the random oracle model. The second depends on 

pseudo-random function and bilinear map, and is 

provable secure in the standard model. Both schemes rely 

on the homomorphic property--aggregating the 

verification proofs into a small value. However, the 

above two methods are not aware of the user’s privacy 

preserving in public audit. 

An effective public PDP scheme should have the two 

important following properties [10]: 

� To allow a TPA to verify the correctness and 

integrity of the data without retrieving a copy of 

the whole data or introducing additional on-line 

burden to the clients. 

� To avoid introducing new vulnerabilities to the 

privacy of the data.


Wang et al. [6, 10] discussed the privacy protection in 

public audit. In their framework, the third-party audit 

protocol with the privacy preserving is independent of 

data encryption. By using homomorphic authenticator 

and random masking, the scheme conceals the content of 

the original data from TPA and TPA can perform 

multiple auditing tasks in a batch manner. 

Erway et al. [11] is the first to support dynamic data 

update by using rank-based verification skip list in cloud 

servers. The most similar PDP scheme to ours was 

proposed by Wang et al. [6], which is also our work’s 

basis. Their scheme supports public data possession 

verification and dynamic data update at the same time. It 

improves the previous PDP models by using the classical 

MHT and enables TPA to complete privacy-preserved 

data integrity check with the support of dynamic data 

update. This method does not require user's real-time 

participation in the cloud, and avoids the leakage of 

user’s privacy. However, it still requires the user to 

calculate initial verification tokens of the files, which is 

not suitable for this paper’s application environment. 

If the TPA takes over a large number of computing 

works of the mobile end-users, first of all, the TPA 

should achieve mutual authentication with the end-users, 

and establish a secure transmission channel on this basis. 

A feasible method is to take the advantage of the trusted 

computing technology, which is a mature technology, but 

has few successful applications applied in the cloud 

computing and few concrete frameworks designed. EMC 

China lab collaborated with Fudan University, Huazhong 

University of Science and Technology, Tsinghua 

University and Wuhan University to carry out a research 

projects on trusted virtual infrastructure, named Daoli 

[12]. The research project is committed to tenants’ 

isolation and protection platform provider away from 

attacks by malicious tenants in multi-tenant cloud 

computing environments. Combination trusted computing 

with virtualization technologies to enhance the security of 

computing platforms, makes cloud service providers 

being able to provide virtual private cloud (VPC) services 

in public cloud. In [13], the authors discussed how to use 

TCB for security enhancements in Xen—a famous 

open-sourced virtual machine monitor can be used in 

cloud computing—and described how this method is used 

to achieve "trusted virtualization" and enhance the 

security of the virtual TPM. They moved the new VM 

creation function to a small trusted VM out of Dom0. 

This method has two main goals: the one is to reduce and 

bound the size of TCB in Xen-based systems, especially 

remove the user space in Dom0 from the TCB, to 

enhance security. Another one is that if we suppose TCB 

were security, the new VM would have maintained the 

same attributes of the security and integrity with physical 

machine. The privacy manager discussed in [14] uses 

TPM to manage privacy keys required in privacy 

protection. Trusted computing technology to the IaaS 

cloud computing systems was introduced in [15]. In 

Eucalyptus, for example, the authors used a Trusted 

Coordinator (TC) (maintained by an external trusted 

entity) to combine unreliable cloud Manager (CM) with a 


number of trusted nodes in order to form its main 

architecture. This framework ensures the safety of the 

customers VM, allows users to verify the IaaS service 

providers and determine whether the services are security 

before they start VM. 

However, about the combination with trusted 

computing and cloud computing, the schemes mentioned 

above did not enhance the security and integrity of the 

cloud storage services with the provable data possession, 

which is a critical measure to enhance the user's 

confidence in using the cloud storage services, especially 

in wireless mobile computing environment. To the best of 

our knowledge, our scheme is the first to explore the 

application of PDP scheme in mobile computing 

environment combined with the trusted computing. 

III. OUR SCHEME 

A. Model Structure 

System participants: as shown in Figure 1, in our 

resource-constrained public provable data possession 

scheme of the cloud storage service, there are three main 

participants: 1)client, i.e. mobile end-user, which has a 

TPM chip and stores data files in the cloud, and expects 

to get trusted storage validation; 2) trusted third party 

auditor (TPA), which is credible for clients and take the 

main file encryption and authentication tasks required 

during the process of the scheme; 3) cloud storage service 

provider (CSP), which has a large capacity of storage and 

provides the users with storage services and the proof of 

data possession when needed. 

In the mobile computing environment, the client’s 

computing and storage capacity is very limited, but it has 

the ability to use its TPM chips to produce and store 

secret keys. The TPA is acted by a service agent located 

between the mobile access point and the gateway of the 

IP network services. The TPA should have high 

performance of computing but limited storage space, 

which is only for small part of the information of the 

clients and the current session message. In addition, it 

should connect securely with the client for providing 

services. Through Internet, the CSP provides 

high-capacity, redundant storage services. Generally 

Fig.1. System structure model 

speaking, the CSP is an unsafe, even malicious entity. 

That is to say, for some financial benefits, it is possible to


read, tamper or delete the user's original files, even to 

forge the proof of the data possession. So the files should 

be encrypted before sending to the CSP. In this paper, we 

assume all parties communicate through secure, reliable, 

authenticated channels for all phases。 

Security goal: In this system, the client and the TPA 

need to establish a secure link. On the basis of security of 

the underlying communication link between the mobile 

terminal and the TPA, firstly, through the Diffie-Hellman 

key exchange protocol, the scheme generates a symmetric 

key for data exchange. Except for some necessary keys 

and random values, the client does not perform additional 

computation works, which are taken over by the TPA. 

The scheme is secure under the random oracle model and 

our security model is based on the one proposed in [6]. 

Design goal: The scheme has three design goals 

shown as followed. 

� public provable data possession: to allow a 

verifier, not just the data owner or users, to have 

the capability to verify the correctness of the 

stored data on demand; 

� stateless verification: to generate the proofs of 

data possession according to the challenge 

produced randomly by the verifier, not to the 

persistent information maintained by some 

entities; 

� resource-constrained mobile environment support: 

to allow end-user just to generate and store the 

keys by utilizing TPM chip instead of doing lots 

of computing works, which are done by the TPA 

now; 

� trusted computing technology applying: to build a 

trusted channel between the TPA and the client, 

then use it to transfer the related messages and 

authorize the TPA to complete those works for 

reduce the client’s workload. 

B. Notation and Preliminaries 

Diffie-Hellman protocol. Diffie-Hellman key 

exchange protocol is one of the most famous schemes in 

cryptography, which mainly utilize the discrete log 

problem to safely exchange a shared symmetric key 

through an unsecured channel. By using this method, the 

end-user and the TPA can share a symmetric key so that 

the data files, asymmetric keys and the information of the 

verification can be transferred safely after being 

encrypted by this key. 

Merkle hash tree. A Merkle Hash Tree (MHT) is a 

well-studied authentication structure [16], which can 

efficiently and securely prove that a set of data blocks are 

undamaged and unaltered. It is constructed as a binary 

tree where the leaves in the Merkle hash tree are the hash 

values of authentic data. During the process of the 

verification, the verifier only needs check whether the 

root value of the tree is tampered. In the archive [6], their 

scheme treats the leaf nodes as the left-to-right sequence 

for getting the positions of error data blocks. For the sake 

of simplicity, in our scheme, MHT tree leaf nodes only 

store the hash of data blocks. 


Bilinear map. A bilinear map is a map e:G1×G1→G2, 

where G1 is a Gap Diffie-Hellman group and G2 is a 

multiplicative cyclic group with a big prime order p. It 

has the following properties: 

� Computable: there exists an efficiently 

computable algorithm for computing the map; 

� Bilinear: for all h1,h2∈G1 and a,b∈Zp, � 

a b 

ab 

e ( h1 

, h2 

) � e( 

h1, 

h2 

) 

; 

Non-degenerate: 

e( 

g, 

g) 

� 1 

, where g is a 

generator of G1. 

C. Some Important Procedures: 

Improved from [4, 6, 17], our secure storage system 

with provable data possession contains the following 

functions: 

KenGen(1 k ) → (pk,sk) : This algorithm takes as input 

initial secure parameter 1 k , and returns the public key pk 

and the private key sk. In our scheme, the end-user 

generates and maintains the keys in TPM chip. 

Encapek(F) → F’: The TPA uses this algorithm to 

encrypt the raw file F with the seal key ek and encode it 

with erasure codes. It returns the sealed file F’. 

SigGen_Clientsk (F’) → Sigsk (H(R)): This algorithm is 

run by end-users. It takes as input the hash value of the 

root of the MHT and outputs its signature as a metadata. 

SigGen_TPA(F’) → Ф: This algorithm is run by TPA. 

It takes as input each data blocks {mi} of the sealed file 

F’, and outputs the signature collection Ф= { σi } on 

{mi}. 

GenProof(chal, F’, Sigsk(H(R)), Ф) → (P): This 

algorithm is run by the storage server. It takes as input the 

verification challenge message “chal” generated by TPA, 

the stored file F’, the metadata signature and the signature 

set Ф, and returns the possession proof P. 

Verify(P, chal) → {TRUE|FALSE}: By run this 

algorithm, according to the random challenge chal, the 

proof P returned from the server and some metadata of 

the end-user, the TPA verify the correctness of the data 

file, and outputs TRUE if the integrity of the file is 

verified as correct, or FALSE otherwise. 

Decapdk(F’) →F: The end-user request to extract a file 

F, and the TPA retrievals the corresponding sealed file F’ 

from the cloud, decodes and decrypts the file to get F 

with the decryption key (dk), then sends F to the 

end-user. 

D. Detailed Implementation 

For implementing the public provable data possession 

of the cloud storage service in our background, first of all, 

a trusted communication channel should be build between 

the TPA and the mobile end-user, which needs mutual 

remote identification. This issue will be discussed more 

detail later in this paper. After mutual identification, by 

utilizing the Diffie-Hellman key exchange protocol, the 

TPA and the end-user negotiate a symmetric key for the 

exchange of other information in this scheme. 

Now we start to present the main idea of our scheme. 

According to the model defined in [4, 6, 10, 17], we 

assume the raw data file F is first encrypted by the key


(ek) and encoded into F’ using erasure codes, then the F’ 

is divided into N blocks {mi}, where mi∈Zp and 1≤i≤N. 

Let e: G1×G1→G2 be a bilinear map, which has a big 

prime order p and a generator g of group G1. Let H: 

{0,1}*→ G1 is a hash function. The main procedure of 

our scheme is as follows: 

Setup: In this phase, we assume the end-user has 

already completed the remote identification with the TPA 

by using TPM and set up a trusted communication 

channel. Through this channel, the phase setup is 

executed as Algorithm 1 described: 

Algorithm 1 Setup: 

1.C→T:g, g α 

2.T→C:g β 

{ F} 

3.C→T: �� 

g 

{ H ( R) 

} , { dk } �� 

4.T→C: �� 

g 

5.C→T:Sigsk (H(R)) = (H(R)) α 

4.T→S : Sigsk (H(R)), F’={mi}, Ф={σi}, where 1≤i≤N, σi 

=[H(mi)u mi ] β 

C represents for the Client, T for the trusted third-party 

agent and S for the cloud storage server 

The first two steps of algorithm 1 are to complete 

Diffie-Hellman key exchange. After this, the client and 

�� 

the TPA share a symmetric key g . Then the client 

encrypts the original file F with this key and sends it to 

the TPA. 

Received the original file F from the client, a pair of 

asymmetric keys (ek, dk) are generated by invoking 

KenGen(*), where (ek) is for encrypting the file and (dk) 

for decrypting after retrieval. By calling Encapek(F), the 

TPA encrypts the file, divides it into small blocks and 

encodes the data blocks with erasure codes. After doing 

these, the TPA gets N blocks {mi} (1≤i≤N) as the stored 

file F’. Then, the TPA calculates H(R), the hash value of 

the root of the MHT, of which the leaves are the hash of 

the corresponding mi. At last of this turn, the TPA sent 

the value H(R) and dk to the client encrypted with the 

�� 

shared key g . 

By running SigGen_Client sk (F’), the client signs H(R) 

and sent the signature of H(R) back to the TPA. The TPA 

calls SigGen_TPA(F’) to calculate the signature 

collections of each blocks of F’, then sends {Sigsk (H(R)), 

F’, Ф} to the cloud storage servers. In algorithm 1, u is a 

element in G1 chosen randomly by the TPA. 

Integrity verification: This phase starts from the client 

or the TPA by sending a verification challenge to the 

cloud storage service provider (CSP). According to the 

challenge, the CSP computes the proof of verification and 

sends it back to the TPA. After verifying the proof, the 

TPA sends the result to the client. The detail process of 

verification is shown as the algorithm 2: 


g 

Algorithm 2 Verify: 

1.T→S:chal={(i, vi)}, where 1≤i≤c 

2.S→T:proof={μ,ω,(H(mi),Ωi), Sigsk (H(R))} , where 1≤i≤c 

3.T→C:Verify(μ,ω, (H(mi),Ωi) ,Sigsk (H(R)) ,chal)={true , false} 

The challenge message “chal” is generated by TPA. 

The TPA chooses c random numbers in the set [1, N] to 

constitute a sequence subset I. For each i∈I, the TPA 

picks a random element vi∈Zp. The challenge message 

“chal” sent to CSP is composed of the number i and the 

corresponding vi. On receiving the “chal”, the CSP runs 

GenProof(chal, F’, Sigsk (H(R)), Ф) to generate the 

proof, which includes the corresponding hash value H(mi) 

of the data blocks {mi} for every i∈I and the additional 

information Ωi for rebuilding the root H(R) of the MHT. 

In addition, the CSP also computes the following two 

values as a part of the proof: 

� 

c 

� � 

i�1 

c 

� 

i�1 

v m � Z 

i 

i 

p 

and 

vi 

� � � i �G1 

. 

After receiving the proof, the TPA runs Verify(*), and 

sends the result back to the client. The main goal of the 

function Verify(*) is to test whether the following two 

equations is correct: 

� 

e(Sig sk (H(R)), g) ? e(H(R), g ) 

……… (1) 

c 

� 

vi 

� �� 

e( 

�, g ) ? e( 

�( H ( mi 

) u , g ) 

i�1 

……. (2) 

If so, the Verify(*) returns TRUE; otherwise FALSE. 

File retrieval: Before retrieval, the client and the TPA 

should negotiate a symmetric session key Ks through 

Diffie-Hellman key exchange protocol. Then, shown as 

algorithm 3, the client sends the decryption key (dk) 

encrypted by Ks to the TPA and the TPA request the CSP 

for extracting the file F’. The CSP sends F’ to the TPA. 

Then the TPA runs Decapdk(F’) to get the raw file F and 

send F to the end-user through a secure communication 

channel. 

Algorithm 3 Retrieval: 

1.C→T:Request(F), {dk}Ks 

2.T→S:Request(F’) 

3.S→T:F’ 

4. T→C:F=Decapdk(F’) 

Noted that, we do not include the process of 

negotiation of the symmetric session key Ks in the above 

description of algorithm 3. 

E. Discussions 

Remote identification: As stated above, the client and 

the TPA use a symmetric key generated through the 

Diffie-Hellman protocol to encrypt the original file, the 

secret information, the hash value of the root of the


Merkle hash tree and some control information 

transferred between them. But before that, both of them 

need remote mutual identification. On one hand, based on 

trusted computing idea, we assume that the TPA is 

trusted in mobile computing environment. That is to say, 

the client does not need to identify the TPA. On the other 

hand, the client is attested by answering the challenge of 

TPA with a message signed by the Attestation Identity 

Key (AIK), which is created and maintained by the TPM 

chip embedded in the client’s device. There are lots of 

schemes to solve this problem. Because this paper 

focuses on the mobile computing, we can adopt the 

scheme proposed by Yang et al. [18] to realize the remote 

identification in wireless environment. 

Verification equation: During the verification phase, if 

the information stored at the cloud has not been tampered, 

replaced or deleted, according to the properties of bilinear 

map, the correctness of Eq. (1) can be elaborated as 

follows: 

e(Sig sk (H(R)), g) 

� 

e((H(R)) 

� 

� e(H(R), g 

� 

, g) 

And the one of equation (2) can be proved as follows: 

c 

� 

i�1 

c 

� 

i�1 

c 

� 

i�1 

� 

v 

e( 

� , g ) � e( 

� , g 

� e( 

� e( 

� e( 

� e( 

c 

� 

i�1 

c 

� 

i�1 

i i 

� 

) 

m � v 

i 

i � 

�( H ( m ) u ) � , g ) 

i 

v � 

i mivi 

� 

�H ( m ) � u � , g ) 

( H ( m ) 

( H ( m ) 

i 

i 

i 

v 

v 

i 

i 

c 

) 

� 

i�1 

) �u 

m v 

� 

) u , g 

i i 

�� 

, g 

Data confidentiality: according the proof in [6], that 

the equation 1 and 2 are satisfied can ensure the data 

stored at the cloud storage servers are correct and intact 

under the random oracle model. In additional, for the 

privacy and the confidentiality, the TPA encrypts the raw 

file with the key (ek) and decrypt with the key (dk) when 

it sends and retrieves the file respectively. The 

asymmetric key pair (ek, dk) is created by the TPA. 

Because the TPA maybe need provider the verification 

service for multi-user in a real circumstance, according to 

the TPM main specification v1.2 [19], in our scheme 

there are two ways to maintain the key pair. One way is 

to use TPM_Bind to bind the key pair and store it on the 

TPA. Another solution is, as shown in algorithm 1, that 

the decryption key (dk), which encrypted by the 

symmetric key shared between the client and the TPA, 

can be sent to the client and sent back to the TPA during 

extracting the file. 

Supporting dynamic data update: the scheme in [6] 

has already supported dynamic data update with the 

operation on the Merkle hash tree, so our scheme also has 


) 

�� 

) 

this property. Because in our mobile computing 

environment, a lot of calculations on mobile device are 

not meaningful, we do not intend to discuss this issue in 

detail. 

IV. SECURITY AND PERFORMANCE ANALYSIS 

A. Security Analysis: 

As mentioned above, an important basis of our scheme 

is building a secure information transferring channel 

between the client and the TPA. The TPA, which we 

assume is trusted, can authenticate the client based on 

hardware TPM chip embedded in the device of the client. 

By using Diffie-Hellman key exchange, the two entities 

can share a symmetric key to ensure the data files and 

signatures transferred between them are more secure. 

Furthermore, the trusted computing technology can also 

help them to avoid Man-in-the-middle attack during the 

process of the Diffie-Hellman key exchange. 

The file transferred between the TPA and the CSP is 

encrypted and encoded file F’, which avoid the CPS 

knowing the content of file and ensure the privacy and 

confidentiality of the raw file. The signature structure of 

MHT is the basis of keeping the integrity of data. That 

the root signature of the MHT is computed by the 

end-user can avoid the leakage of the user’s private key. 

As for the security analysis of the bilinear map in PDP 

schemes, archives [5-6, 10] have exhaustive description, 

so we will not discuss it in detail. 

B. Performance Analysis: 

According algorithm 1 and 2, we can demonstrate the 

overall workload of the computing and storage of each 

parties in our scheme as followed: 

� Mobile terminal: stores the private key α and 

decryption key dk of file; computes the public 

� 

key g and the signature Sigsk(H(R)) of the 

MHT root H(R). 

� 

� TPA: stores the user’s public key g ; 

encodes/decodes, encrypts/decrypts the file, 

computes the data blocks signature collection Ф, 

and verifies the two equations during verification. 

� CSP: stores the signature Sigsk(H(R)), the 

encoded F’ and the Ф; generates the verification 

information μ and ω, and computes the Ωi for 

recovering the MHT. 

It should be noted that we do not include the 

workloads for generating the needed random numbers 

and nonce of every entity in our scheme. 

As described above, through a trusted channel between 

the end-user and the TPA, the heavy works, such as 

encoding (decoding), encryption (decryption) of the file, 

can be moved to the TPA to be done, and the end-user 

only need to generate keys and sign the root of the MHT. 

So our scheme realized the goal: the storage space and 

the computing ability needed for end-user during 

verifying the data possession are as small as possible, 

which is fit for the mobile computing device, such as 

mobile phone and PDA, etc.


By using the MHT idea, the data blocks signatures are 

aggregated into the signature of the root of the MHT to 

verify the data integrity with a minimum of storage space. 

The root signature is put into the cloud server, which 

realized the stateless verification. The MHT can also help 

to implement the dynamic data update as described in [6]. 


When the resource-constrained mobile devices use the 

cloud storage services deployed on traditional IP 

networks, the end-users are most concerned about 

whether the CSP stores their files correctly and dutifully. 

On the basis of the existing researches on PDP and POR, 

a secure storage scheme with public provable data 

possession of the mobile devices in cloud computing is 

proposed in this paper. Through the remote 

authentification of the mobile end-user by using trusted 

computing technology, a trusted third-party agent can 

undertake most computing workload of the client in 

traditional PDP, which makes our scheme be fit for the 

mobile computing environment. Combined with MHT 

and bilinear signature, improved the framework and 

related algorithms from the existing PDP schemes, our 

scheme realized public PDP with the support of dynamic 

data update. The scheme is provable secure under the 

random oracle model as proved in [6]. To the best of our 

knowledge, our scheme is the first to explore the 

application of PDP scheme in mobile computing 

environment combined with the trusted computing. 

Our future works include building a prototype system 

to test the performance of our scheme and exploring the 

application of other PDP framework applied in secure 

storage services of mobile computing environment. 

ACKNOWLEDGEMENT 

The authors wish to thank those reviewers and editors 

of this paper. This work is supported by the Science and 

Technology Support Program of Science and Technology 

Commission of Shanghai Municipality (No. 072712036), 

the National Natural Science Foundation of China (No. 

60803096), the Fundamental Research Funds for the 

Central Universities and Dalian IT teacher’s project. 

REFERENCES 

[1] Deswarte, Y., J.-J. Quisquater, and A. Saidane. “Remote 

integrity checking”. In Proc. of Conference on Integrity and 

Internal Control in Information Systems. 2003. 

[2] Filho, D.L.G. and P.S.L.M. Baretto. “Demonstrating data 

possession and uncheatable data transfer”, In IACR ePrint 

archive. 2006. 

[3] Ateniese, G., et al., “Provable data possession at untrusted 

stores”, in Proceedings of the 14th ACM conference on 

Computer and communications security. 2007, ACM: 

Alexandria, Virginia, USA. p. 598-609. 

[4] Juels, A. and J. Burton S. Kaliski, “Pors: proofs of 

retrievability for large files”, in Proceedings of the 14th ACM 

conference on Computer and communications security. 2007, 

ACM: Alexandria, Virginia, USA. p. 584-597. 


[5] Shacham, H. and B. Waters, “Compact Proofs of 

Retrievability”, in Proceedings of the 14th International 

Conference on the Theory and Application of Cryptology and 

Information Security: Advances in Cryptology. 2008, 

Springer-Verlag: Melbourne, Australia. p. 90-107. 

[6] Wang, Q., et al., “Enabling Public Verifiability and Data 

Dynamics for Storage Security in Cloud Computing”, in 

Computer Security – ESORICS 2009, M. Backes and P. Ning, 

Editors. 2009, Springer Berlin / Heidelberg. p. 355-370. 

[7] Ateniese, G., S. Kamara, and J. Katz, “Proofs of Storage 

from Homomorphic Identification Protocols”, in Advances in 

Cryptology – ASIACRYPT 2009, M. Matsui, Editor. 2009, 

Springer Berlin / Heidelberg. p. 319-333. 

[8] Chang, E.-C., and J. Xu, “Remote Integrity Check with 

Dishonest Storage Server”, in Proceedings of the 13th 

European Symposium on Research in Computer Security: 

Computer Security. 2008, Springer-Verlag: Malaga, Spain. p. 

223-237. 

[9] Johnson, R., et al., “Homomorphic Signature Schemes”, in 

Proceedings of the Cryptographer's Track at the RSA 

Conference on Topics in Cryptology. 2002, Springer-Verlag. p. 

244-262. 

[10] Wang, C., et al., “Privacy-preserving public auditing for 

data storage security in cloud computing”, in Proceedings of the 

29th conference on Information communications. 2010, IEEE 

Press: San Diego, California, USA. p. 525-533. 

[11] C. Chris Erway, A.K., Charalampos Papamanthou, 

Roberto Tamassia, “Dynamic provable data possession”, in 

Proceedings of the 16th ACM conference on Computer and 

communications security. 2009, ACM: Chicago, Illinois, USA. 

p. 213-222. 

[12] Mao, W.B. “Talking About the Cloud Computing”. 2009 

2009-03-03; Available from: 

http://blog.csdn.net/wenbomao/archive/2009/03/03/3952761.as 

px and http://www.daoliproject. org. 

[13] Murray, D.G., G. Milos, and S. Hand, “Improving Xen 

security through disaggregation”, in Proceedings of the fourth 

ACM SIGPLAN/SIGOPS international conference on Virtual 

execution environments. 2008, ACM: Seattle, WA, USA. p. 

151-160. 

[14] Pearson, S., Y. Shen, and M. Mowbray, “A Privacy 

Manager for Cloud Computing”, in Proceedings of the 1st 

International Conference on Cloud Computing. 2009, 

Springer-Verlag: Beijing, China. p. 90-106. 

[15] Santos, N., K.P. Gummadi, and R. Rodrigues, “Towards 

trusted cloud computing”, in Proceedings of the 2009 

conference on Hot topics in cloud computing. 2009, USENIX 

Association: San Diego, California. p. 3-3. 

[16] Cao, T.J., Y.P. Zhang, and C.J. Wang, “Secure Protocols”. 

2009, Beijing, China: BUPT Press. 2. 

[17] Bowers, K.D., A. Juels, and A. Oprea, “Proofs of 

retrievability: theory and implementation”, in Proceedings of 

the 2009 ACM workshop on Cloud computing security. 2009, 

ACM: Chicago, Illinois, USA. p. 43-54. 

[18] Li, Y., M. Jian-Feng, and Z. Jian-Ming, “Trusted and 

anonymous authentication scheme for wireless networks”. 

Journal of China Institute of Communications, 2009. 30(9): p. 

29-35. 

[19] Group, T.C., “TPM Main Specification Level 2 Version 

1.2”, Revision 103. 2007, TCG.


Jian Yang is a lecturer of Dali 

University in China. He was 

responsible for teaching database 

technology and E-commerce in 

Faculty of Computer. He was born on 

Dec in 1976. He received his master 

degree of Engineering in Kunming 

University of Science on May in 2005. 

Now he is a PhD student in Tongji 

University. He has published a dozen of papers on 

computer magazines. His major research interests include 

network security, trusted computing, cloud computing. 

Mr. Yang is a student member of China Computer 

Federation. 

Haihang Wang is a professor and a 

supervisor of Ph.D. student in Tongji 

University at Shanghai. He was born on 

Mar. in 1965 and received his Ph.D. 

degree in Zhejiang University in 1994. 

Now He is engaged in teaching 

E-commerce and Project Management of 

Information System. His research interests 

include intelligent information system, 

network security, E-commerce and supply chain management, 

enterprise information technology, mechatronics and 

automation, production and operations management. He has 

already published 60 papers on international magazines and 

conferences. 


Jian Wang is a lecturer of Henan 

University of Science and Technology in 

China. She earned her master degree in 

Huazhong University of Science & 

Technology in 2005. Now she is a Ph.D. 

student in Tongji University. Her research 

interests include trusted computing and 

network security. 

Chengxiang Tan was born in 1965 and 

earned his Ph.D. in North-west 

Polytechnic University in 1994. He is a 

professor and a supervisor of Ph.D student 

in Tongji University at Shanghai. He is 

engaged in teaching information security, 

digital forensics and the theory of 

information asurance. His major research 

interests include Network and information 

security, wireless and mobile services 

security support, multi-network integration, and digital crime 

investigation and forensic. 

Dingguo Yu was born in 1976, now is 

a PhD candidate of Tongji University, 

China. He received BS degree in 

mathematics in 1998 from Zhejiang 

Normal University, China, and received 

MS degree in computer application 

technology in 2005 from Tongji 

University. His current research interests 

include network & information security 

and mobile computing.


Image Compression Based on Improved FFT 

Algorithm 

Juanli Hu* 

Computer Engineering Department, Zhongshan Polytechnic, Zhongshan, China 

Email: hjlfoxes@163.com 

Jiabin Deng and Juebo Wu # 

Computer Engineering Department, Zhongshan Polytechnic, Zhongshan, China 

# Shenzhen Angelshine Co., LTD, Shenzhen, China 

Email: hugodunne@yahoo.com.cn and wujuebo@gmail.com 

Abstract—Image compression is a crucial step in image 

processing area. Image Fourier transforms is the classical 

algorithm which can convert image from spatial domain to 

frequency domain. Because of its good concentrative 

property with transform energy, Fourier transform has 

been widely applied in image coding, image segmentation, 

image reconstruction. This paper adopts Radix-4 Fast 

Fourier transform (Radix-4 FFT) to realize the limit 

distortion for image coding, and to discuss the feasibility 

and the advantage of Fourier transform for image 

compression. It aims to deal with the existing complex and 

time-consuming of Fourier transform, according to the 

symmetric conjugate of the image by Fourier transform to 

reduce data storage and computing complexity. Using 

Radix-4 FFT can also reduce algorithm time-consuming, it 

designs three different compression requirements of nonuniform 

quantification tables for different demands of 

image quality and compression ratio. Take the standard 

image Lena as experimental data using the presented 

method, the results show that the implementation by Radix- 

4 FFT is simple, the effect is ideal and lower time-consuming. 

Index Terms—Image Compression, Fourier Transform, 

Quantization Table List, Compression Ratio, Coding and 

Decoding 


Image coding is a kind of method by using image 

source coding to achieve data compression, in order to 

ensure the quality of images and try to reduce code rate. 

Through image coding, it can get the goal for saving 

bandwidth or space, and it may also be provided for 

multimedia computer processing [1].With rapid 

development of multimedia and communication 

technology, it requires a higher demands for data storage 

and data transmission, especially in large volumes of 

digital image communication, which greatly restricts the 

development of image communication. Therefore, more 

and more attentions are focused on image compression 

*Corresponding author. 


doi:10.4304/jnw.6.7.1041-1048 

techniques [2]. 

Nowadays, there are many high compression rate 

methods for digital image that can be divided into three 

types: Waveform method, the second-generation coding, 

Fractal coding etc [3]. 

A continuous tone image compression standard JPEG 

[4] is a common tool for static image compression, which 

allows both nondestructive compression and loss 

compression and it is representative of compression wave 

technology. Image is coded by prediction scheme with 

the conservative compression 2:1. In loss compression 

mode, the compression ratio can reach 5-20 times for 

most natural graphics providing better quality in the 

circumstances. However, its main defects are big 

distortion when doing compression with high 

ratio(blocking effect and mosaic noise) and lack of bits 

flow control and weak repair. On the condition that 

compression ration reaches 30-40, it may emerge a 

stronger blocking effect. In accordance with JPEG with 

high compression ratio, many improvement methods are 

proposed to overcome such obstacles like DCT zerotree 

coding and layer type DCT zerotree coding [5]. But in 

high compression ratio, the situation is still block-effect 

fatal weakness [6]. 

Karhunen-Loeve Transform, also known as the 

characteristic vector Transform, principal component 

Transform or Hotelling, is a good way to process image 

with transformation matrix determined by the specific 

statistical characteristics of image(covariance matrix). 

The biggest advantage is the correlation between the 

transform domain can be removed totally, that is, owning 

well decorrelation [7]. Usually, K-L transform is used to 

eliminate the correlation among smaller matrix elements, 

such as to remove the spectrum correlation in many bands 

of remote sensing image compression. In practical 

application, because the transformation matrix is bigger 

and when the processing matrix is bigger, the covariance 

matrix is not easily either. 

For the past few years, since Wavelet Transform has 

the ability and characteristics of local signal analysis in 

time and frequency, it has been widely used in image 

denoising, image reinforcing and image compression.


Wavelet decomposition is a subset of the sub-bands 

decomposition for image, and it provides a type of multiresolution 

representation method. 

The basic ideas for image coding based on wavelet is 

conducting multistage wavelet decomposition to image 

by fast wavelet transform algorithm of Mallat tower, and 

then quantifying each layer of wavelet coefficients and 

coding. JPEG2000 is a new generation of image 

compression international standards developed by 

ISO/IEC organizations. It applies wavelet transform 

(DWT) as the basic algorithm, in combination with 

embedded coding technology, the result can not only 

reach in higher image quality and higher compression 

efficiency, but also meet the needs to the mobile and 

network environment and the interoperability and 

scalability. However, linear filter inherent ringing effect 

will appear if the method in the compression ratio reaches 

about 50. 

Structure Coding is the second generation coding 

technology for image, which is full consideration of the 

human visual physiological and psychological 

characteristics. Its principle is based on the meaning 

element of visual sense to describe images, such as the 

outline and texture. The second generation encoding 

technology can be divided into two categories: directional 

decomposition technique and face outline/characteristics. 

The objective of directional decomposition is to detect 

and express the edge information of image more 

accurately and more effectively, in order to apply proper 

separation and coding. Outline/feature oriented 

technology can make different reflection depending on 

different characteristics according to visual system. It 

extracts the main feature firstly, and then carries out the 

corresponding coding. Such method can give a better 

store for the image edge profile information. Thus, the 

image quality remains a high level when high 

compression ratio. 

In the 1980s, Bamsley and Jacquin put IFS(Iterated 

Function Systems) into image compression. Fractal 

compression introduces the characteristics of self 

similarity to image compression. Through the iteration 

function of fractal image compression system, it realizes 

the segmentation of the original image, and then to map 

each subimage into an iterative function. Sub-images are 

stored by iterative function, the simpler the iterative 

function is, the bigger the compression ratio will be [8]. 

Bamsley put forward Local IFS theory in 1988, which 

solved the problem of the parts and the whole without 

self similarity. After that, many studies indicated that the 

fractal image compression encoding remains very good 

quality when compression ratio is in higher lever (70-80). 

The main problem is the complexity in the phase of 

image coding. Currently, the research on fractal 

technology focuses on the hybrid coding, fractal inverse 

problem and fractal convergence problems with 

improvement, fractal technology and other compression 

techniques (e.g., wavelet transform) [9]. 

Because of the good nature of Fourier transform, it has 

a wide range of applications in the image coding, image 

segmentation, image reconstruction and other areas. DFT 


transform with good energy concentration, due to the 

inconvenient operations and the great amount of 

calculation, hasn’t long been widely used in the image 

compression. For its complex algorithm and timeconsuming 

disadvantages, in this algorithm, the author 

makes use of the fast Fourier transform (FFT) to realize 

the limited distortion coding technology of the image. 

Investigating the feasibility of image compression using 

the Fourier transform, we utilize the conjugate symmetry 

to reduce the data storage. When reducing the timeconsuming, 

and for the different requirements of the 

compression ratio and the image quality, the Radix-4 

algorithm adopts three different quantization tables. 

Through the standard image compression ratio, root mean 

square signal noise ratio and the decoding consumption, 

we can draw that the Fourier transform method of the 

image compression not inferior to the JPEG compression 

system can get an ideal compression. Based on the 

conjugate symmetry of the Fourier transform, we can 

halve the data storage. By means of the Radix-4 Fourier 

transform, we can greatly reduce the time-consuming. 

II. FOURIER SPECTRUM ANALYSIS AND DESIGN OF IMAGE 

In the image processing, it often tends to do 

corresponding transformation for image in converting 

domain when facing to the problems that is complex and 

hard to deal with, in order to concentrate the energy on 

minority transform coefficient. Fourier Transform is a 

classical method to convert image from space domain to 

frequency domain, and it also the foundation of image 

processing titled as the second language for image 

description. It provides another perspective for image 

observation and the image can be transformed into gray 

distribution images to frequency distribution 

characteristics. 

In frequency domain, the more the frequency is, the 

faster the original signal changes. While the original 

signal is less changing as the less frequency. When 

frequency is 0, it means the dc signal has no change. 

Therefore, the size of the frequency reflects the signal 

changes. Most of the energy concentration in the image is 

located on the low dc and regional. Take Lena (512 x 512 

as shown in Fig. 1) for instance, the Fig. 2 shows the 

result of the Fourier transform. It is evidently seen from 

that the energy distribution of Lena image is focused in 

the low frequency part, to lower the frequency with the 

increase. Due to the real input signal, its distribution of 

spectrum is axis-symmetric on 0 1 ,ω ω , so it can only take 

account into the half parts. 

This method executes image partitioning at the 

beginning to divide the data into non overlapping blocks. 

And then each part will be mapped by 2-dimension FFT 

where the coefficients are not related after the 

transformation and the energy of coefficient matrix is 

gathered in low-frequency area. After that, the 

quantization table will be designed to non-uniform 

according to the different requirements, and the 

quantitative results retain the coefficient of low frequency 

part while drop the high frequency coefficient. "Z"


coding is done after the quantized data in this step. 

Finally, the entropy coding is conducted by Huffman 

coding so as to realize image compression. The entire 

process is shown in Fig. 3. 

Figure 1. Lena picture. 

Figure 2. The Fourier spectrum. 

Figure 3. The flowchart of image compression. 

III. MAPPING TRANSFORMATION OF IMAGE BY FFT 

Discrete Fourier Transform requires data discretization 

at the beginning in order to apply into computer 

technology. Discrete Fourier Transform is widely used in 

image processing and digital signal processing. Assume 

f (x) 

is number sequence with N length, and onedimensional 

discrete Fourier transform is defined as 

follows: 

F( 

u) 

N 1 

= ∑ − 

x= 

0 

f ( x) 

exp( 

− j2πux 

) 

N 

where u = 0, 1,... 

N −1 

. As shown in above formula, 

one F ( u) 

needs N times plural multiplication and 

2 

N −1 

times plural additio n; N F (u ) needs N times 

plural multiplication and 

N * ( N −1) 

times plural 

addition. Obviously, the bigger N is, the more calculation 

will 

be. 

Digital image f (m, n) is described as a matrix M rows 

by N columns [f (m, n)] in computer, and the following 


(1) 

definitions are the 2-dimension DFT and inverse 

transform of image matrix respectively. 

F( 

s, 

t) 

N−1 

1 ⎡ 

* ⎢ 

N n= 

0 ⎣ 

1 

M 

tn 

∗exp( 

−j2π 

) 

N 

f ( m, 

n) 

M−1 

= ∑ ∑ 

m= 

0 

sm ⎤ 

f ( m, 

n) 

exp( −j2π 

) 

M 

⎥ 

⎦ 

(2) 

N−1 

1 ⎡ 

* ⎢ 

N t= 

0 ⎣ 

M−1 

1 

sm ⎤ 

F( 

s, 

t) 

exp( −j2π 

) 

M s 0 

M 

⎥ 

= ⎦ (3) 

tn 

* exp( −j2π 

) 

N 

= ∑ ∑ 

Two-dimensional Fourier transform can be seen as two 

times of one-dimensional, namely the Fourier transform 

of image sequence according to the ranks respectively. 

Obviously, the Fourier transform with N points need N * 

N times and N*(N-1) additions. Therefore, the bigger the 

image is the longer consuming the calculation will be and 

it is very important to choose one kind of fast algorithm. 

A. Figures and Tables 

Because the Fourier transform operation contains a lot 

of repetition computation, people studied many fast 

Fourier transform algorithm. 

In 1965, J.W.Cooley and J.W.Turky proposed fast 

Fourier transform algorithm. According to the 

composition of basic wing operation, the algorithm is 

divided into 2-radical, 4-radical, 8-radical, 16-radical 

arbitrary factor FFT algorithm. Currently, 2-radical and 

4-radical are the most widely used. The affiliate 

multiplication and wig of FFT are proportional to 

N log N , where the bigger N can save more 

2 

computation. 

FFT algorithm are as follows: Firstly, by using the 

basic idea of the three characteristics of rotating factors 

W 

kn 

N 

, namely, the periodicity, symmetry and about sex, 

the original N points of DFT long sequence are 

decomposed into two or more short sequences, and 

merging DFT operation fitting for combination. Secondly, 

recombine the DFT of the original sequence after 

calculation of short sequences, in order to improve the 

speed with less computation [10]. These can be divided 

into two kinds of decomposition methods: 

1) Transfer a large Discrete Fourier Transform 

computation into a group of short length, known as 

Decimation-In-Time-FFT (DIT-FFT). 

2) Decompose Fourier series X (k), known as 

Decimation-In- Frequency-FFT (DIF-FFT). 

The basic principle of radix-4 DIT-FFT [11] is as 

follows: Divide FFT with N points into 4 sequences and 

compute the DFT separately. Likewise, turn N/4 points 

into fine particle size, and so on. For multi-points (4m), 

multi-stage decomposition can be established in a similar 

way. For instance, based on radix-4 DIT-FFT, the series x 

(n) with 256 length can be obtained:


x( 

k) 

= X( 

k k k k 

W 

W 

( K 

N 

3 

K0 

4 n3 

N 

3 

3 4 + 

W 

2 

2 

1 

0 

) 

2 

( K1 

4+ 

K0 

) 4 n 

N 

2 

0 

3 

3 2 1 0 

n = 0n 

= 0n 

= 0n 

= 0 

K 4 + K 4+ 

K ) n 

3 

0 1 2 3 

W 

2 

3 

( K2 

4 + K1 

4+ 

K0 

) 4n1 

N 

3 

∑∑∑∑ 

x( 

n n n n ) . 

Whilst the time complexity of radix-2 FFT is: 

3 

. 

2 

1 

0 

(4) 

Figure 4. The flowchart of radix-4 butterfly 

1 

= N log N 

(6) 

2 

m F 2 

In all kinds of FFT algorithms, radix-2 FFT is the 

simplest one but its calculation is more complex than 

using radix-4 FFT. People usually measure a performance 

of the algorithm by the times of addition and 

multiplication. For example, the radix-4 FFT is described 

as: each butterfly is 3 plural by N plural points, having 

log4N levels. Each level has N/4 radix-4 butterflies. 

Generally speaking, it shows the higher cardinal the less 

computation, but to judge whether an algorithm good or 

not is not only to consider the calculation but also the 

complexity. From radix-2 to radix-4, the number of 

multiplication and addition has a big jump and the count 

of radix-4 FFT general reduces about 1/4 by radix-2. 

From 4 to 8, or 8 to 16, the number has not a obvious 

change. Regarding algorithm complexity, radix-2 is 

easiest to control and use while radix-4 is harder to 

control, but radix-4 is the analogy as radix-2. Compared 

with radix-4, the jump of radix-8 and radix-16 is very 

obvious. Considering speed and control complexity, 

radix-4 has the highest realization ratio in FFT. Therefore, 

this paper selects radix-4 FFT algorithm to realize FFT 

processor design. 


The radix-4 butterfly processing is shown as Fig. 4. 

The time complexity of radix-4 FFT is as following 

where there are N/4 FFT 4-points and L levels totally. 

N 3 

mF 3× × ( L −1) 

≈ N log2 

N 

4 8 

= (5) 

B. The feature and data store of conjugate symmetry of 

Fourier transform 

By using Fourier transform, the result is plural number, 

such as matrix 8*8 will be converted to 8*8 plural matrix. 

It seems that the data is double than original one, but the 

fact is that the real data quantity does not increase as a 

result of the conjugate symmetry properties. After DFT 

transform, the matrix 8*8 will be: 

* 

F( s, 

t) 

= F ( 8 − s, 

8 − t) 

Where reflects the cycle around s=4 or t=4 in a space 

spectrum with conjugate symmetry properties as shown 

in Fig. 5. 

Figure 5. The conjugate symmetry of Fourier transform 

(7)


In such process, the real number includes F(0, 0), F(0, 

4), F(4, 0) and F(4, 4), the others is conjugate symmetric. 

F (0, 0) is "dc" component, F(s, 4) and F (4, t) are high 

frequency area and F(4, 4) is the highest. The shaded 

parts are the valid data. 

The method of data store after FFT is closely related to 

quantization table, and this paper adopts the following 

way to save data. Taking image into 8*8 sub-blocks, it 

⎡ F( 

0, 

0) 

⎢ 

⎢ 

F( 

1, 

0). 

R 

⎢F( 

2, 

0). 

R 

⎢ 

⎢F( 

3, 

0). 

R 

⎢ F( 

4, 

0) 

⎢ 

⎢F( 

1, 

4). 

R 

⎢F( 

2, 

4). 

R 

⎢ 

⎢⎣ 

F( 

3, 

4). 

R 

F( 

0, 

1). 

R 

F( 

1, 

0). 

I 

F( 

2, 

0). 

I 

F( 

3, 

0). 

I 

F( 

4, 

1). 

R 

F( 

1, 

4). 

I 

F( 

2, 

4). 

I 

F( 

3, 

4). 

I 

F( 

0, 

1). 

I 

F( 

11, 

). R 

F( 

2, 

1). 

R 

F( 

31, 

). R 

F( 

4, 

1). 

I 

F( 

51, 

). R 

F( 

6, 

1). 

R 

F( 

7, 

1). 

R 

F( 

0, 

2). 

R 

F( 

1, 

1). 

I 

F( 

2, 

1). 

I 

F( 

3, 

1). 

I 

F( 

4, 

2). 

R 

F( 

5, 

1). 

I 

F( 

61, 

). I 

F( 

7, 

1). 

I 

IV. DESIGN AND IMPLEMENT OF QUANTIZATION TABLE 

Quantization is the most important step in the 

compression method. Its function is to map the 

continuous transform coefficient into limited data set. 

Quantitative can be divided into the scalar quantization 

and vector quantization. The scalar quantification can be 

divided into uniform and nonuniform quantification and 

visual quantification in detail. Visual quantification is in 

the process of considering quantification of different 

visual band which has different sensitive degree of 

properties, in order to apply the large quantization step to 

high frequencies. Vector quantization is an image block 

coding method, and the process is to map vector into 

predefined code book where selection of books is 

according to the statistic characteristics before 

quantitative. Vector quantitative is superior to scalar 

quantitative theoretically, but the cost of actual 

implementation is too big. Combined with other methods 

to realize in hardware is the development direction in the 

future. Considering the overall performance of the 

algorithms, the scalar quantitative is adopted in this 

algorithm. 

Subjective evaluation method is to observe an image 

directly and justify the degree of distortion from feeling. 

After that, the evaluation of quality score is given by the 

weighted average scores from all comments and the 

results are the subjective evaluation results, including 

absolute scale and relative scale. 

This evaluation system is consistent with the visual 

perception in relative to the objective evaluation method, 

which is reliable. But it cannot be used in convenient way: 

Firstly, it cannot be used in the process of image coding 

quality evaluation and control. Secondly, the subjective 

assessment of the testee is susceptible to the effects of 

subjective factors, such as age, and deviation, education 

level, personality, cultural background, etc. Thus, this 


can apply symmetry way to compensate the missing 

blocks if they fail to meet the sub-block. If the image is 

true color with 24 bits, RGB should be used to code. 

Then, FFT is carried out for all of the matrix 8*8, and the 

result can be saved half according to conjugate symmetric, 

that is 8*8 real number. 

F( 

0, 

2). 

I 

F( 

1, 

2). 

R 

F( 

2, 

2). 

R 

F( 

3, 

2). 

R 

F( 

4, 

2). 

I 

F( 

5, 

2). 

R 

F( 

6, 

2). 

R 

F( 

7, 

2). 

R 

Figure 6. The matrix 8*8 

F( 

0, 

3). 

R 

F( 

1, 

2). 

I 

F( 

2, 

2). 

I 

F( 

3, 

2). 

I 

F( 

4, 

3). 

R 

F( 

5, 

2). 

I 

F( 

6, 

2). 

I 

F( 

7, 

2). 

I 

F( 

0, 

3). 

I 

F( 

1, 

3). 

R 

F( 

2, 

3). 

R 

F( 

3, 

3). 

R 

F( 

4, 

3). 

I 

F( 

5, 

3). 

R 

F( 

6, 

3). 

R 

F( 

7, 

3). 

R 

F( 

0, 

4) 

⎤ 

F( 

1, 

3). 

I 

⎥ 

⎥ 

F( 

2, 

3). 

I⎥ 

⎥ 

F( 

3, 

3). 

I⎥ 

F( 

4, 

4) 

⎥ 

⎥ 

F( 

5, 

3). 

I⎥ 

F( 

6, 

3). 

I⎥ 

⎥ 

F( 

7, 

3). 

I⎥⎦ 

paper adopts the objective and subjective evaluation 

method of combining ways. 

The scalar qualitative contains uniform quantification 

and nonuniform quantification. The uniform 

quantification is to quantify the range of values for input 

signal by equidistance division. In comparison with 

uniform quantification, the nonuniform quantification has 

two advantages. One is the output of nonuniform 

quantification can obtain higher average signal 

quantization noise power when the input is non-uniform 

probability density. The other is the quantization noise 

power and value is in proportion to signal sampling value. 

Thus, the signal quantitative signal-to-noise ratio can be 

improved. 

This algorithm adopts non-uniform quantification. In 

this process, the low frequency signal has more important 

information, so it requires higher quantitative precision 

than the low energy. Since the high frequency signal is 

not sensitive to eye, the low-frequency coefficient is set 

in a smaller value while the high frequency part set larger. 

In this paper, three varieties of quantization tables are 

utilized to set image data according to the different 

requirements for quality. We quantify the results after 

FFT to control the compression ratio for different image 

quality, as shown in table 1~3. 

TABLE I. 

THE QUANTIZATION TABLE WITH HIGH COMPRESSION RATIO 

8 20 20 32 32 40 40 60 

20 20 32 32 40 40 60 60 

32 32 40 40 60 60 80 80 

40 40 60 60 80 80 80 80 

60 80 80 100 100 100 100 100 

40 40 60 60 80 80 80 80 

60 60 40 40 60 60 80 80 

80 80 32 32 40 40 60 60


TABLE II. 

THE QUANTIZATION TABLE WITH MIDDLE COMPRESSION RATIO 

8 16 16 16 16 32 32 40 

16 16 20 20 32 32 40 40 

16 16 32 32 40 40 60 60 

32 32 40 40 60 60 60 60 

40 40 40 60 60 80 80 100 

32 32 40 40 60 60 60 60 

40 40 32 32 40 40 60 60 

60 60 20 20 32 32 40 40 

TABLE III. 

THE QUANTIZATION TABLE WITH LOW COMPRESSION RATIO 

8 12 12 12 12 20 20 32 

12 12 16 16 20 20 32 32 

12 12 20 20 32 32 40 40 

20 20 32 32 40 40 40 40 

32 32 32 40 40 40 40 60 

20 20 32 32 40 40 40 40 

32 32 20 20 32 32 40 40 

40 40 16 16 20 20 32 32 

This algorithm applies "Z" run length encoding to 

coefficients after quantification and Huffman entropy to 

coding results in order to achieve image compression. 

Entropy coding is a kind of lossless coding and the 

common methods are composed of Run-length encoding 

(RLE), Lempel-Ziv-Welch(LZW), Shannon, Huffman 

and Arithmetic coding. The basic idea of entropy coding 

is using short code to express the signal with larger 

probability and long code for less signal. It gains a shorter 

length in statistic to reduce the average code data of space 

and improve the compression ratio. 

V. EVALUATION ANALYSIS FOR CODE EFFICIENCY AND 

COMPRESSION QUALITY 

The meaning of image quality includes two layers: 

One is to reconstruct the distortion degree of image with 

the original image reconstruction and deviation degree. 

The other is the readability of the image, that is, the 

information from the image who gets. Normally, the eye 

is the information receiver of image, but the visual 

Lena Image 

quality 

Good 

(lossless) 

Bitmap File 

Size (bytes) 

Compressed 

file size 

TABLE IV. 

LENA IMAGE TE ST RESULTS 

Compression 

ratio 

system is very limited and cannot understand the 

distortion degree of image with readability and 

quantitative description. So, image quality assessment 

needs an objective evaluation method besides the 

subjective evaluation methods. 

In objective evaluation methods, 

the mean square error 

(MSE) and mean square signal to noise ratio (SNR) are 

commonly used. SNR is defined as: if the compressed 

image is represented as the superposition of the original 

image and noise, that is 

f ( x, 

y) 

= g ( x, 

y) 

+ e( 

x, 

y) 

(8) 

∑∑ 

∑∑ 

2 

2 

( SNR ) RSM = f ( x, 

y) 

/ e ( x, 

y) 

(9) 

For a digital image compression coding system, we can 

use 

redundant, coding efficiency and compression ratio to 

measure the source characteristics and encoding / 

decoding performance. The compression efficiency of 

image is usually measured with the compression ratio. the 

higher the compression ratio, the greater the image 

compression, and vice versa. In addition, it till needs to 

consider the complexity of the algorithm, including time 

complexity and space complexity. In this paper the 

algorithm coding efficiency is taken as the main 

consideration, and the coding efficiency can be reflected 

through the time-consuming of the encoding and 

decoding. 

It should be noted that the MSE and SNR reflect the 

overall 

differences between the original image and 

reconstructed image, and do not reflect the local 

differences. Sometimes in the same signal to noise ratio, 

visual effects will still have some differences, primarily 

due to the uniformity of the error. In general, if the 

uniformity of the error is high, the visual effect is good, 

otherwise the visual effects is bad. In most cases, we can 

use the PSNR to evaluate the image quality, but 

sometimes the results may be inconsistent with the 

subjective evaluation. But sometimes the outcome may 

have deviation from the results of subjective evaluation. 

This algorithm uses standard image (Lena) as the data 

analysis standard, the results shown in Table 4. Images 

512x512 pixel 24-bit True Color. Compression results 

map, as shown in Fig. 7~ Fig. 10. 

SNR 

Coding time - 

consuming 

(seconds) 

Decoding timeconsuming 

(seconds) 

786486 765279 1.03 Infinite 0.131 0.47 

Better 786486 94588 8.31 20.12 0.280 0.2115 

Medium 786486 77396 10.16 17.88 0.258 0.211 

Lower 786486 62245 12.64 15.71 0.280 0.211 


Test environment: CPU: Pentium® Dual-core 2.5 G; Memory: 2G DDR; OS: Windows XP.


Figure 7. Lossless compression 

Figure 8. Compression ratio by 8.31 




It can be seen that Fourier transform limited 

compression system is better than distortion coding, 

design coding and decoding time of the three. But the 

compression ratio is about 10 and most above 0.25 

seconds with decoding time about 0.2 and visual 

compression ratio of more than 12.64 where the image 

has a slight blocking effect. 

VI. CONCLUSIONS 

Fourier transform has a good transformation energy 

concentration, but for a long time, the results of Fourier 

transform will be plural, leading to the high complexity 

of encoding and long time of encoding/decoding, so can’t 

be widely used. This paper uses Fourier transform 

compression coding system based on 4FFT algorithm. 

Can be seen from Figure 6,7,8,9, image compression is 

better, image compression ratio is at 12.64 and above, a 

slight blocking effect will appear. Can be seen from 

Table 4, when the compression ratio is at about 10 and 

above, the time of compression is between 0.25-0.28 

seconds, with the increase of compression ratio, 

compression time of encoding/decoding change slightly. 

Decoding time is less than the encoding time at about 

0.21 seconds. Fourier transform image compression 

method can also produce ideal compression results. 

According to the conjugate symmetry of Fourier 

transform data, data storage can be reduced by half. 

Through the use of Radix-4 FFT, the algorithm can 

greatly reduce the time-consuming of algorithm. 

Reference to the direction of future research: in the 

compression rate, we can adopt several Fourier transform 

to get more focused information and reduce the 

correlation between pixels; do segmentation coding 

according to the characteristics of spectrum in different 

region; After FFT transform, plural data store the matrix, 

high-frequency information are set into the lower right 

corner of the matrix, low-frequency information are set 

into the top left corner, in order to improve its 

compression ratio. We can use the assembly language in 

FFT module to improve the encoding/decoding speed, or 

use hardware or other fast FFT methods to achieve. The 

biggest problems of Fourier transform and JPEG 

compression method are the severe block effects in the 

high compression ratio. In future work, we should 

consider combining the human visual characteristic to 

compress. 


The authors wish to thank their colleagues and external 

advisors for their help and support, in particular, Shuliang 

Wang (Ph.D. Professor, Wuhan University, China). 

This paper is supported by National 973 

(2007CB310804), National Natural Science Fund of 

China (60743001), Best National Thesis Fund (2005047), 

and Natural Science Fund of Hubei Province (CDB132).


REFERENCES 

[1] W. B. Penneder, J. L. Mitchell. JPEG: Still Image Data 

Compression Standard, Van Nostrand Reinhoid. New York: 

A cademic Press,1993. 

[2] X. C. Zhu, F. Liu, D. Hu. Digital image processing and 

communication. Peking University Press, 2002,108-132. 

[3] J. H. Xu. Image processing and analysis. Beijing: science 

press, 1992.79-86. 

[4] G. A. Rong. Computer image processing. Tsinghua 

university press, 2000.89-98 

[5] P. Kauff, K. Schuur. Shape-adaptive DCT with blockbased 

DC separation and Delta DC correction [J].IEEE 

Trans. Circuits Syst. Video Technol., 1998, 8 (3):237-242. 

[6] M. Bi, S. H. Ong, Y H Ang. Comment on "Shape-adaptive 

DCT for generic coding of video [J].IEEE Trans. Circuits 

Syst. Video Technol.,1996 6(6): 686-688. 

[7] Huazhong university of science and technology department. 

The probability and statistics. Higher Education 

Press,1999.104-128 

[8] O. Egger, P. Fleury, T. Ebrahimi. Shape-adaptive wavelet 

transform for zerotree coding [C]. Proc. Eur. Workshop 

Image Analysis and Coding for TV, HDTV and 

Multimedia Application Rennes France, 1996: 201–208. 

[9] [9] O. Egger. Region representation using nonlinear 

techniques with applications to image and video coding 

[D].Ph.D. dissertation, Swiss Federal Institute of 

Technology (EPFL), Lausanne, Switzerland, 1997. 

[10] X. D Duan, L. Z. Gu. High-performance Radix-4 FFT 

processing Design. Computer Engineering, Vol.34 No.24 , 

2008, 238-243. 

[11] K. Miyase, S. Kajihara. Optimal Scan Tree Construction 

with Test Vector Modification for Test 

Compression[C]//Proc. of IEEE Asian Test Symposium. [S. 

l.]: IEEE Press, 2003. 


Juanli Hu is currently lecturing at the 

Computer Engineering Department of 

Zhong Shan Polytechnic. She has been 

working as Director of Teaching 

(Fundamental Computing Research) 

since 2006. She received her Bachelor 

Degree in Computer Engineering from 

Xi'an University of Technology in 2000, followed by a 

Master Degree in Computer Engineering from Xi'an 

University of Technology in 2005. Her research interests 

include signal, information processing and data mining. 

Some of her research papers are indexed by IEEE CS. 

Jiabin DENG is a lecturer at the 

Computer Engineering Department of 

Zhong Shan Polytechnic. He received 

a Bachelor Degree in Computer 

Engineering from Hubei University 

in 2005, then a Master Degree in 

Computer Engineering from Wuhan 

University in 2007. His research 

interests include artificial intelligence, 

data mining, and complex network. His research papers 

are also indexed by IEEE CS etc. 

Juebo Wu was born in China and has 

obtained B.A. and M.A respectively in 

2005 and 2007 from International 

School of Software, Wuhan University, 

China. At present, he is working for his 

PH.D. candidate of Mapping and 

Remote Sensing in State Key 

Laboratory of Information Engineering 

in Surveying(Wuhan University, China) 

and will graduate in the summer 2010. 

Juebo Wu's primary research is spatial data mining, 

GIS and software engineering, etc. In recent years, he has 

published more than 10 papers indexed by EI/ISTP and 

won two computer software copyright.


Correlative Peak Interval Prediction and Analysis 

of Chaotic Sequences 

Qun Ding 

Electronic Engineering Key Laboratory of Universities in Heilongjiang Province, Heilongjiang University, Harbin, 

China, Email: qunding@yahoo.com 

Lu Wang, Guanrong Chen 

Electronic Engineering Key Laboratory of Universities in Heilongjiang Province, Heilongjiang University, Harbin, 

China, 

Department of Electronic Engineering, City University of Hong Kong, Hong Kong, China 

Abstract—The paper proposes a digital circuit design for 

the logistic-map module used in chaotic stream ciphers, 

analyzes the factors that may affect the output of the 

sequences, and develops a calculation method for estimating 

the output sequential correlative peak interval. With the 

respective tests using different initial values, the values of 

parameter u and the computational precisions, extensive 

experiments have been carried out. A 

formula for calculating correlative peak interval is 

proposed. Moreover, the relationships among precision, 

parameter u and correlative peak interval is provided. To 

ensure the security of the plaintext which is encrypted by 

the output sequence of the logistic-map, a proper precision 

could be chosen according to the formula. It provides a 

theoretic basis for the actual application of the chaos 

cryptology. The basic theory and methods have a significant 

implication on the statistical analysis and practical 

applications of the digital chaotic sequences. A diagram that 

presents the relationship among precision, parameter u and 

correlative peak interval has been generated for analysis. 

Index Terms—discrete chaotic systems, correlative peak 

interval, finite precision, encryption 


Chaos theory has been studied extensively for many 

years, that the analysis of chaotic characteristics and 

practical applications has significant meaning in the 

research of chaos [1,2]. Chaos has many prominent 

features, such as the output of the system is sensitive 

depended on the initial conditions; the output of the 

system has the feature of long-term unpredictability; the 

orbit is irregularity; the output of the system has 

random-like behaviors. Since the properties of the chaos 

are desirable by cryptographic applications, more 

attention has been paid on the research of cryptography 

with chaos. Chaos theory has had many applications on 

the cryptography, such as encrypting still image with 

chaotic maps; generating pseudorandom sequences with 

chaotic sequences instead of the conventional 

m-sequences; using chaotic maps as key generators in 

ciphers design. Particularly, stream cipher based on 


doi:10.4304/jnw.6.7.1049-1056 

Email: wanglu.daily@gmail.com,eegchen@cityu.edu.hk 

chaotic maps can solve some difficult problems in 

nonlinear sequential cipher [3,4. Consequently it provides 

a new approach for information security application by 

increasing the complexity of deciphering. 

Chaotic stream ciphers are desirable for encryption 

devices and secure communication systems. However, 

when a chaotic device or system is realized by a 

computer which the precision is finite, the resultant 

discrete dynamics are different from that of the original 

analog system. Although there are some methods that can 

be used to improve the quality of the discredited chaotic 

systems, such as small-perturbation algorithms or 

multiple cascading chaotic systems [5,6], the finite 

precision of the computer is the main problem in 

application of chaos. Therefore it is difficult to set up an 

output sequence mathematical model which restricts the 

application of chaos. With the improvement of 

computational accuracy and operational speed of 

large-scale integrated circuits, the intrinsic degenerating 

phenomena of various chaotic characteristics can be 

studied more precisely which may promote the 

applications of chaos theory, especially in cryptography 

and secure communications[7,8]. 

II. CHAOS THEORY 

Bifurcation chart is a description of state variant 

based on parameter space. In the bifurcation chart, the 

range of chaotic parameter the process of bifurcation and 

the period window are very clear. The general form of me 

to itself mapping is: 

xn+ 1 f( μ, 

xn), xn R 

= ∈ (1) 

f : I → I is the differentiable function , μ is a 

parameter。If from a certain initial value, the value of 

xn state repeats infinite cycle among p( p≥ 1) 

states, 

and then (2) is a periodic orbit. If the period is 1, it means 

it is a fixed point.


x′ 1, x′ 

2, 

�, x ′ p 

(2) 

According to the method of linear stability, the 

conditions for stable periodic orbits is 

p 

∏ f′ ( μ, 

x′ 

t ) ≤ 1 

(3) 

t= 

1 

and the periodic orbit p is called super-stable. 

We discuss the bifurcation chart with the example of 

logistic map. Logistic map is defined as [9]: 

[ ] [ ] 

xn+ 1 μx(1 x), μ 0, 4 xn 

0,1 

= − ∈ ∈ 

(4) 

from (4) we get the fixed point: 

1 

O: x= 0; A: x= 

1− 

.The stability of fixed point is 

μ 

determined by the gradient f ′ ( x) 

of y = f( x) 

, 

which is: 

f ′ ( x) = μ − 2xμ 

(5) 

Consequently, the stability of fixed points depends 

on the parameters μ .From the behavior of iterative 

equation (4), it relies on the steepness of the parabola 

sensitively which has the same meaning with nonlinearity. 

Therefore, when the parameter μ becomes larger from 

zero, the iterative process (4) has different dynamic 

behavior [10]: 

When 0< μ < 1, 

fix an initial value x0 in [0,1], and 

then iteration process moves to a fixed point quickly 

xn → 0 , due to f ′ (0) = μ < 1, 

so it exist stable fix 

point O. 

When μ = 1 , f ′ (0) = 1 , collapse bifurcation 

occurs. 

When 0< μ ≤ 3= 

μ1 

, there are two fixed points O 

and A. Because of f ′ (0) = μ > 1 , so the point O is 

1 

unstable. For the point of A, f ′ (1 − ) = 2 − μ < 1. 

Therefore the iterative process of initial value x 0 moves 

away from fixed point O and closer to the fixed point A. 

For example, when μ = 2 , after the iteration xn → 0.5 , 

this called period 1, such can be seen from figure 1(a). 

When μ = 3 , as f′ ( A) = 2− μ =− 1 , so it 

generates fork-type bifurcation 

When 3< μ ≤ 1+ 6 = μ2 

, as f ′ (0) = μ > 1 , 

it is still unstable; For the point A, 

1 

f ′ (1 − ) = 2 − μ > 1, 

so the point A changes from 

μ 

stable to unstable, such can be seen from figure 1(b) 


μ 

which is called period 2. 

When 3.449 < μ < 3.545 = μ3 

, the two values of 

period 2 changes unstable again and generates a couple of 

new fixed points, so x n jumps among the four values, 

such can be seen from figure 1(c) which is called period 4. 

Until μ > 3.57 = μ∞ , the time 

series x0, x1, x2, �, xn, 

� are like the random number 

in [0, 1], so it is called chaos, such can be seen from figure 

1(d)[11,12]. 

(a)period 1 Time series 

(b)period 2 Time series 

(c) period 4 Time series


(d) Chaotic Time series 

Figure 1. Different dynamic behavior of logistic map 

Nonlinear equations change the topology structure of 

system trajectory through the above dynamic behavior 

which will cause the overall shape of the system changes 

suddenly and produce the phenomenon of bifurcation [13]. 

This is a necessary process of the generation of chaos, and 

it is also the source that chaos fits cryptographic 

properties. Therefore, it has become a new trend of 

cryptography to design new cryptogram program based on 

chaotic systems. With the large scale application of 

integrated circuit in the chaotic encryption, we need a 

uniform standard to evaluate the models of digital chaotic 

system which still has the characteristic of randomness 

[14]. Through the study we found that although the 

application of chaos cannot fully rely on its features, it 

couldn’t separate its properties from the equation, such as: 

correlation function, bifurcation maps. Consequently, we 

try to construct chaotic and design the characteristics of 

chaotic bifurcation with FPGA hardware platform design. 

Thus, the method of digital chaos has been proposed, 

which will test our security systems and play an important 

role in promoting the application of chaos [15]. 

Ⅲ. CIRCUIT DESIGN FOR LOGISTIC PSEUDORANDOM 

SEQUENCES 

The signal also has the statistical property of the 

correlation that the interval between the relevant peak 

value is equivalent and the interval is usually 

unpredictable. Although there is correlation, as the 

interval between the peak value is large and the circuit is 

simple which is suitable for integrated circuit design, it 

still has practical value in some encryption occasions [16]. 

In this paper, we will define the correlative feature of the 

digital chaotic sequence which is the equivalent interval 

between the peak value as the correlative peak interval. 

The estimate of correlative peak interval is an important 


parameter during the research of the chaotic application. 

The correlative peak interval of the simple 

one-dimensional logistic chaotic sequence is researched 

in this paper. 

The well-known logistic map is (4), if the chaotic 

logistic map is used in a cryptosystem; we have to ensure 

the output sequence is pseudorandom. To ensure the 

logistic map in the region of chaos, the key parameter u 

has to be chosen exactly. 

The circuit design diagram of the new method is 

shown as Figure 2 ,the theory of schematic circuit could 

be seen from [17,18], which has many modules, such as 

operational amplifier (for u), adder, multiplier, and delay 

units. It can generate modulated chaotic output sequences 

via control-shifting and extraction. A 128-bit data 

sequence goes through a data processor which is as the 

initial key .The selector controls the selection. If the SEL 

is 1, then pass the initial value; else if the SEL is 0, RES 

is 0 and EN is 1, then generate a logistic-map output by 

repeated iterations. 

Figure 2. Design diagram of the chaotic logistic-map module 

After the circuit design of logistic chaotic equation 

which is based on FPGA[19], we get the output of the 

simulation that is shown as Figure 3. It could see that 

when the control signals RES、SEL、EN are valid, a total 

of 128 initial key signal which is from Input1 to Input8 

will generate the output signal of the chaotic sequence. 

After a large number of auto-correlation tests, the result 

indicates that there is periodic interval between the peak 

value of the sequence which is shown as Figure 4. The 

correlation peak with the same interval and the interval is 

usually unpredictable. This paper carries out the 

theoretical analysis and calculation on the correlative 

peak interval.


As a key sequence generator, we must first consider 

the selection of the initial key. Through the experiment 

test and analysis, we can see that the statistical 

characteristics of the output sequence are affected by the 

computing accuracy, equation parameters and initial 

values[20]. As the computing precision in the system is 

extremely limited and precision is the main reason for 

degradation of chaotic dynamics, it is not suitable as the 

key input in chaotic encryption algorithm. According to 

the research of chaotic dynamics, when μ is in the 

interval of [3.5699456, 4], Logistic map is in the chaotic 

state and the output sequence is non-periodic and 

non-convergent. However, it can be seen from Lyapunov 

index curve that the interval is not always in chaos state, 

and when μ =4, the map is a full shot in the unit 

interval [0, l] that the chaotic sequence has the 

characteristic of periodicity. Therefore μ can not be the 

initial key input of chaotic encryption .When the initial 

value has a tiny deviation, the orbit will separate with 

exponential speed. So it is impossible to have a 

long-term prediction on the Behavior of the system. Just 

as the chaotic system is sensitive with the initial value, 

when the chaotic system is assigned with different initial 

value, we can get a series different and not related 

chaotic sequence. Therefore, we choose the initial value 

of chaotic systems as the chaotic key input. 


(a) the initial value is 0.118 

Figure 3. Simulation output sequences of the logistic map 

(b) the initial value is 0.1378 

(c) the initial value is 0.216 

Figure 4. Diagram of correlative peak interval 

Ⅳ. THE CORRELATIVE PEAK INTERVAL OF OUTPUT 

SEQUENTIAL 

The correlative peak interval could be got by 

calculating the auto-correlation function. When the 

6 

precision is obtained between 13 and 44, 2× 10 dots are 

used to test the output with different values of u. 

Experiments have confirmed that under the condition of


a fixed computational precision, the parameter u is 

changed. If the parameter u is fixed, the correlative peak 

interval will change according to the change of 

computational precision. The change of the correlative 

peak interval is nonlinear and irregular. In practical 

applications, the parameter u and the precision are fixed 

according to the requirements. An example of a cyclic 

period curve versus the precision is shown as Figure 5. 

The parameter u is fixed which is equal to 3.617. 

Figure 5. A curve changing with the precision (u =3.617) 

Using the least-squares method for curve fitting can 

determine a formula, i.e. a fitting model, f(x). For the 

original data shown in Figure 5, for instance, one may 

construct an exponential fitting model, a piecewise 

fitting model, or a polynomial fitting model. For the first 

model, the iterated value of the exponent diverges; for 

the second model, piecewise fitting is often inconvenient 

to use as a mathematical model (for example, if the test 

data are divided into two parts then the simulation results 

are as shown in Figure 6, therefore, the polynomial 

fitting model is chosen in this investigation. 


(a) Interval function on [13,35] 

(b) Interval function on [36,41] 

Figure 6. Curves obtained by piecewise fitting. 

Linear, quadratic, cubic and quartic polynomial fitting 

have been performed and compared, which is shown 

as Figure 7. 

(a) linear fitting 

(b) quadratic fitting


Figure 7. 

(c) cubic fitting 

(d) quartic fitting 

Polynomial fitting models 

It is clear that the quartic fitting gets the best result in 

approximating the test data. Therefore the quartic fitting 

model is chosen as: 

f ( x) a bx cx dx ex 

2 3 4 

= + + + + (6) 

All the parameters of that equation can be ensured by 

the least square method. 

ϕ( 

a , a , a , a , a ) = ( y − y ) 

0 1 2 3 4 

N 

∑ 

i= 

1 

N 

∑ 

i= 

1 

* 2 

i i 

= ( y−a−ax−ax −ax−ax i 0 1 1i 2 2i 3 3i 4 4i 

To make ϕ(a 0,a 1,a 2,a 3,a 4) 

minimal, the partial 

derivatives of the equation ϕ (a 0,a 1,a 2,a 3,a 4) 

to 

a, �,a 

0 

4 

should be 0. That is: 


) 

2 

(7) 

N ⎧ ∂ϕ 

⎪ =−2 ∑( 

yi −a0 −ax 1 1i −ax 2 2i − ax 3 3i - ax 4 4i) 

= 0 

a0 

i= 

1 

⎪ 

∂ 

⎪ N ∂ ϕ 

⎪ = −2 ∑( 

yi −a0 −ax 1 1i −ax 2 2i − ax 3 3i - ax 4 4i) x1i 

= 0 

⎨∂a1 

i= 

1 

⎪ 

�� 

⎪ 

N ⎪ ∂ ϕ 

⎪ = −2 ∑( 

yi −a0 −ax 1 1i −ax 2 2i −ax 

3 3i - ax 4 4i) x4i 

= 0 

⎩∂a4 

i= 

1 

(8) 

After the readjustment of formula (8), we get the 

equation: 

N N N N 

⎧ 

⎪Na0 

+ ∑x1ia1 + ∑x2ia2 + � + ∑x4ia4 = ∑yi 

i= 1 i= 1 i= 1 i= 1 

⎪ 

N N N N N 

⎪ 

⎪∑x1ia1+ 

∑x1ix1ia1+ ∑x1ix2ia2+ � + ∑x1ix4ia4= ∑x1iyi 

⎨ i= 1 i= 1 i= 1 i= 1 i= 1 

⎪�� 

⎪ 

N N N N N 

⎪ 

⎪∑x 

a + ∑x x a + ∑x x a + � + ∑x x a = ∑x 

y 

⎩ 

(9) 

Formula (9) is called normal equation, which is linear 

equations about a,a, 0 1 �,a4, 

and it could be expressed 

with matrix as: 

4i 1 4i 1i 1 4i 2i 2 4i 4i 4 4i i 

i= 1 i= 1 i= 1 i= 1 i= 1 

⎡ 

⎢ 

⎢ 

⎢ 

⎢ 

⎢ 

� 

⎢ 

⎢ 

⎣ 

� � �� 

⎤ 

⎥ 

⎥ 

⎥ 

⎥ 

⎥ 

⎥ 

⎥ 

⎥ 

⎥⎦ ⎢�⎥ ⎢ ⎥ 

a 

⎡ 

⎢ 

⎢ 

⎢ 

⎢ 

⎢ 

⎢�� 

⎢ 

⎢ 

⎢⎣ 

(10) 

N 

N 

∑ x1i N 

∑ x2i 

N 

�∑ 

x4i N 

∑ yi 

N 

∑ x1i 

i= 1 

i= 1 

N 

2 

∑ x1i i= 1 

i= 1 

� 

i= 1 

N 

�∑ 

x1ix4i i= 1 

⎡a0⎤ ⎢ ⎥ 

a1 

⋅ ⎢ ⎥ = 

i= 1 

N 

∑ x1iyi i= 1 

N 

∑ x 4i 

i= 1 

N 

∑ x4ix1i � 

N 

2 

�∑ 

x4i ⎣ 4 ⎦ N 

∑ x4iyi 

i= 1 i= 1 i= 1 

It can be certified that the coefficient matrix of 

equations (10) is a symmetric positive definite matrix, so 

it exists a unique solution. According to the test data, 

Matlab program is used to derive the coefficients. The 

approximate formula is 

2 3 

4 

f ( x) = 558900 − 103800x+ 69280x − 1977x + 20.41x 

(11) 

Generally, the accurate of the fitting depends on the 

polynomial degree. But if the polynomial degree is too 

high, it requires increasing calculations and causes 

severer oscillations at the two ends of the resultant curve. 

The parameter u of correlative peak interval formula 

chosen to the test data is equal to 3.617. To verify its 

generality, 8 different values of u have been tested and 

analyzed. The precision-u-correlative peak interval 

diagram is shown as Figure 8. 

From this figure, we can see that if the parameter u is 

in the chaotic regions and the computational accuracy is 

higher than 35 the logistic map has a long-correlative 

peak interval. Therefore a good parameter region can 

obtain long-correlative peak interval, which is desirable 

⎤ 

⎥ 

⎥ 

⎥ 

⎥ 

⎥ 

⎥ 

⎥ 

⎥ 

⎥⎦


by many cryptographic applications. 

Figure 8. Diagram of precision-u-correlative peak interval 

The chaotic sequence is non-cycle theoretically, but 

under the situation of finite precision and approximation 

computing, chaotic sequence will become a cycle 

sequence. Another more easily observed phenomenon is 

that after discretization chaos has strong correlation, 

which has a greater threat to the period and will directly 

affects the confidential strength. 

With the relationship between the correlative peak 

interval and the computational precision, we could get 

some long periods sequences within the available limited 

hardware and computing resources. A suitable parameter 

u, precisions and correlative peak interval can effectively 

generate pseudorandom output sequences, which is 

acceptable by encryption devices and secure 

communication systems. 

Ⅴ. SUMMARY 

This paper has implemented and analyzed a new 

design of a logistic-map module by calculating its 

Lyapunov exponent to determine the suitable regions of 

the chaotic parameter u. With the respective tests using 

different initial values, the values of parameter u and the 

computational precisions, extensive experiments have 

been carried out. The correlative peak interval has also 

been tested and analyzed with a fixed initial value, 

alterative parameter u, and different precisions. An 

approximate formula for calculating the correlative peak 

interval is provided. Some conditions for correlative 

peak interval affected by the chaotic sequences are also 

derived. Moreover, through fitting the relation between 

the correlative peak interval and the computational 

precisions, it is possible to define the best regions for 

both correlative peak interval and precisions in the 

design of key sequence generators based on FPGA. 

Finally, a diagram that presents the relationship among 

precision, parameter u and correlative peak interval has 

been generated for analysis. The basic theory and 

methods have a significant implication on the statistical 

analysis and practical applications of the digital chaotic 

sequences. 



This work is supported by the National Natural 

Science Foundation of China (no. 60672011). 

REFERENCES 

[1] F. Dachselt and W. Schwarz. Chaos and cryptography. 

IEEE Transactions on Circuits and Systems – Part I, 2001, 

48(12): 1498-1509. 

[2] M. S. Baptista. Cryptography with chaos. Physics Letters 

A. 1998, 240: 50-54. 

[3] Wei Xiang and Fangqi Chen, Sliding Mode Control 

Strategies for the Hyperchaotic MCK System, ICIC 

Express Letters, vol.3, no.3(A), pp.283-288, 2009. 

[4] Yang Yang, Xiangzhong Bai, Zhenguo Tian and Huan 

Wang, Chaotic Motion in Fluid-solid Interaction 

Problem of the Elastic Cylinder, ICIC Express Letters, 

vol.3, no.3 (A), pp.439-444, 2009. 

[5] Junyan Yi, Gang Yang, Shangce Gao and Zheng Tang, 

Transiently Chaotic Neural Network Based on Switched 

Cooling and Its Application to Maximum Clique 

Problem, International Journal of Innovative Computing, 

Information and Control, vol.5, no.6, pp.1569-1586, 

2009. 

[6] C. Y. Chee and D. L. Xu. Chaotic encryption using 

discrete-time synchronous chaos. Physics Letters A. 

2006, 348(3-6): 284-292. 

[7] Q. Ding, Y. Zu, F. Zang, and X. Peng. Discrete chaotic 

circuit and the property analysis of output sequence. 

International Symposium on Communications and 

Information Technologies. Beijing, China, Oct. 12-14, 

2005, 2: 1009-1012. 

[8] T. Gao and Z. Chen. Image encryption based on a new 

total shuffling algorithm. Chaos, Solitons and Fractals. 

2008, 38: 213-220. 

[9] S. Pan, J. Huang and G. Wang. Modern DSP Technology. 

Xi’an Electronic and Science University Press, Xi’an, 

China. 2003, 57-91. 

[10] W. Zheng. Random Signal Analysis. Harbin Industrial 

Univ ersity Press, Harbin, China, 1999, 66-76. 

[11] K. Li, Y. C. Soh and C. Zhang. A frequency aliasing 

approach to chaos-based cryptosystems. IEEE 

Transactions on Circuits and Systems – Part I. 2004, 

51(12): 2470-2475. 

[12] H. Zhou and X. Ling. Sequence m-perturbation 

implement of finite precision chaotic system. Electronic 

Journals. 1997, 25(7):95-97. 

[13] H. Zhou, J. Yu and X. Ling. The design of chaotic feed 

forward type stream cipher. Electronic Journals. 1998, 

26(1): 98-101. 

[14] R. Huang. Chaos and its Applications. Wuhan University 

Press, Wuhan, China. 2003, 128-138. 

[15] S. Pan, J. Huang and G. Wang. Modern DSP 

Technology. Xi’an Electronic and Science University 

Press, Xi’an, China. 2003, 57-91 

[16] W. Zheng. Random Signal Analysis. Harbin Industrial 

University Press, Harbin, China, 1999, 66-76. 

[17] Q. Ding, J. Pang, J. Fang, and X. Peng. Designing of 

chaotic system output sequence circuit based on FPGA 

and its possible applications in network encryption cards. 

International Journal of Innovative Computing, 

Information and Control, 2007, 3(2): 449-456. 

[18] F. Belkhouche, U. Qidwai, I. Gokcen and D. Joachim. 

Binary image transformation using two-dimensional 

chaotic maps. IEEE International Conference on Pattern 

Recognition. Cambridge, UK, 2004, 4: 823-826. 

[19] C. Y. Chee and D. L. Xu. Chaotic encryption using 

discrete-time synchronous chaos. Physics Letters A.


2006, 348(3-6): 284-292. 

[20] M. I. Sobhy, A-E. R. Shehata. Chaotic algorithms for 

data encryption. IEEE International Conference on 

Acoustics, Speech, and Signal Processing. Salt Lake 

City, UT, USA. 2001, 2: 997-1000. 

Qun Ding, born at Harbin, Heilongjiang 

Province in 1957. She got instrument 

and technology science doctor’s degree 

at Harbin Institute of Technology in 

2007. Now she is the dean of electronic 

engineering college, doctoral director, 

the director of Heilongjiang electronic 

engineering key laboratory, the director 

of Heilongjiang signal and information 

key laboratory, the councilman of 

Heilongjiang communication institute, the panel judge of 

national 863 programs and national nature science fund. Her 

major research field is the security of information and 

encryption communication. 


Lu Wang, born at Harbin, Heilongjiang 

Province in 1984. She got her bachelor’s 

degree in Heilongjiang University at 

Harbin, Heilongjiang Province, China in 

2008. The major of the undergraduate is 

communication engineering. Now she is 

at the graduate stage at Heilongjiang 

University at Harbin, Heilongjiang 

Province, China. The research field is 

the encryption of communication. And 

the major is communication engineering. 

Guanrong Chen,got applied 

mathematics doctor’ degree at Texas 

A&M University in 1987. Now he is the 

chair professor of Hong Kong City 

University and the Changjiang chair 

professor of Peking University. His 

main research field is nonlinear system 

of control theory, dynamics analysis 

research and its application in 

complicated network.


An Energy Efficient Dynamic Clustering Protocol 

based on Weight in Wireless Sensor Networks 

Ming Zhang 1,2 

1 Department of Software,Nanjing University of posts & Telecommunications Nanjing,China. 

Email: lyg690916@163.com 

Suoping Wang 1 

2 Department of Electronic Engineering, Huaihai Institute of Technology, Lian yungang, China 

Email: wangsp@njupt.edu.cn 

Abstract—Nodes in most wireless sensor networks (WSNs) 

are powered by batteries with limited energy. Prolonging 

network lifetime and saving energy are two critical issues 

for WSNs. Clustering is an effective technique to improve 

the energy efficiency and prolong network lifetime of 

wireless sensor networks. In this paper, an energy efficient 

dynamic clustering protocol (EEDCP) based on weight for 

wireless sensor networks is proposed, which is able to 

dramatically prolong network lifetime and save energy. In 

the EEDCP, we introduce the typical energy model to 

compute energy consumption, virtual grid technology to 

construct cluster and a long sleeping state to reduce energy 

consumption. In addition, we use the value of weight to 

measure the size of residual energy instead of voting, which 

can significant reduce the voting times and the number of 

transmitting information. Further, simulation experiments 

are conducted to compare the EEDCP with some wellknown 

clustering algorithms and simulation results show 

that the proposed method overcomes the existing methods in 

the aspects of energy consumption and network lifetime in 

wireless sensor networks. 

Index Terms—Wireless sensor networks (WSNs), Energy 

efficient, Network lifetime, Dynamic clustering, Weight 


Wireless sensor networks (WSNs) have been 

blooming recently, which are being widely used in 

various areas such as reconnaissance, disaster relief, 

intelligent transportation, surveillance, environmental 

monitoring, healthcare, target tracking, and more. WSNs 

are extremely useful to collect information in harsh or 

hostile environment. A WSN has two important and 

interesting characteristics that are different from 

traditional wireless networks [1]. First, after the event 

occurs, multiple sensors nodes (denoted as data source 

nodes) around this event will sense the event, and then 

send the data back to one sensor node (denoted as sink 

node). Hence, communication mode in WSN occurs from 

multiple data source nodes to one data sink node. This is 

a type of multipoint-to-point, rather than the traditional 

Corresponding author:Ming Zhang,E-mail:lyg690916@163.com 


doi:10.4304/jnw.6.7.1057-1064 

point-to-multipoint communication in WSNs[2] . 

A wireless sensor network (WSN) is composed of a 

large number of sensor nodes that are densely deployed 

near an area of interest and are connected by a wireless 

interface. Since each sensor is limited in terms of 

processing capability, wireless bandwidth, battery power 

and storage space, in most applications, it is impossible to 

replenish power resources, a major constraint of WSN 

lifetime is energy consumption. Energy savings 

optimization is thus a major challenge for the success of 

WSNs. Typical tasks of a sensor node in a sensor network 

are to collect data, perform data aggregation, and then 

transmit data. Among these tasks, monitoring and 

transmitting data require much more energy than 

processing it [3]. Therefore, in wireless sensor network, a 

significant focus has been put on increasing energy 

efficiency [4]. Generally, there are two basic approaches 

to the problem of saving energy in WSN. The first one is 

scheduling some sensor nodes to go into an active mode 

while enabling the other sensor nodes to go into a lowpower 

sleep mode [5,6]. The second approach is to select 

the optimization routing algorithm, eliminating redundant 

energy consumption. 

Hence, proper energy efficient dynamic routing 

protocols should be designed to increase the lifetime of 

the network greatly. In this paper, an energy efficient 

dynamic clustering protocol (EEDCP) based on weight is 

proposed for wireless micro sensor networks to facilitate 

the achievement of low energy dissipation. From the 

simulation results, it is illustrated that the EEDCP 

achieves an order of magnitude increase in system 

lifetime when compared to the general – purpose 

approaches. Moreover, for a given quality, the overall 

residual is reduced by an order of magnitude. 

The rest of this paper is organized as follows. Section 

II gives the detailed related work done. Section III 

presents the system model for our architecture, such as 

network model, energy model and node state transition 

model. Section IV gives the detailed description of the 

algorithm and structure. Section V gives the experimental 

results and section Ⅵconcludes the paper. 

II. RELATED WORK


There are many energy-saving routing algorithms used 

in wireless sensor networks, and new ideas for routing are 

announced in recent years. In this section, we review 

some of the most effective algorithm. 

In low-energy adaptive clustering hierarchy LEACH) 

[7], the authors discuss an energy efficient algorithm. 

Various algorithms developed after that is based on this 

algorithm. In order to determine the cluster head, LEACH 

uses randomization technique Gossiping [8] is the 

improvements of flooding algorithm (Flooding), it can 

effective resolve the implosion and information overload 

problem which lead to energy loss, but it can not solve 

some of the data overlap and too much delay, but it can 

not balance the energy of nodes. Hybrid energy-efficient 

distributed clustering (HEED)[19] is based on LEACH 

thinking, the important difference is the choice of cluster 

head and cluster head formation. In PEGASIS (Power- 

Efficient Gathering in Sensor Information Systems), 

author tried to foster the past technique [10]. This new 

mechanism is a chain-based power efficient protocol 

based on LEACH [11]. It assumes that each node must 

know location information about all other nodes at first. 

PEGASIS starts with the farthest node from the base 

station. The chain can be constructed easily by using a 

greedy algorithm. The chain leader aggregates data and 

forwards it to the base station. In order to balance the 

overhead involved in communication between the chain 

leader and the base station, each node in the chain takes 

turn to be the leader. 

A clustering-based routing protocol called base station 

controlled dynamic clustering protocol (BCDCP)[12], 

which utilizes a high energy base station to set up cluster 

heads and perform other energy-intensive tasks, can 

noticeably enhance the lifetime of a network. United 

voting dynamic cluster routing algorithm based on 

residual-energy in wireless sensor networks (UVDC)[13], 

which periodically selected cluster head according to 

residual energy among the nodes located in the event area, 

so the voting cost of UVDC is gigantic and the large 

redundant nodes will waste limited energy. Sensor 

protocols for information via negotiation (SPIN)[14] is 

also improved the flooding algorithm, before transferring 

data, it only transmit data to needed neighbor nodes 

which using meta-data to reduce redundant information to 

save energy consumption. Directed diffusion (DD)[11] 

periodic automatic forms the enhanced path, because of 

node energy and topology changes, the enhanced path 

will be different in different period, the most of data from 

the source to the cluster is transmitted by the enhanced 

path, thus reduce the energy consumption of nonenhanced 

nodes. And GBR (Gradient-Based Routing) is 

also proposed as a variant of directed diffusion [15]. The 

key idea in GBR is to memorize the number of hops 

when the interest in diffused through the whole network. 

In GBR, three different data dissemination techniques 

have been discussed [11] (i) Stochastic Scheme, where a 

node picks one gradient at random when there are two or 

more next hops that have the same gradient, (ii) Energybased 

scheme, where a node increases its height when its 

energy drops below a certain threshold, so that other 


sensors are discouraged from sending data to that node, 

and (iii) Stream-based scheme, where new streams are 

not routed through nodes that are currently part of the 

path of other streams. The main objective of these 

schemes is to obtain a balanced distribution of the traffic 

in the network, thus increasing the network lifetime. 

Threshold sensitive energy efficient sensor network 

protocol (TEEN)[16] is designed for responsive 

applications, it determine whether to send data by setting 

up a reasonable soft and hard threshold to compare with 

the monitoring data .It only transmit the interest 

information to users to effectively reduces the network 

traffic and thus reduce network energy consumption. 

Guojun Wang, et al., have proposed a local updatebased 

routing protocol in WSNs with a mobile sink [17]. 

The protocol proposed by the authors saves the energy for 

sensor networks and makes the sink keep continuous 

communications to sensors by confining the destination 

area into a local area for updating the sink location 

information as the sink moves. Hayoung Oh and Kijoon 

Chae have presented a sensor routing scheme [18], EESR 

(Energy-Efficient Sensor Routing) that provides energyefficient 

data delivery from sensors to the base station. 

Their scheme divides the area into sectors and locates a 

manager node to each sector. 

Besides these algorithms mentioned above, there exist 

several other algorithms [19], such as: Soro et al. [20] 

proposed an unequal clustering size model for network 

organization, which can lead to more uniform energy 

dissipation among cluster head nodes, thus increasing 

network lifetime. Ye et al. [21] proposed a clustering 

algorithm, which achieves good cluster head distribution 

with no iteration and introduces a weighted function for 

the plain node to make a decision for joining a proper 

cluster. 

Ⅲ.SYSTEM MODEL 

A. Network model 

In this paper, we consider the wireless sensor networks 

where N nodes in field A are homogenous and energy 

constrained and the sensor network has the following 

properties [22]: 

(1) This network is a static densely deployed network. 

It means a large number of sensor nodes are densely 

deployed in a two-dimensional geographic space, forming 

a network and these nodes do not move any more after 

deployment. 

(2) There exists only one Sink node, which is deployed 

at a fixed place outside the WSNS. 

(3) The energy of sensor nodes cannot be recharged. It 

means sensor node will die if its energy be exhausted. 

(4) Sensor nodes are location-aware, i.e. sensor node 

can get its location information through other 

mechanisms such as GPS or position algorithms (in order 

to describe the position of node uses (Xi,Yj) ,the Sink 

node as (Xsink,Ysink)). 

(5) The radio power can be controlled, i.e., a node can 

vary its transmission power depending on the distance to 

the receiver [23].


In table 1,we give the symbols definition in this paper. 

TABLE 1. 

Symbols 

NOTATION 

Definition 

N The set of sensor nodes 

x,y Coordinates of the node 

di,j The distance of node i and node j 

Etx The consumption energy of node I transmitted a 

packet. 

Erx The consumption energy of node I receive a 

packet 

Er(i) The residual energy of node i 

Ep The consumption of processing in cluster head 

Ti Listening timer 

Ts Sleeping timer 

Tj Sensing timer 

r(i) The number of active node I receives voting 

information 

s(i) The number of active node I sends voting 

information 

v(t) The total number of sending and receiving 

voting information in head 

Gi The ith virtual grid in a cluster 

k The size of a packet 

Weight The value of residual energy divided by Er(i)for 

each active node 

M,N The number of virtual grid in a cluster 

First, we use the virtual grid ideas to divide the field A 

into many same square, namely, there are many clusters, 

and each node can directly communicate with other nodes 

in a cluster. Then the cluster was divided into M×N small 

area (the value of M,N are determined by the cluster’s 

size, assume that there are M×N grid in a cluster, each 

grid is named Gk(k=1.. M×N). 

Figure 1. Virtual grid model 

Fig.1 shows the virtual grid ideas, in order to 

conveniently describe, we suppose that the value of M 

and N are equal to three, each small square call as a 

virtual grid, nodes are randomly distributed into this 

virtual grid, such as CH1 has 9 virtual grid, namely, G1, 

G2, G3,G4,G5,G6,G7,G8 and G9,we suppose virtual grid 

G5 as a cluster head grid, and the red pentagram as the 

cluster head., for arbitrary adjacent virtual grid G1 and 

G2,each node in G1 can communicate with all nodes in 


G2, and vice versa. In a cluster, the red dot as the 

working node in each grid and each node can 

communicate with cluster head, and we suppose the 

number of simultaneous working node is one in a virtual 

grid (red dot), other nodes are sleeping (black dot). In 

order to guarantee the network normal working and 

prolong network lifetime, one sleeping node in a virtual 

grid will be awaken at the right time so as to instead of 

the energy-exhausted node or disabled node [24]. 

B. Energy model 

We adopt a simplified power model of radio communication 

in document [25], namely, in order to send a k-bit 

packet information and the sending distance is di,j, the 

sending energy consumption is 

ETx( k, d) = Eelec × k + ε amp × k× d× d 

(1) 

The distance of node I and node j is di,j: 

2 

2 

| d i, j | = ( xi − xj) + ( yi − yj) 

The receiving energy consumption is 

(2) 

ERx( k) = Eelec × k 

(3) 

Where Eelec is the energy/bit consumed by the sender 

amp 

and receiver electronics, J/bit, Eelec=50nJj/bit,. ε is 

amp 

the J/(bit × m2), ε =100pJ/bit/m2.we commonly 

assume that the sending distance and d2 is directly 

proportional for shorter distance, while the sending 

distance and d4 is directly proportional for longer 

distance, so we can see the directly sending to long 

distance is consumed more energy than multi-hop 

sending. 

But the differentiation from the document [26], we 

consider the processing consumption in order to 

proximity real scene, the energy consumption of cluster 

head is Ep 

m 

E ( k , m ) = 1 / 3 × E × k 

∑ 

P elec i 

i = 1 

So the residual energy of cluster head is: 

n1 

n2 

Er() i = Er() i − ∑ ETx( kn, dn) −∑ERx( kl) −Ep( 

k, m), n1, n2∈N n = 1 

l= 

1 

(4) 

(5) 

Where n1,n2 are the cluster head respectively sending 

and receiving times before time Ti. 

The residual energy of ordinary node is: 

Er() i = Er() i −ETx( kn, dn) −ERx( 

kl) 

(6) 

C. Node state transition model 

The energy dissipation in wireless sensor networks has 

three models: sensor model; procession model; wireless 

radio model [27]. In order to maximum lifetime and 

minimum routing, nodes in the EE-MLMR have various 

operation modes with different levels of activation and, 

thus, different levels of energy consumption. We put 

forward the new state conversion model which have flag


of valve which depend on the EPGR state model. In this 

model, each node has six operation modes [28]: mode 1: 

sleeping-sensing off and radio off; mode 2: sensing - 

sensing on and radio off; mode 3: receiving sensing on 

and radio receiving; mode 4: transmitting -sensing on and 

radio transmitting; mode 5: listening - sensing on; mode 

6:long sleeping- sensing off and radio off forever, no 

responding. 

Figure 2. The transition model os sensor nodes 

Where Ts is a sleeping timer, Ti is listening timer, Tj is 

sensing timer. 

The Fig. 2 shows the ‘‘commands’’ performed along 

the path (state transition) between states [29]. It means 

that whenever a node changes its state-based energy 

dissipation model to current state it performs tests and 

actions until the new state is reached. “sleeping’’ 

determines whether the node will sleep or not; The 

‘‘receiving’’ test depends also on the characteristics of 

the event. Its value is influenced by the degree of 

cooperation needed by the application. The ‘‘sensing’’ 

test is called only if there is no event in the area of the 

node. If no event happens, this test will depend on the 

degree of coverage needed by the application. In the 

“transmitting” state, if flag=0,then sensing off and radio 

off and node convert to long sleeping; if flag=1,then 

convert to the transmitting. The “listening” test 

determines whether a new sensing event is present; the 

long sleeping denotes the node never respond any event. 

“Timer’’ is an action that starts a timer. The outcome of 

each test depends on a probabilistic parameter associated 

with the test. These transitions try to capture the behavior 

of a sensor node, specially in terms of energy 

consumption. 

Ⅳ.ENERGY EFIICIENT DYNAMIC CLUSTERING 

PROTOCOL BASED ON WEIGHT 

At any time, only one node within a virtual grid stays 

active to be a coordinator, while the others fall into 

sleeping mode. Doing this significantly reduces the 

energy consumption because nodes in the idle state spend 

much more energy as compared with the sleeping state. 

In our protocol, we use weight as the selection criteria 

for new cluster head, when the residual energy of the 

cluster head is lower than threshold, the EEDCP will be 

implemented. First, it compute the total residual 


energy(Et) of all active nodes from the cluster head 

member table, then respectively compute the weight of all 

active nodes as shown in formula (7). Second, we select 

the minimal weight active node as the new cluster head 

and inform all member nodes and the cluster head 

neighbor. All received information nodes will reply 

related information and update its table. Lastly, the old 

cluster head will select a new active node replace him and 

then goto long sleeping. 

weight( i) 

= 

k 

∑ (7) 

j = 1 

E r ( j) 

E r ( i ) 

Where k is the number of active nodes in the 

cluster,Er(i) is the residual energy of active node. 

The energy efficient dynamic voting cluster (EEDCP) 

based on weight has four steps: Initialization, active node 

selection, dynamic clustering based on weight phase, 

sensing and sending. When the residual energy of cluster 

head is lower than threshold, dynamic clustering based on 

weight will happens, namely, cluster head will initiate 

clustering process from the active node. if new cluster 

head is come into being, clustering is formed. The 

detailed process described in the followings: 

Step1:Initialization: the whole area was divided into 

many same squares, namely, there are many clusters, and 

each node can directly communicate with other nodes in a 

cluster. Then the cluster was divided into M×N small 

area (the value of M,N are determine by the cluster’s size, 

there are M×N grid in a cluster, each grid is named 

Gk(k=1.. M × N). The first cluster head is randomly 

selected from the active nodes and each active node has a 

neighbor table (as shown in table 2.)to record its member 

information and all directly connect active node ID and 

energy information, then cluster head will send 

information to other cluster head, so as to form a cluster 

neighbor table(as shown in table 3.). 

Step2:Sensing and sending: when there is a event, the 

active node will collect event information, then compute 

its residual energy, if Er is lower than threshold then goto 

active node selection; else it will send event information 

and the residual energy to the cluster head. 

Step3:Active node selection: for each grid, if its active 

node residual energy is lower than threshold, the active 

node will select a new node as the new active node from 

its neighbor table which has maximum residual energy, 

and send the new active node ID and energy information 

to cluster head and its directly connect active nodes. 

Step4:Dynamic clustering based on weight: Cluster 

head is responsible for receiving data, gathering data, 

sending data to next top, computing its residual energy 

and maintaining its all table. After some time, if the 

residual energy of cluster head is lower than threshold, 

the cluster head will implement the dynamic clustering 

process. First, it compute the total residual energy of all 

active nodes and the weight (is equal to the total residual 

energy of all active nodes in the cluster divided by the 

residual energy of each active node.) of each active nodes; 

Second, old head will use active node selection and select 

a new active node and update its member table; Third, old 

head will select a minimal weight active node as the new


cluster, if there are more than one active node has the 

same weight, it will randomly select one as the new 

cluster head, then, the cluster head will inform its 

member nodes and neighbor nodes about the new cluster 

head, such as: ID,(xi,yi) etc, all active node in this cluster 

and the neighbor nodes received this information will 

update its table and record the new cluster head and 

replay a ACK to old head; Fourth, the old head will send 

the information to the new cluster head, including its 

member table and neighbor table. Lastly, the old cluster 

head goes to sleep. 

TABLE 2. 

THE ACTIVE NODE STRUCTURE 

Name Propetries 

ID Identification of nodes 

xi,yi Position od node i 

Er(i) Residual energy of node i 

flag 0-cluster head;1-node itself;2-member node;3neighbor 

node 

state 0-active;1-sleeping 

TABLE 3. 

THE CLUSTER HEAD STRUCTURE 

Name Properties 

ID Identification of nodes 

xi,yi Position od node i 

Er(i) Residual energy of node i 

flag 0-cluster head;1-node itself;2-member node;3neighbor 

node 

weight The value of node residual energy divided by total 

residual energy 

state 0-active;1-sleeping 

For example,at a time, if cluster head (such as G5) 

residual energy is lower than threshold in CH1 and other 

active node residual energy is show as Fig.3. 

Figure 3. The residual energy of active nodes in cluster CH1 

Fig.4 is the voting relationship and the voting 

results .In the Fig.4,s(i) is the times of active node i 

sending information and r(i) is the times of active node i 

receiving information, we can see that G2 has the 

maximum ballot, namely, r(G2)=5(the residual energy of 

G2 is 0.8 and is the maximum),so the new cluster head is 

G2.but in this voting process, the total times of voting is 

20,namely, all active nodes have sent 20 times voting 

information and the receiving voting information times is 

20, the total times is 4o, namely v(t)=40,in the voting 


process, each active node has consumed sending energy, 

receiving energy and computing energy, so the total 

consumption energy is large. 

At a time,the active nodes residual energy in the 

cluster CH1 as shown in Fig.5,the result is shown in Fig.6 

using volting algorithm.From Fig.6,we can see that there 

are two active nodes G2 and G8 have the same ballot, 

namely, r(G2) and r(G8) are equal to 5 ,in this case, 

generally randomly select one of the active nodes as the 

new cluster head which the maximum residual energy 

active node may not be the new cluster head, such as if 

we select G2 as the new cluster, we find the maximum 

residual energy active node is G8,so it can not ensure the 

residual energy active node must be as the new cluster 

head which lead to a decline in the network lifetime. 

Figure 4. The voting relation and the voting times 

Figure 5. The residual energy of active nodes in cluster CHi 

But using our proposed method, we can use minimal 

times to generate the new cluster head, significantly save 

energy and ensure the residual energy active node as the 

new cluster head, for example, for the same case as 

shown in Fig.3,the total computing times is 10 times (one 

is compute the total energy, the others are compute the 

weight of each active node)and each active nodes does 

not require any messages to send and receive in this 

process, the old head can choosees a new cluster head 

G2 as shown in Fig.7,and we can see that the w(2) is the 

minimum, namely, G2 is the maximum energy active 

node. At the same circumstances in Fig.5,using our 

method, we can see from the Fig.8 that the minimal 

weight active node(G8) is the only new cluster head , so


we can reduce the energy consumption in the new cluster 

head generation, always select the maximum active node 

as the new head and extend the network lifetime. 

Figure 6. The voting relation and the times of sending and 

receiving for each active node. 

Figure 7. The weight of each active node using EEDCP 

Figure 8. The weight of each active node using EEDCP 


Ⅴ.SIMULATION RESULTS 

A. Simulation Parameters: 

We have implemented our proposed protocol in NS-2 

(ver. 2.31). We considered a 600 node random network 

deployed in an area of 360 X 360 m. Initially the nodes 

are placed randomly in the specified area. The only Sink 

node is assumed to be situated 100 meters away from the 

above specified area. At the same time, we considered 

specified area is divided into 90 X 90 m square area 

called cluster and each cluster is divided into 30 X 30 m 

area called virtual grid. Obviously, the first set of cluster 

heads are taken randomly. The initial energy of all the 

nodes assumed as 5 joules. The radio range is varies from 

30m to 120m.Each data packet has 64 bytes, and the 

others are 36 bytes long. Summary of parameters and 

defined values are shown in Table 4. 

TABLE 4. 

SIMULATION PARAMETERS AND VALUES 

Simulation parameters value 

N(total nodes) 600nodes 

A(network size) 360×360 m 

Cluster size 90×90 m 

Virtual grid size 30×30 m 

Number of sink 1 

Eelec 50nJ/bit 

ε 

amp 

0.0013pJ/bit/m 2 

Data packet size 64 bytes 

Other packet size 32 bytes 

Simulation times 150 seconds 

Threshold energy 0.2w 

E(i i i l ) 5J l 

B. Experimental results and analysis 

From the diagram of Fig.9, we can see that there are 

considerable differences on the average energy 

consumption among the three algorithms. EEDCP has the 

minimum energy consumption, and with the increase of 

number nodes in each cluster, the consumption is slowly 

increase, For UVDC, because of there are large redundant 

nodes and the consumption of voting is gigantic, so the 

average energy consumption is rapidly increase. For 

LEACH, because of randomly rotating the role of a 

cluster head among all the nodes, the average energy 

consumption is approximate linear increase. 

. 

Figure 9. The average energy consumption of LEACH, UVDC 

and EEDCP


In this second series of experiments, we compare the 

three energy efficient clustering algorithm LEACH, 

UVDC and EEDCP with regard to the network lifetime, 

when the number of nodes in a cluster from 15 to 40. The 

LEACH algorithm that does never take energy into 

account and always randomly rotating the role of a 

cluster head among all the nodes. Simulation results are 

illustrated in Figure 10, assuming that the initial energy 

of the nodes is uniformly as 5J 

Figure 10. Comparison of network lifetime with LEACH, UVDC 

and EEDCP 

As expected, LEACH provides the smallest network 

lifetime. This shows that the random selection of the 

cluster head is not sufficient to save energy. UVDC 

provides better results than LEACH, but in the voting 

cluster head process, UVDC consumed large energy in 

voting and sending information. The main conclusion of 

these experiments is that EEDCP significantly 

outperforms LEACH and UVDC whatever the number of 

nodes in a cluster. Moreover, EEDCP prolongs the 

network lifetime of 21% compared with LEACH for a 

different number of nodes in a cluster. Notice that in the 

same conditions, UVDC prolongs the network lifetime of 

only 6%. 

Ⅵ.CONCLUSIONS In WSNs, it is significant to prolong network lifetime 

so that more data can be collected by the sensor(s) to 

transmit to sink node. It is well known that, efficiently 

use of energy is critical for network lifetime. Although 

some routing algorithms like voting dynamic cluster 

routing algorithm based on residual-energy (UVDC) can 

dynamic clustering, they usually place too heavy burden 

of voting information in cluster which consumed large 

valuable energy. 

In this paper, an energy efficient dynamic clustering 

protocol (EEDCP) based on weight for wireless sensor 

networks is proposed, which is bale to dramatically 

prolong network lifetime and save energy. In the EEDCP, 

we introduce the typical energy model to compute energy 

consume, virtual grid technology to construct the cluster 

and a long sleeping state to reduce energy consumption. 

In addition, we use the value of weight to measure the 

size of residual energy instead of voting ,which can 

significant reduce the voting times and the number of 

transmitting information. Further, simulation experiments 

are conducted to compare the EEDCP with some well- 


known clustering algorithms and simulation results show 

that the proposed methods overcomes the existing 

methods in the aspects of energy consumption and 

network lifetime in wireless sensor networks. 


We acknowledge the support of University Natural 

Science Foundation of Jiangsu Province (No. 

08KJD520003),we sincerely thank the anonymous 

reviewers for their constructive comments and 

suggestions. 

REFERENCES 

[1] Frank Yeong-Sung Lin , Hong-Hsu Yen and Shu-Ping Lin, 

A Novel Energy-Efficient MAC Aware Data Aggregation 

Routing in Wireless Sensor Networks, Sensors,2009(9), 

1518-1533. 

[2] Krishnamachari, B.; Estrin, D.; Wicker, S. Modeling Data- 

Centric Routing in Wireless Sensor Networks. USC 

Computer Engineering Technical Report CENG 02-14, 

2002. 

[3] Pottie, G.J.; Kaiser, W.J. Wireless integrated network 

sensors. Commun. ACM 2000, 43, 51-58. 

[4] Qin Wang, WoodwardYang, “Energy Consumption Model 

for Power Management in Wireless Sensor Networks”. 

IEEE Communications Society,2007.pp:142-151. 

[5] Cardei, M.; Du, D.Z. Improving wireless sensor network 

lifetime through power aware organization. Wirel. Netw. 

2005, 11, 333-340. 

[6] Carle, J.; Simplot, D. Energy-efficient area monitoring by 

sensor networks. IEEE Comput. 2004, 37, 40-46. 

[7] Heinzelman W, Chandrakasan A, Balakrishnan H. Energy 

efficient communication protocol for wireless microsensor 

networks Proceedings of the 33rd Annual Hawaill 

International Conferenc On System Sciences, Jan 4-7, 

2000, Maui, HI, USA. Los Alamitos CA, USA: IEEE 

Computer Society, 2000: 223. 

[8] Santi P , Simon J . Silence is Golden with Hi}gh 

Probability : Maintaining a Connected Backbone in 

Wireless Sensor Networks.In Proceeding of 1st European 

Workshop on WirelessSensor Networks(EWSN 2004), 

Jan 2004,106—121. 

[9] Sungju Lee, Jangsoo Lee , Hongjoong Sin,et. Al, An 

Energy-Efficient Distributed Unequal Clustering Protocol 

for Wireless Sensor Networks, Proceedings of World 

Academy of Science,Engineering and Technology volume 

36 december 2008 issn 2070-3740. 

[10] Manjeshwar A , Agrawal D . TEEN : A Protocol for 

Enhanced Efficiency in Wireless Sensor Networks . In 

Proceeding of the lth International Workshop on Parallel 

and Distributed Computing Issues in Wireless Networks 

and Mobile Computing’01.2001:23-27. 

[11] Intanagonwiwat C,Govindan R,Estdn D,et a1.Directed 

Diffusion for Wireless Sensor Networking [J].IEEE/ 

ACM Transactions on Networking,2003,1 1(1):2-16. 

[12] Muruganathan S D, Ma DCF, Bhasin PI, et al. A 

centralized energy-efficient routing protocol for wireless 

sensor networks. IEEE Communications Magazine, 

2005,43(3): 8 – 13. 

[13] Guo Bin,Li Zhe.United voting dynamic cluster routing 

algorithm based on residual-energy in wireless sensor 

networks.Journal of Electronics & Information Technology. 

2007,29(12).pp:3006-3010.


[14] Heinzelman W,chandrakasan A,BalakrishnanH.An 

Application-Specific Protocol Architecture for Wireless 

Microsensor Networks.IEEE Transactions on Wireless 

Communications,October 2002,1(4):660—670. 

[15] C. Schurgers and M. B. Srivastava, “Energy Efficient 

Routing in Wireless Sensor Networks,” in proc. IEEE 

Military Communications Conf. vol. 1, pp. 357-361. 2001. 

[16] Manjeshwar A , Agrawal D . TEEN : A Protocol for 

Enhanced Efficiency in Wireless Sensor Networks . In 

Proceeding of the lth International Workshop on Parallel 

and Distributed Computing Issues in Wireless Networks 

and Mobile Computing’01.2001:23-27. 

[17] Xinyun Fan,Fubao Wang etc,Wireless sensor network 

routing protocols. Computer Measurement & Control, 

2005,1 3 (9) :1010-1013. 

[18] LindseyS,Raghavendra C . PEGASlS : Power-Efficient 

Gathering in Sensor Information Systems.In Proceeding 

of the IEEE Aerospace Conference . Montana : IEEE 

Aerospace and Electronic Systems Society,2002:1 125- 

l 130. 

[19] Guojun Wang , Tian Wang, Weijia Jia, Minyi Guo, Hsiao- 

Hwa Chen, Mohsen Guizani “Local Update-Based Routing 

Protocol in Wireless Sensor Networks with Mobile Sinks” 

This full text paper was peer reviewed at the direction of 

IEEE Communications Society subject matter experts for 

publication in the ICC 2007 proceedings. 

[20] Hayoung Oh and Kijoon Chae “An Energy-Efficient 

Sensor Routing with low latency, scalability for Smart 

Home Networks” International Journal of Smart Home, 

Vol. 1, No. 2, July, 2007. 

[21] Chan H, Perrig A (2004) ACE: An emergent algorithm for 

highly uniform cluster formation. In:Proceedings of the 

first European workshop on sensor networks (EWSN), 

2004 3. 

[22] Ye M, Li CF, Chen GH, Wu J (2004) EECS: an energy 

efficient clustering scheme in wireless sensor networks. In: 

Proceedings of the IEEE international workshop on 

strategies for energy efficiency in ad hoc and sensor 

networks (IWSEEASN’05), April 2004. 

[23] Ming Liu · Jiannong Cao · Yuan Zheng,et.al. An energyefficient 

protocol for data gathering and aggregation in 

wireless sensor networks, J Supercomput (2008) 43: 107– 

125. 

[24] Heinzelman WR, et al(2002) An application—specific 

protocol architecture for wireless microsensor networks. 

IEEE Trans Wireless Commun 1(4):660–670. 

[25] ZHOU Si-Wang, LIN Ya-Ping etc. A Wavelet Data 

Compression Algorithm Using Ring Topology for Wireless 

Sensor Networks.Journal of Software. 2007.18(3).669-680. 

[26] Gandham S R,Dawande M,Prakash R,etc.Energy efficient 

schemes for wireless sensor networks with multiple mobile 

base station[A].In:GLOBECOM 2003,IEEE 

Comunnications Society [C].SanFrancisco,USA:2003.377- 

381. 

[27] D.Braginsky,D.Estrin,Rumor routing algorithm for sensor 

networks, WSNA’ 02, Atlanta, GA, September, 2002. 

[28] V.Rodoaplu and T.H.Meng, Minimum energy mobile 

wireless networks, IEEE J.Select.Areas Communi., vol.17, 

no 8,pp.1333-1334,1999. 

[29] Ming Zhang, Suoping Wang et.al, An Novel Energy- 

Efficient Minimum Routing Algorithm (EEMR) in 

Wireless Sensor Networks, WICOM2008,135-438. 


Ming Zhang is an associate 

professor with the School of 

Electronic Engineering, Huaihai 

Institute of Technology, Lian 

yungang,China.He received his 

Master degree in Computer Science 

and Technology from Soochow 

University, Jiangsu, China in 2002. 

Since 2006, he has been pursuing his 

Dr. degree in the Department of Software at Nanjing 

University of Posts & Telecommunications. His current 

research interests include wireless sensor networks, 

wireless networks and software technology. 

Suoping Wang is currently as 

Prof/Head of Wujiang School, 

Nanjing University of Posts & 

Telecommunications,Nanjing, Chaina. 

He graduated from Department of 

Radio, Tsinghua University, Beijing, 

China in 1970.He received his 

Master degree in Department of Communications and 

electronic Engineering,Nanjing University of posts & 

Telecommunications,Nanjing, China in 1981.His research 

focuses on Real-time Systems, Wireless Network, 

Network Communication Theory and Technology.


Performance of UWB Systems with Direct- 

Sequence Bipolar Pulse Amplitude Modulation 

and RAKE Reception over IEEE 802.15.3a 

Channel 

Jingjing Wang 1, 2 Hao Zhang 2 

1 College of information Science & Technology, Qingdao University of Science & Technology, Qingdao, China 

2 Department of Electrical Engineering, Ocean University of China, Qingdao, China 

Email: kathy1003@163.com 

Abstract—Direct-Sequence Pulse Amplitude Modulation 

(DS-PAM) has been widely proposed for Ultra-Wideband 

(UWB) communication systems because it provides better 

performance with low computational complexity. UWB 

signals suffer from severe multi-path interference when 

employed in indoor fading environments. But using RAKE 

reception can make use of the rich multi-path of UWB 

systems to improve system performance. In this paper we 

present the performance of a RAKE receiver employing 

maximal ratio combining (MRC) in a DS UWB system with 

BPAM modulation. Performance in a practical multi-path 

fading Channel (IEEE 802.15.3a Channel) is considered to 

analyze the performance of DS-PAM UWB systems with 

different RAKE receivers. The bit error rate (BER) of 

ARake, PRake, and SRake over DS-BPAM UWB systems is 

simulated. The results indicate that ARake has the best 

performance, SRake is better than PRake when the number 

of fingers is the same. 

Index Terms—Performance, Ultra-Wideband, Direct- 

Sequence, Pulse Amplitude Modulation, IEEE 802.15.3a, 

RAKE Receiver 


In wireless communications, electromagnetic waves 

with an instantaneous bandwidth greater than 25% of the 

center operating frequency or an absolute bandwidth of 

1.5 GHz or more are referred to as Ultra-Wideband 

(UWB) signals[1][2]. The basic concept of UWB is to 

transmit and receive an extremely short duration burst of 

radio frequency (RF) energy to implement high data rate 

transmission. UWB is a promising technology for future 

high speed wireless communication systems. Pulse 

amplitude modulation (PAM), pulse position modulation 

(PPM) and on/off keying (OOK) modulation are the most 

commonly used modulation schemes in UWB systems. 

PPM modulation uses the precise collocation of impulses 

in time to convey information, while PAM and OOK use 

the impulse amplitude for this purpose. UWB systems 

with PAM and PPM modulation have been extensively 

This work is supported by Outstanding Youth Fund of Shandong 

province under Grant no.JQ200821 and New Century Educational 

Talents Plan of Chinese Education Ministry under Grant no. NCET-08- 

0504 


doi:10.4304/jnw.6.7.1065-1071 

investigated [3][4][5][6]. 

Wireless communications systems typically operate in 

multi-path fading channels. In addition to the direct path 

signal (if present), many reflected path signals can arrive 

at the receiver with different delays and attenuations, 

resulting in fading and inter symbol interference. 

Employing a RAKE receiver is an efficient means of 

overcoming these effects to achieve better performance 

[7][8]. Actually, one of the advantages of broadband 

wireless communication systems such as code division 

multiple access (CDMA) and UWB is the capability of 

utilizing multi-path signals to improve system 

performance and capacity. Since UWB systems can 

resolve many paths they are rich in multi-path diversity, 

so the use of RAKE diversity combining can be very 

effective. Considering the reasons given above, a RAKE 

receiver is an essential component of future UWB 

communication systems [9][10]. 

The performance of PPM and PAM with Multiple 

Receive antennas has been investigated [11][12] [13]. 

And [14] presented the Performance of UWB systems 

with PPAM and RAKE Reception. But most conclusions 

are derived over additive white Gaussian channels and 

Rayleigh, Ricean fading channels. IEEE 802.15.3a 

channel which could express the actual indoor fading 

channel well is neglected. 

Compared with TH-PPM, DS-PAM provides a lower 

BER for the same Signal to Noise Ratio (SNR) and the 

computational complexity. Compared with PPAM, DS- 

PAM provides a less complicated hardware. 

In this paper, we consider a DS-BPAM UWB system 

over IEEE 802.15.3a channel model with a RAKE 

receiver, and the performance with different RAKE 

receivers is analyzed. 

The remainder of the paper is organized as follows. In 

Section II, UWB system model is described. Section III 

introduces the channel model and three primary 

parameters that are important to characterize multi-path. 

Section IV introduces the IEEE802.15.3a channel model. 

Section V presents the error probability analysis of a DS- 

BPAM UWB system with a RAKE receiver over IEEE 

802.15.3a channel. The performance of a DS-BPAM 

UWB system with different RAKE receivers is examined 

and some conclusions are given in Section VI.


b 

Code 

Repetition 

Coder 

(Ns, 1) 

Binary 

+1 Series 

II. UWB SYSTEM MODEL 

Transmission 

Coder 

C is a binary code 

Figure 1 shows the transmission scheme for a PAM- 

DS-UWB signal [15]. Input data sequence is a binary 

sequence = (…, b0, b1, bk, bk+1,). And each bit is 

repeated Ns times by the first module, Code Repetition 

Coder. Then a new binary sequence is generated 

which is (…, b0, b0 , …, b0, b1, b1, …, b1, …, be, be, …, be ). 

The second module transforms the sequence to a 

sequence which only includes positive and negative 

element. The transmission encoder applies a binary code 

composed of ± and period Np to the binary sequence d 

which equal to a . The sequence d enters the PAM 

modulator which generates a sequence of unit pulses 

whose position is . Then, the output of the modulator 

enters the pulse shaper filter and the result s 

b 

* 

a 

* 

a 

a 

1 

⋅c 

jTS 

(t) 

is 

transmitted. 

The PAM-DS-UWB signal s(t) 

at the output of the 

transmitter can be expressed as 

T 

a* 

s( t) 

ETX 

d j p( 

t − jTS 

) 

= ∑ ∞ 

j= 

−∞ 

Where S is the frame time, i.e. average pulse 

repetition period, p(t) is the energy-normalized waveform 

of the basic pulse and E is the transmitted energy per 

TX 

pulse. E is assumed to be 1 in Figure 1. 

TX 

At the receiver, the received signal is PAM 

demodulated. After detecting procedure in demodulator, 

DS code which is identical to that utilized at the 

transmitter is employed to recover the transmitted 

sequences. Then, the output of code repetition decoder is 

decoded to estimate original data sequence. 

III. CHANNEL MODEL 

For a UWB system with multi-path fading, the 

following discrete impulse response of the channel is 

considered: 

∑ − L 1 

l= 

0 

(1) 

h( t) 

= a δ ( t −τ 

) 

(2) 

τ 

where al 

is the channel gain for the l-th path, l is the 

delay for the l-th path, L is the number of resolvable 

paths, and δ (⋅) 

is the Dirac delta function. If the relative 

delay of two paths is less than a pulse width, they cannot 

be identified by a RAKE receiver. Thus we 

τl −τk ≥Tp 

assume , ∀l ≠ k T p , where is the width of p(t). 


l 

l 

a 

The primary parameters that are important to 

characterize multi-path: the total multi-path gain, the root 

mean square delay spread, and the power delay profile. 

The following sub-sections describe in more detail each 

of these components. 

A. The Total Multi-Path Gain 

The total multi-path gain G measures the total amount 

of energy collected over the N received pulses when a 

pulse with unitary energy is transmitted. The G parameter 

can be determined as follows: 

∑ − L 1 

al 

l= 

0 

2 

G = 

(3) 

Given the G value, the impulse response can be written: 

Where 0 1 ,..., L− 

∑ − L 1 

l = 0 

h ( t ) = G α δ ( t − τ ) 

(4) 

l 

α α are the energy-normalized channel 

gain parameters verifying: 

∑ − L 1 

l= 

0 

2 

l 

l 

α = 1 

(5) 

Note that G ≤1 

and is related to the attenuation 

suffered by the transmitted pulses during propagation. In 

multi-path environments, G decreases with distance 

according to the following law: 

G0 

G = (6) 

γ 

d 

Where G is the reference value for power gain evaluated 

0 

at d=1 m and γ is the exponent of the power or energy 

attenuation law. The G0 

value can be evaluated as 

follows: 

0 / 10 

10 A − 

G = (7) 

0 

Where A (in dB) represents the path loss at a reference 

0 

PAM 

Modulator 

Figure1. Transmission scheme for a PAM-DS-UWB signal. 

d 

d=a.c 

Pulse 

Shaper 

p(t) 

distance d0=1 m, that is, 10 log 10( 

/ ) , is the 

E E 

A = 

E 

0 TX RX 0 

s( t) 

d j p( 

t − jTS 

) 

= ∑ ∞ 

j= 

−∞ 

RX 0 

energy of a single pulse at d0. Values for both and 0 

are suggested in [16] for different propagation 

environments: A 47dB 

0 and = 1. 

7 

= γ for a LOS 

environment, and A 51dB 0 and = 3. 

5 

= environment. 

γ for a NLOS 

A γ


B. The Root Mean Square Delay Spread 

τ 

rms 

= 

L−1 

2 

∑τ 

l 

l= 

0 

G 

a 

2 

l 

⎛ 

⎜ 

− ⎜ 

⎜ 

⎜ 

⎝ 

L−1 

∑ 

τ a 

l 

l= 

0 

Equation (8) measures the effective duration of the 

channel impulse response. It is a fundamental parameter 

for evaluating the presence of Inter Symbol Interference 

(ISI) at the receiver. If the time interval separating two 

pulses is smaller than τ rms , ISI is present. 

C. The Power Delay Profile(PDP) 

The Power Delay Profile (PDP) of an impulse response 

given by Equation (4) is a graphical representation that 

shows time of arrival of the different contributions versus 

received power. Time of arrival of a generic path is 

usually indicated relative to the LOS contribution, which 

has a time of arrival fixed at 0. 

G 

2 

l 

2 

⎞ 

⎟ 

⎟ 

⎟ 

⎟ 

⎠ 

IV. THE IEEE802.15.3A CHANNEL MODEL 

The IEEE 802.15.3a channel model is based on the 

Saleh-Valenzuela (S-V) model in [17], where the impulse 

response is composed of exponentially decaying signal 

clusters to model the dense multi-path components. The 

UWB indoor channel model is then [18] 

− L 1 K ( l) 

∑∑ 

l= 

0 k = 0 

(8) 

h( t) 

= X α δ ( t −T 

−τ 

) 

(9) 

l, 

k 

where 

α } are the coefficients of the k-th multi-path 

{ l, k 

contribution of the n-th cluster, 

T 

{ l } is the time of arrival of the l-th cluster, 

{ τ lk }is the delay of the k-th multi-path contribution 

within the l-th cluster, 

{X} is a log-normal random variable representing the 

amplitude gain of the channel. 

The proposed model describes four different 

measurement environment named CM1, CM2, CM3 and 

CM4 as shown in Table 1. CM1 describes a line-of-sight 

(LOS) channel with a distance from transmitter to 

receiver less than 4 meters. CM2 describes a non-LOS 

channel with the same range (0-4m). CM3 describes a 

non-LOS channel for distances between 4 and 10m. CM4 

describes an extreme NLOS multi-path channel. 

The parameters are defined as: 

Λ, inter-cluster (cluster) average arrival rate; 

λ , intra-cluster (ray) average arrival rate; 

Γ , inter-cluster (cluster) average decay rate; 

γ , intra-cluster (ray) average decay rate; 

σ ξ , average cluster lognormal standard deviation; 

σ ζ , average ray lognormal standard deviation; 

σ g , channel amplitude gain standard deviation. 


l 

lk 

TABLE I. 

PARAMETERS FOR IEEE802.15.3A CHANNEL MODEL 

Channel 

Model 

CM 1 

Λ λ Γ σ ξ σ ζ σ g 

LOS 

(0-4m) 

CM 2 

0.0233 2.5 7.1 4.3 3.3941 3.3941 3 

NLOS 

(0-4m) 

CM 3 

0.4 0.5 5.5 6.7 3.3941 3.3941 3 

NLOS 

(4-10m) 

CM 4 

0.0667 2.1 14 7.9 3.3941 3.3941 3 

Extreme 

NLOS 

0.0667 2.1 24 12 3.3941 3.3941 3 

V. ERROR PROBABILITY ANALYSIS OF DS- 

BPAM UWB SYSTEM OVER IEEE802.15.3A 

CHANNEL MODEL WITH RAKE RECEIVER 

In multi-path fading channels, many reflected path 

signals can arrive at the receiver with different delays and 

attenuation, resulting in fading and inter-symbol 

interference. Employing a RAKE receiver is an efficient 

means of overcoming these effects. 

m 1(t) 

r(t) τ + Ts 

Z1 … 

m 2(t) 

… 

m N(t) 

∫ 

τ 

dt 

correlator 

τ + T s 

∫ 

τ 

τ + T s 

∫ 

τ 

d t 

dt 

Z 2 

Z N 

ω 1 

ω 2 

… 

ω N 

Detector 

Figure 2. The structure of a RAKE receiver with N correlators. 

Estimated 

symbol 

The typical structure of a RAKE receiver is shown in 

Figure 2, which consists of a series of correlators and a 

detector. The correlators or matched filters are also called 

fingers. Each RAKE finger is matched to a particular 

multi-path component in order to combine them 

coherently. A reference or template signal matched to the 

incoming received signal is used by the RAKE receiver. 

Each finger of the RAKE receiver uses a delayed version 

of the template signal to match the delay to a specific 

multi-path component. In order to enable symbol-rate 

sampling, the received signal is correlated with a symbollength 

template signal, and the output of the correlator is 

sampled once per symbol. If the receiver uses all L 

received paths, it is called All-RAKE (ARake)[19]. 

However, the number of multi-path components that can 

be utilized in a typical RAKE combiner is limited by 

power issues, design complexity, and the quality of the 

channel estimation. Thus, in practice, only a subset of the 

resolved multi-path components is used, giving rise to the 

γ 

Z


Partial RAKE (PRake) and Selective RAKE (SRake) 

receivers which have a limited number of fingers. The 

PRake receiver uses the first M arriving paths out of the L 

resolvable multi-path components, while SRake searches 

for the M best paths [19]. 

The transmitted symbol can be simply described as 

formula (1).The combining algorithm commonly used in 

RAKE receivers is maximal ratio combining (MRC) 

which is known to maximize the signal-to-noise ratio 

(SNR) for diversity channels [20]. 

The received signal over IEEE 802.15. 3a channel can 

be simply described as 

∑ − L 1 

l= 

0 

r( t) 

= α s( 

t −τ 

) + σ n( 

t) 

where n(t) is white noise. 

The template signal is 

where 

l 

L 1 

= ∑ − 

m ( t) 

ω m( 

t −τ 

) 

R 

j= 

0 

m( t τ j ) = d j p( 

t − jTS 

−τ 

j ) 

l 

j 

n 

j 

(10) 

(11) 

− (12) 

and ωj is the RAKE combining weight of j-th branch, L 

is the number of RAKE receiver branches. For BPAM, d 

∈{0,1}. ω = [ω0, ω1, … , ωL−1] are the RAKE combining 

weights. If MRC technique is used, the amplitudes of the 

received multi-path components (MPCs) are estimated 

and used as weighing vector ω in each finger. In case of 

ARake, the combining weights are chosen as ω = α, 

where α = [α0, α1, … , αL−1] are the fading coefficients of 

the channel. If the set of indices of the M best fading 

coefficients with largest amplitude is denoted by S, then 

the combining weights ω of an SRake are chosen as [21], 

⎧α 

l, l∈S ω = ⎨ 

⎩ 0, l∉S , (13) 

Similarly, for PRake using the first M multi-path 

components, the weights of MRC combining are given by 

[20], 

⎧αl 

, l = 0, …, M −1 

ω = ⎨ 

(M≤L) (14) 

⎩ 0, l = M, …L−1 The output of combiner is 

τ + T s 

Z = ∫ r( t) m ( t) dt , (15) 

τ 

Assuming a perfect match of the received signal with 

the template signal, zero inter-frame and inter-symbol 

interference, and symbol rate sampling at the output of 

RAKE fingers, then Equation (15) can be rewritten in 

discrete time as 


R 

∑ − L 1 

b 

l= 

0 

Z = E ω α + n 

(16) 

where Eb is the energy per bit. 

+∞ 

σ n R 

−∞ 

n= ∫ n() t m () t dt 

l 

l 

(17) 

n is the noise at the output of the correlator which is 

2 

approximately distributed as n ~ N(0, σn 

). 

To determine the BER at the output of the RAKE, the 

output SNR needs to be evaluated. From Equation (15), 

the approximate signal energy and the noise variance at 

the output of RAKE are evaluated as 

( 

2 

) = 

L−1 

b( ∑ 

l= 

0 

l 

2 

l) 

, (18) 

E signal E ω α 

2 

= 

L−1 

2 

n ∑ 

l= 

0 

E( noise ) σ ωl 

2 

. (19) 

In case of BPAM, for a given SNR per Bit γb, the 

approximate expression of BER conditioned on a 

particular channel realization is given by [22], 

⎛ L−1 

2 

Eb( 

ωα 

⎞ 

l 0 l l 

Pe| α ( γ b) 

Q( SNR) Q⎜ 

∑ ) 

= 

= ≈ ⎟ 

⎜ 2 L−1 

2 

⎜ 

⎟ 

σn ω ⎟ 

⎝ ∑l= 

0 l ⎠ 

(20) 

where Q(.) is the standard Q function. 

However, to obtain the error probabilities when 

channel fading coefficients α are random, we must 

average the Pe(γb) over the probability density function 

of γb [21], 

∞ 

P = P( γ ) p( γ ) dγ 

. (21) 

∫ 

b 

e e b b 

0 

By evaluating the probability distribution function of 

output SNR, average BER can be obtained. It is difficult 

to obtain a closed-form expression of (21). However, this 

average can be evaluated numerically, or by employing 

Monte-Carlo simulations. 

Figure 3 shows the equivalent RAKE receiver structure 

with BPAM based on discrete-time channel models. 

The adoption of a Rake considerably increases the 

complexity of the receiver. This complexity increases 

with the number of multi-path components analyzed and 

combined before decision, and can be reduced by 

decreasing the number of components processed by the 

receiver. However, a deduction of the number of paths 

leads to a decrease of energy collected by the receiver. It 

is important to catch a good compromise between the two 

elements.


L−1 

∑ 

rt () = αst ( − τ ) + nt () 

l l 

l= 

0 

m(t) 

τ + Ts 

∫ 

τ 

dt 

Correlator 

t = kΔt 

Detector 

ωN-1 ZN-1 ω2 ω1 

Z Z 2 

1 

Estimate 

Bits 

Figure 3. Equivalent RAKE receiver structure with BPAM 

VI. NUMERICAL RESULTS AND CONCLUSIONS 

In this section, the performance of a DS-BPAM UWB 

system with RAKE receiver over an IEEE 802.15.3a 

channel is presented. 

Figure 4. Bit error rate for DS-BPAM UWB over CM1 of IEEE 

802.15.3a channel with different RAKE receivers. 

Figure 4 shows the bit error rate for DS-BPAM over 

CM1 of IEEE 802.15.3a channel with different RAKE 

receivers. This shows that there is almost a 1dB gain with 

an ARake receiver over an SRake receiver (the number of 

fingers S=5) at a BER of 10 -2 . Also, there is almost a 

1.5dB gain with a SRake receiver over a PRake receiver 

(the number of fingers S=L=5) at a BER of 10 -2 . 

Figure 5 shows the bit error rate for DS-BPAM UWB 

over a CM2 of IEEE 802.15.3a channel with different 

RAKE receivers. In this case, there is almost a 1dB gain 

with an ARake receiver over a 5 finger SRake receiver at 

a BER of 10 -1 . Also, there is about a 1.7dB gain with an 

SRake receiver over a PRake receiver (S=L=5) at a BER 

of 10 -1 . 

Figures 6 and 7 show the bit error rates for DS-BPAM 

UWB over CM3 and CM4 of IEEE 802.15.3a channel 

with different RAKE receivers respectively. It is 

obviously that the bit error rates of a DS-BPAM UWB 

system have the similar trend over CM3 and CM4 

channels with CM1 and CM2. 

Z N 

Δ t 

Δt 

ω N 

Zout Z>0 b=0; 

= jT + N Δt 

Z


Figure 8. Bit error rate for DS-BPAM UWB over CM1-CM4 of IEEE 

802.15.3a channel with ARAKE receiver 


802.15.3a channel with SRAKE(S=5) receiver 


802.15.3a channel with PRAKE (P=5) receiver 

From Figure 4-7, the conclusion can be drawn that All- 

RAKE (ARake) receiver has the best performance, 

because it uses all the multi-path components which the 

receiver can identify. With the same number of fingers, 

SRake has a better performance than PRake. Because 

SRake has a selection process, the complexity of channel 

estimation and channel tracking is same with ARake, but 


the number of branches is smaller than that of ARake. 

PRake only considers the first arrival components and 

without having a selection process. For all types of 

RAKE reception, the bigger number of fingers, the better 

performance will be achieved. 

Figure 8 to 10 compare the BER performance of DS- 

BPAM UWB over the CM1 to CM4 of IEEE 802.15.3a 

channel with ARAKE, SRAKE(S=5), PRAKE (P=5) 

receiver, respectively. As expected, the system 

performance over a CM1 channel is the best, while the 

BER performance of DS-BPAM deteriorates sharply 

when signals are transmitted over a CM4 channel. 


The authors would like to thank the anonymous 

reviewers for their constructive comments and questions 

that greatly improved the paper. 

REFERENCES 

[1] OSD/DARPA Ultra-Wideband Radar Review Panel, 

“Assessment of Ultra-Wideband Technology,” Defense 

Advanced Research Projects Agency, July, 1990. 

[2] J.D. Taylor, Introduction to Ultra-Wideband Radar 

Systems, CRC Press, 1995. 

[3] R.A. Scholtz, “Multiple access with time-hopping impulse 

modulation,” Proc. IEEE Military Commun. Conf., pp. 11– 

14, Oct. 1993. 

[4] M.Z. Win and R.A. Scholtz, “Impulse radio: How it 

works,” Proc. IEEE Commun. Letts., vol. 2, pp. 36–38, 

Feb. 1998. 

[5] F. Ramirez-Mireles and R.A. Scholtz, “System 

performance analysis of impulse radio modulation,” Proc. 

IEEE Radio and Wireless Conf., pp. 67–70, Aug. 1998. 

[6] F. Ramirez-Mireles and R.A. Scholtz, “Multiple-access 

performance limits with time hopping and pulse position 

modulation,” Proc. IEEE Military Commun. Conf., pp. 

529–533, Oct. 1998. 

[7] J.G. Proakis, Digital Communications, 3rd Ed., McGraw 

Hill, New York, NY, 1995. 

[8] R. Price and P.E. Green Jr., “A communication technique 

for multi-path channels,” Proc. IRE, vol. 46, pp. 555–570 

Mar. 1958. 

[9] B. Mielczarek, M.O. Wessman and A. Svensson, 

“Performance of coherent UWB Rake receivers with 

channel estimators,” Proc. IEEE Vehic. Tech. Conf., pp. 

1880–1884, Oct. 2003. 

[10] S. Imada and T. Ohtsuki “Pre-RAKE diversity combining 

for UWB systems in IEEE 802.15 UWB multi-path 

channel,” Proc. Int. Workshop on Ultra Wideband Systems, 

pp. 236–240, May 2004. 

[11] H. Zhang and T. A. Gulliver, “Performance and capacity of 

PAM and PPM UWB time-hopping multiple access 

communications with receive diversity,” EURASIP J. 

Applied Signal Processing, pp. 306-315, Mar. 2005. 

[12] H. Zhang and T. A. Gulliver, “Performance and capacity of 

PAM and PPM UWB systems with multiple receiver 

antennas,” Proc. IEEE Pacific Rim Conf. on Commun., 

Computers and Signal Processing, pp. 740-743, Aug. 2003. 

[13] H. Zhang and T. A. Gulliver, “Capacity of time-hopping 

PPM and PAM UWB multiple access communications 

over indoor fading channels,” EURASIP J. Wireless


Commun. and Networking, vol. 2008, [online]. 

Available:http://www.hindawi.com/journals/wcn/2008/273 

018.html [accessed: Sept.12, 2009] 

[14] Wei Li, T. Aaron Gulliver and Hao Zhang, “Performance 

of ultra-wideband transmission with pulse position 

amplitude modulation and rake reception” 2005 

IEEE/ACES International Conference on Wireless 

Communications and Applied Computational 

Electromagnetics,pp.1-4 

[15] Benedetto, M. G. D. and Giancola, G. 2004 Understanding 

Ultra Wide Band Radio Fundamentals. Prentice Hall, New 

Jersey. 

[16] Ghassemzadeh,S.S., L.J. Greenstein, A. Kavčić, T. 

Sveinsson, and V. Tarokh, “an empirical indoor path loss 

model for uwb channels”, Journal of Communication and 

Networks, VOL.5, pp.303-308, Dec. 2003 

[17] J. G. Proakis Digital Communications, 4th ed. 

Boston:McGraw-Hill,2001. 

[18] J. Foerster, ed., Channel modeling sub-committee report 

final, IEEE 802.15 Working Group for Wireless Personal 

Area Networks(WPANs), IEEE P802.15-02/490r1-SG3a, 

Feb.2003. 

[19] D. Cassioli, M. Z.Win, F. Vatalaro, and A. F. Molisch, 

“Performance of low-complexity RAKE reception in a 

realistic UWB channel,” Proc. IEEE Int. Conf. Commun., 

pp. 763-767, May 2002. 

[20] S. Tantikovit, A. U. H. Sheikh, and M. Z. Wang, 

“Combining schemes in RAKE receiver for low spreading 

factor long-code W-CDMA systems,” IEE Elect. Letts., vol. 

36, no. 22, pp. 1872–1874, Oct. 2000. 

[21] S. Gezici, H. Kobayashi, H. V. Poor, and A. F. Molisch, 

“Performance evaluation of impulse radio UWB systems 

with pulse-based polarity randomization in asynchronous 

multi-user environments,” Proc. IEEE Wireless Commun. 

and Networking Conf., pp. 908-913, Mar. 2004. 

[22] H. Hashemi, “Impulse Response Modeling of Indoor 

Radio Propagation Channels,” IEEE JSAC, Vol. 11, No. 7, 

Sept. 1993,pp,967-968 

Jingjing Wang was born in Anhui, China, in 1975. She 

received her B.S. degree in industrial automation from 

Shandong University, Jinan, China, in 1993, the M.Sc. degree 

from control theory and control engineering, Qingdao 

University of Science & Technology, Qingdao, China in 2002. 

From 1997 to 1999, she was the assistant engineer of Shengli 

Oilfield, Dongying, China. From 2002 to now, she is an 

associate professor at the College of information Science & 

Technology, Qingdao University of Science & Technology. Her 

research interests include 60GHz wireless communication, and 

ultra wideband radio systems. 

Hao Zhang was born in Jiangsu, China, in 1975. He 

received his B.S. degree in telecom engineering and industrial 

management from Shanghai Jiaotong University, Shanghai, 

China, in 1994, the M.B.A. degree from New York Institute of 

Technology, Old Westbury, NY, in 2001, and the Ph.D. degree 

in electrical and computer engineering from the University of 

Victoria, Victoria, BC, Canada, in 2004. 

From 1994 to 1997, he was the Assistant President of ICO 

(China) Global Communication Company, Beijing, China. He 

was the Founder and CEO of Beijing Parco Company, Ltd., 

Beijing, China, from 1998 to 2000. In 2000, he joined Microsoft 

Canada, Vancouver, BC, as a Software Engineer, and was Chief 

Engineer at Dream Access Information Technology, Victoria, 

BC, Canada, from 2001 to 2002. He is currently an Adjunct 


Assistant Professor with the Department of Electrical and 

Computer Engineering, University of Victoria. His research 

interests include ultra wideband radio systems, MIMO wireless 

systems, and spectrum communications.


Data Accuracy Estimation for Spatially 

Correlated Data in Wireless Sensor Networks 

under Distributed Clustering 

Jyotirmoy Karjee , H.S Jamadagni 

Centre for Electronics Design and Technology, Indian Institute of Science, Bangalore, India 

kjyotirmoy@cedt.iisc.ernet.in, hsjam@cedt.iisc.ernet.in 

Abstract—Objective-The main purpose of this paper is to 

construct a distributed clustering algorithm such that each 

distributed cluster can perform the data accuracy at their 

respective cluster head node before data aggregation and 

transmit the data to the sink node.Design 

approach/Procedure – We investigate that the data are 

spatially correlated among the sensor nodes which form the 

clusters in the spatial domain. Due to high correlation of 

data, these clusters of sensor nodes are overlapped in the 

spatial domain. To overcome this problem, we construct a 

distributed clustering algorithm with non-overlapping 

irregular clusters in the spatial domain. Then each 

distributed cluster can perform data accuracy at the cluster 

head node and finally send the data to the sink node. 

Findings- Simulation result shows the associate sensor nodes 

of each distributed cluster and clarifies their data accuracy 

profile in the spatial domain. We demonstrate the 

simulation results for a single cluster to verify that their 

exist an optimal cluster which give approximately the same 

data accuracy level achieve by the single cluster. Moreover 

we find that as the distance from the tracing point to the 

number of sensor node increases the data accuracy 

decreases. Design Limitations – This model is only applicable 

to estimate data accuracy for distributed clusters where the 

sensed data are assumed to be spatially correlated with 

approximately same variations. Practical implementation – 

Measure the moisture content in the distributed agricultural 

field. Inventive/Novel idea- This is the first time that a data 

accuracy model is performed for the distributed clusters 

before data aggregation at the cluster head node which can 

reduce data redundancy and communication overhead. 

Index Terms—Wireless sensor networks, distributed 

clusters, data accuracy, spatial correlation 


Wireless sensor network has made a drastic change in 

communications for the last several years. One of the 

vital tasks of wireless sensor network is to sense or 

measure the physical phenomenon of data such as 

measurement of humidity, temperature, seismic event etc 

from the environment [1]. Physical phenomenon of data 

is measured or sense by a device called sensor nodes 

which are capable to sense, process and communicate the 

data through out the network. Since most of the data are 

spatially correlated [2] among them, the sensor nodes 

form clusters in the sensor field to reduce data collection 

cost [3]. According to literature survey, LEACH [4] gives 

a clear idea about how dynamically cluster and cluster 


doi:10.4304/jnw.6.7.1072-1083 

head are created according to a priori probability. Finally 

cluster head aggregate all the data and send it to the sink 

node. Similarly SEP [5] demonstrates the formation of 

cluster in heterogeneous sensor networks. Since data 

correlation in wireless sensor networks shows Gaussian 

distribution with zero mean, literature [6] shows the 

spatial correlation among data is high in sensor networks 

but it lags the practical implementation of analyzing the 

correlated data for transmitting the packets for 

communication. Literature [7] proposes a grid based 

spatial correlation clustering method where the entire 

cluster is equipped in a grid sensor field. However this 

type of model rarely happens in an original scenario in 

wireless sensor networks. Moreover literature [8] 

proposes a disk-shaped circular cluster where sensor 

nodes are grouped into disjoint sets each managed by a 

designated cluster head which lags the practical shape of 

a cluster. As most cases the cluster formation are 

irregular in shape for the spatial domain. Hence in this 

paper we propose a foundation of distributed clustering 

algorithm which is much more practical than the previous 

work done in the spatial domain. In our model, we 

propose a spatially correlated distributed irregular non 

overlapping cluster formation in the spatial domain. 

These distributed irregular cluster formation in the spatial 

domain is much more practical model in original scenario 

than the previous literature discussed above. 

Most of the work done till today is based upon the 

fact that the sink node or the base station is responsible 

for estimating the data accuracy for physically sensed 

data by sensor nodes [9, 10, 11] .Therefore it is applicable 

for one hop communication where the raw data are 

sensed and measured by the sensor nodes and directly 

transmitted to the sink node. Again we propose a model 

[12] for data accuracy where we have considered two hop 

communications in which physical phenomenon of 

sensed data is transmitted via intermediate node called 

cluster head (CH)[18]node. But in this paper we propose 

a distributed clustering algorithm where each cluster can 

perform data accuracy at their respective CH node and 

finally send the data to the sink node. Each distributed 

cluster is responsible for sensing and measuring the 

physical phenomenon of data in the sensor region. 

The main goal of this paper is to estimate data 

accuracy for each distributed cluster before data 

aggregation [19] at their respective CH node which can 

reduce the data redundancy and communication


overhead. However to the best understanding of the 

authors, there is no work done so far on verifying the data 

accuracy for distributed cluster before data aggregation 

[21, 22] at their respective CH node. Since from the 

literature survey we have seen that most of the work done 

till today is that data from cluster of sensor nodes directly 

send to CH node for aggregation without verifying its 

accuracy. Hence it is important that the most precise or 

accurate data send by the distributed cluster can 

aggregate at their respective CH node before transmitting 

to the sink node and not aggregating all the redundant 

data at CH node. The data send by each distributed 

cluster should first verify its accuracy level at their 

respective CH node then only the data get aggregates and 

finally send to the sink node. Since CH node verifies the 

data accuracy for their respective distributed cluster, it 

may reduce the power consumption and increase the 

lifetime of the networks. 

Another important reason for estimating data accuracy 

for each distributed cluster before data aggregation at 

their respective CH node, if some of the sensor nodes in 

the distributed cluster get malicious [20]. If some of the 

sensor nodes become malicious in the distributed cluster, 

then it can sense and read inaccurate data. These 

inaccurate data send by the malicious nodes gets 

aggregated with the other correct data results in 

inaccurate (incorrect) data aggregation at the CH node of 

their respective cluster and finally send to the sink node. 

This may increase the power consumption, data 

redundancy and communication overhead in the 

distributed network. It results very high or low variations 

of the estimated data accuracy value compare to the 

actual variations of estimated data accuracy value at the 

CH node. Hence to overcome this problem, it is important 

to estimate the data accuracy at CH node for distributed 

cluster before data aggregation and send the accurate data 

to the sink node. In our model we assume that the sensed 

data are spatially correlated with approximately the same 

variations in each distributed cluster and the sensor nodes 

are appropriate to sense the correct data. We verify 

estimated data accuracy with approximately same 

variations at the CH node for each distributed cluster. 

In our model, each distributed cluster is responsible to 

sense the physical phenomenon of data such as moisture 

content of soil in the sensor region. Once the data 

accuracy is processed by CH node for each distributed 

cluster, it transmits the estimated accurate data to the sink 

node. From the literature survey, it is clear that only the 

sensor nodes are responsible to sense the physical 

phenomenon of data and not the sink node. But in our 

model not only sensor nodes are responsible to sense the 

physical phenomenon of data but the CH node can also 

do the sensing phenomenon in each distributed cluster. 

We investigate how each distributed cluster can sense the 

physical phenomenon of data to estimate the data 

accuracy in the sensor field. Literature [9, 13] has given 

some approaches regarding jointly sensing nodes which 

gives an idea about how the raw data is sensed by the 

jointly sensing nodes and how the number of jointly 

sensing nodes affects the data accuracy. However they 


address this problem if only sensing nodes are 

responsible to retrieve physical phenomenon of data 

where they investigate to find a proper number and 

positions of jointly sensing nodes. But in our model, we 

consider both the sensor nodes and the CH node which 

forms each distributed cluster in the sensor field are 

sensing the physical phenomenon such as humidity or 

moisture content of the soil. Since we verify data 

accuracy for each distributed cluster in the sensor field, 

there exit an optimal cluster which gives approximately 

the same data accuracy level achieve by each cluster. 

Rest of the paper is given as follows. In section II, 

we construct a data correlation model for sensor nodes in 

spatial domain. These data correlation can give rise to 

overlapping of clusters in the sensor region. Hence to 

overcome this problem, we propose a distributed 

clustering algorithm with non overlapping irregular 

clusters in the spatial domain. Then we perform data 

accuracy for each distributed cluster at CH node before 

data aggregation in the sensor region. In section III, we 

verify simulation results for distributed clusters. We 

demonstrate results how each distributed cluster are 

formed with their respective associate nodes and their 

data accuracy. Then we show the performance model of a 

single cluster with respect to data accuracy. Finally we 

conclude our work in section IV. 

II. SYSTEM MODEL 

In this section, sensor nodes deployment strategies are 

done where the sensor nodes form distributed clusters 

which are capable to perform data accuracy in the spatial 

domain. We propose an algorithm for distributed clusters 

which perform data accuracy at the cluster head node 

where the data are spatially correlated and finally send 

the data to the sink node. Let a set of sensor nodes are 

deterministically deployed uniformly over a sensor region 

Z. These set of sensor nodes forms the cluster head nodes 

[18] for the distributed clusters equipped with additional 

energy resource [5]. Since CH node perform the data 

accuracy for the respective distributed clusters, we set the 

CH node with additional energy resource and distributed 

deterministically in the sensor field. Again another set of 

sensor nodes are randomly deployed over the sensor 

region Z and are called normal nodes [5]. Normal nodes 

form the distributed cluster along with their respective 

CH node which can sense and measure the spatially 

correlated data and estimate the data accuracy at the CH 

node.CH node has more energy resource than the normal 

nodes because CH nodes has to estimate the data 

accuracy for the cluster. Thus CH nodes and normal 

nodes form the total set of sensor nodes represented as L 

with Z ⊆ R 2 where ||L|| can be represented as total 

number of sensor nodes. They are capable for sensing and 

measuring the spatially correlated data in the sensor 

region Z. For example, we measure the moisture content 

of soil at different locations of sensor region Z. Generally 

there are much more variations in measurement of 

moisture content at different locations in the sensor field. 

Some places the water (or moisture) content in the soil


are more than other different places where the water 

(moisture) content is less. Thus there are variations of 

monitoring the measurement of moisture content in the 

soil at different places in the sensor region Z. 

A. Data Correlation for sensor nodes in Spatial Domain 

We consider reference values for higher concentration 

of moisture content at different places of sensor region Z. 

Suppose the reference values are called tracing points 

[20] and can be represented as S i where i=1, 2, 3…n are 

the number of tracing points at different locations in the 

sensor field with higher variations .The tracing points can 

be located at the different places of sensor field where the 

moisture content is high. For example, water (or 

moisture) content in the soil can be higher at different 

locations of the sensor field. It is considered as reference 

values for tracing points at different locations in sensor 

the region Z. Although the data are spatially correlated in 

the sensor region, there are variations in measurement for 

concentratation of data (moisture content) at different 

places in sensor region Z. The higher concentratation of 

data has higher variations with respect to lower variations 

of data at different places. In spatial domain, data 

correlation depends upon the distance between the tracing 

points to the sensor nodes and the distance between 

jointly sensing nodes [13]. Thus we have two points to 

note in our work. Firstly, data correlation decreases as the 

distance between the tracing points (or reference values) 

to the sensor nodes increases. Secondly, data correlation 

decreases as the distance between jointly sensing nodes 

increases. Thus data correlation is more when the sensor 

nodes are close to each other. 

Since these tracing points has higher concentratation 

of moisture content with higher variations, the sensor 

nodes can sense the higher variation of tracing points (or 

reference values )at different locations in sensor field . 

There may be higher or lower variations of data 

(moisture) measurement in spatial domain where the data 

are spatially correlated in the sensor field. Thus if the 

distance from the tracing point to the sensor nodes 

increases, the variations of the data correlation also get 

decreases. 

We represent a single tracing point where S i for i=1 

sensed by the sensor nodes Si and Sj where they sense and 

do measurement over a window frame of time T to 

capture the continuous data sample with Si={ si1 , si2, si3, 

……..sin } and Sj={sj1 , sj2, sj3, ……..sjn} respectively. The 

data correlation is strong when the tracing point is sensed 

by the sensor nodes Si and Sj located near to each other. 

The data correlation decreases as sensor node Si and Sj are 

far apart from tracing point. We compute the mean of the 

sampled data of sensor nodes as follows 

_ n 1 

Si= ∑ sik 

n k = 1 

and 

_ n 1 

Sj= ∑ sjk 

n k = 1 

Variance of the sample data collected by nodes Si and Sj 

can be given as 

n _ 

1 

2 

var( S ) = ( s −Si) 

− ∑ (1) 

i ik 

n 1 k = 1 


And 

n 

_ 

1 

2 

var( S ) = ( s −S 

j ) 

− ∑ (2) 

j jk 

n 1 k = 1 

The covariance is given as 

n _ _ 

1 

cov( Si, Sj) = ( sik −Si)( sjk −S 

j) 

( n −1) k= 

1 

∑ (3) 

The correlation coefficient ( ρ S , S) 

for correlation 

i j 

between data sensed by the sensor nodes Si and Sj for the 

tracing points can be given by 

cov( Si, S j) 

ρ Si, S = 

j var( S ).var( S ) 

ρ 

i j 

1 

n _ _ 

∑( 

sik −Si)( sjk−Sj) ( n−1) 

k= 

1 

Si, Sj n _ 1 2 

∑( sik −Si) n−1k= 1 

n _ 1 

2 

∑( 

sjk −Sj) 

n−1k= 

1 

= 

⎡ ⎤⎡ ⎤ 

⎢ ⎥⎢ ⎥ 

⎣ ⎦⎣ ⎦ 

The equation-no 4 shows the data correlation 

coefficient for nodes Si and Sj in the spatial domain. 

Similarly from the co-variance model [16], we get the 

correlation coefficient ( ρ Si, S) 

for the data in spatial 

j 

domain. 

2 2 

[ i, j] = cov[ i, j] = σ i [ i, j] = σ i. 

ρ[ 

i, j] 

S S 

(4) 

ES S S S corrS S S S 

cov[ S , S ] E[ S , S ] 

ρ[ 

Si, Sj] 

= = (5) 

σ σ 

i j i j 

2 2 

i 

S 

i 

S 

Again from the power exponential model [16,17], we 

get the correlation coefficient function between node Si 

(xi, yi) and node Sj (xj, yj) as follows 

ρ[ 

S , S ] e 

⎛ d ⎞ 

−⎜ ⎟ 

θ2 

⎝θ1⎠ = (6) 

i j 

We define a threshold τ which can determine whether 

the data are spatially correlated among the sensor nodes 

to trace the higher variations of data (called as tracing 

points) in the spatial domain. θ is called a ‘Range 

1 

parameter’ which controls how fast the spatially 

correlated data decays with the distance. θ is called a 

2 

‘Smoothness parameter’ which controls the geometrical 

properties of wireless sensor field. 

If ρ[ S , S ] ≥ τ , Data are strongly correlated in 

i j 

spatial domain for nodes Si and Sj . 

If ρ[ Si, S j] 

< τ , Data are weakly correlated in spatial 

domain for nodes Si and Sj. 

From equation no. (4), (5) and (6), we can derive the 

correlation coefficient ρ [ Si, S j] 

of data for nodes Si and 

Sj represented as follows:


θ2 

⎛d⎞ cov[ S , S ] − 

i j 

⎜ 

θ 

⎟ 

1 

ρ[ 

S 

i 

, S 

j 

] = = e 

⎝ ⎠ 

2 

σ 

Si 

(7) 

When the data are strongly correlated for nodes Si and 

Sj in the spatial domain we have 

θ2 

⎛ d ⎞ 

cov[ Si, S ] −⎜ ⎟ 

j ⎝θ1⎠ [ Si, Sj] e 

2 

σ i 

S 

ρ = = ≥τ 

From the equation no (8), we can derive as following 

or 

or 

or 

θ2 

⎛ d ⎞ 

−⎜ ⎟ 

⎝θ1⎠ ≥ 

e τ 

2 

d 

log( ) 

1 

θ 

⎛ ⎞ 

−⎜ ⎟ ≥ τ 

⎝θ⎠ 2 

d 1 

log 

1 

θ 

⎛ ⎞ ⎛ ⎞ 

⎜ ⎟ ≤ ⎜ ⎟ 

⎝θ⎠ ⎝τ⎠ d 

2 

2 2 ⎛ 1 

θ ⎛ ⎞⎞ 

≤ θ 2 

1 log ⎜ ⎟ 

(8) 

⎜ 

τ 

⎟ 

⎝ ⎝ ⎠⎠ 

(9) 

where the Euclidean distance between the node Si (xi, yi) 

and node Sj (xj, yj) as follows 

d = ( x − x ) + ( y − y ) 

2 2 2 

i j i j 

Put the value of 

2 

d in equation no. (9) ,we get 

2 2 2 ⎛ 1 

θ ⎛ ⎞⎞ 

( x ) ( ) 2 

i − xj + yi − yj ≤θ1⎜log⎜ ⎟ 

τ 

⎟ 

⎝ ⎝ ⎠⎠ 

(10) 

Compare equation no. (10) with equation of circle 

with cluster head at the centre with the radius of the 

cluster r, we get 

2 2 2 

( x − x ) + ( y − y ) = r 

(11) 

i j i j 

From equation no. (10) and (11) , we get 

r 

θ 

θ ≤ 2 

1 log ⎜ ⎟ 

2 2 

2 

⎛ ⎛1⎞⎞ ⎜ 

τ 

⎟ 

⎝ ⎝ ⎠⎠ 

(12) 

The equation no. (12), shows the relation between the 

radius of the cluster and the threshold value of spatially 

correlated data. The radius of the cluster depends upon 

the threshold value τ , θ1 andθ 2 .If the value of threshold 

τ increases, the radius of the cluster from the CH node 

located at the centre of the cluster get decreases. So we 

have taken the appropriate value of θ 1 , θ2 and the 

threshold valueτ to maintain a good correlation of data 

between sensor nodes for the clusters. 


2 

B. Distributed Cluster Formation in Spatial Domain 

We consider a square field of area with Z=Z1 x Z2 

where the cluster head (CH) node are deterministically 

deployed uniformly and the normal nodes are deployed 

randomly in the sensor field Z which form the distributed 

cluster. Since the number of cluster head node deployed 

in the sensor region is known, we get the same number of 

clusters as the number of cluster head nodes. We are 

interested in measuring the moisture content profile in 

each cluster embedded in the sensor field Z. Thus we 

assume that every cluster has a single tracing point. Every 

cluster in the sensor field is responsible for sensing and 

measuring the physical phenomenon of data for the 

tracing point value. The highly correlated data among the 

sensor nodes and the CH node forms the cluster. The CH 

node located at the centre of each cluster performs the 

estimation of data accuracy and finally send the data to 

the sink node. The number of tracing points is equal to 

the number of cluster head nodes. Hence in our model 

numbers of sensor (normal) nodes are considered to be 

more than the number of cluster head nodes. 

=Tracing point in each distributed cluster 

=Cluster head node in each distributed cluster 

Figure 1: Overlapping clusters in sensor region 

Thus in the square sensor field Z ,every cluster are 

embedded in the sensor field which are capable to sense 

their respective tracing point (to measure the high 

variation of correlated data ) distributed uniformly as 

shown in Figure-1.Thus the known number of clusters 

formed in the sensor region Z can be represented as N as 

follows 

⎛⎢Z1 ⎥⎢Z 2 ⎥ ⎢Z1 ⎥⎢Z 2 ⎥⎞ 

⎜⎢ 1 1 

2r ⎥⎢2r ⎥ 

+ 

⎢ 

+ + ≤ 

2r ⎥⎢2r ⎥⎟ 

Number of 

⎝⎣⎦⎣ ⎦ ⎣ ⎦⎣ ⎦⎠ 

⎛ ⎡ Z1 ⎤⎡Z 2 ⎤ ⎡Z1 ⎤⎡Z 2 ⎤⎞ 

Clusters ≤ ⎜ ⎢ 

1 1 

2r ⎥⎢2r ⎥ 

+ 

⎢ 

+ + 

2r ⎥⎢2r ⎥⎟ 

⎝ ⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥⎠ 

Since the sensor field is square, Z1=Z2=W


⎢ ⎥ 

⎢ ⎥ 

⎢ 2 

⎥ 

⎢ W W ⎥ 

+ + 1 

⎢ ⎛ 2 ⎞ θ 1 

⎥ 

⎢ 2θ 1 

2 2 θ 2 

⎛ ⎞ 

⎜ ⎛ ⎛ ⎞⎞ 

log ⎥ 

θ 

1 

log ⎟ 1 ⎜ ⎟ 

⎢ ⎜ ⎜ ⎜ ⎟ 

τ 

τ 

⎟ ⎟ 

⎝ ⎠ ⎥ 

⎢ ⎜ ⎝ ⎝ ⎠⎠ 

⎟ 

⎣ ⎝ ⎠ 

⎥⎦ 

≤ Number _ of 

⎡ ⎤ 

⎢ ⎥ 

⎢ ⎥ 

⎢ W 2 

W 

⎥ 

Clusters≤ 

⎢ + + 1 ⎥ 

⎢ ⎛ 2 ⎞ θ 

2 

1 

2 θ 

⎛ ⎞ ⎥ 

⎢ ⎜ 2 ⎛ ⎛ 1 ⎞ ⎞ ⎟ θ log 

2 θ log 

1 ⎜ ⎟ 

τ 

⎥ 

⎢ ⎜ 1 ⎜ ⎜ ⎝ ⎠ 

τ ⎟ ⎟ ⎟ 

⎝ ⎝ ⎠ ⎠ 

⎥ 

⎢ ⎜ ⎟ 

⎢ ⎝ ⎠ 

⎥ 

(13) 

The equation no. (13) shows the relation between the 

number of clusters and the threshold used for data 

correlation. If the threshold increases, the number of 

clusters with in the sensor field will get increases and 

vice versa. Thus we should choose appropriate threshold 

for clusters to perform data correlation in the spatial 

domain. Since the data are spatially correlated among the 

sensor nodes, there exist overlapping of clusters in the 

sensor region Z as shown in Figure 1. Equations no (12) 

and (13) derives how the clusters are overlapped among 

them in the sensor region Z. Hence it is important to find 

a distributed algorithm for clusters that can separate out 

the clusters from each other in the sensor region. 

Overlapping of cluster can sense the same correlated data 

among the sensor nodes and send the overlapped data to 

the sink node. It is like utilizing the same resource among 

the sensor nodes .Hence it leads to wastage of energy 

resource among the clusters and increases the data 

redundancy. Here we propose a distributed algorithm for 

cluster to overcome this problem for spatially correlated 

data and form non-overlapped irregular clusters in the 

sensor region Z. 

______________________________________________ 

Algorithm I: Distributed clustering algorithm for 

spatially correlated data in sensor field Z. 

______________________________________________ 

• Let U be the set of cluster head (CH) nodes 

deterministically deployed uniformly in sensor 

region Z. 

• Let V be the set of sensor (normal) nodes 

randomly deployed in sensor region Z. 

• Let d(a,b) be the Euclidian distance between 

node a and b. 

• Let dv be the distance from node v to the nearest 

CH node. 

• Initialize dv= ∞ 

• Initialize CHv=0 

• for v ∈V 

• for u ∈U 

• if d(v,u)


Node 

2 

Here we consider the mathematical analysis for data 

accuracy for the single cluster in the sensor region Z. 

Thus every cluster distributed in the sensor region can 

verify its data accuracy before data aggregation at the CH 

node. Once this procedure is being done by CH nodes for 

all the distributed clusters, they send the data to the sink 

node. Each sensor node i in the distributed cluster M can 

measure and observe the physically sensed data Si for 

tracing point S with observation noise Ni. Hence the 

observation and measurement made by the sensor node i 

in a given cluster is given by 

X = S + N where i ∈ M (14) 

i i i 

The sensor node i can sense and measure the observe 

sample Xi and transmits Xi to cluster head node sharing 

wireless additive white Gaussian noise (AWGN) channel 

[9,14]. Hence the observation and measurement received 

by the CH node from other sensor nodes in the cluster 

with transmission noise N over the AWGN channel is 

ti 

given by 

Node 

1 

Node 

m-1 

S 

Y = X + N 

2 2 t2 

Y = X + 

1 1 

N 

t1 

X 

CH 

= S 

CH 

+ N 

CH 

Y = X + 

m−1 m−1 

N 

tm 

Figure 2: Data accuracy model for distributed cluster 

Y = X + N = S + N + N 

i i ti i i ti 

Where i ∈ M and i ∉CH (15) 

We adopt uncoded transmission with finite number of 

sensor nodes for optimal point-to-point transmission [10] 

and consider the encoding power constraint value P, the 

measured value received by the CH are given by 

P 

(16) 

Z = Y = α( 

S + N + N ) 

i 2 2 2 i i i ti ( σ + σ + σ ) 

Si Ni Nti 

where i ∈ M and i ∉CH 

P 

and α = 2 2 2 

( σ + σ + σ ) 

Si Ni Nti 

CH node can sense and measure the tracing point S by 

finding the estimate of each physical phenomenon Si for 

node i. We take minimum mean square estimation 

(MMSE) for optimal decoding phenomenon [15] for 

uncoded transmission .CH node can find the MMSE for 

sensing and measuring the physical phenomenon Si 


−1 

CH 

extracted by sensor node i with observed sample Zi 

represented as 

ˆ 

E[ S Z ] 

S = Z 

i i 

i 2 

E[ Z ] i 

i 

where i ∈ M and i ≠ CH (17) 

Since the sensor node i can sense and measure the 

physical phenomenon Si of S , we take independent 

identically distributed (i.i.d) Gaussian random variable 

2 

2 

σ i.e E[S]=0 , var[S]= σ 

S 

S 

with zero mean and variance 

for tracing points . Similarly for sensing and measuring 

2 

phenomenon of Si, we assume E[Si]=0 , var[Si]= σ . 

We also have taken the observation noise Ni and 

transmission noise N ti with an independent identically 

distributed Gaussian random variable with variances 

2 

σ 

2 

, σ respectively with zero means. 

Ni Nti 

Hence E[Ni]=0 , 

E[ N ]=0,var[Ni]= 2 

σ ,var[ N ]= 2 

σ respectively. 

t i 

Thus, 

N i 

2 

= i i Si ESZ [ ] ασ 

t i 

N ti 

2 2 2 2 2 

= + + 

i Si Ni Nti 

EZ [ ] α ( σ σ σ ) 

Thus the estimation of ˆ Si is given by 

2 

σ Si = + + 

i 2 2 2 i i ti 

( σ + σ + σ ) 

Si Ni Nti 

Sˆ ( S N N ) 

2 

Si + 

2 

σ Si 2 

Ni + 

2 

Nti 

β = i 

( σ σ σ ) 

where i ∈ M and i ∉ CH 

Si 

(18) 

for 0< β


Thus the estimation of ˆ SCH is given by 

Where β 

σ 

Sˆ ( S N ) 

CH 

= 

2 

S CH 

2 2 

( σ + σ 

SCH NCH 

) 

CH 

+ 

CH 

CH 

= 

σ 

2 

SCH 2 

S CH 

+ 

2 

N CH 

( σ σ ) 

(21) 

for 0< β


0 at d= ∞ . We have taken power exponential model [17] 

i.e. 

PE . 

V i, j 

K ( d ) = 

e 

θ 

−( 

d 2 

i, j/ 

θ1) 

, θ θ 1 2 

> 0; ∈ (0, 2] θ is the 

1 

‘Range parameter’ and θ is the ‘Smoothness parameter’. 

2 

Using (15) and (23) in (25), we perform the normalized 

data accuracy with spatial correlation model for every 

distributed cluster in the sensor region given as follows: 

1 ⎡ M −1 

θ 

( / ) 2 

θ ⎤ 

−d ( / ) 2 

( ) (2 ( Si , ) 

θ 

1 

−d 

D M 1) 2 

( SCH , ) 

θ 

= ⎢β 1 

A 

i ∑ e − + βCHe 

⎥ 

m ⎢⎣ ⎥ 

i = 1 

⎦ 

⎡ ⎤ 

1 ⎢ ⎥ 

− 

M−1M−1 θ M−1 

( / ) 2 

θ 

( / ) 2 

2 ⎢ββ ( −d (,) ij 

θ 

1 

−d 

i 1) (2 

( CHi ,) 

θ 

1 

i∑∑e −+ βCH βi∑e + βCH) 

⎥ 

⎢ ⎥ 

⎣ i= 1 j≠ i i= 

1 

⎦ 

m 

Node 

1 

ρ( 

S , S ) 

CH 

1 

ρ( 

S , S ) 

i j 

ρ( 

S , S ) 

CH 

CH 

Node 

2 

ρ( 

SS , ) 

(26) 

The equation no. (26) shows that the normalized data 

accuracy D ( M ) for each cluster depends upon m sensor 

A 

nodes and factors i β and β respectively. Since we get 

CH 

a normalized data accuracy at each CH node for each 

cluster, we construct a spatial correlation model given by 

equation no. (26) for each individual distributed cluster in 

the sensor region. The spatial correlation model for each 

distributed cluster can be explained as follows: 

� Each sensor node i can sense a tracing point 

S in each distributed cluster where i ∈ M 

and i ∉CH node 

� CH node itself can sense the tracing point S 

in each distributed cluster. 

CH 

ρ( 

S , S ) 

i j 

ρ( 

S , S ) 

ρ( 

S , S ) 

i j 

CH 

ρ( 

SS , ) 

1 

ρ( 

SS , ) 

2 

Node 

3 

ρ( 

SS , ) 

3 

= Data sensed by node i from point event 

= Spatial data correlation between node i ,j 

= Data transmitted to the CH node 

= Data transmitted to sink node 

Figure 3: Spatial correlation model for distributed cluster 


2 

Sink 

3 

S 

� A spatial correlation between node i, j in 

each distributed cluster where i,j ≠ CH 

node. 

� Each sensor node i transmits the sensed data 

to the CH node in each distributed cluster 

where i ∈ M and i ∉ CH. 

Thus each distributed cluster formed in the sensor 

region has different set of sensor nodes. Hence each 

cluster can perform the normalized data accuracy at the 

CH node before data aggregation. The purpose of 

verifying the data accuracy for each cluster is to confirm 

that the most accurate data send by m set of sensor nodes 

can aggregate at the CH node rather than aggregating all 

the redundant data at the CH node. To visualize the 

correlation model for distributed cluster, we take an 

example where m=4 sensor nodes and out of m sensor 

nodes one node is chosen as a CH node as shown in 

Figure 3. Once we estimate the data accuracy at the CH 

node for each distributed cluster, the most accurate data 

get aggregated and finally send to the sink node. 

III. SIMULATION RESULTS 

In the first simulation setup , twenty five CH nodes are 

deterministically deployed uniformly and hundred sensor 

(normal) nodes are deployed randomly in a wireless 

sensor field of 120 m X 120 m based sensor topology as 

shown in Figure 1 . Each CH node performs the data 

accuracy for their respective cluster. Hence each cluster 

can sense and measure a single tracing point randomly 

located in each cluster region. Once each cluster can 

sense and measure their respective tracing point, it 

performs the data accuracy at CH node and finally 

transmits the data to the sink node. 

#CH Associated Nodes (Normal Nodes) Data 

Nodes 

Accuracy 

CH1 2 21 59 78 

0.837847 

CH2 1 6 7 8 13 35 43 76 92 

93 

0.843960 

CH3 11 17 22 69 84 98 0.866797 

CH4 10 46 62 

CH5 4 40 58 87 

CH6 15 25 32 33 53 73 81 

CH7 36 41 57 61 74 80 83 

CH8 13 19 28 31 49 85 95 

CH9 9 29 37 38 

0.694458 

0.833017 

0.820673 

0.862045 

0.793657 

0.882088 

CH10 20 23 63 75 77 79 88 91 97 0.857425 

CH11 44 51 66 86 99 

CH12 45 50 55 89 

CH13 5 18 24 47 48 52 82 

CH14 27 30 34 39 71 100 

0.820772 

0.809979 

0.813055 

0.756650


CH15 26 60 72 

CH16 70 

CH17 65 94 

CH18 68 90 

0.787127 

0.714302 

0.854421 

0.873163 

CH19 42 0.705224 

CH20 56 0.759352 

CH21 67 0.730805 

CH22 12 0.799681 

CH23 96 0.739846 

CH24 14 16 0.894157 

CH25 54 64 0.843685 

Table 1: Data Accuracy for each distributed cluster 

According to our proposed algorithm-I discuss 

previously, each CH node can form the cluster with their 

associated sensor nodes. Once the sensor nodes are 

associated with each CH node, they form distributed 

clusters in the sensor region Z. Thus twenty five CH 

nodes can form twenty five individual non-overlapping 

distributed clusters. Each distributed cluster can perform 

the data accuracy at their respective CH node as shown in 

Table-1. Similarly in the second simulation set up as 

shown in Table-2, we perform hundred runs for each CH 

nodes associated with their respective sensor nodes and 

find their average data accuracy for each cluster. 

#CH 

Nodes 

Average 

Data Accuracy 

#CH 

Nodes 

Average 

Data Accuracy 

CH1 0.8494 CH14 0.7327 

CH2 0.8731 CH15 0.7778 

CH3 0.8765 CH16 0.9662 

CH4 0.8734 CH17 0.8001 

CH5 0.8468 CH18 0.7706 

CH6 0.8401 CH19 0.9662 

CH7 0.8364 CH20 0.8111 

CH8 0.7975 CH21 0.8135 

CH9 0.9033 CH22 0.9736 

CH10 0.8615 CH23 0.9047 

CH11 0.7942 CH24 0.8343 

CH12 0.8171 CH25 0.8352 

CH13 0.7796 

Table 2: Average Data Accuracy for each distributed cluster 


In the third simulation set up, we take a single circular 

cluster of m=4 sensor nodes which can sense and 

measure a tracing point. We put m sensor nodes in a 

deployed circular cluster and a tracing point S located at 

the centre of the deployed circular cluster. i.e dS,i (where 

i=1,2,3)and dS,CH are equidistance as shown in the Figure- 

4. Here we have fixed the number of m sensor nodes and 

vary the distance from the tracing point S to m sensor 

nodes. As we increase the radius of the deployed circular 

cluster for dS,i and dS,CH with same proportion , D ( M ) A 

decreases i.e. the distance from the tracing point S to the 

m sensor nodes increases as shown in Figure 5. We put 

θ = {50,100} and θ =1 for our statistical data 

1 

2 

performance for the normalized data accuracy DA( M ) . 

Data Accuracy D A (M) 

Node 

3 

0.96 

0.94 

0.92 

0.9 

0.88 

0.86 

0.84 

0.82 

0.8 

CH 

Node 

Node 

2 

Node 

1 

Figure 4: Deployed sensor nodes in circular cluster topology 

S 

m=4 ,θ 1 =50 

m=4,θ 1 =100 

0.78 

1 2 3 4 5 6 7 8 9 10 

Radius of the deployed circle 

Figure 5: Data accuracy versus radius of the circular cluster 

In the fourth simulation setup, the distance from the 

tracing point S to m sensor nodes is fixed in the deployed 

circular cluster of radius =5 metre. We increase the 

number of sensor nodes with a fixed distance from the 

tracing point S i.e we increase m sensor nodes with fixed 

deployed circular cluster of radius 5 metre. At first, we 

put m=2 (one CH node and one sensor node) which


shows that the data accuracy is very poor with its value in 

between 0.6 to 0.75 for θ ={50,100,200,400}.The reason 

1 

is that there is only one sensor node which shows that the 

third condition of spatial correlation model given in 

section II(D) doesn’t satisfies the DA( M ) at the CH node . 

But if we put m=3 (one cluster head and two sensor 

nodes), there is a drastic improvement of DA( M ) since all 

the conditions for spatial correlation model are satisfied. 

The Figure-6 also shows that five to eight nodes are 

sufficient to perform the D ( M ) for the cluster, if the 

A 

distance from tracing point to m sensor nodes with 

deployed circular cluster of radius is 5 metre. 

For the simplicity of our model, we perform the fifth 

simulation set up where we have simulated a wireless 

sensor field (900 metre 2 ) of 5m X 5m grid based single 

cluster topology with a fixed tracing point (S) at the 

centre and a CH node on the corner edge with 47 sensor 

nodes distributed uniformly in the grid based cluster 

topology as shown in Figure 7.Our assumptions is that 

cluster of m sensor nodes are in the sensing range of the 

tracing point (S). Initially we put m=4(one cluster head 

node and three sensor nodes located at the four extreme 

corner of sensor field).We verified that D ( M = 4) is 

A 

0.6333 when θ =50 as shown in Figure 8. If we increase 

1 

θ = 400, then D ( m = 4) =0.911. 

1 

Data Accuracy D A (M ) 

1 

0.95 

0.9 

0.85 

0.8 

0.75 

0.7 

0.65 

A 

θ 1 =50 

θ 1 =100 

θ 1 =200 

θ 1 =400 

2 3 4 5 6 7 8 9 10 11 12 

Number of Sensor Nodes 

Figure 6: Data accuracy versus number of sensor nodes in a 

cluster 

This shows that θ control as how fast the spatially 

1 

correlated data decays with distance between sensor 

nodes and the tracing point. Hence it is always suitable to 

take the value of θ large for large sensor field to get 

1 

DA( M ) in an efficient way. Now we increase cluster of m 

sensor nodes with increment of four sensor nodes every 

time concentrating towards tracing point till m sensor 

nodes are able to sense and measure the tracing point S in 


the region. As we increase the sensor nodes, the data 

accuracy DA( M ) also get increases. Hence for 900 metre 2 

sensor field, 15 to 20 sensor nodes are sufficient to give 

DA( M ) of 0.944 for θ =400 and 1 

DA( M ) remains 

approximately constant still we increase the number of 

sensor nodes for the cluster. We plot in the Figure-8 for 

the DA( M ) versus node density for a cluster. Node density 

is defined as the number of sensor nodes per unit area in a 

single cluster. Hence it is needless to choose so many 

sensor nodes to achieve data accuracy for the cluster in 

sensor field to sense and measure a tracing point. 

Data Accuracy D A (M ) 

1 

0.95 

0.9 

0.85 

0.8 

0.75 

0.7 

0.65 

CH node 

θ 1 =50 

θ 1 =100 

θ 1 =200 

θ 1 =400 

0 0.01 0.02 0.03 

Node density 

0.04 0.05 0.06 

Figure 8: Data accuracy vs. node density in a single cluster 

In the sixth simulation setup, we take a single cluster of 

m sensor nodes randomly deployed in a region (30 X 30 

= 900 metre 2 ) that sense and measure a tracing point. We 

fix the tracing point at x,y (15,15) coordinate and CH 

node at x,y (0,0) coordinate with 99 sensor nodes 

S 

Figure 7: Sensor nodes deployed in grid topology


randomly deployed in the region. For each run we verify 

DA( M ) with respect to randomly deployed cluster of m 

sensor nodes. Finally we verify for 100 runs and find the 

average DA( M ) for the cluster of m sensor nodes. Figure 9 

shows that if the value of θ =400, 1 DA( M ) is 0.944 for 10 

to 15 sensor nodes. If we continuously increase the 

number of sensor nodes the DA( M ) remains approximately 

same. Hence it is useless to deploy sensor nodes beyond 

15 sensor nodes because 10 to 15 sensor nodes are 

sufficient to give approximately the same DA( M ) for the 

cluster with 1 θ =400. Again if we constantly increasesθ , 1 

average DA( M ) also get increases for the cluster of m 

sensor nodes. But after certain approximate value of θ 1 

the DA( M ) remains approximately constant for the cluster. 

If we continuously increase the value of θ the average 

1 

DA( M ) remains approximately constant since it achieve the 

saturation level in the cluster. Finally the output graph 

shows distortion in the signal due to additive white 

Gaussian noise components. 

Average Data Accuracy D A (M) 

1 

0.95 

0.9 

0.85 

0.8 

0.75 

0.7 

0.65 

0.6 

θ 1 =50 

θ 1 =100 

θ 1 =200 

θ 1 =400 

0.55 

0 10 20 30 40 50 60 70 80 90 100 

Number of Sensor Nodes 

Figure 9: Average data accuracy versus number of sensor nodes in a 

single cluster 

Since the data are spatially correlated in the sensor 

region, we propose a distributed algorithm with non 

overlapping irregular cluster for the spatially correlated 

data in the sensor region. Each distributed cluster can 

perform DA( M ) before data aggregation at their respective 

CH node. Hence it is important to sense and measure the 

most appropriate (accurate) data send by each distributed 

cluster at the CH node rather than aggregating all the 

redundant data at their respective CH node. Thus it can 

reduce the data redundancy. Since the data accuracy is 

performed by each distributed cluster, we verified from 

the simulation results that there exists a minimal set of 

sensor nodes with optimal cluster which is sufficient to 


give approximately the same DA( M ) as achieved by the 

each distributed cluster. Therefore the time complexity 

done at each CH node of respective distributed cluster for 

aggregating the most accurate data send by their 

respective optimal cluster will be less. Thus we find an 

optimal cluster from each distributed cluster which can 

reduce the data redundancy and communication 

overhead. 

In the fifth simulation setup, a grid based single cluster 

is formed where we deployed m=48 sensor nodes 

uniformly. We examine that 15 to 20 nodes are sufficient 

to perform DA( M ) =0.944 for 1 θ =400 in 900 metre2 cluster 

region. Similarly in sixth simulation setup a cluster with 

m=100 sensor nodes are randomly deployed in 900 

metre 2 region and we get 10 to 15 sensor nodes are 

sufficient to perform DA( M ) =0.944 for θ =400. Therefore 

1 

it is unnecessary to choose so many sensor nodes in 900 

metre 2 region as DA( M ) remains approximately same as it 

achieve the saturation level still we increase m sensor 

nodes in the cluster. Hence we have P minimal set of 

sensor nodes with optimal cluster which is sufficient to 

give approximately the same DA( M ) by M set of sensor 

nodes in each distributed cluster as shown by Venn 

diagram in Figure 10. 

Figure 10: Venn diagram for optimal cluster in each distributed cluster 

IV . CONCLUSIONS 

In this paper we investigate that the data are spatially 

correlated among sensor nodes and form clusters in the 

sensor region. Since the data are highly correlated in the 

spatial domain, the sensor nodes form regular 

overlapping clusters among them in the sensor region. 

Overlapping of cluster can sense and measure the same 

correlated data among the clusters. Thus to overcome this 

situation, we constructed a distributed clustering 

algorithm with data accuracy model .We perform data 

accuracy for each distributed cluster. We find that the 

most accurate data send by the distributed cluster can 

aggregate at the CH node rather than aggregating all the 

redundant data at their respective CH node. We 

demonstrate by simulation that the data accuracy for a 

single cluster depend on number of sensor nodes and their 

exist an optimal cluster which is adequate to sense and 

measure the tracing point to perform approximately the 

same data accuracy level achieve by single cluster. 

Finally we conclude that the data accuracy performed for 

each distributed cluster can reduce the data redundancy 

and communication overhead. 

M 

P


REFERENCES 

[1] I.F Akyuildz ,W.Su , Y. Sankarasubramanian and E. 

Cayirci, “A survey on sensor Networks ”,IEEE 

Communcations Magazine, vol .40 , pp.102-114 ,Aug 

2002. 

[2] S.S. Pradhan , K. Ramchandran ,“Distributed Source 

Coding : Symmetric rates and applications to sensor 

networks”, in procecding of the data compressions 

conference 2000,pp.363-372. 

[3] A. Abbasi and M. Younis , “A survey on clustering 

algoirthms for wireless sensor networks ” Computer 

communications , vol-30 , n-14-15 ,pp-2826-2841, 2007. 

[4] W.B Heinzelman , Anantha P. Chandrakasan, “An 

Application Specific Protocl Architecture for Wireless 

Microsensor Networks ” , IEEE transactions on wireless 

communications , vol. no 4 , Oct-2002. 

[5] Georgios Smaragdakis, Ibrahim Matta ,Azer Bestavros , “ 

SEP : A stable Electon Prtocol for cluster hetrerogenous 

wireless sensor networks ” 

[6] Chongqing Zhang , Binguo wang , Shen Fang , Zhe Li , “ 

Clustering Algorithms for wireless sensor networks using 

spatial data correlation ”, International conference on 

information and Automation , pp-53-58 ,june 2008. 

[7] Zhikui chen , Song Yang , Liang Li and Zhijiang Xie , “ A 

clustering Approximation Mechinism based on Data 

Spatial Correlation in Wireless sensor Networks “, 

Proceedings of the 9 th international conferenses on 

wireless telecommunication symposium -2010. 

[8] Ali Dabirmoghaddam , Majid Ghaderi , Carey Williamson 

, “Energy Efficient Clustering in wireless Sensor 

Networks with spatially correlated dara “ IEEE infocom 

2010 proceedings. 

[9] Kang Cai, Gang Wei and Huifang Li,“Information 

Accuracy versus Jointly Sensing Nodes in Wireless 

Sensor Networks” IEEE Asia Pacific conference on 

curcuit and systems 2008 ,pp.1050-1053. 

[10] M.Gastpar, M. Vetterli, “ Source Channel Communication 

in Sensor Networks ”, Second International Workshop on 

Information Processing in Sensor Networks (IPSN’2003). 

[11] Varun M.C,Akan O.B and I.F Akyildiz, “ Spatio- 

Temporal Correlation : Theory and Applications Wireless 

Sensor Networks” , Computer Network Journal (Elsevier 

Science ), vol. 45 , pp.245-259 , june 2004. 

[12] Jyotirmoy karjee , H.S Jamadagni , “Data Accuracy 

Estimation for Cluster with Spatially Correlatd Data in 

Wireless Sensor Networks ” ,to be published in the 

proccedings ICISCI-2011, Harbin ,China 

[13] Huifang Li, Shengming Jiang ,Gang Wei ,“Information 

Accuracy Aware Jointly Sensing Nodes Selection in 

Wireless Sensor Networks ”,MSN 2006 , LNCS 4325 , 

pp.736-747. 

[14] T.J. Goblick ,“ Theoritical Limitions on the transmission 

of data from analong sources”,IEEE Transaction Theory , 

IT-11 (4) pp.558-567 ,1965. 

[15] V.Poor ,“ An Introduction to Signal Detection and 

Estimation ”,Second edition , Springer ,Berlin 1994. 

[16] J.O. Berger , V.de Oliviera and B.Sanso ,“ Objective 

Bayesian Anylysis of Spatially correlated data 

”J.Am.Statist. Assoc. Vol-96,pp.1361-1374,2001. 

[17] De Oliveira V, Kedan B and Short D.A , “ Bayesian 

predication of transformed Gaussian random fields” 

Journal of American statistical Association 92, pp.1422- 

1433. 

[18] L.Guo , F chen , Z Dai , Z. Liu „”Wireless sensor network 

cluster head selection algorithm based on neural 


networks” , PP-258-260 , International conference on 

Machine vision and human machine interference, 2010. 

[19] T.Minming , N Jieru , W Hu , Liu Xiaowen “ A data 

aggregation Model for underground wireless sensor 

network” Vol-1, pp-344-348 , WRI world congress on 

computer science and information engineering, 2009 . 

[20] Jyotirmoy karjee , Sudipto Banerjee , “ Tracing the 

Abnormal Behavior of Malicious Nodes in MANET ”, 

Fourth International conference on wireless 

communications , networking and Mobile Computing 

,pp-1-7 Dalian-china -2008 . 

[21] C.Y. cho , C.L Lin , Y.H Hsiao , J S wang , K.C yong “ 

Data aggegation with spatially correlated grouping 

Techninques on cluster based WSNs” , SENSORCOMM 

,pp-584-589, venice- 2010. 

[22] Shirshu Varma , Uma shankar tiwary , “ Data Aggregation 

in Cluster based wireless sensor Networks ”Proceedings 

of the first International confernce on Intelligent human 

computer interaction , page-391-400 , part-5 , 2009. 

Jyotirmoy Karjee received his B.E 

(Electronics), M.E (Information 

Technology) specialization in Network 

Security in 2003 and 2005 respectively. 

He worked in Prakriti Inbound Pvt. Ltd 

as a software engineer for a year and 

worked as a lecturer in Sikkim Manipal 

Institute of Technology, Sikkim till 

2008. He is currently pursuring his 

Ph.D degree at Centre for Electronics 

Design and Technology, Indian Instutute of Science, Bangalore. 

His current research interests include data accuracy estimation 

and data aggregation in wireless sensor networks. 

Prof. H.S Jamadagni received his 

M.E and Ph.D degree in Electrical & 

Communication Engineering from 

Indian Institute of Science ,Bangalore. 

Currently He is the professor at Centre 

for Electronics Design and 

Technology, Indian Institute of 

Science. He is one of the main 

coordinators for the intel higher 

education program and was the key mentors for various intel 

workshops in india. His current research work includes in the 

areas of embedded systems, VLSI for wireless networks and 

wireless sensor networks.


Networking as a Service: a Cloud-based Network 

Architecture 

Tao Feng, Jun Bi, Hongyu Hu and Hui Cao 

Network Research Center, Tsinghua University 

Department of Computer Science, Tsinghua University 

Tsinghua National Laboratory for Information Science and Technology (TNList), Beijing, China 

fengt09@mails.tsinghua.edu.cn, junbi@tsinghua.edu.cn, huhongyu@cernet.edu.cn and cao-h06@mails.tsinghua.edu.cn 

Abstract—With the rapid development and integration of 

the Internet, wireless communication network and the 

Internet of Things, the Internet faces many challenges as a 

bearer network: a large volume of information exchange, 

multi-level QoS and smoothly switching multiple access 

protocols. The Internet should be able to provide a variety 

of network capacities in a more dynamic and on-demand 

way, not just limited network resource provision through 

virtualization. The elastic network is expected to adapt to 

network changes by enabling network protocols selection 

and combination dynamically. Cloud computing illustrates a 

new Internet-based model of IT resources (hardware, 

software, data) provision, delivery and consumption as a 

service. Therefore, networking as a service can provide 

guaranteed quality of service and good quality of experience 

to users who do not care about any network configuration 

and network management. In this paper, we propose a novel 

idea of networking as a service by combining the service 

provision model of cloud computing with the openness of the 

network protocol. The related conception and stakeholders 

of networking as a service is depicted. Cloud-based network 

architecture is design to present the provision, delivery and 

consumption of networking as a service and discuss the key 

features of cloud-based network. Finally, a prototype of 

cloud-based network is implemented by extending 

OpenFlow architecture. 

Index Terms—network capacity; networking as a service; 

cloud computing; network architecture 


With the rapid development and integration of the 

Internet, wireless communication network and the 

Internet of Things, the Internet faces many challenges as 

a bearer network in the future: a large volume of 

information exchange, multi-level QoS, smoothly 

switching multiple access protocols, mobility and 

management. The design philosophy of the current 

Internet [12] limits the flexibility of the network 

architecture to meet new requirements. For instance, the 

end-to-end argument proposes that a network simply 

forwards packets between end-systems while complex 

data processing function is implemented on the end- 

This is the extended version of our paper at ICISCI’10. 

Corresponding author: Tao Feng, fengt09@mails.tsinghua.edu.cn 


doi:10.4304/jnw.6.7.1084-1090 

systems [33]. At present, the Internet is expected to 

handle more complex and customized forwarding 

capacity in the company of more and more mobile 

devices connected and new client/server paradigm of 

cloud computing. Even the current Internet provides a lot 

of new features or services that go beyond forwarding in 

order to deal with more tussles [13], such as 

heterogeneous network resources, personalized delivery 

service, trust network access and low-cost network 

maintenance, and this shift towards more network 

capacities will continue. 

New network features or services are difficult to be 

introduced into the current network because a network 

protocol is locked in a vendor device. Network service 

should be decoupled from specific data transport 

technologies so that new features or services can be 

deployed freely. In addition, the Internet should be able to 

provide a variety of network capacities in a more 

dynamic and on-demand way, not just limited network 

resource provision through virtualization [7]. The elastic 

network is expected to adapt to network changes by 

enabling network protocols selection and combination 

dynamically. Cloud computing [1] illustrates a new 

Internet-based model of IT resources (hardware, 

software, data) provision, delivery and consumption as a 

service. Therefore, network capacity on demand can 

provide guaranteed quality of service and good quality of 

experience to users who do not care about any network 

configuration and network management. 

There will be little place either for static network 

configurations like the current network stack or for 

manual optimization and tuning as enforced at the interlayer 

boundaries of the current network in such a 

dramatically dynamic network operational circumstance. 

On the opposite end, the revolution that removes such 

constraints and at the same time maximizes the ondemand 

capacity of network will play a key role in the 

evolution towards future network. 

In this paper, we propose a novel idea of networking as 

a service by combining the service provision model of 

cloud computing with the openness of the network 

protocol. The related conception and stakeholders of 

networking as a service is depicted. Cloud-based network 

architecture is design to present the provision, delivery 

and consumption of network protocol as a service and


discuss the key features of cloud-based network. Finally, 

a prototype of cloud-based network is implemented by 

extending OpenFlow architecture. 

II. RELEATED WORK 

The design of the current Internet architecture has been 

rethinking and some architectural principles for new 

Internet architectures have been proposed for some years 

on structuring a new generation of network protocols 

[14], adding mechanisms to the core of the Internet [15] 

and exploring specific architectural issues [16]. The trend 

of the Internet towards a commercial development has 

changed the underlying hypotheses of trust and economic 

incentives [17]. 

Several new features have been proposed and 

implemented in the Internet with the development of the 

Internet, which was not considered in the initial design of 

network architecture. Since the Internet does not support 

the dynamic deployment of new protocols, on-demand 

composition of network protocols and pay-as-you-go 

business model of network capacity, these features 

needed to be added as special processing functions to 

future network. Some of the current research gives a ray 

of hope for network capacity on demand. The following 

list highlights such features and functions: 

A. The openness of the network protocol 

The openness of the network protocol refers to the 

future that the introduction, deployment and operation of 

network protocol or service can be achieved on minimum 

cost by standardizing network interfaces and enhancing 

interoperability of network protocol. The openness of the 

network protocol is a prerequisite for network protocol as 

a service. The realization of the openness future makes 

the network protocol custom and flexible adoption 

according to different application scenarios and user 

requirements. Open architectures and analogous work on 

the openness of networks have contributed ideas for the 

programming behavior of a node [29, 30] and flow-driven 

modification of the data plane services [31]. To manage 

the complexity of new protocol in the network, a working 

group of IETF has attempted to define Open Pluggable 

Edge Services (OPES) [19]. In such architecture, a set of 

data flow operations that are implemented on nodes 

throughout the network can be specified in end-systems. 

Currently, there are two ways to achieve the openness of 

the network protocol in the control plane: the out-box 

openness and the in-box openness. OpenFlow [2] is one 

of the implementations of the out-box openness. 

OpenFlow provides a way to control network device by 

network protocols running outside of a network device. 

OpenFlow achieves a variety of network behaviors on the 

switch by controlling the flowtable such as routing, 

firewall, and so on. On the other hand, the JUNOS SDK 

[3] is another way to open network protocol. The JUNOS 

SDK enables developers to innovate on top of JUNOS 

and Juniper Networks platforms, so developers can 

create, deploy, and validate innovative applications 

tailored to specific needs. 


B. The modularity of the network protocol 

Modularity is central tenets in the design and 

implementation of hardware and software system. In the 

paper of [16], the modularity of the network architecture 

is defined that breaks a network system into parts, 

normally to permit independent construction and 

replacement, reuse of parts, and so on. Early works on 

modular protocols have provided some solutions on 

protocol decomposition, configurable frameworks and 

process model. [8] proposed an x-kernel environment and 

mechanisms for communication between microprotocols. 

[9] is a configurable communication 

framework that provides a runtime platform of protocols 

consisting of standard, reusable services. [10] proposed a 

process-per-protocol model, a process or thread 

shepherds a message through the protocol stack. Some 

work [11] has proved the success of a modular platform 

with widespread deployment for reconfiguration of the 

entire data plane of a router system. Active networks [18] 

provided a powerful and very general approach to module 

packet processing function. 

C. Service-oriented network protocol and network 

architecture 

The research on service-oriented network protocol 

composition and network architecture enables dynamic 

adjustment of network features possible according to the 

requirement of users and applications. Service-oriented 

network architecture provides mechanisms for composing 

custom protocol stack [23], such as the SILO architecture 

[22]. The key technologies on service-oriented network 

protocol include abstraction of network service, network 

protocol composition and service path selection. A 

number of previous research projects have addressed 

some general thinking about how to specify a network 

service. [20] emphasize specifically on the middle boxes 

in the network such as traversing firewalls and network 

address translators, it can be seen as a step towards 

managing connections involving general services. [21, 

35] provided a more general method that specifies 

services very similar to pipeline abstractions. Service 

socket [28] is a user-level abstraction that has 

implemented some networking applications and services 

in networks. Some approaches focus on the 

decentralization of service composition. The SpiderNet 

project [24] provides the ability of service composition 

by a decentralized approach in P2P networks. A similar 

research about service composition [25] discuss the 

challenges that how to support for the service compositio 

on top of the Internet Indirection Infrastructure (i3). In 

[26, 27], the path selection is also done in a distributed 

manner, but end-systems and other entities along the path 

may specify specific service requirements. 

D. The appropriate mechanism of the provision, delivery, 

and consumption of network protocol 

The appropriate mechanism of the provision, delivery, 

and consumption of network protocol is becoming an 

important foundation for network protocol as a service. A 

network protocol developed by JUNOS SDK developers 

needs to deploy and run on each network device of


Juniper. The development can gain development fees 

based on a software license. Cloud computing is different 

from JUNOS SDK. The devices and services are 

centralized deployment and running in data centers. 

Service developers can publish and sell their own 

software services to the cloud service provider. 

III. NETWORKING AS A SERVICE 

A. Conceptualisation 

Networking as a service refers to a new Internet-based 

model that communication service provider (CSP) can 

deliver network protocols on-demand and reliably to the 

user based on SLA. The service consumer can use the 

service as pay as you go and achieve a good quality of 

experience. 

From the perspective of service, the abstraction of 

network function and the layer of network protocol stack 

will be re-organized and divided into three layers: service 

specification, network capacity, network behavior. In this 

vision, network service in different abstract forms will 

regard as middle ground for the continuous resolution of 

tussles between providers and users. At the top level of 

abstraction the service specification which defines data 

transmission parameters of user information needs to 

satisfy end user requirements. At the intermediate level 

of abstraction network capacity in accord with service 

specification is set up with network protocol composition. 

The dynamics of network capacity construction provides 

a utility function for the composition of horizontal service 

across the network and vertical service within a node. In 

this case, it is the service that utilizes the network and 

drives the customization of network capacity. As we 

move to lower levels this customization process is 

mapped to network behavior, access technologies and 

resource management policies such as forwarding, 

filtering, dropping, and so on. 

Figure 1. Cloud-based network. 

Cloud-based network (CBN), as shown in Fig. 1, is a 

form of implementation of networking as a service, which 

learns from concepts and ideas of cloud computing and 

service-oriented architecture. CBN provides the ability to 


deploy and run network protocols in the cloud, configure 

network resources and compose network protocol 

dynamically according to the user's service requirements 

and SLA, accordingly generate network control rules to 

manipulate forwarding behavior of a network device. 

CBN transforms network protocol to network service 

with zero-configuration [4] and zero-maintenance for 

network users. Each CSP may set up a CBN or the 

federation of CBNs to serve for the costumers. 

Protocol service instance (PSI) is a set of network 

protocols corresponding to each service requirement. PSI 

is the minimum unit of a network service in CBN. 

B. Stakeholders in Cloud-based Network 

In the current Internet business model, network-related 

stakeholders consist of end users, communication service 

providers and network equipment providers formed. In 

this case, end users pay for communications service 

providers to apply for network access services by 

communications service providers. Communications 

service providers are regard as a “pipeline” manager 

since it is almost impossible that communication service 

providers can deploy a new protocol to provide 

customized or value-added network services because the 

protocols are embedded in the device by network 

equipment providers. The business model of the Internet 

will be changed with the emergence of networking as a 

service. Communications service providers will enhance 

the ability to control the network. The role of network 

equipment providers will be subdivided. Users will apply 

for appropriate network services according to their need 

and consume the service in the way of pay-as-you-go 

with business development, thereby reduce the cost of 

network investment and maintenance. 

Communication Service Providers: Communication 

service providers are responsible for the management and 

maintenance of network-based cloud and the provision of 

a guaranteed quality of network services to service 

consumers. CSP can gain service revenue from service 

consumers according to the period, quality, quantity and 

scale of network service. 

Protocol Developers: Protocol developers can develop 

various network protocols with API and specification of 

CBN. After passed a test, a network protocol can be 

published to the CBN. Protocol developers can gain 

license fees from CSP according to the scale of 

deployment and frequency of running of the network 

protocols. 

Network Equipment Providers: In the CBN, the 

various components and interfaces of network devices 

will be standardized. Thus, the function of network 

equipment providers will be refined and divided into 

network components providers and network equipment 

integrators. Network component providers will focus on 

improving the performance, capacity of network 

components, while network equipment integrators will 

focus on improving the stability and reliability of network 

equipment composed by network components provided 

by network component providers.


Figure 2. The reference architecture of Cloud-based network 

Network Service Consumers: Network service 

consumers, including personal and business users, 

purchase network services in a “pay as you go” model. 

They do not need to purchase expensive network 

equipment, do not need to care about the network 

configuration and maintenance. They do only need to put 

forward the requirements of network services according 

to the development of the business. And then use it. 

IV. CLOUD-BASED NETWORK REFERENCE ARCHITECTURE 

Cloud-based network reference architecture, as shown 

in Fig. 2, is divided into four layers: network resource 

pool, network operation interface, network runtime 

environment and network protocol service. Network 

operation interface is implemented in network device to 

manipulate network resource. Network runtime 

environment is a platform for network protocol 

deployment and operation. Network protocol service 

generates control rules and call for network operation 

interface to control and manage network resource. 

A. Network Resource Pool 

Network resource pool (NRP) is the network resource 

such as ports, bandwidth, queue, address, which can be as 

a basic service related with packet forwarding. Examples 

are Amazon EC2 for IP address and bandwidth 

assignment. Instead of some components of raw network 

hardware, NRP typically offers the combination of these 

resources as a service through unified configuration and 

management. 

B. Network Operation Interface 

Network operation interface (NOI) is open and 

standardized API in order to configure and manage NRP. 

NOI provides three types of operating functions: 

parameter configuration for the network resource, 

forwarding control and event report. Parameter 


configuration function provides to set or get the max or 

minimum bandwidth limit, the numbers of queues, and IP 

address of a port, which can construct a user-oriented 

network topology. Forwarding control function provides 

abilities to output, drop, and filter packets according to 

the rules. Event report function provides an alarm or trap 

information when some network resource is down or 

overload. 

C. Network Runtime Environment 

Each of protocol set is called protocol service instance 

(PSI) which can be set up and running as a plug-in in 

network runtime environment (NRE). There is always a 

daemon running as a default PSI that provides a basic 

network layer protocol such as IP. NRE is responsible 

for billing, resource allocation, assessment, interconnect 

and reliability assurance for each PSI. 

Scheduling: According to the network service requests 

and current SLA state of the user, the scheduling function, 

firstly, will search related network resources and network 

protocols. If the requests are satisfied, the scheduling 

function will reconfigure network resource properties and 

generates PSI to control the rule of packet forwarding by 

issuing to the network equipment. Meanwhile, the 

scheduling function will inform the user the service is 

working and start the billing. 

Plug-in: Plug-in function enables network protocols to 

deploy, start, stop, upgrade and uninstall without reboot 

the system, just like OSGi, which is a module system and 

service platform for the Java programming language that 

implements a complete and dynamic component model, 

something that does not exist in standalone Java/VM 

environments. Thus, network protocol in PSI can be 

dynamically adjusted and smoothly switched by Plug-in 

feature. 

Pricing: To regard network capacity as a service, there 

will be a new billing model instead of bit-per pricing or 

online-time pricing: network capacity-based pricing and 

flow-per pricing. Pricing function enables to calculate the 

cost of service consumers according to the amount of 

involved network protocols and run time of each PSI. 

Evaluation: Evaluation function provides the ability to 

monitor the operational status and network resource 

usage of PSI. The evaluation feature can determine 

whether the service provided by the PSI match the SLA 

through checking the service request. 

Interconnection: Interconnection function provides a 

communication mechanism and interface between multi 

PSIs. A PSI can send and receive the status information 

of a protocol through an interconnection interface such as 

JSON when it needs access or negotiates some 

information of protocols in another PSI. Eventually, 

network clouds can be interconnected by the multi PSIs 

interconnection. 

Migration: Migration function provides the mobility of 

network services in the PSI level against network failures 

and high load. The PSI migration process includes: PSI 

state capture, marshaling PSI state and PSI service 

relocation.


D. Network Protocol Service 

With the openness of network, a variety of new 

network protocol will be designed and implemented. 

How to identify and manage the new network protocols is 

a new problem in the future. Network protocol service 

consists of three functions: the description, management 

and composition of network protocol. 

Service Description of Network Protocol: Service 

description of network protocol is a structured language 

such as XML that provides a model for describing the 

capabilities of a network protocol. The NRE can choose 

appropriate network protocols to set up a PSI that can 

meet the user’s demands. Therefore, a network protocol 

should be able to accurately express properties and 

forward capacities of network protocol. 

Service Lifecycle Management of Network Protocol: 

The feature of service lifecycle management of network 

protocol provides the functions to manage network 

protocol versions, registration, certification and licensing. 

Service Composition of Network Protocol: The feature 

of service composition of network protocol provides the 

ability to generate new, more powerful network protocol 

service by composing protocols with different functions. 

The service composition of network protocols may learn 

from context aware service composition [5] and semantic 

web service composition [6]. 

V. IMPLEMENTATION 

OpenFlow is an ideal way to build a network cloud. 

OpenFlow [36] is an open standard that allows network 

researchers to run experimental protocols in production 

network. It provides an open protocol to program the 

flowtable in a network device. The protocols 

implemented in a server outside control network devices 

by OpenFlow protocol, which is embedded in a device 

currently. It is in the process of being implemented by 

major switch vendors and used today by universities to 

deploy innovative networking technology. Thus, 

openness of network protocol in OpenFlow provides the 

possibility of network protocols as a service. 

NOX [34] is an open-source OpenFlow controller 

intended to simplify the development of software for 

controlling or monitoring networks composed of 

OpenFlow switches. Programs written within NOX 

(using either C++ or Python) have flow-level control of 

the network. This means that they can determine which 

flows are allowed on the network and the path they take. 

In addition, NOX provides abilities to access to the 

network state including the network topology and the 

location of all detected hosts. 

Apache Hadoop is an open source distributed 

processing framework. The framework split dataset into 

manageable blocks in order to compute large datasets. It 

is in charge of the whole process by launching protocol 

instances, processing the protocol messages across many 

machines where the protocol is physically deployed and, 

at the end, aggregating the set of forwarding rules output 

into a final result [32]. 


Figure 3. An initial prototype of Cloud-based network. 

In our laboratory, an initial prototype of Cloud-based 

network, as shown in Fig. 3, has been implemented by 

extending OpenFlow architecture. The implementation is 

divided into two levels: controllers cloud plan and 

network data plan. Controllers cloud plan provides the 

functions of NRE and NPS in Cloud-based Network 

Reference Architecture. A master controller with Apache 

Hadoop is responsible for distributing the data stream to 

three slave servers with different protocols. The LAMP 

on the master controller is responsible for network 

protocol registration and lookup. Xen deployed on each 

slave server make multi slave server as a slave server 

cluster. Network data plan consists of six OpenFlow 

switches to receive control information of flows from the 

master server in controllers cloud plan. 

VI. FUTURE WORK 

In this article, we have demonstrated the conception 

and role of on-demand provision of network capacity in 

order to achieve networking as a service. We have 

designed a novel future network architecture leaned from 

cloud computing and service-oriented architecture: cloudbase 

network. 

Based on the cloud-based network, we have built a 

prototype by extending the OpenFlow architecture and 

virtualization technology to verify on-demand provision 

of network capacity. 

Moving forward, there are some future works in cloudbased 

network architecture. First of all, based on the 

above prototype implementation, we will evaluate the 

performance and latency of network protocol as a service 

in the cloud-based network. And we will research an 

accurate expression of the demand for network services. 

A formal network protocol description language will be 

designed to configure network services automatically 

according to an accurate expression of network service 

requirements. Then, we will extend the current prototype 

implementation to multi-clouds interconnection by 

designing a cloud interconnection communication


protocol. In the multiple clouds based network prototype, 

we will research the capability of service optimization to 

provide a service consumer with the nearest service 

delivery. Finally, we will research the migration ability of 

PSI in multiple clouds to improve the reliability of the 

cloud-based network. 


This work was supported by National Science 

Foundation of China under Grant 61073172, Program for 

New Century Excellent Talents in University, and 

National Basic Research Program ("973" Program) of 

China under Grant 2009CB320501. 

REFERENCES 

[1] NIST Definition of Cloud Computing v15, 

csrc.nist.gov/groups/SNS/cloud-computing/cloud-defv15.doc 

[2] N. McKeown, T. Anderson, H. Balakrishnan,G. Parulkar, 

L. Peterson, J. Rexford, S. Shenker, and J. Turner. 

Openflow: enabling innovation in campus networks. 

SIGCOMM Comput. Commun. Rev.,38(2):69–74, 2008. 

[3] J. Kelly, W. Araujo, and K. Banerjee, “Rapid service 

creation using the JUNOS SDK,” in ACM PRESTO, 2009. 

[4] S. Cheshire and D. H. Steinberg, “Zero Configureation 

Networking: The Definitive Guide,” O’Reilly, 2006,. 

[5] Vukovic M Context Aware Service Composition. PhD 

thesis, University of Cambridge, 2006 

[6] RAO, J. Semantic Web service composition via logicbased 

program synthesis. PhD thesis. Department of 

Computer and Information Science, Norwegian University 

of Science and Technology, 2004. 

[7] N. M. K. Chowdhury and R. Boutaba, “A Survey of 

Network Virtualization”, Technical Report, David R. 

Cheriton School of Computer Science, University of 

Waterloo, Waterloo, Ontario, Canada, Tech. Rep. CS- 

2008-25, Oct 2008. 

[8] N. Hutchinson, L. Peterson, The x-kernel: An architecture 

for implementing network protocols, IEEE Transactions on 

Software Engineering 17 (1991) 64–76. 

[9] M. Zitterbart, B. Stiller, A. Tantawy, A model for flexible 

high-performance communication subsystems, IEEE 

Journal on Selected Areas in Communications 11 (1993) 

507–518. 

[10] D. Schmidt, T. Suda, Transport system architecture 

services for high-performance communications systems, 

IEEE Journal on Selected Areas in Communications 11 

(1993) 489–506. 

[11] R. Morris, E. Kohler, J. Jannotti, M. Kaashoek, The click 

modular router, SIGOPS Operating Systems Review 33 

(1999) 217–231. 

[12] D. D. Clark, “The design philosophy of the DARPA 

internet protocols,” in Proc. of ACM SIGCOMM 88, 

Stanford, CA, Aug. 1988, pp. 106–114. 

[13] D.D. Clark, J. Wroclawski, K.R. Sollins, and R. Braden, 

“Tussle in Cyberspace: Defining Tomorrow’s Internet,” 

Proc. ACM SIGCOMM 2002, pp. 347-356. 

[14] D. D. Clark and D. L. Tennenhouse, “Architectural 

considerations for a new generation of protocols,” in Proc. 

of ACM SIGCOMM 90, Philadelphia, PA, Sept. 1990, pp. 

200–208. 

[15] M. S. Blumenthal and D. D. Clark, “Rethinking the design 

of the internet: the end-to-end arguments vs. the brave new 


world,” ACM Transactions on Internet Technology, vol. 1, 

no. 1, pp. 70–109, 2001. 

[16] D. Clark, K. Sollins, J. Wroclawski, D. Katabi, J. Kulik, X. 

Yang, B. Braden, T. Faber, A. Falk, V. Pingali, M. 

Handley, and N. Chiappa, “New Arch: future generation 

internet architecture,” Tech. Rep., Dec. 2003. 

[17] S. Shenker, D. Clark, D. Estrin, and S. Herzog, “Pricing in 

computer networks: reshaping the research agenda,” 

SIGCOMM Computer Communication Review, vol. 26, 

no. 2, pp. 19–43, 1996. 

[18] Tennenhouse, D. L., and Wetherall, D. J. Towards active 

network architecture. ACM SIGCOMM Computer 

Communication Review 26, 2 (Apr. 1996), 5–18. 

[19] Barbir, A., Reinaldo, P., Chen, R., Hofmann, M., and 

Hilarie, O. An architecture for open pluggable edge 

services (OPES). RFC 3835, Network Working Group, 

Aug. 2004. 

[20] Guha, S., and Francis, P. An end-middle-end approach to 

connection establishment. In SIGCOMM ’07: Proceedings 

of the 2007 conference on Applications, technologies, 

architectures, and protocols for computer communications 

(Kyoto, Japan, Aug. 2007), pp. 193–204. 

[21] Keller, R., Ramamirtham, J., Wolf, T., and Plattner, B. 

Active pipes: Program composition for programmable 

networks. In Proc. of the 2001 IEEE Conference on 

Military Communications (MILCOM) (McLean, VA, Oct. 

2001), pp. 962-966. 

[22] Rudra Dutta, G. N. R., Baldine, I., Bragg, A., and 

Stevenson, D. The SILO architecture for services 

integration, control, and optimization for the future 

internet. In Proc. of IEEE International Conference on 

Communications (ICC) (Glasgow, Scotland, June 2007), 

pp. 1899–1904. 

[23] Baldine, I., Vellala, M., Wang, A., Rouskas, G., Dutta, R., 

and Stevenson, D. A unified software architecture to 

enable cross-layer design in the future internet. In Proc. of 

Sixteenth IEEE International Conference on Computer 

Communications and Networks (ICCCN) (Honolulu, HI, 

Aug. 2007). 

[24] Gu, X., Nahrstedt, K., and Yu, B. SpiderNet: An integrated 

peer-to-peer service composition framework. In Proc. of 

Thirteenth IEEE International Symposium on High- 

Performance Distributed Computing (HPDC) (Honolulu, 

HI, June 2004), pp. 110–119. 

[25] Lakshminarayanan, K., Stoica, I., and Wehrle, K. Support 

for service composition in i3. In Proc. of the 12th annual 

ACM international conference on Multimedia (New York, 

NY, Oct. 2004), pp. 108–111. 

[26] Huang, X., Ganapathy, S., and Wolf, T. A scalable 

distributed routing protocol for networks with data-path 

services. In Proc. of 16th IEEE International Conference 

on Network Protocols (ICNP) (Orlando, FL, Oct. 2008). 

[27] Fu, X., Shi, W., Akkerman, A., and Karamcheti, V. CANS: 

composable, adaptive network services infrastructure. In 

Proc. of the 3rd USENIX Symposium on Internet 

Technologies and Systems (USITS) (San Francisco, CA, 

Mar. 2001), pp. 135–146. 

[28] Schmitt, M., Acharya, A., and Ibel, M. Service Sockets: A 

uniform user-level interface for networking applications. 

Tech. Rep. TRCS99-39, University of California, Santa 

Barbara, Santa Barbara, CA, Dec. 1999. 

[29] D.J. Wetherall, J. Guttag, and D.L. Tennenhouse, “ANTS: 

A Toolkit for Building and Dynamically Deploying 

Network Protocols,” Technical Report, MIT, 1997, in Proc. 

OPENARCH’98. 

[30] T. Meyer, L. Yamamoto, C. Tschudin, An artificial 

chemistry for networking, in: Bio-Inspired Computing and


Communication, First Workshop on Bio-Inspired Design 

of Networks (BIOWIRE 2007), pp. 45–57. 

[31] D. S. Alexander, W. A. Arbaugh, M. W. Hicks, P. Kakkar, 

A. D. Keromytis, J. T. Moore, C. A. Gunter, S. M. Nettles, 

and J. M. Smith, “The SwitchWare active network 

architecture,” IEEE Network, vol. 12, pp. 29–36, Aug. 

1998. 

[32] VARIA, J.2009. Cloud Architectures. Amazon Web 

Services. 

[33] T. Wolf, “Service-centric end-to-end abstractions in nextgeneration 

networks,” in Proc. of Fifteenth IEEE 

International Conference on Computer Communications 

and Networks (ICCCN), Arlington, VA, Oct. 2006, pp. 79– 

86. 

[34] NOX: An OpenFlow Controller. http://noxrepo.org/wp/. 

[35] S. Shanbhag and T. Wolf, “Implementation of end-to-end 

abstractions in a network service architecture,” in Fourth 

Conference on emerging Networking EXperiments and 

Technologies (CoNEXT), Madrid, Spain, 2008 

[36] The OpenFlow Switch Consortium. 

http://www.openflowswitch.org/ 

Tao Feng was born in Shandong, China, in 1979. He received 

the M.S. degree in communication engineering in 2006. He is 

currently working towards the Ph.D. degree in network 

technology form Tsinghua University, Beijing, China. 

He has been selected for the APAN’31 fellowship program. 

His research interests include future network, data center 

network, QoS and network management. 


Jun Bi received the B.S., M.S., and Ph.D. degree in computer 

science from Tsinghua University, Beijing, China. His 

dissertation studied Internet routing protocols and high 

performance routers and won the best dissertation award from 

Tsinghua University. 

From 1999 to 2000, he was a postdoctoral scholar of High 

Speed Network Department in Bell Laboratories Research, 

Lucent Technologies, New Jersey, USA. From 2000 to 2003, he 

was a research scientist of Bell Labs Research Communication 

Science Division and Bell Labs Advanced Communication 

Technologies Center. His research interests include Next 

Generation Internet Architecture and Protocols, High 

Performance Routers/Switches, Source Address Validation, 

Internet Routing, IPv4/IPv6 Transition, etc. 

Prof. Jun Bi is a full professor and director of Network 

Architecture & IPv6 Research Division, Network Research 

Center of Tsinghua University. 

Hongyu Hu was born in Hubei, China, in 1976. She received 

Ph.D.degree in Beijing Institute of Technology, Beijing, China. 

She current is a Post Doctor in Network Research Center, 

Tsinghua University, Beijing, China. 

Her current research interests include future Internet, QoS 

routing and IP multicast. 

Hui Cao was born in Shandong, China, in 1980. She received 

the M.S. degree in communication engineering in 2006. She is 

currently working towards the Ph.D. degree in trust computing 

technology form Tsinghua University, Beijing, China. 

Her research interests include trust computing, social 

network and complex network.

Aims and Scope. 

Call for Papers and Special Issues 

Journal of Networks (JNW, ISSN 1796-2056) is a scholarly peer-reviewed international scientific journal published monthly, focusing on theories, 

methods, and applications in networks. It provide a high profile, leading edge forum for academic researchers, industrial professionals, engineers, 

consultants, managers, educators and policy makers working in the field to contribute and disseminate innovative new work on networks. 

The Journal of Networks reflects the multidisciplinary nature of communications networks. It is committed to the timely publication of highquality 

papers that advance the state-of-the-art and practical applications of communication networks. Both theoretical research contributions 

(presenting new techniques, concepts, or analyses) and applied contributions (reporting on experiences and experiments with actual systems) and 

tutorial expositions of permanent reference value are published. The topics covered by this journal include, but not limited to, the following topics: 

• Network Technologies, Services and Applications, Network Operations and Management, Network Architecture and Design 

• Next Generation Networks, Next Generation Mobile Networks 

• Communication Protocols and Theory, Signal Processing for Communications, Formal Methods in Communication Protocols 

• Multimedia Communications, Communications QoS 

• Information, Communications and Network Security, Reliability and Performance Modeling 

• Network Access, Error Recovery, Routing, Congestion, and Flow Control 

• BAN, PAN, LAN, MAN, WAN, Internet, Network Interconnections, Broadband and Very High Rate Networks, 

• Wireless Communications & Networking, Bluetooth, IrDA, RFID, WLAN, WMAX, 3G, Wireless Ad Hoc and Sensor Networks 

• Data Networks and Telephone Networks, Optical Systems and Networks, Satellite and Space Communications 

Special Issue Guidelines 

Special issues feature specifically aimed and targeted topics of interest contributed by authors responding to a particular Call for Papers or by 

invitation, edited by guest editor(s). We encourage you to submit proposals for creating special issues in areas that are of interest to the Journal. 

Preference will be given to proposals that cover some unique aspect of the technology and ones that include subjects that are timely and useful to the 

readers of the Journal. A Special Issue is typically made of 10 to 15 papers, with each paper 8 to 12 pages of length. 

The following information should be included as part of the proposal: 

• Proposed title for the Special Issue 

• Description of the topic area to be focused upon and justification 

• Review process for the selection and rejection of papers. 

• Name, contact, position, affiliation, and biography of the Guest Editor(s) 

• List of potential reviewers 

• Potential authors to the issue 

• Tentative time-table for the call for papers and reviews 

If a proposal is accepted, the guest editor will be responsible for: 

• Preparing the “Call for Papers” to be included on the Journal’s Web site. 

• Distribution of the Call for Papers broadly to various mailing lists and sites. 

• Getting submissions, arranging review process, making decisions, and carrying out all correspondence with the authors. Authors should be 

informed the Instructions for Authors. 

• Providing us the completed and approved final versions of the papers formatted in the Journal’s style, together with all authors’ contact 

information. 

• Writing a one- or two-page introductory editorial to be published in the Special Issue. 

Special Issue for a Conference/Workshop 

A special issue for a Conference/Workshop is usually released in association with the committee members of the Conference/Workshop like 

general chairs and/or program chairs who are appointed as the Guest Editors of the Special Issue. Special Issue for a Conference/Workshop is 

typically made of 10 to 15 papers, with each paper 8 to 12 pages of length. 

Guest Editors are involved in the following steps in guest-editing a Special Issue based on a Conference/Workshop: 

• Selecting a Title for the Special Issue, e.g. “Special Issue: Selected Best Papers of XYZ Conference”. 

• Sending us a formal “Letter of Intent” for the Special Issue. 

• Creating a “Call for Papers” for the Special Issue, posting it on the conference web site, and publicizing it to the conference attendees. 

Information about the Journal and Academy Publisher can be included in the Call for Papers. 

• Establishing criteria for paper selection/rejections. The papers can be nominated based on multiple criteria, e.g. rank in review process plus 

the evaluation from the Session Chairs and the feedback from the Conference attendees. 

• Selecting and inviting submissions, arranging review process, making decisions, and carrying out all correspondence with the authors. 

Authors should be informed the Author Instructions. Usually, the Proceedings manuscripts should be expanded and enhanced. 

• Providing us the completed and approved final versions of the papers formatted in the Journal’s style, together with all authors’ contact 

information. 

• Writing a one- or two-page introductory editorial to be published in the Special Issue. 

More information is available on the web site at http://www.academypublisher.com/jnw/.

(Contents Continued from Back Cover) 

An Energy-Efficient Communication Protocol for Wireless Sensor Networks 

Fengjun Shang 

Robust Cross-layer Design of Wireless Multimedia Sensor Networks with Correlation and 

Uncertainty 

Lei You and Chungui Liu 

The E-Commerce Model of Health Websites: An Integration of Web Quality, Perceived Interactivity, 

and Web Outcomes 


A New Method of Time-frequency Synthesis of Harmonic Signal Extraction from Chaotic 

Background 

Erfu Wang, Zhifang Wang, Jing Ma, and Qun Ding 

Provable Data Possession of Resource-constrained Mobile Devices in Cloud Computing 

Jian Yang, Haihang Wang, Jian Wang, Chengxiang Tan, and Dingguo Yu 

Image Compression Based on Improved FFT Algorithm 

Juanli Hu, Jiabin Deng, and Juebo Wu 

Correlative Peak Interval Prediction and Analysis of Chaotic Sequences 

Qun Ding, Lu Wang, and Guanrong Chen 

REGULAR PAPERS 

An Energy Efficient Dynamic Clustering Protocol Based on Weight in Wireless Sensor Networks 

Ming Zhang and Suoping Wang 

Performance of UWB Systems with Direct-Sequence Bipolar Pulse Amplitude Modulation and 

RAKE Reception over IEEE 802.15.3a Channel 

Jingjing Wang and Hao Zhang 

Data Accuracy Estimation for Spatially Correlated Data in Wireless Sensor Networks under 

Distributed Clustering 

Jyotirmoy Karjee and H.S Jamadagni 

Networking as a Service: a Cloud-based Network Architecture 

Tao Feng, Jun Bi, Hongyu Hu, and Hui Cao 

999 

1009 

1017 

1025 

1033 

1041 

1049 

1057 

1065 

1072 

1084

Journal of Networks - Academy Publisher

You also want an ePaper? Increase the reach of your titles

Delete template?

Save as template?