Download - Academy Publisher

ISBN 978-952-5726-09-1 (Print)
Proceedings of the Second International Symposium on Networking and Network Security (ISNNS ’10)
Jinggangshan, P. R. China, 2-4, April. 2010, pp. 227-229
Emotion Recognition of EMG Based on BP
Neural Network
Xizhi Zhu
School of Physical & Electronics Engineering, Taizhou University, Taizhou, China
zxztzc@126.com
Abstract— This paper presents the emotion recognition of
BP neural network classifier. Experiment analyzes wavelet
transform of surface Electromyography (EMG) to extract
the maximum and minimum multi-scale wavelet coefficients
firstly. And then we enter the two kinds of structural
feature vector classifier for emotion recognition. The
experimental results show that both classifiers shows a
effective recognition on the four kinds of emotions, joy,
anger, sadness, pleasure, etc. Also the experimental results
show that feature vector extracted by wavelet transform can
characterize emotional patterns, and it has a stronger
emotional recognition effect than the traditional classify
method.
Index Terms—Surface Electromyography(EMG)Signal;
Emotional Recognition; Wavelet Transform; BP Neural
Network
I. INTRODUCTION
The study of emotion recognition has an important
significance in understanding human emotions in the role
of human intelligence. In daily life, human intelligence
not only shows in the normal rational thinking and logical
reasoning ability, but also in the normal emotional
capabilities. In computer science, this ability to promote
the establishment of a friendly man-machine interface is
of great significance. As the deepening of affective
computing, the request for emotion recognition
[1, 2]
technology will be correspondingly enhanced.
The MIT media experiment team led by Professor
Picard collected a 20-day physiological signal from eight
kinds of emotions of an actor when performing
deliberately, to extract the value of demographic
characteristics using way of Fisher Projection [2] . At last
they get 83% of the emotion recognition rate [3] . In
Augsburg University, Germany, Johannes Wagner and
others got approximately 80% of the emotion recognition
rate on a subject's evoked in the music under joy, anger,
sadness, pleasure of the four kinds of EMG signal by the
use of physical K-nearest neighbor, linear discriminant
function and multi-layer in 25 days [4] . Among them,
Yang Ruiqing and Liu Guangyuan led the feature
selection by the use of discrete binary particle swarm
optimization (BPSO), proving that the method of BPSO
on physiological signal for feature selection is feasible [5] .
Also Niu Xiaowei and Liu Guangyuan selected the most
representative characteristics of the corresponding optimal
combination of emotional states by using the genetic
algorithms, proving that the optimization problem, the
genetic feature selection algorithm being used to identify
© 2010 ACADEMY PUBLISHER
AP-PROC-CS-10CN006
227
the choice of the combinational optimal portfolio of the
physiological signal recognition optimal emotional
characteristics, is feasible [6] .
In this paper, the method of Wavelet Transform is
used in surface EMG aiming for non-stationary features of
surface EMG signal in order to extract more effective,
reliable, robust signal characteristics. This will help
improve the recognition rate of the surface EMG. Also
this paper uses the surface EMG signal with objective data
for six-scale decomposition of surface EMG with the
method of wavelet transform and extract the maximum of
multi-scale wavelet coefficients, constructing 14-
dimensional feature vector, then we input into the BP
neural network for emotion pattern recognition.
II. RECOGNITION METHODS OF EMOTIONAL MOTION
BP neural network is a multi-level error feedback
network proposed by Rumelhart and Mc Clelladn in 1985.
It uses the difference between the actual output and the
desired output to correct the connection weight of the
network and the threshold of each node in each layer
from back to front. BP neural network contains an input
layer, a middle layer (hidden layer) and an output layer.
There is a full connectivity between the upper and lower
layers and no connections between neurons in each layer.
For the input signal, it needs to spread towards to hidden
layer nodes and transformed by the function, then
transmit the input signal of hidden layer nodes to the
output layer nodes. Usually, the transfer function of BP
neural network is Sigmoid Type differentiable function,
which can achieve arbitrary non-linear mapping between
the input and output, so BP network has been widely
applied in pattern recognition, function approximation
and other areas [7, 8] .
The three nodes of the BP network is represented as:
Input layer Hidden layer Output layer
Figure 1. BP neural network structure