Additional Material, Journal of Imaging Science - Society for Imaging ...
Additional Material, Journal of Imaging Science - Society for Imaging ...
Additional Material, Journal of Imaging Science - Society for Imaging ...
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
JIST<br />
Vol. 51, No. 4<br />
July/August<br />
2007<br />
<strong>Journal</strong> <strong>of</strong><br />
<strong>Imaging</strong> <strong>Science</strong><br />
and Technology<br />
imaging.org<br />
<strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology
Editorial Staff<br />
Melville Sahyun, editor<br />
sahyun@infionline.net<br />
Donna Smith, production manager<br />
dsmith@imaging.org<br />
Editorial Board<br />
Philip Laplante, associate editor<br />
Michael Lee, associate editor<br />
Nathan Moroney, associate editor<br />
Mitchell Rosen, color science editor<br />
David S. Weiss, associate editor<br />
David R. Whitcomb, associate editor<br />
JIST papers are available <strong>for</strong> purchase<br />
at www.imaging.org and through<br />
ProQuest. They are indexed in<br />
INSPEC, Chemical Abstracts, <strong>Imaging</strong><br />
Abstracts, COMPENDEX, and ISI:<br />
<strong>Science</strong> Citation Index.<br />
Orders <strong>for</strong> subscriptions or single<br />
copies, claims <strong>for</strong> missing numbers,<br />
and notices <strong>of</strong> change <strong>of</strong> address<br />
should be sent to IS&T via one <strong>of</strong> the<br />
means listed below.<br />
IS&T is not responsible <strong>for</strong> the accuracy<br />
<strong>of</strong> statements made by authors and<br />
does not necessarily subscribe to their<br />
views.<br />
Copyright ©2007, <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong><br />
<strong>Science</strong> and Technology. Copying<br />
<strong>of</strong> materials in this journal <strong>for</strong> internal<br />
or personal use, or the internal or personal<br />
use <strong>of</strong> specific clients, beyond<br />
the fair use provisions granted by the<br />
US Copyright Law is authorized by<br />
IS&T subject to payment <strong>of</strong> copying<br />
fees. The Transactional Reporting Service<br />
base fee <strong>for</strong> this journal should be<br />
paid directly to the Copyright Clearance<br />
Center (CCC), Customer Service,<br />
508/750-8400, 222 Rosewood Dr.,<br />
Danvers, MA 01923 or online at<br />
www.copyright.com. Other copying<br />
<strong>for</strong> republication, resale, advertising or<br />
promotion, or any <strong>for</strong>m <strong>of</strong> systematic<br />
or multiple reproduction <strong>of</strong> any material<br />
in this journal is prohibited except with<br />
permission <strong>of</strong> the publisher.<br />
Library <strong>of</strong> Congress Catalog Card<br />
No. 59-52172<br />
Printed in the USA.<br />
<strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and<br />
Technology<br />
7003 Kilworth Lane<br />
Springfield, VA 22151<br />
www.imaging.org<br />
info@imaging.org<br />
703/642-9090<br />
703/642-9094 fax<br />
Manuscripts should be sent to the<br />
postal address above as describe at<br />
right. E-mail PDF and other files as requested<br />
to dsmith@imaging.org.<br />
Guide <strong>for</strong> Authors<br />
Scope: The <strong>Journal</strong> <strong>of</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology (JIST) is dedicated to the advancement <strong>of</strong> imaging science knowledge, the<br />
practical applications <strong>of</strong> such knowledge, and how imaging science relates to other fields <strong>of</strong> study. The pages <strong>of</strong> this journal are<br />
open to reports <strong>of</strong> new theoretical or experimental results, and to comprehensive reviews. Only original manuscripts that have not<br />
been previously published nor currently submitted <strong>for</strong> publication elsewhere should be submitted. Prior publication does not refer<br />
to conference abstracts, paper summaries, or non-reviewed proceedings, but it is expected that <strong>Journal</strong> articles will expand in scope<br />
the presentation <strong>of</strong> such preliminary communication. Please include keywords on your title and abstract page.<br />
Editorial Process/Submission <strong>of</strong> Papers <strong>for</strong> Review: All submitted manuscripts are subject to peer review. (If a manuscript appears better<br />
suited to publication in the <strong>Journal</strong> <strong>of</strong> Electronic <strong>Imaging</strong>, published jointly by IS&T and SPIE, the editor will make this recommendation.)<br />
To expedite the peer review process, please recommend two or three competent, independent reviewers. The editorial staff, will<br />
take these under consideration, but is not obligated to use them.<br />
Manuscript Guidelines: Please follow these guidelines when preparing accepted manuscripts <strong>for</strong> submission.<br />
• Manuscripts should be double-spaced, single-column, and numbered. It is the responsibility <strong>of</strong> the author to prepare a succinct,<br />
well-written, paper composed in proper English. JIST generally follows the guidelines found in the AIP Style Manual, available<br />
from the American Institute <strong>of</strong> Physics.<br />
• Documents may be created in Micros<strong>of</strong>t Word, WordPerfect, or LaTeX/REVTeX.<br />
• Manuscripts must contain a title page that lists the paper title, full name(s) <strong>of</strong> the author(s), and complete affiliation/address <strong>for</strong><br />
each author. Include an abstract that summarizes objectives, methodology, results, and their significance; 150 words maximum.<br />
Provide at least four key words.<br />
• Figures should con<strong>for</strong>m to the standards set <strong>for</strong>th at www.aip.org/epub/submitgraph.html.<br />
• Equations should be numbered sequentially with Arabic numerals in parentheses at the right margin. Be sure to define symbols<br />
that might be confused (such as ell/one, nu/vee, omega/w).<br />
• For symbols, units, and abbreviations, use SI units (and their standard abbreviations) and metric numbers. Symbols, acronyms,<br />
etc., should be defined on their first occurrence.<br />
• Illustrations: Number all figures, graphs, etc. consecutively and provide captions. Figures should be created in such a way that<br />
they remain legible when reduced, usually to single column width (3.3 inches/8.4 cm); see also<br />
www.aip.org/epub/submitgraph.html <strong>for</strong> guidance. Illustrations must be submitted as .tif or .eps files at full size and 600 dpi;<br />
grayscale and color images should be at 300 dpi. JIST does not accept .gif or .jpeg files. Original hardcopy graphics may be sent<br />
<strong>for</strong> processing by AIP, the production house <strong>for</strong> JIST. (See note below on color and supplemental illustrations.)<br />
• References should be numbered sequentially as citations appear in the text, <strong>for</strong>mat as superscripts, and list at the end <strong>of</strong> the document<br />
using the following <strong>for</strong>mats:<br />
• <strong>Journal</strong> articles: Author(s) [first/middle name/initial(s), last name], “title <strong>of</strong> article (optional),” journal name (in italics), ISSN<br />
number (e.g. <strong>for</strong> JIST citation, ISSN: 1062-3701), volume (bold): first page number, year (in parentheses).<br />
• Books: Author(s) [first/ middle name/initial(s), last name], title (in italics), (publisher, city, and year in parentheses) page reference.<br />
Conference proceedings are normally cited in the Book <strong>for</strong>mat, including publisher and city <strong>of</strong> publication (Springfield, VA, <strong>for</strong> all<br />
IS&T conferences), which is <strong>of</strong>ten different from the conference venue.<br />
• Examples<br />
1. H. P. Le, Progress and trends in ink-jet printing technology, J. <strong>Imaging</strong> Sci. Technol. 42, 46 (1998).<br />
2. E. M. Williams, The Physics and Technology <strong>of</strong> Xerographic Processes (John Wiley and Sons, New York, 1984) p. 30.<br />
3. Gary K. Starkweather, “Printing technologies <strong>for</strong> images, gray scale, and color,” Proc. SPIE 1458: 120 (1991).<br />
4. Linda T. Creagh, “Applications in commercial printing <strong>for</strong> hot melt ink-jets,” Proc. IS&T’s 10th Int’l. Congress on Adv. In<br />
Non-Impact Printing Technologies (IS&T, Springfield, VA 1994) pp. 446-448.<br />
5. ISO 13655-1996 Graphic technology: Spectral measurement and colorimetric computation <strong>for</strong> graphic arts images (ISO,<br />
Geneva), www.iso.org.<br />
6. <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology website, www.imaging.org, accessed October 2003.<br />
Reproduction <strong>of</strong> Color: Authors who wish to have color figures published in the print journal will incur color printing charges.<br />
The cost <strong>for</strong> reproducing color illustrations is $490 per page; color is not available to those given page waivers, nor can color page<br />
charges be negotiated or waived. Authors may also choose to have their figures appear in color online and in grayscale in the printed<br />
journal. There is no additional charge <strong>for</strong> this, however those who choose this option are responsible <strong>for</strong> ensuring that the captions<br />
and descriptions in the text are readable in both color and black-and-white as the same file will be used in the online and<br />
print versions <strong>of</strong> the journal. Only figures saved as TIFF/TIF or EPS files will be accepted <strong>for</strong> posting. Color illustrations may be<br />
also submitted as supplemental material <strong>for</strong> posting on the IS&T website <strong>for</strong> a flat fee <strong>of</strong> $100 <strong>for</strong> up to five files.<br />
Website Posting <strong>of</strong> Supplemental <strong>Material</strong>s: Authors may also submit additional (supplemental) materials related to their articles<br />
<strong>for</strong> posting on the IS&T Website. Examples <strong>of</strong> such materials are charts, graphs, illustrations, or movies that further explain the<br />
science or technology discussed in the paper. Supplemental materials will be posted <strong>for</strong> a flat fee <strong>of</strong> $100 <strong>for</strong> up to five files. For<br />
each additional file, a $25 fee will be charged. Fees must be received be<strong>for</strong>e supplemental materials will be posted. As a matter <strong>of</strong><br />
editorial policy, appendices are normally treated as supplemental material.<br />
Submission <strong>of</strong> Accepted Manuscripts: Author(s) will receive notification <strong>of</strong> acceptance (or rejection) and reviewers’<br />
reports. Those whose manuscripts have been accepted <strong>for</strong> publication will receive correspondence in<strong>for</strong>ming them <strong>of</strong> the issue <strong>for</strong><br />
which the paper is tentatively scheduled, links to copyright and page charge <strong>for</strong>ms, and detailed instructions <strong>for</strong> submitting accepted<br />
manuscripts. A duly signed transfer <strong>of</strong> copyright agreement <strong>for</strong>m is required <strong>for</strong><br />
publication in this journal. No claim is made to original US Government works.<br />
Page charges: Page charges <strong>for</strong> the <strong>Journal</strong> is $80/printed page. Such payment is<br />
not a condition <strong>for</strong> publication, and in some circumstances page charges are<br />
waived. Requests <strong>for</strong> waivers must be made in writing to the managing editor prior<br />
to acceptance <strong>of</strong> the paper and at the time <strong>of</strong> submission.<br />
Manuscripts submissions: Manuscripts should be submitted both electronically<br />
and as hardcopy. To submit electronically, send a single PDF file attached to an e-<br />
mail message/cover letter to jist@imaging.org. To submit hardcopy, mail 2 singlespaced,<br />
single-sided copies <strong>of</strong> the manuscript to: IS&T. With both types <strong>of</strong> submission,<br />
include a cover letter that states the paper title; lists all authors, with complete<br />
contact in<strong>for</strong>mation <strong>for</strong> each (affiliation, full address, phone, fax, and e-mail); identifies<br />
the corresponding author; and notes any special requests. Unless otherwise<br />
stated, submission <strong>of</strong> a manuscript will be understood to mean that the paper has<br />
been neither copyrighted, classified, or published, nor is being considered <strong>for</strong><br />
publication elsewhere. Authors <strong>of</strong> papers published in the <strong>Journal</strong> <strong>of</strong> <strong>Imaging</strong><br />
<strong>Science</strong> and Technology are jointly responsible <strong>for</strong> their content. Credit <strong>for</strong> the<br />
content and responsibility <strong>for</strong> errors or fraud are borne equally by all authors.<br />
JOURNAL OF IMAGING SCIENCE AND TECH-<br />
NOLOGY ( ISSN:1062-3701) is published bimonthly<br />
by The <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology,<br />
7003 Kilworth Lane, Springfield, VA 22151. Periodicals<br />
postage paid at Springfield, VA and at<br />
additional mailing <strong>of</strong>fices. Printed in Virginia,<br />
USA.<br />
<strong>Society</strong> members may receive this journal as part <strong>of</strong><br />
their membership. Forty-five dollars ($45.00) <strong>of</strong><br />
membership dues are allocated to this subscription.<br />
IS&T members may refuse this subscription by written<br />
request. Domestic institution and individual nonmember<br />
subscriptions are $195/year or $50/single<br />
copy. The <strong>for</strong>eign subscription rate is $205/year.<br />
For online version in<strong>for</strong>mation, contact IS&T.<br />
POSTMASTER: Send address changes to JOURNAL<br />
OF IMAGING SCIENCE AND TECHNOLOGY,<br />
7003 Kilworth Lane, Springfield, VA 22151.
JIST<br />
Vol. 51, No. 4<br />
July/August<br />
2007<br />
<strong>Journal</strong> <strong>of</strong><br />
<strong>Imaging</strong> <strong>Science</strong><br />
and Technology ®<br />
Feature Article<br />
283 Improved Calibration <strong>of</strong> Optical Characteristics <strong>of</strong> Paper by an Adapted<br />
Paper-MTF Model<br />
Safer Mourad<br />
General Papers<br />
293 Gloss Granularity <strong>of</strong> Electrophotographic Prints<br />
J. S. Arney, Ling Ye, Eric Maggard, and Brian Renstrom<br />
299 Forensic Examination <strong>of</strong> Laser Printers and Photocopiers Using Digital<br />
Image Analysis to Assess Print Characteristics<br />
J. S. Tchan<br />
310 Moiré Analysis <strong>for</strong> Assessment <strong>of</strong> Line Registration Quality<br />
Nathir A. Rawashdeh, Daniel L. Lau, Kevin D. Donohue,<br />
and Shaun T. Love<br />
317 Analysis <strong>of</strong> the Influence <strong>of</strong> Vertical Disparities Arising in Toed-in<br />
Stereoscopic Cameras<br />
Robert S. Allison<br />
328 Improved B-Spline Contour Fitting Using Genetic Algorithm <strong>for</strong> the<br />
Segmentation <strong>of</strong> Dental Computerized Tomography Image Sequences<br />
Xiaoling Wu, Hui Gao, Hoon Heo, Oksam Chae, Jinsung Cho, Sungyoung Lee,<br />
and Young-Koo Lee<br />
337 Colorimetric Characterization Model <strong>for</strong> Plasma Display Panel<br />
Seo Young Choi, Ming Ronnier Luo, Peter Andrew Rhodes, Eun Gi Heo,<br />
and Im Su Choi<br />
348 Real-Time Color Matching Between Camera and LCD Based on 16-bit<br />
Lookup Table Design in Mobile Phone<br />
Chang-Hwan Son, Cheol-Hee Lee, Kil-Houm Park, and Yeong-Ho Ha<br />
360 Solving Under-Determined Models in Linear Spectral Unmixing <strong>of</strong><br />
Satellite Images: Mix-Unmix Concept (Advance Report)<br />
Thomas G. Ngigi and Ryutaro Tateishi<br />
continued on next page<br />
imaging.org<br />
<strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology
IS&T BOARD OF DIRECTORS<br />
President<br />
Eric G. Hanson<br />
Department Manager<br />
Hewlett Packard Company<br />
Immediate Past President<br />
James R. Milch Jim<br />
Director Research & Innovation Labs.<br />
Carestream Health, Inc.<br />
Executive Vice President<br />
Rita H<strong>of</strong>mann<br />
Chemist, R&D Manager<br />
Il<strong>for</strong>d <strong>Imaging</strong> Switzerland GmbH<br />
continued from previous page<br />
368 Color Shift Model-Based Segmentation and Fusion <strong>for</strong> Digital Aut<strong>of</strong>ocusing<br />
Vivek Maik, Dohee Cho, Jeongho Shin, Donghwan Har, and Joonki Paik<br />
380 Error Spreading Control in Image Steganographic Embedding Schemes Using<br />
Unequal Error Protection<br />
Ching-Nung Yang, Guo-Jau Chen, Tse-Shih Chen, and Rastislav Lukac<br />
386 In Situ X-ray Investigation <strong>of</strong> the Formation <strong>of</strong> Metallic Silver Phases During the<br />
Thermal Decomposition <strong>of</strong> Silver Behenate and Thermal Development <strong>of</strong><br />
Photothermographic Films<br />
B. B. Bokhonov, M. R. Sharafutdinov, B. P. Tolochko, L. P. Burleva, and D. R. Whitcomb<br />
Conference Vice President<br />
Robert R. Buckley Rob<br />
Research Fellow<br />
Xerox Corporation<br />
Publication Vice President<br />
Franziska Frey<br />
Assist. Pr<strong>of</strong>., School <strong>of</strong> Print Media<br />
Rochester Institute <strong>of</strong> Technology<br />
Secretary<br />
Ramon Borrell<br />
Technology Strategy Director<br />
Hewlett Packard Company<br />
Treasurer<br />
Peter D. Burns<br />
Principal Scientist<br />
Carestream Health, Inc.<br />
Vice Presidents<br />
Choon-Woo Kim<br />
Inha University<br />
Laura Kitzmann<br />
Marketing Dev. & Comm. Manager<br />
Sensient <strong>Imaging</strong> Technologies, Inc.<br />
Michael A. Kriss<br />
MAK Consultants<br />
Ross N. Mills<br />
CTO & Chairman<br />
imaging Technology international<br />
IS&T Conference Calendar<br />
For details and a complete listing <strong>of</strong> conferences, visit www.imaging.org<br />
Digital Fabrication Processes Conference<br />
September 16–September 20, 2007<br />
Anchorage, Alaska<br />
General chair: Ross Mills<br />
NIP23: The 23rd International Congress on<br />
Digital Printing Technologies<br />
September 16–September 20, 2007<br />
Anchorage, Alaska<br />
General chair: Ramon Borrell<br />
IS&T/SID’s Fifteenth Color <strong>Imaging</strong><br />
Conference cosponsored by SID<br />
November 5–November 9, 2007<br />
Albuquerque, New Mexico<br />
General chairs: Jan Morovic<br />
and Charles Poynton<br />
Electronic <strong>Imaging</strong><br />
IS&T/SPIE 20th Annual Symposium<br />
January 26–January 31, 2008<br />
San Jose, Cali<strong>for</strong>nia<br />
General chairs: Nitin Sampat<br />
CGIV 2008: The Fourth European Conference<br />
on Color in Graphics, Image and Vision<br />
June 10–13, 2008<br />
Terrassa, Spain<br />
General chair: Jaume Pujol<br />
Archiving 2008<br />
June 24–27, 2008<br />
Bern, Switzerland<br />
General chair: Rudolf Gschwind<br />
Jin Mizuguchi<br />
Pr<strong>of</strong>essor, Yokohama National Univ.<br />
David Weiss<br />
Scientist Fellow, Eastman Kodak<br />
Company<br />
Chapter Director<br />
Franziska Frey – Rochester<br />
Patrick Herzog – Europe<br />
Takashi Kitamura – Japan<br />
Executive Director<br />
Suzanne E. Grinnan<br />
IS&T Executive Director<br />
ii J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
<strong>Journal</strong> <strong>of</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology® 51(4): 283–292, 2007.<br />
© <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology 2007<br />
Improved Calibration <strong>of</strong> Optical Characteristics <strong>of</strong> Paper<br />
by an Adapted Paper-MTF Model<br />
Safer Mourad <br />
Empa, Swiss Federal Laboratory <strong>for</strong> <strong>Material</strong>s Testing and Research, Laboratory <strong>for</strong> Media Technology,<br />
Dübendorf, Switzerland<br />
E-mail: safer.mourad@empa.ch<br />
Abstract. The calibration <strong>of</strong> color printers is highly influenced by<br />
optical scattering. Light scattered at microscopic level within printed<br />
papers induces a blurring phenomenon that affects the linearity <strong>of</strong><br />
the tone reproduction curve. The induced nonlinearity is known as<br />
optical dotgain. Engeldrum and Pridham analyzed its impact on<br />
printing, using Oittinen’s light scattering model. They determined the<br />
scattering and absorption coefficients based on spectral measurements<br />
<strong>of</strong> solid patches only. Their calibration achieves good independence<br />
<strong>of</strong> any printing irregularities. However, the microscopic<br />
knife-edge measurements <strong>of</strong> Arney et al. showed that the model<br />
overestimates the influence <strong>of</strong> the absorption coefficient. Unlike<br />
Oittinen’s model, we directly approach the laterally scattered light<br />
fluxes. This is achieved by an extended three-dimensional Kubelka-<br />
Munk model. We describe how to determine our coefficients using<br />
measurements <strong>of</strong> mere solid patches, which allows us to decouple<br />
the optical dot gain from other printing influences. Our improved<br />
model successfully corrects the observed overestimation and is able<br />
to predict Arney’s microscopic measurements. © 2007 <strong>Society</strong> <strong>for</strong><br />
<strong>Imaging</strong> <strong>Science</strong> and Technology.<br />
DOI: 10.2352/J.<strong>Imaging</strong>Sci.Technol.200751:4283<br />
INTRODUCTION<br />
The appearance <strong>of</strong> halftone images is determined by optical<br />
scattering in paper, which greatly affects the calibration <strong>of</strong><br />
color printers. Optical scattering is the reason <strong>for</strong> the<br />
nonlinearity <strong>of</strong> the tone reproduction curve known as optical<br />
dot gain. This paper describes an improved and easy<br />
characterization method <strong>for</strong> optical dot gain.<br />
Engeldrum and Pridham 1 analyzed the effects <strong>of</strong> optical<br />
dot gain on printing using Oittinen’s scattering model. 2 This<br />
estimates the lateral extent <strong>of</strong> the scattering effect based on<br />
the Kubelka-Munk analysis <strong>of</strong> mere vertical radiant fluxes<br />
originally proposed <strong>for</strong> uni<strong>for</strong>m paints. 3 Engeldrum and<br />
Pridham estimated the values <strong>of</strong> the classical Kubelka-Munk<br />
fitting coefficients <strong>of</strong> light absorption and scattering by spectral<br />
measurements <strong>of</strong> printed solid patches, i.e., fulltone<br />
single-color patches printed with full area coverage. They<br />
applied the coefficients obtained to Oittinen’s model and<br />
predicted the optical lateral point spread function. Using<br />
simple solid patches <strong>of</strong>fers the significant advantage <strong>of</strong> attaining<br />
calibration independent <strong>of</strong> any halftone printing irregularities.<br />
However, Arney et al. compared microscopic<br />
<br />
IS&T Member<br />
Received Dec. 23, 2005; accepted <strong>for</strong> publication Jan. 28, 2007.<br />
1062-3701/2007/514/283/10/$20.00.<br />
knife-edge measurements <strong>of</strong> different paper substrates with<br />
the numerical results <strong>of</strong> Oittinen and Engeldrum’s model<br />
and observed overestimated influence <strong>of</strong> the absorption coefficient<br />
on the width <strong>of</strong> the predicted spread function. 4,5<br />
In contrast to Oittinen’s model, our approach intrinsically<br />
accounts <strong>for</strong> scattered lateral light fluxes. 6,7 Our concept<br />
extends the classical Kubelka-Munk model to threedimensional<br />
space and analyses the balances <strong>of</strong> diffuse light<br />
fluxes across the six faces <strong>of</strong> an elementary paper volume<br />
cube. Since the lateral fluxes are explicitly considered, the<br />
extended model improves the discrimination between the<br />
scattering and absorption coefficients. After a brief review <strong>of</strong><br />
the related background, we introduce the model used as a<br />
mathematical tool and demonstrate how to determine its<br />
fitting coefficients without using any microscopic device. We<br />
then present a few application results and compare them<br />
with the findings <strong>of</strong> Arney et al. 5<br />
BACKGROUND<br />
Nowadays, color printers are calibrated based on measurements<br />
<strong>of</strong> color patches that encompass the whole color<br />
gamut. Usually the patches are arranged within standardized<br />
color wedges and printed according to a known input set <strong>of</strong><br />
device-dependent sampling points. Together with the measured<br />
colors, the values <strong>of</strong> the sampling points constitute the<br />
printer’s tone transfer function. Common color management<br />
systems require these transfer functions in a tabulated<br />
<strong>for</strong>m called the calibration pr<strong>of</strong>ile. In practice, the tone<br />
transfer function <strong>of</strong> printers is highly nonlinear and needs to<br />
be measured on a large number <strong>of</strong> at least one thousand<br />
printed color sampling points. The measured color response<br />
deviations from linearity are referred to as dot gain and are<br />
induced by two distinct effects. The first effect is called mechanical<br />
or physical dot gain and arises due to the nonlinear<br />
response <strong>of</strong> the reproduction process. It leads, e.g., to differences<br />
between the intended dot size and the dot size actually<br />
printed. The second effect is the optical dot gain 8 , also called<br />
Yule-Nielsen effect 9 , and is generally caused by the phenomenon<br />
<strong>of</strong> scattered light within the paper substrate. It induces<br />
microscopically a blurring phenomenon, which is the reason<br />
<strong>for</strong> the optical dot gain. Both dot gains have similar influences<br />
on the printed halftone images and hence on the tone<br />
reproduction curve. There<strong>for</strong>e, it is hard to characterize their<br />
distinct impact by means <strong>of</strong> mere macroscopic reflectance<br />
283
Mourad: Improved calibration <strong>of</strong> optical characteristics <strong>of</strong> paper by an adapted paper-MTF model<br />
measurements <strong>of</strong> reproduced color wedges. However, controlling<br />
and improving the image quality <strong>of</strong> color prints ultimately<br />
requires an analytical understanding <strong>of</strong> both effects<br />
in detail. In this publication we consider only the optical dot<br />
gain.<br />
In order to assess the optical effect, several authors have<br />
proposed and used direct microanalytical measurements.<br />
5,10–13 These approaches have a microscopic device in<br />
common, which projects a focused image <strong>of</strong> either a knife<br />
edge, 10 an isolated illumination point, 12 or a sinusoidal<br />
pattern 11 on top <strong>of</strong> the paper. Using a microspectrophotometer<br />
allows the visible optical effect <strong>of</strong> the scattered light to<br />
be characterized by spectral measurements <strong>of</strong> the spatial distribution<br />
<strong>of</strong> the reflected light. However, use <strong>of</strong> such a microscopic<br />
device is not always af<strong>for</strong>dable, especially from the<br />
perspective <strong>of</strong> printer manufacturers or graphic arts bureaux.<br />
We propose to simplify these ef<strong>for</strong>ts by introducing a<br />
calibration technique based on our light scattering and color<br />
halftone model. 6,7<br />
The analysis <strong>of</strong> the optical properties <strong>of</strong> paper can be<br />
considered as a special application <strong>of</strong> light scattering in turbid<br />
media. 14 From a fundamental point <strong>of</strong> view, light scattering<br />
can be derived from Maxwell’s equations; see, <strong>for</strong> instance,<br />
Ishimaru. 15 Concerning this approach, however, the<br />
same author also states: “... its drawback is the mathematical<br />
complexities involved, and its usefulness is limited.” [Ref.<br />
16, p. 2210]. On the other hand, transport theory directly<br />
models the transport <strong>of</strong> radiant power through turbid media.<br />
Because <strong>of</strong> its experimental con<strong>for</strong>mity, transport theory<br />
is preferred in a large number <strong>of</strong> applications. The pragmatic<br />
success is thereby emphasized and the approximating character<br />
<strong>of</strong> the solutions is accepted, which is especially the case<br />
<strong>for</strong> the simple and popular variants like the Kubelka-Munk<br />
two-flux theory, see, e.g., Ref. 17. In this tradition our color<br />
prediction model follows an engineering approach <strong>for</strong> printing<br />
applications, particularly in connection with the control<br />
and calibration <strong>of</strong> color halftone printers where “engineering”<br />
stands <strong>for</strong> a balance between simplicity <strong>of</strong> use and accuracy<br />
<strong>of</strong> prediction.<br />
Other halftone color prediction models are a continuing<br />
subject <strong>of</strong> past and current investigations. Emmel, 18,19 <strong>for</strong><br />
example, presents a recent survey and introduces a novel<br />
mathematical framework <strong>for</strong> spectral predictions <strong>of</strong> color<br />
halftone prints. The framework uses a global analytical approach<br />
based on matrix algebra that unifies most <strong>of</strong> the<br />
classical color prediction models. It is an efficient and intuitive<br />
model employing overall probabilities <strong>of</strong> photons entering<br />
and emerging through particular inking levels. However,<br />
the probabilistic description 20 used is “taken throughout the<br />
full sample area,” 19 which leads to a halftone-independent<br />
model. This makes it necessary to recalibrate the model’s<br />
coefficients <strong>for</strong> every different halftone technique used <strong>for</strong><br />
printing.<br />
As an alternative to Emmel’s approach, our proposed<br />
halftone prediction model 7 is particularly intended to meet<br />
the requirements <strong>of</strong> a halftone-dependent characterization <strong>of</strong><br />
digital printing devices. In order to be adaptable to arbitrary<br />
Figure 1. Diagram <strong>of</strong> an upper paper section with a light path scattered<br />
between the entry point x,y and the exit point x,y.<br />
halftone schemes we chose a numerical convolution approach<br />
using a separate optical modulation transfer function<br />
(MTF). The MTF model is founded on a three-dimensional<br />
extension <strong>of</strong> the Kubelka-Munk approach derived by analyzing<br />
multivariate partial differential equations with common<br />
computer algebra systems. 21 The derived extension approximates<br />
the scattered lateral light within semi-isotropic substrates.<br />
Although the one-dimensional Kubelka-Munk<br />
theory has methodological weaknesses, 17,22 the inaccuracy <strong>of</strong><br />
the predictions <strong>for</strong> the applications considered is limited.<br />
Moreover, it can be expected that the three-dimensional extension<br />
also increases the prediction accuracy. Like the underlying<br />
theory, the current approach relates the light propagation<br />
characteristics to a few substrate-dependent scattering<br />
and absorption coefficients. Specular and internal reflections<br />
at the interfaces and transmittances are considered as<br />
boundary conditions. Furthermore, the scattering concept<br />
can easily be extended to brightened fluorescent media as<br />
shown in Ref. 7.<br />
MTF-BASED SPATIAL-SPECTRAL HALFTONE<br />
PREDICTION MODEL<br />
In this section we describe the model used <strong>for</strong> light propagation<br />
in printed paper without elaborating the mathematical<br />
derivation. We are especially interested in an optical halftone<br />
model as a function <strong>of</strong> several easy predictable<br />
parameters such as scattering or partial reflection coefficients.<br />
The basis <strong>of</strong> the proposed approach 7 is to relate the<br />
reflected (reemitted) image and the local impact <strong>of</strong> the<br />
spread light to what is known as the point spread function<br />
(PSF) <strong>of</strong> paper. 10,23 The PSF models the scattered light intensity<br />
by the probability density h R x−x,y−y <strong>of</strong> a photon<br />
that enters the substrate at location x,y to exit at<br />
x,y; see Figure 1. Let x,y be the inner transmittance <strong>of</strong><br />
the print layer at location x,y, wherex,y= ink if<br />
x,y is covered by ink and x,y=1 otherwise, then the<br />
light reflected at point x,y <strong>of</strong> a halftone print is given by 23<br />
Rx,y = x,yh R x − x,y − yx,ydxdy.<br />
For simplicity, we ignore <strong>for</strong> the moment any specular or<br />
partial internal reflections at the paper interfaces. Usually,<br />
the computational ef<strong>for</strong>t <strong>of</strong> calculating the convolution integral<br />
is reduced by applying a two-dimensional Fourier trans<strong>for</strong>m<br />
to Eq. (1). Application <strong>of</strong> the commonly known convolution<br />
theorem replaces the integral operation by simple<br />
multiplication in the Fourier domain, yielding<br />
1<br />
284 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Mourad: Improved calibration <strong>of</strong> optical characteristics <strong>of</strong> paper by an adapted paper-MTF model<br />
Figure 2. Scheme <strong>of</strong> the paper bulk with the considered inner partial<br />
reflections and the illuminating flux i 0 .<br />
Rx,y = x,yF −1 †H R ,Fx,y‡,<br />
where F denotes the Fourier trans<strong>for</strong>m and F −1 its inverse<br />
[see the Appendix section and Bracewell 24 <strong>for</strong> details]. The<br />
variables (,) are the lateral frequencies in the Fourier domain<br />
and H R , is the MTF <strong>of</strong> paper, i.e., the Fourier<br />
trans<strong>for</strong>m <strong>of</strong> the PSF. Fx,y is the Fourier trans<strong>for</strong>m <strong>of</strong><br />
the inner transmittance <strong>of</strong> the print layer.<br />
We now extend Eq. (2) <strong>for</strong> practical purposes and incorporate<br />
additional surface refractive corrections in a similar<br />
way to Saunderson. 25 To begin with, we first supplement<br />
the specularly reflected fraction ap x,y at the upper side<br />
ink-air interface. (We use the following subscripts: ap: air to<br />
paper, pa: paper to air, and pb: paper to backing). Secondly,<br />
we distinguish between the incident transmitted fraction<br />
ap x,y and the emerging fraction pa x,y because <strong>of</strong> different<br />
illuminating and viewing geometries, yielding<br />
Rx,y = ap x,y + pa x,yF −1 †H R ,F ap x,y‡.<br />
In Eq. (3), the multiplication with the outward transmittance<br />
pa x,y represents the passage <strong>of</strong> the emerged light<br />
through the ink layer be<strong>for</strong>e being captured by a sensor. On<br />
the other hand, the multiplication <strong>of</strong> the MTF with the Fourier<br />
trans<strong>for</strong>med inward transmittance F ap x,y accounts<br />
<strong>for</strong> the spreading effect induced by the scattered light. Its<br />
propagation through the inner paper bulk is constrained at<br />
the bottom by the inner partial reflection pb and at the top<br />
interfaces by pa ; see Figure 2. We distinguish between both<br />
reflectances because the scattering bulk is usually faced by<br />
different media during measurements, as will be seen later.<br />
Note that Eq. (3) considers the halftone structures only<br />
through both inner transmittance factors ap x,y and<br />
pa x,y, i.e., we assume the MTF H R , to be independent<br />
from the local halftone structure. Consequently, we<br />
consider the situation depicted in Fig. 2, and the inner partial<br />
reflectances pa and pb are chosen halftone independently.<br />
Finally, with respect to multi-ink prints, an additional<br />
model is required <strong>for</strong> the calculation <strong>of</strong> the spectral inner<br />
transmittance <strong>of</strong> the overprinted ink layers. As such, Beer-<br />
Lambert’s multiplication <strong>of</strong> transmittances is a frequently<br />
chosen and simple approach. 26 More complex alternatives<br />
may be considered when dealing with fluorescent inks 19 or<br />
with ink penetration effects. 27<br />
In microscopic image analyses, Eq. (3) proves very useful.<br />
In this section we outline how to determine a mathematical<br />
expression <strong>of</strong> the MTF H R ,. The bulk <strong>of</strong> papers<br />
usually consists <strong>of</strong> a fiber network which deflects<br />
passing light rays in arbitrary directions. 28 This behavior is<br />
2<br />
3<br />
responsible <strong>for</strong> the scattering properties <strong>of</strong> the paper and,<br />
there<strong>for</strong>e, <strong>for</strong> its MTF. As mentioned earlier, the phenomenological<br />
Kubelka-Munk (KM) approach 3,29 is successfully<br />
used in application fields involving familiar light propagation<br />
problems. However, the KM approach considers only<br />
two diffuse radiant fluxes, one in the direction <strong>of</strong> incidence<br />
and the other in the opposite direction. In other words, the<br />
spatial distribution <strong>of</strong> the re-emitted light cannot be accounted<br />
<strong>for</strong> by using the results <strong>of</strong> the traditional KM<br />
theory. There<strong>for</strong>e, we approached the MTF by extending the<br />
KM theory to a three-dimensional space 7 in a similar way to<br />
the two-dimensional extension proposed by Berg. 30<br />
For an infinitesimal volume cube <strong>of</strong> the substrate, six<br />
light fluxes along and opposed to the coordinate axes x, y,<br />
and z are considered. The fluxes are non-negative by definition<br />
and are specified <strong>for</strong> −x,y and 0zD,<br />
where D is the thickness <strong>of</strong> the paper sheet. The study is<br />
confined to temporal steady state analyses <strong>of</strong> samples with a<br />
given absorption 0 that are illuminated with a light<br />
source <strong>of</strong> finite power. Hence the fluxes are expected to decay<br />
greatly in both lateral directions as x +y →. The<br />
correctness <strong>of</strong> this assumption was already shown by the<br />
measurements <strong>of</strong> Yule and others see, e.g., Ref. 10. Now, let<br />
us consider the intensity <strong>of</strong> the downward incident light flux<br />
along the vertical propagation direction z and call it i. The<br />
theory <strong>of</strong> KM is based on the assumption that the fractional<br />
amount <strong>of</strong> light lost by absorption between z and z+dz is<br />
given by dz, where the absorption density corresponds to<br />
the absorption coefficient K <strong>of</strong> the original KM theory. Also<br />
scattering decreases the considered intensity but, in our case,<br />
we consider different kinds <strong>of</strong> scattering densities: b and l ,<br />
where b is the back scattered intensity and l is the intensity<br />
scattered laterally to the initial direction. Accordingly,<br />
after passing dz the flux i is reduced by dz, where the<br />
coefficients obey<br />
= +4 l + b .<br />
We call the modeled scattering behavior semi-isotropic,<br />
where isotropic stands <strong>for</strong> the symmetry related to x,y,z and<br />
semi indicates l b . The absorbed light is lost to the system,<br />
but back-scattered light from the upward propagating<br />
flux is added to i. A quarter <strong>of</strong> the laterally scattered light<br />
from each lateral flux is also added to i. Applying the same<br />
reasoning to each <strong>of</strong> the remaining five fluxes we obtained a<br />
system <strong>of</strong> six coupled linear partial differential equations 7<br />
(PDE) that are familiar to the type <strong>of</strong> the seven-flux model<br />
<strong>of</strong> Yoon et al. 31 [Remark: According to the design <strong>of</strong> these<br />
PDEs, the incident light flux primarily propagates along the<br />
coordinate axes after an initial scattering event. However, in<br />
our applications, the resulting rotational asymmetry reaches<br />
numerically at most about one percent. 7 We disregard this<br />
asymmetry in favor <strong>of</strong> the obvious application advantages<br />
presented below].<br />
In order to derive the paper’s MTF, we used the generalized<br />
two-dimensional Fourier trans<strong>for</strong>m. This allows the<br />
PDE system to be reduced to a pair <strong>of</strong> equations similar to<br />
the original ordinary KM differential equations in z but with<br />
4<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 285
Mourad: Improved calibration <strong>of</strong> optical characteristics <strong>of</strong> paper by an adapted paper-MTF model<br />
added dependencies on the spatial frequencies and . 7 For<br />
the PDE system at hand, the generalized two-dimensional<br />
Fourier trans<strong>for</strong>m is appropriate, since the fluxes are defined<br />
on the real plane R 2 and decay greatly as x +y →. 24,32,33<br />
Solving the system <strong>of</strong> trans<strong>for</strong>med PDEs at z=D yields the<br />
spectral reflectance MTF<br />
where<br />
H R , = Fh R x,y = A <br />
B <br />
,<br />
A = a 12 + a 21 − c pb e cD − a 12 + a 21 + c pb e −cD ,<br />
5.1<br />
5.2<br />
B =−a 21 + c + a 12 pa + pb + a 21 − c pa pb e cD + a 21<br />
− c + a 12 pa + pb + a 21 + c pa pb e −cD . 5.3<br />
The coefficients a 12 =a 12 , a 21 =a 21 , and c=c depend on<br />
the lateral frequencies , and are given by Eqs.<br />
(6.1)–(6.6):<br />
where<br />
2 2<br />
c =a 21 − a 12 , 6.1<br />
a 12 =<br />
a 21 =<br />
s 3<br />
s 1<br />
,<br />
s 2<br />
s 1<br />
,<br />
6.2<br />
6.3<br />
s 1 = − b 2 −2 l + b +2 l + b +4 2 2 − b 2 <br />
2 + 2 +16 4 2 2 ,<br />
s 2 = − b 2 −2 l + b +2 l + b −4 l 2 <br />
+4 2 − b + b −2 l 2 2 + 2 <br />
6.4<br />
+16 4 2 2 , 6.5<br />
s 3 = − b 2 −2 l + b b +2 l + b −4 l 2 <br />
+4 2 − b b + b −2 l 2 2 + 2 <br />
+16 4 b 2 2 . 6.6<br />
Likewise, the microspectral transmittance distribution is<br />
modeled by<br />
Tx,y = pa x,yF −1 †H T ,F bp x,y‡,<br />
with the spectral transmittance MTF H T ,<br />
7<br />
H T , = Fh T x,y =− 2c <br />
B <br />
.<br />
Here, bp x,y describes the fraction <strong>of</strong> light transmitted into<br />
the substrate through the bottom layer. Equation (7) implies<br />
that the optical spreading has no direct observable effect on<br />
transmittance measurements <strong>of</strong> single-side, upward-oriented<br />
printed paper sheets. This is consistent with microscopic<br />
transmittance images published by Koopipat et al. 13<br />
The two spatial <strong>for</strong>mulas, Eqs. (3) and (7), are the foundation<br />
used in predicting the spectral reflectance <strong>of</strong> arbitrary<br />
halftone prints presented and discussed in the Model Application<br />
and Model Discussion sections below. The following<br />
section considers the calibration <strong>of</strong> the optical parameters.<br />
MODEL CALIBRATION<br />
Our next objective is to determine the optical dot gain <strong>for</strong><br />
arbitrary halftones and dithering frequencies from a few<br />
macroscopic spectral measurements. More precisely, we need<br />
to calculate the isolated optical dot gain in order to distinguish<br />
between the different effects leading to printing<br />
nonlinearities. In order to meet this requirement, Eq. (3)<br />
describes the spectral reflectance Rx,y as a function <strong>of</strong> the<br />
bulk parameters D, , l , and b together with the surface<br />
refractive coefficients ap , ap , pa , pa , and pb . Un<strong>for</strong>tunately,<br />
only the paper thickness, D, is directly measurable<br />
with common instruments. There<strong>for</strong>e, we determine the remaining<br />
parameters in such a way as to best match the calculated<br />
results to the measured spectra <strong>of</strong> a small set <strong>of</strong> test<br />
patches. In order to avoid any printing irregularity and to<br />
increase the accuracy <strong>of</strong> the parameter estimation, the test<br />
patches are chosen to be as unambiguous as possible. In<br />
particular, we chose only solid patches <strong>of</strong> the primary<br />
colors—in our case cyan, magenta, yellow, and black prints,<br />
abbreviated to CMYK—plus a sample <strong>of</strong> paper-white (W).<br />
For this work, we consider only single side prints. As well as<br />
avoiding the printing irregularities, the uni<strong>for</strong>mity <strong>of</strong> the<br />
solid patches also reduces the matrix multiplication <strong>of</strong> Eq.<br />
(3) to a simple scalar multiplication. This is because<br />
Fx,y<strong>of</strong> a uni<strong>for</strong>m patch is only different from zero at<br />
zero frequencies ,0,0. Hence, <strong>for</strong> a solid patch,<br />
Eqs. (3) and (7) reduce to<br />
R solid = ap + pa H R 0,0 ap ,<br />
T solid = pa H T 0,0 bp .<br />
8<br />
9<br />
10<br />
In other words, the <strong>for</strong>m <strong>of</strong> the PSF h R x,y is not involved<br />
in the calibration process.<br />
Traditionally, the KM theory scattering and absorption<br />
coefficients K and S are determined using two distinct reflectance<br />
measurements: one measurement over a black<br />
backing and another over a white backing, both backings <strong>of</strong><br />
known reflectances. However, in our case, more measurements<br />
are required because we need to determine not only<br />
the scattering and absorption coefficients but also the inner<br />
transmittance <strong>of</strong> the primary inks used in addition to the<br />
286 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Mourad: Improved calibration <strong>of</strong> optical characteristics <strong>of</strong> paper by an adapted paper-MTF model<br />
Figure 3. Considered configurations <strong>of</strong> the calibration measurements <strong>of</strong> a paper sheet printed on one side<br />
with a solid patch. Upper left: Spectral reflectance on a known black backing. Upper right: Spectral reflectance<br />
on a known white backing. Lower left: Regular spectral transmittance. Lower right: Flipped spectral<br />
transmittance.<br />
refractive coefficients. There<strong>for</strong>e, we propose measuring the<br />
spectral reflectance and transmittance <strong>of</strong> each <strong>of</strong> the five<br />
solid test set patches (CMYK/W) in each <strong>of</strong> the four configurations<br />
depicted in Figure 3. These comprise the usual<br />
reflectance with a black and a white backing <strong>of</strong> known<br />
reflectances R b in addition to the transmittance <strong>of</strong> the<br />
samples in both a regular configuration and a flipped-over<br />
configuration yielding a total <strong>of</strong> nineteen different spectra.<br />
The spectral measurements were carried out using a<br />
standard R/T spectrophotometer (Gretag-Macbeth<br />
SpectroScan T). Its geometry <strong>of</strong> illumination (45°, collimated<br />
annular) and <strong>of</strong> viewing (0°) contradicts, strictly spoken,<br />
the general KM assumption <strong>of</strong> diffuse illumination and<br />
measurement. Nevertheless, according to Kubelka, 34 as we<br />
examine highly scattering paper substrates, we expect the<br />
geometric discrepancies at most to scale the determined values<br />
<strong>of</strong> the scattering and absorption coefficients by a constant<br />
factor. In our application we disregard its effect as long<br />
as the geometry <strong>of</strong> the measuring equipment is kept unchanged<br />
between calibration and prediction. The geometry<br />
<strong>of</strong> each measurement configuration also affects the assumed<br />
surface refractive corrections that are mutually connected by<br />
functional relations. We make these relations explicit by<br />
first-order approximations and introduce additional common<br />
measurement coefficients. 35 The considered surface refractive<br />
coefficients are illustrated in Figure 4. For simplicity,<br />
we assume the individual inks to have equal refractive indices,<br />
which may, however, deviate from the bulk refractive<br />
index <strong>of</strong> the unprinted paper. <strong>Additional</strong>ly, we disregard any<br />
wavelength dependency <strong>of</strong> the refractive indices. In order to<br />
list the refractive coefficients <strong>for</strong> each <strong>of</strong> the printed and<br />
unprinted cases, we hereafter identify each coefficient with<br />
one <strong>of</strong> the subscripts I or , respectively. (With regard to<br />
halftone patches, we vary these coefficients in accordance<br />
with the ink coverage).<br />
We begin with the situation <strong>of</strong> measuring the reflectance<br />
<strong>of</strong> the samples and consider the recorded fraction ap<br />
<strong>of</strong> the specular reflection s :<br />
apI = K s sI , ap = K s s , 11<br />
where K s amounts to the fraction captured by the instrument’s<br />
field <strong>of</strong> view. 36 Next, the incident transmission ap is<br />
determined by the transmitted fraction s =1− s in addition<br />
to what is known as the internal transmittance <strong>of</strong> the upper<br />
side layer [Ref. 26 p. 30] thus we approximate<br />
apI = 1− sI , ap =1− s . 12<br />
Similarly, we approximate the coefficient <strong>for</strong> the outward<br />
transmittance through the upper interface<br />
paI = 1− iI , pa =1− i . 13<br />
In Eq. (13), 1− i represents the fraction <strong>of</strong> the light transmitted<br />
diffusely from inside the scattering substrate. It thus<br />
differs from the fraction 1− s used in Eq. (12), which accounts<br />
<strong>for</strong> the collimated incidence. Accordingly, the amount<br />
<strong>of</strong> internal light reflected diffusely at the upper interface is<br />
Figure 4. Refraction and transmission coefficients at the interfaces <strong>of</strong> the<br />
paper sheet.<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 287
Mourad: Improved calibration <strong>of</strong> optical characteristics <strong>of</strong> paper by an adapted paper-MTF model<br />
Figure 5. Spectral calibration per<strong>for</strong>mance <strong>for</strong> the nonbrightened double-calandered APCO paper. The solid<br />
lines plot the spectral predictions and the crosses mark the corresponding spectral measurements. The paper<br />
has a thickness <strong>of</strong> about 99 m.<br />
paI = iI , pa = i . 14<br />
We continue at the bottom <strong>of</strong> the paper and consider the<br />
internal fraction <strong>of</strong> light reflected diffusely at the bottom<br />
interface. Similar to Eq. (14), it is determined by the internal<br />
reflection i plus the transmitted part through the bottom<br />
layer that is reflected back into the substrate at the backing<br />
pbI = iI + 1− iI 1− sI 2 R b ,<br />
pb = i + 1− i 1− s R b .<br />
15<br />
b = l = .<br />
17<br />
As already pointed out, we estimate the phenomenological<br />
model parameters as the set <strong>of</strong> values which minimizes<br />
the deviation between the nineteen measured spectra<br />
and their calculated predictions. We obtain the parameter<br />
estimates by using common least square minimization routines<br />
such as the lsqcurvefit <strong>of</strong> MATLAB. This routine<br />
solves nonlinear data-fitting problems in the least square<br />
sense. 37 In our case, the routine finds the parameters X<br />
sought that minimize the error to the measured spectra<br />
S meas <br />
Here, is squared because <strong>of</strong> the double light passage<br />
according to Beer-Lambert’s law. 26 Finally, <strong>for</strong> the case <strong>of</strong><br />
measuring the transmittance <strong>of</strong> the samples, the part transmitted<br />
from beneath the sample into the scattering substrate<br />
is approximated to<br />
1<br />
min<br />
X 2 S solidX − S meas 2 2 = 1 R solidi X<br />
2 i<br />
− R measi 2 + 1 T solidj X − T measj 2 ,<br />
2 j<br />
18<br />
bpI = 1− sI , bp =1− s . 16<br />
These relations reduce the multitude <strong>of</strong> refractive model parameters<br />
to K s , s , i <strong>for</strong> the printed and unprinted cases in<br />
addition to <strong>of</strong> each primary ink. Together with the<br />
scattering and absorption coefficients, six scalar coefficients<br />
and seven spectral coefficients are obtained that may be estimated<br />
using the nineteen available spectral measurements.<br />
However, in order to simplify the optimization process further,<br />
we limit the model to the isotropic case assuming both<br />
scattering coefficients to be equal, hence<br />
where R solid X and T solid X are the spectral reflectance<br />
and transmittance <strong>of</strong> the samples calculated with the<br />
parameters fixed to X according to Eqs. (9) and (10), respectively.<br />
The routine used finds the coefficients so that the<br />
solution is always bounded within an appropriately chosen<br />
range <strong>for</strong> each parameter <strong>of</strong> X.<br />
CALIBRATION RESULTS<br />
An example <strong>of</strong> the obtained calibration results is illustrated<br />
in Figure 5. The depicted spectra were obtained from a rep-<br />
288 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Mourad: Improved calibration <strong>of</strong> optical characteristics <strong>of</strong> paper by an adapted paper-MTF model<br />
Figure 6. Fitted estimates <strong>of</strong> the spectral absorption and scattering coefficients<br />
underlying the data <strong>of</strong> Fig. 5.<br />
Figure 8. From bottom to top reflects the dependency <strong>of</strong> the optical dot<br />
gain <strong>of</strong> the APCO paper on increasing screen frequencies. The simulated<br />
halftone technique is a conventional, circular dot screen illustrated below<br />
the chart.<br />
al. 40 <strong>for</strong> a critical review. Originally, the MD model was proposed<br />
in a densitometric <strong>for</strong>m due to limitations <strong>of</strong> the instruments.<br />
We use its spectral <strong>for</strong>m, which reads<br />
Figure 7. Fitted estimates <strong>of</strong> the spectral internal transmittances C ,<br />
M , Y , and K underlying the data <strong>of</strong> Fig. 5.<br />
resentative calibration <strong>of</strong> a non-brightened doublecalandered<br />
APCO paper 38 printed by a common, four-color<br />
xerographic desktop printer. The average calibration accuracy<br />
achieved 1E * 94 , which is in the order <strong>of</strong> the colorimetric<br />
discrimination per<strong>for</strong>mance <strong>of</strong> the human eye. Figure<br />
6 shows the fitted spectral model coefficients <strong>of</strong> absorption<br />
and scattering obtained. In this experiment, the estimates<br />
<strong>for</strong> the inner reflectance and transmittance coefficients<br />
<strong>of</strong> the printed interfaces are i =10%, s =1.25% and<br />
K s =35%. These scalar values, and particularly i differ from<br />
usual reports in related research. 36,39 In our opinion, this<br />
discrepency is mainly due to the different underlying approaches<br />
<strong>of</strong> accounting <strong>for</strong> the interactions between the partial<br />
reflections and the scattered differential light fluxes. Finally,<br />
Figure 7 plots the estimated internal transmittances<br />
<strong>of</strong> the separate CMYK inks.<br />
MODEL APPLICATIONS<br />
Given a calibrated parameter set <strong>of</strong> a particular paper-ink<br />
combination, such as those derived in the previous section,<br />
the spectral prediction Eqs. (3) and (7) allow the expected<br />
microimages <strong>of</strong> arbitrary halftone prints on that paper to be<br />
simulated. In this section, after recalling the basics <strong>of</strong> dot<br />
gain calculations, we apply the proposed approach and analyze<br />
optical dot gain predictions <strong>for</strong> idealized halftone prints.<br />
Common dot gain calculations <strong>of</strong> arbitrary halftones are<br />
based on the model <strong>of</strong> Murray-Davies (MD); see Wyble et<br />
R = 1−a t R g + a t R t ,<br />
19<br />
where R is the predicted reflectance spectrum, R g is<br />
the reflectance spectrum <strong>of</strong> the bare substrate and R t is<br />
the spectral reflectance spectrum <strong>of</strong> the color at full area<br />
coverage. The linear interpolation variable a t commonly refers<br />
to the theoretical area coverage <strong>of</strong> the predicted halftone<br />
print, i.e., the dot area <strong>of</strong> the binary image actually sent to<br />
the printer. Usually, the MD model overestimates the measured<br />
spectral reflectances R meas since it disregards the<br />
combined optical and mechanical dot gain effects as introducedintheBackground<br />
section. The overestimation gives<br />
rise to introduction <strong>of</strong> the effectiveareacoveragea eff , which<br />
refers to an estimated value that best fits the calculated reflectance<br />
R to the measured spectrum R meas . For a<br />
single wavelength, typically chosen at minimum reflectance,<br />
a eff is given by using Eq. (19)<br />
a eff = R meas − R g<br />
R t − Rg , 20<br />
with suppressed wavelength notation <strong>for</strong> simplicity. This allows<br />
the dot gain to be defined as<br />
= a eff − a t .<br />
21<br />
Use <strong>of</strong> Eq. (3) allows estimation <strong>of</strong> the optical dot-gain<br />
o <strong>of</strong> any given binary halftone image Ix,y with a dot area<br />
equal to a s . In this case o is given by<br />
o = a eff − a s = Rx,y − R g<br />
R t − R g<br />
− a s . 22<br />
In Eq. (22), Rx,y is the spatial average <strong>of</strong> the reflectance<br />
image predicted using Eq. (3) and replaces the measured<br />
reflectance R meas <strong>of</strong> Eq. (20). An illustrative example is seen<br />
in Figure 8 which depicts optical dot-gain predictions <strong>for</strong> the<br />
APCO paper 38 against the area coverage a s <strong>for</strong> various screen<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 289
Mourad: Improved calibration <strong>of</strong> optical characteristics <strong>of</strong> paper by an adapted paper-MTF model<br />
Figure 9. Calculated optical dot gain curves at medium area coverage<br />
against the screen ruling with respect to two typically used halftone<br />
screens. The plots were predicted using the data <strong>of</strong> the APCO paper.<br />
Figure 11. Comparison <strong>of</strong> our LSF-predictions with normalized measurement<br />
points <strong>of</strong> a different but apparently similar paper type published by<br />
Arney et al. 5 The dots depict Arney’s measurements. The dashes and the<br />
solid line plot the obtained predictions <strong>for</strong> the ColorCopy and the APCO<br />
paper, respectively.<br />
particular the impact <strong>of</strong> the optical dot-gain’s variations<br />
around the commonly used ruling <strong>of</strong> 60 lpcm 150 lpi.<br />
In this region, any slight shift <strong>of</strong> the screen ruling obviously<br />
induces a non-negligible error arising solely from the optical<br />
dot gain.<br />
Figure 10. Calculated optical dot gain curves <strong>for</strong> two different types <strong>of</strong><br />
<strong>of</strong>fice paper plotted against screen ruling. The conventional circular dot<br />
screen was used. The dashes and the solid line plot the obtained predictions<br />
<strong>for</strong> the ColorCopy paper and the APCO paper, respectively.<br />
frequencies (so-called ruling). The type and magnitude <strong>of</strong><br />
the optical dot gain curves shown are close to analyses presented,<br />
e.g., by Gustavson 8 or by Yang [Ref. 41 p. 201], which<br />
apply a simple Gaussian-like function to approach the shape<br />
<strong>of</strong> the paper PSF.<br />
Clearly, <strong>for</strong> calculating the halftone reflectances, the<br />
simplifying assumption underlying Eq. (9) no longer holds<br />
true and the coefficients become spatially dependent. Hence<br />
we numerically evaluate the spatial <strong>for</strong>mula, Eq. (3), by using<br />
the standard inverse two-dimensional fast Fourier trans<strong>for</strong>m<br />
(iFFT). 24 Thereby we sample the coefficients <strong>of</strong> concern<br />
at spatial high resolution frequency grids , with a<br />
sampling frequency exceeding the Nyquist rate <strong>of</strong> the halftone<br />
structures, i.e., the coefficients are filled out at each<br />
x i ,y j according to the simulated binary image Ix i ,y j with<br />
the calibrated values <strong>of</strong> section Calibration Results.<br />
Another application example is given in Figure 9, which<br />
illustrates the influence <strong>of</strong> different types <strong>of</strong> halftone screens<br />
on the magnitude <strong>of</strong> the optical dot gain at medium area<br />
coverage <strong>of</strong> a s =50%. With the proposed simulation, it is<br />
also convenient to explore the effect <strong>of</strong> changing the type <strong>of</strong><br />
paper on the optical dot gain. For the case <strong>of</strong> the conventional<br />
circular dot halftone screen, Figure 10 shows the optical<br />
dot gain prediction <strong>for</strong> the APCO paper compared with<br />
results obtained with a similarly calibrated parameter set <strong>for</strong><br />
the ColorCopy <strong>of</strong>fice paper. 42 These graphs demonstrate in<br />
MODEL COMPARISON AND DISCUSSION<br />
To explore the per<strong>for</strong>mance <strong>of</strong> our microscopic predictions,<br />
we consider results <strong>of</strong> edge-trace measurements published by<br />
Arney et al. 5 For this purpose, we compare our prediction<br />
results with their published data <strong>of</strong> a measured line spread<br />
function (LSF) and the corresponding MTF <strong>of</strong> an <strong>of</strong>fice copy<br />
paper <strong>of</strong> a similar type to those analyzed in our laboratory.<br />
Figure 11 depicts a direct comparison <strong>of</strong> Arney’s measurements<br />
and our prediction <strong>of</strong> a microscopic line edge observation<br />
derived by using Eq. (3). The comparison <strong>of</strong> the<br />
MTF, the normalized modulus <strong>of</strong> the digital Fourier trans<strong>for</strong>m<br />
<strong>of</strong> the same data is illustrated in Figure 12. Both figures<br />
demonstrate good agreement between our predictions and<br />
the microdensitometric measurements. Note that the model<br />
parameters were calibrated using spectral measurements <strong>of</strong><br />
solid patches only.<br />
A further comparison arises from Arney’s discussion <strong>of</strong><br />
the influence <strong>of</strong> Kubelka-Munk’s absorption coefficient K on<br />
Oittinen and Engeldrum’s model. Oittinen, 2 as well as<br />
Engeldrum and Pridham, 1 suggested a simplified relationship<br />
between the Kubelka-Munk theory and the MTF <strong>of</strong><br />
paper. According to their approach, the MTF <strong>of</strong> paper approximates<br />
the vertical derivative <strong>of</strong> the classical KM<br />
function. 3 It depends on the thickness <strong>of</strong> the sample D, the<br />
scattering coefficient S, and the absorption coefficient K.<br />
However, with edge-trace measurements Arney et al. showed<br />
that K has only a small influence on the width <strong>of</strong> the MTF <strong>of</strong><br />
the paper compared to S and D. 4 Figure 13 illustrates this<br />
finding and compares the results <strong>of</strong> our model with those<br />
obtained by Oittinen and Engeldrum’s model. In particular,<br />
the figure plots the scalar measure k p proposed by Arney,<br />
which is equal to the inverse <strong>of</strong> the frequency at which the<br />
normalized MTF loses its half magnitude. The plot demon-<br />
290 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Mourad: Improved calibration <strong>of</strong> optical characteristics <strong>of</strong> paper by an adapted paper-MTF model<br />
our model shows good agreement. Moreover, the calibrated<br />
model agrees well with the experimental insight <strong>of</strong> Arney et<br />
al. 4 that the optical absorption coefficient <strong>of</strong> paper has an<br />
insignificant effect on the MTF width <strong>of</strong> paper.<br />
Figure 12. Comparison <strong>of</strong> the modulus <strong>of</strong> the digital Fourier trans<strong>for</strong>m <strong>of</strong><br />
the data presented in Fig. 11.<br />
ACKNOWLEDGMENTS<br />
The author would like to thank Pr<strong>of</strong>essor R. D. Hersch,<br />
École Polytechnique Fédérale de Lausanne (EPFL), Switzerland,<br />
<strong>for</strong> the continuing and supporting discussions. The<br />
constructive discussions with K. Simon, M. Vöge, and P.<br />
Zolliker are also gratefully acknowledged. Part <strong>of</strong> the investigation<br />
was financed by the Swiss Innovation Promotion<br />
Agency (grant KTI/CTI 6498.1 ENS-ET).<br />
Appendix<br />
We used the following <strong>for</strong>m <strong>of</strong> the two-dimensional Fourier<br />
trans<strong>for</strong>m F, <strong>of</strong> a two-dimensional function fx,y<br />
<br />
F, fx,ye =−<br />
−2ix+y dx dy,<br />
23<br />
and the inverse two-dimensional Fourier trans<strong>for</strong>m<br />
<br />
fx,y F,e =−<br />
2ix+y d d,<br />
24<br />
Figure 13. Comparing the involvement <strong>of</strong> the absorption coefficient K <strong>of</strong><br />
Oittinen and Engeldrum’s model with that <strong>of</strong> the absorption coefficient <br />
<strong>of</strong> our model obtained <strong>for</strong> the ColorCopy paper. The comparison is carried<br />
out using Arney’s data and his proposed k p scalar metric.<br />
strates the involvement <strong>of</strong> Oittinen and Engeldrum’s absorption<br />
coefficient K shown by Arney 5 and the more realistic<br />
influence <strong>of</strong> our absorption coefficient .<br />
CONCLUSIONS<br />
Our prediction model <strong>of</strong>fers a new way <strong>of</strong> minimizing the<br />
ef<strong>for</strong>t <strong>of</strong> characterizing and calibrating color halftone printers<br />
and extends the computational framework <strong>of</strong> controlling<br />
color printers online. The model used computes high resolution<br />
spectral color images <strong>of</strong> arbitrary halftone prints on<br />
common <strong>of</strong>fice paper and newsprint types <strong>of</strong> paper. Light<br />
scattering effects are accounted <strong>for</strong> by relating the appearance<br />
characteristics to a few substrate-dependent fitting coefficients.<br />
The main advantage <strong>of</strong> the approach is to uncouple<br />
the calibration <strong>of</strong> the scattering and absorption<br />
coefficients <strong>of</strong> the paper from printing irregularities without<br />
using a microscopic device. We base the calibration on reflectance<br />
measurements <strong>of</strong> solid patches because they are<br />
affected much less by the mechanical dot-gain than halftone<br />
patches are. Thereby, the calibration achieves a good independence<br />
from the printing process and requires merely a<br />
common spectrophotometer <strong>for</strong> reference measurements.<br />
Comparing published microspectral knife-edge measurements<br />
<strong>of</strong> Arney et al. 5 with corresponding simulations <strong>of</strong><br />
see Ref. 24 <strong>for</strong> details.<br />
REFERENCES<br />
1 P. G. Engeldrum and B. Pridham, “Application <strong>of</strong> turbid medium theory<br />
to paper spread function measurements”, Proc.-TAGA 1, 339–352<br />
(1995).<br />
2 P. Oittinen, “Limits <strong>of</strong> microscopic print quality”, in Advances in<br />
Printing <strong>Science</strong> and Technology (Pentech, London, 1982), Vol. 16, pp.<br />
121–138.<br />
3 P. Kubelka and F. Munk, “Ein Beitrag zur Optik der Farbanstriche”, Z.<br />
Tech. Phys. (Leipzig) 12, 593–601 (1931).<br />
4 J. S. Arney, E. Pray, and K. Ito, “Kubelka-Munk theory and the<br />
Yule-Nielsen effect on halftones”, J. <strong>Imaging</strong> Sci. Technol. 43, 365–370<br />
(1999).<br />
5 J. S. Arney, J. Chauvin, J. Nauman, and P. G. Anderson, “Kubelka-Munk<br />
theory and the MTF <strong>of</strong> paper”, J. <strong>Imaging</strong> Sci. Technol. 47, 339–345<br />
(2003).<br />
6 S. Mourad, P. Emmel, K. Simon, and R. D. Hersch, “Prediction <strong>of</strong><br />
monochrome reflectance spectra with an extended Kubelka-Munk<br />
model”, Proc. IS&T/SID Tenth Color <strong>Imaging</strong> Conference (IS&T,<br />
Springfield, VA, 2002) pp. 298–304.<br />
7 S. Mourad, Color Prediction <strong>for</strong> Electrophotographic Prints on Common<br />
Office Paper, Ph.D. Thesis, École Polytechnique Fédérale de Lausanne,<br />
Switzerland, 2003, http://diwww.epfl.ch/w3lsp/publications/colour.<br />
8 S. Gustavson, “Color gamut <strong>of</strong> halftone reproduction”, J. <strong>Imaging</strong> Sci.<br />
Technol. 41, 283–290 (1997).<br />
9 J. A. C. Yule and W. J. Nielsen, “The penetration <strong>of</strong> light into paper and<br />
its effect on halftone reproduction”, Proc.-TAGA, 65–76 (1951).<br />
10 J. A. C. Yule, D. J. Howe, and J. H. Altman, “The effect <strong>of</strong> the<br />
spread-function <strong>of</strong> paper on halftone reproduction”, Tappi J. 50,<br />
337–344 (1967).<br />
11 S. Inoue, N. Tsumura, and Y. Miyake, “Analyzing CTF <strong>of</strong> print by MTF<br />
<strong>of</strong> paper”, J. <strong>Imaging</strong> Sci. Technol. 42, 572–576 (1998).<br />
12 S. Gustavson, Dot Gain in Colour Halftones, Ph.D. Thesis, Dept. <strong>of</strong><br />
Electrical Engineering, Linköping University, Sweden, 1997.<br />
13 C. Koopipat, N. Tsumura, and Y. Miyake, “Effect <strong>of</strong> ink spread and<br />
optical dot gain on the MTF <strong>of</strong> ink jet image”, J. <strong>Imaging</strong> Sci. Technol.<br />
46, 321–325 (2002).<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 291
Mourad: Improved calibration <strong>of</strong> optical characteristics <strong>of</strong> paper by an adapted paper-MTF model<br />
14 B. Philips-Invernizzi, D. Dupont, and C. Cazé, “Bibliographical review<br />
<strong>for</strong> reflectance <strong>of</strong> diffusing media”, Opt. Eng. (Bellingham) 40,<br />
1082–1092 (2001).<br />
15 A. Ishimaru, Wave Propagation and Scattering in Random Media<br />
(Academic Press, New York, 1978), Vols. I&II.<br />
16 A. Ishimaru, “Diffusion <strong>of</strong> light in turbid material”, Appl. Opt. 28,<br />
2210–2215 (1989).<br />
17 B. Hapke, “Kubelka-Munk theory: What’s wrong with it”, in Theory <strong>of</strong><br />
Reflectance and Emittance Spectroscopy (Cambridge University Press,<br />
Cambridge, UK, 1993), Chap. 11.<br />
18 P. Emmel, Modèles de prédiction couleur appliqués à l’impression jet<br />
d’encre, Ph.D. thesis, École Polytechnique Fédérale de Lausanne, 1998,<br />
http://diwww.epfl.ch/w3lsp/publications/colour.<br />
19 P. Emmel, “Physical models <strong>for</strong> color prediction”, in Digital Color<br />
<strong>Imaging</strong>, Handbook (The Electrical Engineering and Applied Signal<br />
Processing Series) (CRC Press, Boca Raton, FL, 2003), pp. 173–238.<br />
20 J. S. Arney, “A probability description <strong>of</strong> the Yule-Nielsen effect, I”, J.<br />
<strong>Imaging</strong> Sci. Technol. 41, 633–636 (1997).<br />
21 Wolfram Research, Inc. MATHEMATICA ® . http://www.wolfram.com.<br />
22 L. Yang and S. J. Miklavcic, “Revised Kubelka-Munk theory. III. A<br />
general theory <strong>of</strong> light propagation in scattering and absorptive media”,<br />
J. Opt. Soc. Am. A 22, 1866–1873 (2005).<br />
23 F. R. Ruckdeschel and O. G. Hauser, “Yule-Nielsen effect in printing: a<br />
physical analysis”, Appl. Opt. 17, 3376–3383 (1978).<br />
24 R. N. Bracewell, The Fourier Trans<strong>for</strong>m and its Applications, 3rd ed.<br />
(McGraw-Hill, New York, 2000).<br />
25 J. L. Saunderson, “Calculation <strong>of</strong> the color <strong>of</strong> pigmented plastics”, J. Opt.<br />
Soc. Am. 32, 727–736 (1942).<br />
26 G. Wyszecki and W. S. Stiles, Color <strong>Science</strong>, 2nd ed. (John Wiley & Sons,<br />
Inc., New York, 1982).<br />
27 L. Yang, R. Lenz, and B. Kruse, “Light scattering and ink penetration<br />
effectsontonereproduction”,J.Opt.Soc.Am.A18, 360–366 (2001).<br />
28 G. Kortüm, Reflectance Spectroscopy: Principles, Methods, Applications<br />
(Springer-Verlag, Berlin-Heidelberg, 1969).<br />
29 P. Kubelka, “New contributions to the optics <strong>of</strong> intensely light-scattering<br />
materials. Part II: Nonhomogenous layers”, J. Opt. Soc. Am. 44, 330–335<br />
(1954).<br />
30 F. Berg, Isotrope Lichtstreuung in Papier - Neue Überlegungen zur<br />
Kubelka-Munk-Theorie, Ph.D. Thesis, Technische Hochschule<br />
Darmstadt, 1997.<br />
31 G. Yoon, A. J. Welch, M. Motamedi, and M. C. J. Van Gemert,<br />
“Development and application <strong>of</strong> three-dimensional light distribution<br />
model <strong>for</strong> laser irradiated tissue”, IEEE J. Quantum Electron. 23,<br />
1721–1733 (1987).<br />
32 M. J. Lighthill, Introduction to Fourier Analysis and Generalised<br />
Functions (Cambridge University Press, Cambridge, UK, 1980).<br />
33 D. G. Duffy, Trans<strong>for</strong>m Methods <strong>for</strong> Solving Partial Differential<br />
Equations (CRC Press, Boca Raton, FL, 1994).<br />
34 P. Kubelka, “New contributions to the optics <strong>of</strong> intensely light-scattering<br />
materials Part. I”, J. Opt. Soc. Am. 38, 448–457 (1948).<br />
35 D. B. Judd and G. Wyszecki, Color in Business, <strong>Science</strong> and Industry, 3rd<br />
ed. (John Wiley & Sons, Inc., New York, 1975).<br />
36 F. R. Clapper and J. A. C. Yule, “The effect <strong>of</strong> multiple internal<br />
reflections on the densities <strong>of</strong> half-tone prints on paper”, J. Opt. Soc.<br />
Am. 43, 600–603 (1953).<br />
37 MathWorks. MATLAB Optimization Toolbox. Consider especially<br />
lsqcurvefit and fmincon, http://www.mathworks.com.<br />
38 ISO-2846-1. Graphic technology—Colour and transparency <strong>of</strong> ink sets <strong>for</strong><br />
four-colour-printing, 1st ed. (ISO, Geneva, 1997). The nonbrightened<br />
APCO II/II paper is specified in Annex A.<br />
39 D. B. Judd, “Fresnel reflection <strong>of</strong> diffusely incident light”, J. Res. Natl.<br />
Bur. Stand. 29, RP1504 (1942).<br />
40 D. R. Wyble and R. S. Berns, “A critical review <strong>of</strong> spectral models<br />
applied to binary color printing”, Color Res. Appl. 25, 4–19 (2000).<br />
41 L. Yang, S. Gooran, and B. Kruse, “Simulation <strong>of</strong> optical dot gain in<br />
multichromatic tone reproduction”, J. <strong>Imaging</strong> Sci. Technol. 45, 198–204<br />
(2001).<br />
42 Neusiedler, Color Copy paper. http://www.neusiedler.com(2000).<br />
Brightened, highly opaque <strong>of</strong>fice paper.<br />
292 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
<strong>Journal</strong> <strong>of</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology® 51(4): 293–298, 2007.<br />
© <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology 2007<br />
Gloss Granularity <strong>of</strong> Electrophotographic Prints<br />
J. S. Arney and Ling Ye<br />
Rochester Institute <strong>of</strong> Technology, Rochester, New York 14623<br />
E-mail: arney@cis.rit.edu<br />
Eric Maggard and Brian Renstrom<br />
Hewlett-Packard Co., Boise, Idaho 83714<br />
Abstract. The random variation in gloss <strong>of</strong>ten observed in images<br />
produced in electrophotographic printers has been examined by an<br />
analytical technique that combines the capabilities <strong>of</strong> a microdensitometer<br />
with a goniophotometer. The technique is called<br />
microgoniophotometry and measures both the spatial and the angular<br />
distribution <strong>of</strong> the specular component <strong>of</strong> reflected light. The<br />
analysis provides in<strong>for</strong>mation about the spatial variation <strong>of</strong><br />
specularly reflected light at all angles through which the specular<br />
light is reflected, not just at the equal/opposite angle at which gloss<br />
is traditionally measured. The results <strong>of</strong> this analysis have lead to an<br />
optical model <strong>of</strong> the random spatial variation in gloss. The results<br />
indicate that dry toner is typically not completely fused and can be<br />
described as a surface composed <strong>of</strong> two distinct regions. These two<br />
regions differ in the extent <strong>of</strong> fusing that has occurred, as manifested<br />
by their differences in specular reflectance characteristics.<br />
The difference in reflectance is manifested primarily in their different<br />
angular distributions <strong>of</strong> specular light and also in their spatial<br />
frequency. © 2007 <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology.<br />
DOI: 10.2352/J.<strong>Imaging</strong>Sci.Technol.200751:4293<br />
INTRODUCTION<br />
A bidirectional reflectance distribution function (BRDF) is a<br />
useful way to characterize the angular distribution <strong>of</strong> specular<br />
light reflected from materials. 1–8 Moreover, one would<br />
expect the BRDF to be a necessary part <strong>of</strong> a complete instrumental<br />
characterization <strong>of</strong> visual attributes <strong>of</strong> gloss. 9 In<br />
addition to the angular distribution <strong>of</strong> the specular light, the<br />
spatial distribution <strong>of</strong> the specular light may also play a role<br />
in visual gloss. 10,11 As illustrated in Figure 1, gloss in electrophotographic<br />
prints is not always spatially uni<strong>for</strong>m. Indeed,<br />
spatial variations in gloss take many <strong>for</strong>ms. Artifacts<br />
such as streaking and banding are <strong>of</strong>ten observed in high<br />
gloss prints, and differential gloss involves differences in<br />
gloss between bordering regions <strong>of</strong> different color. The current<br />
report focuses on gloss granularity, which is the random<br />
gloss variation across a printed surface. Gloss granularity is<br />
illustrated in Fig. 1 with samples A and B showing different<br />
degrees <strong>of</strong> gloss granularity.<br />
Granularity analysis is an analytical technique that<br />
evolved during the 20th century to characterize silver halide<br />
photographic film. 12 The typical microdensitometer was an<br />
optical microscope with a fixed aperture and an electronic<br />
light detector. The film sample was scanned under the microscope<br />
and a trace <strong>of</strong> irradiance versus location was recorded.<br />
This technique is called microdensitometry. Currently,<br />
a microdensitometry scan may be per<strong>for</strong>med more<br />
easily by a s<strong>of</strong>tware routine applied to a digital image captured<br />
with a camera and appropriate microscope optics. 13,14<br />
Several reports have been published on the application <strong>of</strong><br />
microdensitometry techniques to the analysis <strong>of</strong> gloss<br />
granularity. 10,11,15 All <strong>of</strong> these techniques involve detection <strong>of</strong><br />
light at the specular angle (equal/opposite angle) while scanning<br />
across the surface <strong>of</strong> the sample. The current work<br />
extends this analytical technique to a measurement <strong>of</strong> the<br />
entire BRDF (goniophotometry) scanned spatially across the<br />
surface <strong>of</strong> a printed sample (microdensitometry). This analytical<br />
technique is called microgoniophotometry.<br />
THE MICROGONIOPHOTOMETER<br />
The microgoniophotometer has been described in detail in<br />
previous reports and is summarized in Figure 2. 1,2,16–18 The<br />
print sample is wrapped around a cylinder, and this presents<br />
all sample angles from −90° to +90° to the camera. The<br />
sample is illuminated with a linear light source placed at an<br />
angle <strong>of</strong> 20° from the camera. This places a bright specular<br />
line at the half angle, =10° between the camera and the<br />
source. Two images captured with this system are illustrated<br />
in Figure 3.<br />
As illustrated in Fig. 3, the specular component <strong>of</strong> the<br />
reflected light maintains its polarization and is observed only<br />
<br />
IS&T Member<br />
Received Jan. 3, 2007; accepted <strong>for</strong> publication Feb. 3, 2007.<br />
1062-3701/2007/514/293/6/$20.00.<br />
Figure 1. Examples <strong>of</strong> A rough and B smooth gloss granularity in<br />
electrophotographic prints produced by two different printers using different<br />
toners and fusing conditions.<br />
293
Arney et al.: Gloss granularity <strong>of</strong> electrophotographic prints<br />
Figure 5. BRDF <strong>of</strong> vs and BGDF <strong>of</strong> vs generated from Fig. 4.<br />
Curves are normalized to 1.00 at the peak value in order make a<br />
comparison.<br />
Figure 2. Schematic illustration <strong>of</strong> the microgoniophotometer. A linear<br />
polarizer is placed in front <strong>of</strong> the line light source, and another polarizer,<br />
called the analyzer, is in front <strong>of</strong> the camera.<br />
Figure 3. Images captured with the analyzer in front <strong>of</strong> the camera parallel<br />
to and perpendicular to the polarization direction <strong>of</strong> the light source<br />
polarizer.<br />
Figure 6. Micr<strong>of</strong>acets <strong>of</strong> the surface are randomly oriented at different tilt<br />
angles. If the facet tilt results in an equal/opposite angle between the<br />
camera and the light source, then light enters the camera. Otherwise the<br />
specular light misses the camera. A piece <strong>of</strong> shattered automobile window<br />
glass is a macroscopic illustration <strong>of</strong> bilevel gloss granularity.<br />
sample. A plot <strong>of</strong> the mean value versus tilt angle, vs , is<br />
a bidirectional reflectance distribution function, BRDF. A<br />
plot <strong>of</strong> the standard deviation versus tilt angle, vs , isa<br />
bidirectional granularity distribution function, BGDF. (See<br />
Figure 5.) It is the granularity <strong>of</strong> the specular light at each<br />
angle on the BRDF.<br />
Figure 4. The difference image A-B shows only the specularly reflected<br />
light. The mean, , and the standard deviation, , <strong>of</strong> the specular light is<br />
determined at each column in the image.<br />
in the image with parallel polarizers. Both the crossed and<br />
the parallel polarizers capture the same amount <strong>of</strong> diffuse,<br />
randomly polarized light. The difference image, (A-B) in<br />
Figure 4, shows only the specular light.<br />
The horizontal location <strong>of</strong> each column in the difference<br />
image (A-B) corresponds to a tilt angle, , on the print<br />
A FACET MODEL OF SPECULAR GRANULARITY<br />
Johansson, Béland, and MacGregor have introduced a model<br />
<strong>of</strong> specular reflection called the micr<strong>of</strong>acet model, 10,11 and<br />
the micr<strong>of</strong>acet model has been applied to the problem <strong>of</strong><br />
synthetic scene generation in computer graphics. 19 The<br />
micr<strong>of</strong>acet model assumes the surface that reflects the specular<br />
light can be described as a set <strong>of</strong> small facets, each at a<br />
randomly tilted angle, as illustrated schematically in<br />
Figure 6. The only facets that will deliver light to the camera<br />
are those facets tilted exactly to produces an equal/opposite<br />
294 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Arney et al.: Gloss granularity <strong>of</strong> electrophotographic prints<br />
Figure 7. Example <strong>for</strong> a sample <strong>of</strong> solid black toner printed by a typical<br />
electrophotographic printer. The solid line is Eq. 3, and the points are<br />
from experimental measurements <strong>of</strong> and 2 over the range −50° <br />
50°.<br />
angle between the source and the camera. Otherwise the<br />
light misses the camera. The result would be expected to be<br />
the bilevel image <strong>of</strong> specular glints, as illustrated in Fig. 6.<br />
The line light source used in the microgoniophotometer<br />
is assumed to be infinite in the direction colinear with the<br />
cylinder so that a facet tilt in the orthogonal direction, ,<br />
always directs light to the camera. There<strong>for</strong>e, the BRDF measured<br />
with the microgoniophotometer should be a direct<br />
measure <strong>of</strong> the random distribution <strong>of</strong> facet tilt angles in the<br />
direction. By normalizing the area under the BRDF, vs<br />
, to unity, the probability density function, P, <strong>for</strong> the<br />
random tilt angles, , can be <strong>for</strong>med as shown in Eqs. (1).<br />
The value <strong>of</strong> P at each angle, , is a measure <strong>of</strong> the fraction<br />
<strong>of</strong> the surface that contains facets at exactly angle :<br />
90<br />
K =−90<br />
d and P = <br />
K . 1<br />
Each facet that is at the correct specular angle delivers<br />
light at irradiance I to the camera. All other facets produce<br />
an irradiance <strong>of</strong> I=0. The result is irradiance I at the facet<br />
location projected onto the camera sensor plane. This bilevel<br />
set <strong>of</strong> facets should produce an average value and a standard<br />
deviation given by Eqs. (2) and (3). Note from Eq. (1) that<br />
the area under the BRDF ( vs ) is an experimental measure<br />
<strong>of</strong> the irradiance, I=K,<br />
= P · I, where I = K, 2<br />
2 = P · 1−P · I 2 .<br />
In order to test the facet model quantitatively, experimental<br />
measurements <strong>of</strong> 2 versus were carried out <strong>for</strong><br />
twenty samples <strong>of</strong> solid black (single toner) produced by<br />
different printers with different toners and different fusing<br />
conditions on different substrates. Values <strong>of</strong> P were calculated<br />
from with Eq. (1), and the data was plotted as 2<br />
versus P·1−P. Figure 7 is an example <strong>for</strong> a typical solid<br />
black toner printed by laser EP. The measured values <strong>of</strong> 2<br />
were much lower than predicted, and the data do not show<br />
the linearity <strong>of</strong> Eq. (3). Thus the facet model illustrated in<br />
Fig. 6 does not provide a complete, quantitative rationale <strong>for</strong><br />
the measured data.<br />
3<br />
Figure 8. The blurring effect <strong>of</strong> the camera pixels projected onto the<br />
surface facets.<br />
AN EXPANDED FACET MODEL<br />
It is not surprising that the experimentally measured values<br />
<strong>of</strong> 2 are lower than predicted. Equation (3) is based on the<br />
facets as if they were measured with infinite resolution.<br />
However, there is no reason to expect the surface facets to be<br />
large relative to the size <strong>of</strong> the camera pixels projected onto<br />
the surface. Indeed, if the camera pixels are larger than the<br />
facet size, the camera image will blur the image through a<br />
convolution with the effective aperture <strong>of</strong> the camera pixels.<br />
This is illustrated in Figure 8. The effect can be described<br />
quantitatively by modifying Eq. (3) with a blurring factor, k,<br />
as shown in Eq. (4):<br />
2 = P · 1−P · I 2 · k 2 .<br />
The nonlinearity observed in Fig. 7 requires additional<br />
modification <strong>of</strong> the facet model. Figure 9 suggests a modification<br />
based on the microstructure <strong>of</strong> the facets. Visual inspection<br />
<strong>of</strong> the printed samples in specular light indicates<br />
that the samples have a variety <strong>of</strong> different microstructures.<br />
Moreover, visual inspection <strong>of</strong> many samples suggests that<br />
the microstructures may be described as a population <strong>of</strong> two<br />
types <strong>of</strong> surfaces; one with well fused toner and the other<br />
with more poorly fused toner. This model is illustrated schematically<br />
in Figure 10.<br />
These two regions would be expected to contribute to<br />
the overall measured BRDF and granularity <strong>of</strong> the sample.<br />
This is described in Eqs. (5)–(7), where P a and P b are the<br />
probability density functions <strong>for</strong> the distribution <strong>of</strong> surface<br />
tilt angles in the two regions illustrated in Fig. 10, a and b<br />
are the rms granularity characteristic <strong>of</strong> the two regions, and<br />
F is the fraction <strong>of</strong> the surface that is region (a). Note that<br />
Eq. (7) reduces to Eq. (3) <strong>for</strong> P a =P b :<br />
P = F · P a + 1−F · P b ,<br />
2 = F · a2 · I 2 + 1−F · b2 · I 2 ,<br />
4<br />
5<br />
6<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 295
Arney et al.: Gloss granularity <strong>of</strong> electrophotographic prints<br />
Figure 11. P normalized BRDF versus angle <strong>for</strong> a typical solid black<br />
printed by laser EP. The solid line is experimental data. The dotted line is<br />
the model <strong>of</strong> Eqs. 5, 9, and10 with s a =5.1°, s b =12.7°, and<br />
F=0.3.<br />
Figure 9. Closeup <strong>of</strong> the specular band <strong>for</strong> experimental samples 1 and<br />
2.<br />
Figure 12. BRGF versus angle <strong>for</strong> a typical solid black printed by<br />
laser EP. The solid line is experimental data. The dotted line is the model<br />
<strong>of</strong> Eq. 8 with k a =0.95 and k b =0.20.<br />
Figure 10. Schematic illustration <strong>of</strong> partial fusing <strong>of</strong> toner.<br />
2 = F · P a · 1−P a · I 2 + 1−F · P b · 1−P b · I 2 .<br />
Equation (7) needs to be adjusted to account <strong>for</strong> the<br />
aperture effect <strong>of</strong> the camera pixels, as described above.<br />
However, one might expect the pixel aperture effect, the constant<br />
k in Eq. (4), not to be the same <strong>for</strong> the two regions.<br />
Thus we write Eq. (7). Equations (5)–(8) represent an expanded<br />
facet model <strong>of</strong> specular reflections:<br />
2 = F · P a · 1−P a · I 2 · k a<br />
2<br />
7<br />
Figure 13. Example <strong>for</strong> a sample <strong>of</strong> solid black toner printed by a typical<br />
laser EP printer. The solid line is Eq. 3, and the points are from experimental<br />
measurements <strong>of</strong> and 2 over the range −50° 50°.<br />
+ 1−F · P b · 1−P b · I 2 · k 2 b . 8<br />
By combining Eqs. (5), (9), and (10), the BRDF can be<br />
modeled by adjusting the parameters, s a , s b , and F to achieve<br />
APPLYING THE EXPANDED FACET MODEL<br />
the best fit with the experimental data. Figure 11 shows the<br />
In order to model the BRDF and BGDF, the two individual result <strong>for</strong> one <strong>of</strong> the printed samples. The model parameters<br />
PDF functions P a and P b are needed. These functions were s a , s b , and F were adjusted to achieve the minimum rms<br />
assumed to be normal distributions described by Eqs. (9) deviation from the experimental data.<br />
and (10):<br />
Equation (8) has two additional parameters, k a and k b ,<br />
that must be adjusted to model the BGDF, versus .<br />
1<br />
P a = e<br />
s a<br />
−2 2<br />
/2s a,<br />
9<br />
Figure 12 shows the minimum rms deviation between the<br />
2 model and the data, and Figure 13 shows the corresponding<br />
plot <strong>of</strong> versus P. The model provides a rationale <strong>for</strong> the<br />
1<br />
significant deviation from linearity predicted by Eq. (3).<br />
P b = e<br />
s b<br />
−2 2<br />
/2s b. 10<br />
2<br />
296 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Arney et al.: Gloss granularity <strong>of</strong> electrophotographic prints<br />
Figure 14. Examples <strong>of</strong> differences in behavior observed and modeled <strong>for</strong> other solid samples <strong>of</strong> black toner<br />
from different printers. Model parameters <strong>for</strong> s a , s b , F, k a , and k b are also shown.<br />
SPATIAL SIGNIFICANCE OF PARAMETERS k a AND k b<br />
Figure 14 illustrates the behavior <strong>of</strong> three additional samples<br />
<strong>of</strong> solid black toner printed by different electrophotographic<br />
printers. The differences in behavior are more easily observed<br />
by plotting versus P·1−P. The solid lines show<br />
the models that best fit the data, and the modeled values <strong>of</strong><br />
s a , s b , F, k a , and k b are also shown. From an analysis <strong>of</strong> 15<br />
samples <strong>of</strong> black toner produced in different printers, this<br />
behavior appears to be representative <strong>of</strong> typical electrophotographic<br />
samples.<br />
The physical meanings <strong>of</strong> parameters a , b , and F are<br />
indicated in the diagram <strong>of</strong> Fig. 10. In all cases s a s b , which<br />
suggests that the range <strong>of</strong> surface tilt angles in region (a) is<br />
less than the range <strong>of</strong> angles in region (b). This is reasonable<br />
if the toner in region (a) is more thoroughly fused than<br />
region (b). The fraction F in every case is less than 0.5,<br />
which suggests that there is less <strong>of</strong> the smooth region (a)<br />
than <strong>of</strong> the more rough region (b).<br />
The physical meaning <strong>of</strong> the parameters k a and k b is less<br />
obvious. In every case k a k b . This suggests the effect <strong>of</strong> the<br />
pixel aperture convolution with the facet size has more <strong>of</strong> a<br />
blurring effect in the rough region (b) than in the smooth<br />
region (a). A possible rationale <strong>for</strong> this observation may be<br />
that the rough region (b) is also a higher frequency region.<br />
The low pass filtering effect <strong>of</strong> the pixel aperture would indeed<br />
be expected to have a have a larger effect on the higher<br />
frequency region (b) than the lower frequency region (a).<br />
Thus k a and k b provide spatial in<strong>for</strong>mation about the gloss<br />
granularity in addition to the magnitude parameters s a and<br />
s b .<br />
As a check <strong>of</strong> the interpretation <strong>of</strong> k a and k b as indices<br />
<strong>of</strong> relative spatial frequency, the (A) image illustrated in<br />
Fig. 3 was low-pass filtered with a Gaussian kernel <strong>of</strong> radius<br />
R. Values <strong>of</strong> R were selected over the range R=0 (no filtering)<br />
to R=20 m. Each image was analyzed to extract experimental<br />
values <strong>of</strong> and as described above, and from<br />
fitting the model to each data set, values <strong>of</strong> the model parameters<br />
were determined as described above. The results<br />
are shown in Figure 15. As one would expect, the smoothing<br />
kernel had only a small effect on the width parameters, s a<br />
and s b . However, the values <strong>of</strong> k a and k b declined significantly,<br />
with k a decreasing much more than k b .<br />
Figure 15. Values <strong>of</strong> s a , s b , k a ,andk b <strong>for</strong> a printed sample <strong>of</strong> black toner<br />
analyzed through low pass filters <strong>of</strong> radius 0R20 m.<br />
DISCUSSION<br />
The behavior shown in Fig. 15 is consistent with the interpretation<br />
<strong>of</strong> k a and k b as noise attenuation factors related to<br />
the low pass filtering effect <strong>of</strong> the effective pixel aperture and<br />
the assumption that facets in the smooth region (a) are<br />
larger (lower frequency) than those in less well fused regions<br />
(b). The smaller facets in region (b) are low pass filtered to a<br />
larger extent than those in region (b) by the pixel aperture<br />
effect, so k b k a . Further filtering by the added Gaussian<br />
filters lowers both k a and k b , as expected, and they approach<br />
the same values <strong>for</strong> extreme low-pass filtering R=20 m.<br />
As discussed in a previous report, the width <strong>of</strong> the<br />
BRDF is an inverse index <strong>of</strong> traditional gloss. 13 A narrow<br />
curve correlates with a high gloss reading. In the current<br />
work, it appears that fused toner can be interpreted in terms<br />
<strong>of</strong> two spatial regions that differ in the degree <strong>of</strong> fusing. The<br />
well fused region has a narrow BRDF, indicated by the value<br />
<strong>of</strong> s a , and the poorly fused region has a broader BRDF indicated<br />
by s b . The magnitude <strong>of</strong> the rms deviation <strong>of</strong> gloss,<br />
called gloss granularity, is indicated by the values <strong>of</strong> k a and<br />
k b . As is typical <strong>of</strong> granularity indices, their magnitude is<br />
dependent on the effective spatial aperture <strong>of</strong> measurement.<br />
In this case that spatial aperture is the area <strong>of</strong> a camera pixel<br />
projected onto the surface. The range <strong>of</strong> behaviors <strong>of</strong> k a and<br />
k b observed in these experiments indicates that gloss granularity<br />
has a significant spatial frequency component that remains<br />
to be examined in future research.<br />
REFERENCES<br />
1 J. S. Arney and Hung Tran, “An inexpensive micro-goniophotometry<br />
you can build”, Proc. IS&T’s PICS Conference on Digital Image Capture,<br />
Reproduction, and Image Quality (IS&T, Springfield, VA, 2002) pp.<br />
179–182.<br />
2 J. S. Arney, H. Hoon, and P. G. Anderson, “A micro-goniophotometer<br />
and the measurement <strong>of</strong> print gloss”, J. <strong>Imaging</strong> Sci. Technol. 48, 458<br />
(2003).<br />
3 J. M. Bennett and L. Mattsson, Introduction to Surface Roughness and<br />
Scattering, 2nd ed. (Optical Soc. <strong>of</strong> America, Washington, DC, 1999),<br />
Chap. 3.<br />
4 J. C. Stover, Optical Scattering, Measurement and Analysis (McGraw Hill,<br />
NY, 1990).<br />
5 I. Nimer<strong>of</strong>f, “Two-parameter gloss methods”, J. Res. Natl. Bur. Stand.<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 297
Arney et al.: Gloss granularity <strong>of</strong> electrophotographic prints<br />
58(3), 127 (1957).<br />
6 Standard Practice <strong>for</strong> Angle Resolved Optical Scatter Measurements on<br />
Specular or Diffuse Surfaces, Standard Procedure No. E 1392-96,<br />
American <strong>Society</strong> <strong>for</strong> Testing and <strong>Material</strong>s, 1996.<br />
7 I. Nimer<strong>of</strong>f, “Analysis <strong>of</strong> goniophotometric reflection curves”, J. Res.<br />
Natl. Bur. Stand. 48(6), 441 (1952).<br />
8 H. Rothe and D. Huester, “Application <strong>of</strong> circular and spherical statistics<br />
<strong>for</strong> the interpretation <strong>of</strong> BRDF measurements”, Proc. SPIE 3141(02), 13<br />
(1997).<br />
9 M. Colbert, S. Pattanaik, and J. Krivanek, “BRDF-shop: creating<br />
physically correct bidirectional reflectance distribution functions”, IEEE<br />
Comput. Graphics Appl. 26(1), 30 (2006).<br />
10 P.-Å. Johansson, “Optical homogeneity <strong>of</strong> prints”, doctoral thesis, KTH,<br />
Royal Institute <strong>of</strong> Technology, Stockholm, Sweden, 1999.<br />
11 M.-C. Béland, “Gloss variation <strong>of</strong> printed paper: relationship between<br />
topography and light scattering”, doctoral thesis, KTH, Royal Institute <strong>of</strong><br />
Technology, Stockholm, Sweden, 2001.<br />
12 R. E. Swing, An Introduction to Microdensitometry (SPIE Optical<br />
Engineering Press, Bellingham, WA, 1998).<br />
13 J. S. Arney, P. G. Engeldrum, and H. Zeng, “An Expanded Murray-<br />
Davies model <strong>of</strong> tone reproduction in halftone imaging”, J. <strong>Imaging</strong> Sci.<br />
Technol. 39, 502 (1995).<br />
14 J. S. Arney, C. Scigaj, and P. Mehta, “Linear color addition in halftones”,<br />
J. <strong>Imaging</strong> Sci. Technol. 45, 426 (2001).<br />
15 Y. Kipman, P. Mehta, K. Johnson, and D. Wolin, “A new method <strong>of</strong><br />
measuring gloss mottle and micro-gloss using a line-scan CCD camera<br />
based imaging system”, Proc. IS&T’s NIP17 (IS&T, Springfield, VA,<br />
2001) p. 714.<br />
16 J. S. Arney, L. Ye, and S. Banach, “Interpretation <strong>of</strong> gloss meter<br />
measurements”, J. <strong>Imaging</strong> Sci. Technol. 50(6), 567 (2006).<br />
17 J. S. Arney, P. G. Anderson, G. Franz, and W. Pfeister, “Color properties<br />
<strong>of</strong> specular reflections”, J. <strong>Imaging</strong> Sci. Technol. 50(3), 228 (2006).<br />
18 J. S. Arney, L. Ye, J. Wible, and T. Oswald, “Analysis <strong>of</strong> paper gloss”, J.<br />
Pulp Pap. Sci. 32(1), 19 (2006).<br />
19 M. Ashikhmin, S. Premoze, and P. Shirley, “A micr<strong>of</strong>aceted-based BRDF<br />
generator,” Proc. SIGGRAPH (ACM Press, NY, 2000) pp. 65–74.<br />
298 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
<strong>Journal</strong> <strong>of</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology® 51(4): 299–309, 2007.<br />
© <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology 2007<br />
Forensic Examination <strong>of</strong> Laser Printers and Photocopiers<br />
Using Digital Image Analysis to Assess Print<br />
Characteristics<br />
J. S. Tchan<br />
MATAR Research Group, London College <strong>of</strong> Communication, Elephant and Castle,<br />
London SE1 6SB, England<br />
E-mail: j.tchan@lcc.arts.ac.uk<br />
Abstract. The work in this paper describes a method that can assist<br />
the process <strong>of</strong> print identification with respect to the printing<br />
machine that produced it. The method used high spatial resolution<br />
and low-noise digital image analysis to measure the sharpness, intensity<br />
and size characteristics <strong>of</strong> individual text characters. The<br />
relative variations <strong>of</strong> these variables were used to identify the machine<br />
that produced the print under examination. The results<br />
showed that three machines could be distinguished and one <strong>of</strong><br />
these machines also showed differences in the print produced when<br />
the toner cartridge was changed. © 2007 <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong><br />
and Technology.<br />
DOI: 10.2352/J.<strong>Imaging</strong>Sci.Technol.200751:4299<br />
Received Oct. 12, 2006; accepted <strong>for</strong> publication Mar. 30, 2007.<br />
1062-3701/2007/514/299/11/$20.00.<br />
INTRODUCTION<br />
A reason why it is frequently not feasible to prosecute counterfeiters<br />
and document fraudsters is due to the difficulty in<br />
establishing links between the counterfeiters or fraudsters<br />
and their printing equipment. This is because <strong>of</strong> the extremely<br />
wide range <strong>of</strong> both cheap and expensive laser and<br />
ink jet printing machines available. This problem is exacerbated<br />
by the fact that new varieties <strong>of</strong> printing machine are<br />
being commercially produced on a continual and frequent<br />
basis.<br />
Also both laser and ink jet printers, which make up<br />
most <strong>of</strong> the cheap <strong>of</strong>fice printing market, have disposable<br />
ink and toner cartridges, or the cartridges can be refilled.<br />
This can prevent valid chemical analysis <strong>of</strong> any similarities or<br />
differences in chemical compositions from the various ink<br />
and toner cartridges. Also chemical analysis is a process that<br />
requires the destruction <strong>of</strong> part <strong>of</strong> the evidence.<br />
Methods which involve microscopy 1 are frequently used<br />
by <strong>for</strong>ensic scientists to determine the production source <strong>of</strong><br />
digital print. The linking <strong>of</strong> a document to a digital printer<br />
in these cases usually involves analyzing variables such as ink<br />
or toner overspray and assessing alignment, spacing and<br />
copy distortion.<br />
Investigations have been carried out by Oliver and<br />
Chen, 2 Tchan, Thompson, and Manning, 3 and Tchan 4,5 using<br />
digital image analysis to link documents to printing machines.<br />
Oliver and Chen have studied the relationship between<br />
the raggedness <strong>of</strong> print and text character distortion<br />
<strong>of</strong> different printers. Tchan, Thompson, and Manning have<br />
taken a similar approach but have also used neural networks<br />
to link the contrast, noise and edge characteristics <strong>of</strong> printed<br />
text to the printing machine that produced it. These attempts<br />
at using digital image analysis however could only<br />
provide a positive test <strong>for</strong> a small range <strong>of</strong> printing machines<br />
and do not account <strong>for</strong> the effect <strong>of</strong> replaceable toner and<br />
ink cartridges.<br />
The analysis <strong>of</strong> the actual shapes <strong>of</strong> the text characters<br />
produced by different print engines using digital image<br />
analysis is another possible way <strong>of</strong> fingerprinting printing<br />
machines, according to Tchan. 6 However, this method not<br />
only suffers from the influence <strong>of</strong> replaceable ink and toner<br />
components distorting the results <strong>of</strong> the analysis, but other<br />
drawbacks as well. First, processing huge amounts <strong>of</strong> image<br />
data is time-consuming due to the large number <strong>of</strong> fonts<br />
and their range <strong>of</strong> sizes from many different makes and<br />
models <strong>of</strong> printing machines. Secondly, the problem <strong>of</strong> ink<br />
spread on different kinds <strong>of</strong> paper or humidity conditions in<br />
ink jet printing processes distorts the shapes <strong>of</strong> text characters.<br />
An identification methodology <strong>for</strong> fingerprinting printing<br />
machines recently considered is to use a technique called<br />
ESDA 7 (Electrostatic Detection Apparatus). ESDA has been<br />
employed to detect and evaluate roller pressure marks on<br />
paper. 8 These pressure marks are due to the interaction between<br />
the paper feeder rollers and the paper substrate. The<br />
original application was <strong>for</strong> the detection <strong>of</strong> pen impressions<br />
several layers down in a notepad. It works by charging paper<br />
surfaces with a high voltage. If toner particles are applied to<br />
the charged paper surface, imperceptible pen marks due to<br />
writing on a piece <strong>of</strong> paper placed above this sheet may be<br />
revealed. This means <strong>for</strong> example, in a notepad, writing can<br />
be read from sheets many layers down from the top sheet<br />
that has the actual writing. As the ESDA technique has been<br />
shown to detect weak imperceptible pressure marks from<br />
pens, it might be able to detect pressure marks from printing<br />
rollers.<br />
If the pressure marks can be detected, then to link a<br />
document to the printer that produced it would require<br />
comparing the width and spacing characteristics <strong>of</strong> the roll-<br />
299
Tchan: Forensic examination <strong>of</strong> laser printers and photocopiers using digital image analysis…<br />
Table I. List <strong>of</strong> the three printing machines and the printing machine with a change <strong>of</strong><br />
toner cartridge used in the investigation.<br />
Printing System<br />
Canon IR 3570 Photocopier<br />
HP 1200 Laser Jet Printer<br />
HP 4250 Laser Jet Printer Toner Cartridge 1<br />
HP 4250 Laser Jet Printer Toner Cartridge 2<br />
ers <strong>of</strong> the machine in question. However, detecting the pressure<br />
marks may not always be possible <strong>for</strong> the following<br />
reasons. First, the pressure exerted by the rollers is weak,<br />
generally much weaker than pen pressure. If the pressure<br />
marks are present, they will sometimes be difficult to detect,<br />
even with the most sophisticated digital image processing<br />
systems. Secondly, assuming that the pressure marks are detectable,<br />
heavy handling can destroy them.<br />
A similar technique that has been recently explored concerns<br />
ink jet printers and indentations on the paper after a<br />
printed sheet has been fed through the machine. 9 These indentations<br />
are caused by the spoke wheels that feed the paper<br />
through the ink jet printer and are most perceptible on<br />
moist parts <strong>of</strong> the paper caused by wet ink. Due to the heavy<br />
pressure imparted by the spoke wheels, the indentations can<br />
be seen using optical microscopy without the aid <strong>of</strong> ESDA.<br />
However not all ink jet printers have spoke wheels so this<br />
technique does not apply to all ink jet printing systems.<br />
Another type <strong>of</strong> method that has been considered <strong>for</strong><br />
the identification <strong>of</strong> laser printers exploits imperfections in<br />
the print known as banding. Banding imperfections are lines<br />
across the printed page when smooth print is required. 10<br />
The effect has been attributed to the following two causes.<br />
First, fine banding due to the imbalance <strong>of</strong> the rotor component<br />
<strong>of</strong> the polygon mirror or mechanical weaknesses <strong>of</strong><br />
the laser scanning unit. Secondly, rough banding caused by<br />
unsteady motion <strong>of</strong> the photoconductor drum or the fuser<br />
unit.<br />
Mikkilineni et al. 11 have devised a system that uses a<br />
scanner to analyze relative texture differences on a printed<br />
page caused by banding effects. This system has shown that<br />
9 out <strong>of</strong> a set <strong>of</strong> 10 laser printing machines were successfully<br />
identified.<br />
The method described in this paper is an alternative<br />
method <strong>of</strong> measuring banding effects in laser printers and<br />
photocopiers. Instead <strong>of</strong> scanning images and analyzing the<br />
relative texture <strong>of</strong> text characters, it uses a high resolution<br />
and low-noise digital image analysis system to measure the<br />
following variables in printed text. These variables are sharpness,<br />
intensity, and size. The following section describes the<br />
methodology and the experimental setup involved.<br />
EXPERIMENTAL PROCEDURE<br />
When a completely black page was printed out on a photocopier,<br />
two different laser printers and on one <strong>of</strong> the two<br />
laser printers with a different toner cartridge, Table I, the<br />
following effects in Figure 1 were seen. These are sketches <strong>of</strong><br />
Figure 1. Sketches <strong>of</strong> the lines produced by the four different printing<br />
samples used in the investigation.<br />
the banding lines seen with an indication <strong>of</strong> their dimensions<br />
and separations. It was observed that some <strong>of</strong> the horizontal<br />
lines from the HP 4250 <strong>for</strong> the same toner cartridge<br />
were not in fixed positions and the lines were not always<br />
equal in number <strong>for</strong> different printed sheets. It is unknown<br />
whether some <strong>of</strong> the lines were random or followed a complex<br />
pattern since further investigation <strong>of</strong> this effect has yet<br />
to be made. However, <strong>for</strong> this part <strong>of</strong> the investigation, only<br />
a confirmation <strong>of</strong> the existence <strong>of</strong> the banding effects that<br />
are a common feature <strong>of</strong> digital printers was required.<br />
The experimental procedure can also be separated into<br />
three distinct stages. In the first stage the banding effect was<br />
observed <strong>for</strong> a test page that was entirely covered in solid<br />
toner as stated above. In stage two, a test page <strong>of</strong> the same<br />
text character was produced and physical differences in the<br />
print were investigated <strong>for</strong> each printing machine used. This<br />
was completed using high-resolution digital image analysis.<br />
In stage three, a page <strong>of</strong> ordinary text was produced and<br />
patterns in the text were again investigated using highresolution<br />
digital image analysis. The process is illustrated in<br />
Figure 2, the flow chart below, and will be discussed in<br />
greater detail later.<br />
The effects <strong>of</strong> banding on printed text were investigated<br />
using a high-resolution digital image analysis system, which<br />
has been built specifically to analyze the print. Figure 3 illustrates<br />
how the camera was attached to the stand and how<br />
the lighting source was attached to the camera.<br />
300 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Tchan: Forensic examination <strong>of</strong> laser printers and photocopiers using digital image analysis…<br />
Figure 2. The three different stages <strong>of</strong> the investigation.<br />
Figure 4. The three stages required to make the necessary measurements<br />
<strong>for</strong> the analysis.<br />
Table II. A summary <strong>of</strong> the measurements taken.<br />
Measurements<br />
Variable<br />
1 Location <strong>of</strong> the peak maximum <strong>for</strong> the print<br />
region, Fig. 5.<br />
2 The value <strong>of</strong> the peak maximum <strong>for</strong> the print<br />
region, Fig. 5.<br />
3 The text character area is calculated by counting<br />
the number <strong>of</strong> pixels below an arbitrarily selected<br />
threshold between the nonimage and image<br />
peaks, Fig. 5.<br />
4 Integration <strong>of</strong> the peak area <strong>for</strong> the print region<br />
divided by the text character area to determine<br />
the average image intensity, Fig. 5.<br />
Figure 3. The camera, lens, lighting system, and stand.<br />
The camera employed in the investigation was a<br />
Hamamatsu C4742-95 camera. This camera had a Peltier<br />
cooled CCD chip to increase the signal to noise ratio. The<br />
camera was attached firmly to a camera stand that weighs<br />
approximately 30 kg. The lighting unit was firmly screwed<br />
onto the lens <strong>of</strong> the camera. The lighting unit consisted <strong>of</strong> a<br />
circular array <strong>of</strong> red LEDS. The LEDS were connected to a<br />
laboratory power supply with low ripple.<br />
Figure 4 shows in block diagram <strong>for</strong>m the different<br />
hardware and s<strong>of</strong>tware components <strong>of</strong> the image analysis<br />
system. The image data from the camera was digitized to<br />
8 bit resolution using Matrox Mil s<strong>of</strong>tware. The data was<br />
subsequently analyzed using Visual Basic that was compatible<br />
with Matrox Mil. Algorithms were developed using Visual<br />
Basic Active X language that could per<strong>for</strong>m the following<br />
computations on individual text characters.<br />
Four measurements were taken from the text characters.<br />
They related to the image sharpness, intensity and size <strong>of</strong> the<br />
print under investigation. These are summarized in Table II.<br />
The image window size, Figure 5, was 480 by 483 and had a<br />
tonal resolution <strong>of</strong> 256.<br />
Measurement 1 is the position <strong>of</strong> the peak in standard<br />
8 bit gray scale <strong>for</strong> the tonal distribution <strong>of</strong> the printed region.<br />
The location <strong>of</strong> this peak has the lower gray scale<br />
value, in this case at about 55, Fig. 5. The other peak corresponds<br />
to the tonal distribution <strong>of</strong> the unprinted white paper.<br />
The position <strong>of</strong> the peak <strong>for</strong> the white paper used in the<br />
investigation is located at just over 100, Fig. 5. Such a low<br />
value is due to the arbitrarily low lighting exposure chosen<br />
<strong>for</strong> the CCD camera. Attempting to stretch the distance between<br />
the two peaks <strong>for</strong> the black print and white paper<br />
regions, by increasing the exposure too much, can sometimes<br />
reduce the precision <strong>of</strong> the system.<br />
Measurement 2 is the height <strong>of</strong> the peak from the tonal<br />
distribution <strong>of</strong> the printed region.<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 301
Tchan: Forensic examination <strong>of</strong> laser printers and photocopiers using digital image analysis…<br />
Figure 6. The two sets <strong>of</strong> test targets that were used to produce the results<br />
in this investigation.<br />
Figure 5. An illustration <strong>of</strong> the data acquired from the image analysis<br />
system.<br />
Measurement 3 is the area <strong>of</strong> the printed text character,<br />
this is calculated by counting the number <strong>of</strong> pixels below a<br />
fixed arbitrary gray scale level.<br />
Measurement 4 averages the overall intensity <strong>of</strong> the<br />
printed region by integrating the total intensity from the<br />
printed region and dividing it by the number <strong>of</strong> pixels from<br />
the text character below the fixed arbitrary level chosen <strong>for</strong><br />
measurement 3. Figure 5 and Table II illustrate and summarize<br />
the measurements made. The measurements were taken<br />
individually and sequentially <strong>for</strong> each text character by<br />
manual alignment under the image window.<br />
Figure 6 shows the two print samples used in the investigation.<br />
They were both printed on the same batch <strong>of</strong> standard<br />
laser printer paper in all cases. The font was Times New<br />
Roman and the font size was 22 pts. 22 pts is a large-sized<br />
font; it was used because it facilitated relatively quick measurements<br />
to show that the method is viable.<br />
A test sheet was produced that consisted <strong>of</strong> a series <strong>of</strong><br />
the letter “W” in Times Roman font and 22 points in size.<br />
There were 17 W’s across the page and 30 down the page.<br />
The selection <strong>of</strong> the letter “W” was arbitrary. However, the<br />
size was important to facilitate ease <strong>of</strong> measurement when<br />
recording the data using the digital image analysis system.<br />
Figure 7. In the case <strong>of</strong> the normal text page a mask was required to<br />
eliminate the effect <strong>of</strong> adjacent text characters.<br />
In the case <strong>of</strong> the page <strong>of</strong> normal text, Fig. 6 on the<br />
right, a cardboard mask, Figure 7, was required to shield the<br />
effects <strong>of</strong> nearby letters influencing the readings since unwanted<br />
parts <strong>of</strong> letters appeared in the image window. The<br />
test set <strong>of</strong> “W’s” did not have this effect since the spacing <strong>of</strong><br />
the “W’s” was designed to eliminate the requirement <strong>for</strong> a<br />
mask. Figure 8 shows the “e’s” that were analyzed in the page<br />
<strong>of</strong> normal text.<br />
RESULTS<br />
First, the accuracy <strong>of</strong> the system was established by assessing<br />
variations in the intensity and area measurements <strong>for</strong> a<br />
302 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Tchan: Forensic examination <strong>of</strong> laser printers and photocopiers using digital image analysis…<br />
Figure 8. The letter “e’s” that where selected from the test page <strong>for</strong> the<br />
classification <strong>of</strong> the two toner cartridges <strong>of</strong> the HP 4250.<br />
single text character sampled 50 times. A maximum value<br />
from the mean <strong>of</strong> ±0.05% was observed when the mask was<br />
not employed and ±0.03% when the mask was used <strong>for</strong> the<br />
intensity measurements. These error values were tripled in<br />
size <strong>for</strong> the area measurements. This text character print<br />
sample was also used to check <strong>for</strong> any substantial drift in the<br />
system at the beginning <strong>of</strong> each day over a period <strong>of</strong> about<br />
20 days that the measurements were taken. Substantial drift,<br />
which could be caused by small change in the LED voltage,<br />
lens settings or focus, did not occur over the period <strong>of</strong> data<br />
collection. This was probably due to the mechanical robustness<br />
<strong>of</strong> the optical system and the quality <strong>of</strong> the LED illumination<br />
system.<br />
Secondly, two W test pages from the HP 4250 printer,<br />
one from each <strong>of</strong> the two toner cartridges were analyzed.<br />
This, as in all <strong>of</strong> this investigation, required individual sequential<br />
manual alignment and subsequent measurement<br />
from each text character on a line <strong>of</strong> text. Figure 9 shows<br />
how the size <strong>of</strong> the letter “W” <strong>for</strong> line 1 and 10 <strong>of</strong> the grid<br />
changes down the page <strong>for</strong> the HP 4250 printer using the<br />
two different toner cartridges. The consistent patterns indicate<br />
that the digital image analysis system has recorded<br />
meaningful results.<br />
Thirdly, a demonstration was made <strong>of</strong> how the four<br />
printing samples could be distinguished using the W template.<br />
This was achieved by using three adjacent W test<br />
sheets from each <strong>of</strong> the four printing samples in a print run<br />
<strong>of</strong> 10. The reported results in this section used only the first<br />
horizontal and vertical lines <strong>of</strong> the W test sheets because <strong>of</strong><br />
Figure 9. A comparison <strong>of</strong> the text character size down line 1 and 10 <strong>for</strong> the two HP 4250 toner cartridges.<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 303
Tchan: Forensic examination <strong>of</strong> laser printers and photocopiers using digital image analysis…<br />
Figure 10. A comparison <strong>of</strong> the text character area down line 1 <strong>of</strong> the W test pages. Left, <strong>for</strong> 3 adjacent<br />
pages; right, their average.<br />
the time-consuming nature <strong>of</strong> recording the measurements<br />
and associated time constraints <strong>of</strong> the researcher. The time<br />
problem was only discovered during the experimental phase<br />
and centered on alignment difficulties <strong>of</strong> the text characters<br />
in the image window. The left-hand graphs <strong>of</strong> Figures 10–13<br />
show the individual data from the three sheets and on the<br />
right-hand side their averages.<br />
It was shown that the peak size, the average intensity<br />
and the size measurements yielded useful in<strong>for</strong>mation <strong>for</strong><br />
the classification process. The position <strong>of</strong> the image peak<br />
304 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Tchan: Forensic examination <strong>of</strong> laser printers and photocopiers using digital image analysis…<br />
Figure 11. A comparison <strong>of</strong> the text character area across line 1 <strong>of</strong> the W test pages. Left, <strong>for</strong> 3 adjacent<br />
pages; right, their average.<br />
remained constant at either 54 or 55 and provided no useful<br />
classification data. It does however provide a useful check on<br />
the stability <strong>of</strong> the illumination levels throughout the period<br />
when the measurements were taken. The experimental results<br />
using the W template show that all four printing<br />
samples could be differentiated by a combination <strong>of</strong> the<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 305
Tchan: Forensic examination <strong>of</strong> laser printers and photocopiers using digital image analysis…<br />
Figure 12. A comparison <strong>of</strong> the relative average intensity across the W test pages, left, <strong>for</strong> 3 adjacent pages;<br />
right, their average.<br />
analysis <strong>of</strong> the area variation <strong>of</strong> the text characters down the<br />
page and the average intensity, peak size and area variations<br />
<strong>of</strong> the text characters across the page, Figs. 10–13.<br />
Finally, the masking technique was employed, Fig. 7,<br />
with the printed page <strong>of</strong> normal text shown in Figs. 6 and 8<br />
to find differences from the different printing examples, in<br />
this case from the HP 4250 printer when the toner cartridge<br />
was changed. Figure 14 shows a shallow well or an inverted<br />
sharp spike at position 15 <strong>for</strong> toner cartridge 1 and an<br />
inverted sharp spike <strong>for</strong> toner cartridge 2 at position 14. This<br />
result was obtained from a print run <strong>of</strong> 201. The graphs,<br />
Fig. 14, show the results from sheets 1 <strong>of</strong> 101 and 201. The<br />
306 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Tchan: Forensic examination <strong>of</strong> laser printers and photocopiers using digital image analysis…<br />
Figure 13. A comparison <strong>of</strong> the text character peak intensities across the W test pages. Left, <strong>for</strong> 3 adjacent<br />
pages; right, their average.<br />
measurements were also sampled <strong>for</strong> other sheets in the<br />
print run at 20 sheet intervals (numbers 1, 21, 41, 61, 81,<br />
101, 121, 141, 161, 181, and 201 in total) and were<br />
100% consistent in the fact that the inverted sharp spike at<br />
position 14 only appears <strong>for</strong> all print samples from toner<br />
cartridge 2 and not at all <strong>for</strong> toner cartridge 1.<br />
CONCLUSIONS<br />
The results <strong>of</strong> the investigation thus far demonstrate the potential<br />
<strong>of</strong> the method <strong>for</strong> the <strong>for</strong>ensic analysis <strong>of</strong> print, both<br />
in linking a machine to a particular document and to show<br />
whether a document has been tampered with.<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 307
Tchan: Forensic examination <strong>of</strong> laser printers and photocopiers using digital image analysis…<br />
Figure 14. The area differences using the page <strong>of</strong> text <strong>for</strong> the HP 4250 laser printer <strong>for</strong> the two toner<br />
cartridges.<br />
It has been shown that a small number <strong>of</strong> toner based<br />
printing systems can be classified using high-resolution image<br />
analysis to measure the relative changes in the physical<br />
properties <strong>of</strong> individual text characters both across and<br />
down pages <strong>of</strong> printed text. Even at this stage <strong>of</strong> its development<br />
the system has potentially useful <strong>for</strong>ensic applications.<br />
Also, these results correlate with the work carried out by<br />
Mikkilineni et al. on the measurement <strong>of</strong> surface texture by<br />
scanning the text characters <strong>of</strong> laser printers.<br />
In this investigation more work is required on smaller<br />
text characters. In particular positive results are required on<br />
font sizes <strong>of</strong> 10 since this is typical <strong>for</strong> documents. If limitations<br />
in the hardware become apparent or the banding<br />
signatures become weaker when smaller font sizes are considered<br />
then a larger set <strong>of</strong> text characters could be analyzed<br />
statistically to try to overcome the limitations.<br />
However, the further work stated above requires assistance<br />
from better measurement and analysis techniques with<br />
greater precision. This is due to the labor intensive nature <strong>of</strong><br />
the experimental work. Statistical analysis techniques such as<br />
moving averages or autocorrelation analysis can enhance the<br />
data, thereby reducing the volume <strong>of</strong> data required from the<br />
308 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Tchan: Forensic examination <strong>of</strong> laser printers and photocopiers using digital image analysis…<br />
print samples <strong>for</strong> accurate classification. The current masking<br />
method is difficult to carry out in practice <strong>for</strong> a large<br />
number <strong>of</strong> small text characters and needs an alternative. A<br />
possible solution to this problem is to use or develop character<br />
recognition s<strong>of</strong>tware that can automatically isolate and<br />
classify individual text characters.<br />
The suggestions given <strong>for</strong> further work indicate, in the<br />
longer term, since data acquisition is time-consuming, an<br />
automated system that uses a high-resolution and fast scanning<br />
device is needed. This method is expensive to develop<br />
and implement but will greatly improve the volume <strong>of</strong> print<br />
that can be processed in a given time, and should enable<br />
machines to be examined more both quickly and more accurately<br />
from smaller font sized print.<br />
REFERENCES<br />
1 B. S. Lindbolm, and R. Gervais, Scientific Examination <strong>of</strong> Questioned<br />
Documents (Taylor and Francis, Boca Raton, FL, 2006).<br />
2 J. Oliver, and J. Chen, “Use <strong>of</strong> signature analysis to discriminate digital<br />
printing technologies”, Proc. IS&T’s NIP18 (IS&T, Springfield, VA, 2002)<br />
pp. 218–222.<br />
3 J. S. Tchan, R. C. Thompson and A. Manning, “The use <strong>of</strong> neural<br />
networks in an image analysis system to distinguish between laser prints<br />
and their photocopies”, J. <strong>Imaging</strong> Sci. Technol. 44(2), 132–144 (2000).<br />
4 J. S. Tchan, “Classifying digital prints according to their production<br />
process using image analysis and artificial neural networks”, Proc. SPIE<br />
3973, 105–116, (2000).<br />
5 J. S. Tchan, “The development <strong>of</strong> an image analysis system that can<br />
detect fraudulent alterations made to printed images”, Proc. SPIE 5310,<br />
151–159 (2004).<br />
6 J. S. Tchan, “Forensic analysis <strong>of</strong> print using digital image analysis”,<br />
Proc. SPIE 5007, 61–72 (2003).<br />
7 J. Levinson, Questioned Documents: A Lawyer’s Handbook (Academic<br />
Press, London, 2001).<br />
8 G. M. Laporte, “The use <strong>of</strong> an electrostatic detection device to identify<br />
individual and class characteristics on documents produced by printers<br />
and copiers-A preliminary study”, J. Forensic Sci. 49(3), 610–620 (2004).<br />
9 Y. Akao, K. Kobayashi, and Y. Seki, “Examination <strong>of</strong> spur marks found<br />
on inkjet printed documents”, J. Forensic Sci., 50(4), 915–923 (2005).<br />
10 J. You, “Banding reduction in an electrophotographic printer”, J.<br />
<strong>Imaging</strong> Sci. Technol. 49(6), 635–640 (2005).<br />
11 A. K. Mikkilineni, P. Chiang, G. N. Ali, G. T. C. Chiu, J. P. Allebach, and<br />
E. J. Delp, “Printer identification based on graylevel co-occurrence<br />
features <strong>for</strong> security and <strong>for</strong>ensic applications”, Proc. SPIE 5681,<br />
430–440 (2005).<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 309
<strong>Journal</strong> <strong>of</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology® 51(4): 310–316, 2007.<br />
© <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology 2007<br />
Moiré Analysis <strong>for</strong> Assessment <strong>of</strong> Line<br />
Registration Quality<br />
Nathir A. Rawashdeh, Daniel L. Lau and Kevin D. Donohue <br />
University <strong>of</strong> Kentucky, ECE, 453 F. Paul Anderson Tower, Lexington, Kentucky 40506-0046<br />
E-mail: nathir@ieee.org<br />
Shaun T. Love<br />
Lexmark International, Inc., 740 W. New Circle Rd., Lexington, Kentucky 40550<br />
Abstract. This paper introduces objective macro and micro line<br />
registration quality metrics based on Moiré interference patterns<br />
generated by superposing a lenticular lens grating over a hardcopy<br />
test page consisting <strong>of</strong> high-frequency Ronchi rulings. Metrics <strong>for</strong><br />
macro and micro line registration are defined and a measurement<br />
procedure is described to enhance the robustness <strong>of</strong> the metric<br />
computation over reasonable variations in the measurement process.<br />
The method analyzes low frequency interference patterns,<br />
which can be scanned at low resolutions. Experimental measurements<br />
on several printers are presented to demonstrate a comparative<br />
quality analysis. The metrics demonstrate robustness to small<br />
changes in the lenticular lens and grating superposition angle. For<br />
superposition angles varying between 2° and 5°, the coefficients <strong>of</strong><br />
variance <strong>for</strong> the two metrics are less than 5%, which is small enough<br />
<strong>for</strong> delineating between test patterns <strong>of</strong> different print quality.<br />
© 2007 <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology.<br />
DOI: 10.2352/J.<strong>Imaging</strong>Sci.Technol.200751:4310<br />
INTRODUCTION<br />
Image quality analysis is an important component in the<br />
development and operation <strong>of</strong> various digital imaging technologies,<br />
such as displays, scanners and printers. To produce<br />
visually pleasing images, devices must be designed to minimize<br />
defects, such as problems related to color registration<br />
and line quality. An efficient way <strong>of</strong> measuring imaging defects<br />
is through the use <strong>of</strong> special test targets, which are<br />
designed to test the limits <strong>of</strong> the respective imaging technology.<br />
Analysis based on test target results can be used to track<br />
and minimize image defects during the development phase.<br />
This paper presents a method <strong>for</strong> analyzing printed line<br />
quality by analyzing the Moiré patterns resulting from the<br />
superposition <strong>of</strong> a test pattern, consisting <strong>of</strong> finely spaced<br />
lines, and an array <strong>of</strong> cylindrical lenses <strong>of</strong> similar spacing.<br />
Other approaches to line quality attributes <strong>for</strong> hardcopy<br />
output include blurriness, raggedness, stroke width, darkness,<br />
contrast, fill, and registration. 1–3 The test targets <strong>for</strong><br />
these measures consist <strong>of</strong> a printed black line on a white<br />
background. The quality attributes are then quantified<br />
through measurements from the printed line. Blurriness<br />
measures the average transition length from light to dark,<br />
<br />
IS&T Member<br />
Received Jan. 5, 2007; accepted <strong>for</strong> publication Mar. 1, 2007.<br />
1062-3701/2007/514/310/7/$20.00.<br />
and raggedness measures the geometric distortion <strong>of</strong> the<br />
line’s edge from its ideal shape. Line width is the average<br />
stroke width measured from either edge along a direction<br />
normal to the line under analysis. Line darkness measures<br />
the mean line density, which can vary due to voids <strong>for</strong> example.<br />
The contrast attribute captures the relationship between<br />
the darkness <strong>of</strong> the line and that <strong>of</strong> its surrounding<br />
field by measuring the mean reflectance factors. Contrast<br />
can vary due to blurring, extraneous marks, haze, or substrate<br />
type. Fill refers to the appearance <strong>of</strong> darkness within<br />
the inner boundary <strong>of</strong> the line. One example <strong>of</strong> line registration<br />
is the color registration <strong>of</strong> the CMYK components in<br />
an inkjet printer. If the same line is printed once with each<br />
color, then ideally, all four color lines should collapse into<br />
one, and any consistent increase in line width would indicate<br />
position errors, or mis-registration, <strong>of</strong> one or more ink<br />
components. 3<br />
This paper introduces new metrics that differ from previous<br />
line quality attribute measures in that they are directly<br />
based on the printer’s ability to create fine detailed lines.<br />
While this metric may be influenced by measures such as<br />
raggedness and blur, its use <strong>of</strong> fine details makes it unique<br />
relative to previous measures. The measurement method involves<br />
the analysis <strong>of</strong> low frequency Moiré patterns that<br />
change according to small changes in the test patterns. The<br />
test pattern consists <strong>of</strong> finely spaced parallel lines, which an<br />
imperfect printer reproduces with some line placement (or<br />
registration) errors. The parallel lines are no longer uni<strong>for</strong>mly<br />
spaced in this case, and this is reflected in the resulting<br />
Moiré line shape. Moiré patterns are used as a nondestructive<br />
analysis tool in Moiré interferometry. For this<br />
method a photographic grid is printed on the surface <strong>of</strong> a<br />
material under investigation and is irradiated by coherent<br />
light. The interfering fringes (Moiré patterns) can indicate<br />
the presence <strong>of</strong> local stress and de<strong>for</strong>mation <strong>for</strong> in-plane<br />
displacement. 4,5 Moiré interferometry techniques have the<br />
advantage <strong>of</strong> being able to analyze a broad range <strong>of</strong> engineering<br />
materials in small analysis zones at high spatial resolution<br />
and sensitivity. This work extends the principles <strong>of</strong><br />
Moiré interferometry to assess line registration quality by<br />
analyzing the Moiré patterns produced by the superposition<br />
310
Rawashdeh et al.: Moiré analysis <strong>for</strong> assessment <strong>of</strong> line registration quality<br />
Figure 1. Moiré pattern <strong>of</strong> spacing P and direction <strong>for</strong>med by the<br />
angled superposition <strong>of</strong> a Ronchi ruling and a lenticular grating.<br />
<strong>of</strong> a lenticular grating on a printed Ronchi-ruling test pattern<br />
to characterize underlying printed line registration<br />
errors.<br />
The lenticular grating consists <strong>of</strong> a plastic sheet that is<br />
smooth on one side and holds an array <strong>of</strong> parallel cylindrical<br />
lenses or prisms on the other side. The printed test pattern is<br />
a Ronchi ruling (a rectangular spatial wave linear grating)<br />
with a similar line spacing as the lenticular grating. This<br />
quality assessment approach lends itself to automation, since<br />
the lenticular grating is thin enough to be fixed to the glass<br />
surface <strong>of</strong> a flatbed scanner and does not interfere with its<br />
automatic document feeder mechanism. Since only the<br />
shape <strong>of</strong> the Moiré lines is used <strong>for</strong> analysis, it is sufficient to<br />
use a relatively inexpensive scanner (or scan faster), because<br />
high-resolution detail and tone reproduction accuracy are<br />
not crucial. This paper presents the underlying equations<br />
affecting the critical details <strong>of</strong> the Moiré patterns, describes a<br />
procedure <strong>for</strong> robust measurement and computation <strong>of</strong><br />
macro and micro line quality metrics, and presents results<br />
<strong>for</strong> several printers. Measurements are analyzed and compared<br />
to a visual assessment <strong>of</strong> line quality based on a magnified<br />
view <strong>of</strong> the Ronchi pattern created with a high resolution<br />
scanner.<br />
The text is organized as follows. The Moiré Model section<br />
describes the Moiré line model and discusses normalization<br />
techniques and ranges <strong>of</strong> superposition angles <strong>for</strong><br />
robust measurements. The Line Registration Quality Measurements<br />
and Metrics section describes the measurement<br />
procedure and computation <strong>of</strong> the macro and micro line<br />
registration metrics. The Results and Analysis section presents<br />
measurement results from three different printers and<br />
analyzes measurement variability and quality assessment. Finally,<br />
the Conclusion section summarizes results and presents<br />
conclusions.<br />
MOIRÉ MODEL<br />
Figure 1 illustrates the Moiré fringe pattern produced by the<br />
superposition <strong>of</strong> a (printed) linear grid <strong>of</strong> spacing P 0 , and a<br />
lenticular grating <strong>of</strong> spacing P 1 at an angle . The Moiré<br />
lines are produced by the lenticular lenses intersecting with<br />
the individual lines <strong>of</strong> the Ronchi ruling. Only two lenses are<br />
illustrated in this figure; however, a sheet consisting <strong>of</strong> many<br />
lenticular lenses produces extended patterns <strong>of</strong> Moiré lines.<br />
Figure 2. Photograph <strong>of</strong> a printed horizontal Ronchi ruling test pattern at<br />
a small angle with a superimposed lenticular lens grating. Resulting Moiré<br />
patterns are dark vertical curved lines.<br />
The Moiré line spacing, as shown in Fig. 1, is related to the<br />
superposition parameters by 6,7<br />
P =<br />
P 0 P 1<br />
P 0 2 + P 1 2 −2P 0 P 1 cos .<br />
The angle <strong>of</strong> the Moiré lines with the base <strong>of</strong> the lenticular<br />
sheet is given by 6,7<br />
tan =<br />
P 1 sin<br />
P 0 − P 1 cos .<br />
An actual Moiré pattern from such a sheet is shown in Figure<br />
2. The Moiré lines deviate from straight lines due to<br />
printer imperfections. The Ronchi rule pattern was printed<br />
with a 0.4233 mm spacing, and the lenticular lens sheet consisted<br />
<strong>of</strong> lenses with a spacing <strong>of</strong> 0.630 mm (40 lenses per<br />
inch) and the lenses had a magnification factor <strong>of</strong> 1.505<br />
(making the effective Ronchi line spacing equal 0.637 mm).<br />
Fluctuations in the printed line spacing, P 0 , result from line<br />
registration errors and create deviations in the Moiré line<br />
angle , according to Eq. (2).<br />
The sensitivity <strong>of</strong> the resulting Moiré line direction<br />
angle to the superposition is shown in Figure 3, which<br />
plots Eqs. (1) and (2) as functions <strong>of</strong> . For the 10° interval<br />
shown, the Moiré line spacing decreases from around<br />
80 mm to 2.5 mm. Both P and exhibit relatively little<br />
change <strong>for</strong> greater than 4°. In practical implementations,<br />
the superposition angle cannot be precisely controlled. So<br />
ensuring that these changes do not significantly affect the<br />
metric is a critical issue to the usefulness <strong>of</strong> this method.<br />
There<strong>for</strong>e, selecting an around 4° reduces the impact <strong>of</strong><br />
small changes in the superposition angle. This results in<br />
multiple low-frequency Moiré lines over the test pattern <strong>for</strong><br />
robust analysis.<br />
1<br />
2<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 311
Rawashdeh et al.: Moiré analysis <strong>for</strong> assessment <strong>of</strong> line registration quality<br />
where x is the additive displacement term <strong>for</strong> underlying<br />
line spacing P 0 , and x is the deviation from Moiré angle<br />
. The angular deviations vary over the printed pattern<br />
based on changes in the underlying test pattern.Without loss<br />
<strong>of</strong> generality, this direction is denoted as a function <strong>of</strong> a<br />
single variable x. To separate the deviation terms, a Taylor<br />
series can be applied to the cotangent function and expanded<br />
about . After higher-order terms are dropped (assuming<br />
small deviations), Eq. (3) results in<br />
x− x sin2 <br />
P 1<br />
<br />
sin + cot + cot<br />
− P 0csc<br />
csc 2 P 1 csc 2 ,<br />
where terms in the parenthesis are constant over x and relate<br />
to a constant <strong>of</strong>fset on the Moiré angle . They are subtracted<br />
out in the estimation procedure. The effective gain<br />
term that scales the line position deviations to Moiré pattern<br />
angle deviations is given by<br />
4<br />
Figure 3. Plots <strong>of</strong> Moiré fringe spacing P and direction as a function <strong>of</strong><br />
the superposition angle between a Ronchi ruling and a lenticular grating.<br />
Circles indicate manual measurements; solid lines are plots <strong>of</strong> Eqs.<br />
1 and 2.<br />
Also included in the plots <strong>of</strong> Fig. 3, are actual measurements<br />
<strong>of</strong> the Moiré line spacing and angle <strong>for</strong> five values <strong>of</strong><br />
. The measurements were made by manually setting the<br />
superposition angle and using a ruler to visually measure<br />
the resulting Moiré spacing, and a protractor to measure the<br />
Moiré direction ø. The resulting measurements agreed well<br />
with Eqs. (1) and (2) as can be seen by the measurement<br />
marker on the graphs <strong>of</strong> Fig. 3. For the measurement system<br />
proposed in this work, the lenticular grating is <strong>of</strong> high precision,<br />
while the actual superposition angle may also be variable<br />
depending on the mechanics used to load the test sheet.<br />
The following equations show the critical parameters relating<br />
the underlying line registration to the Moiré pattern parameters<br />
used in the measurement. From this derivation, a<br />
normalization step is presented to reduce the sensitivity <strong>of</strong><br />
metrics computed from the Moiré pattern to variations in<br />
parameters <strong>of</strong> the measurement system.<br />
A relationship between changes in the underlying line<br />
spacing and change in the angle <strong>of</strong> the Moiré pattern can be<br />
seen from taking the reciprocal <strong>of</strong> Eq. (2) and adding deviation<br />
terms to the test pattern line spacing and Moiré pattern<br />
angles to obtain<br />
cot + x = P 0 + x − P 1 cos<br />
P 1 sin<br />
= P 0 + x<br />
P 1 sin − cot,<br />
3<br />
g m =− sin2 <br />
P 1 sin ,<br />
where the gain/sensitivity is determined by the lenticular<br />
grid spacing P 1 and superposition angle . An alternate<br />
derivation <strong>of</strong> g m can be obtained directly through the ratio<br />
<strong>of</strong> the root-mean-square (rms) deviations <strong>of</strong> and P 0 . This<br />
would eliminate the <strong>of</strong>fset (zero-order) term <strong>of</strong> Eq. (4) and<br />
allow the gain factor g m to be computed directly from the<br />
derivatives <strong>of</strong> Eq. (3) with respect to and P 0 . The gain<br />
factor g m in this case is simply /P 0 .<br />
Since the deviations, x, will be extracted from the<br />
Moiré patterns and used <strong>for</strong> characterization, the sensitivity<br />
to becomes an issue <strong>for</strong> consistent measurements (small<br />
changes in , <strong>for</strong> near zero, can result in large changes in<br />
the gain). This variability can be significantly reduced by<br />
dividing the measured angle deviation by the measured distance<br />
between the moiré lines, if the effective Ronchi pattern<br />
line spacing is close to that <strong>of</strong> the lenticular grid. With P 1<br />
equal to P 0 , Eq. (1) can be simplified using the half angle<br />
<strong>for</strong>mula to show the Moiré line spacing is related to by<br />
P 1<br />
5<br />
P =<br />
2 sin/2 . 6<br />
For small (as is the case here), sin approximately equals<br />
(in radians). Thus, by applying this approximation to Eqs.<br />
(5) and (6), the normalized gain becomes<br />
ḡ m = g m<br />
P − sin2 <br />
P 1<br />
2<br />
. 7<br />
This equation shows that the repeatability <strong>of</strong> the measurement<br />
is enhanced through this normalization. The dominant<br />
scale factor controlling the gain on the angular displacement<br />
is now primarily dependent on the lenticular gird spacing,<br />
which can be precisely controlled and does not change with<br />
the superposition angle. The next section describes the extraction<br />
<strong>of</strong> x and P <strong>for</strong>m the scanned Moiré patterns, and<br />
the development <strong>of</strong> the metrics based on the normalization<br />
described by Eq. (7).<br />
312 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Rawashdeh et al.: Moiré analysis <strong>for</strong> assessment <strong>of</strong> line registration quality<br />
Figure 4. Line misregistration in 8.5 mm long vertical slices <strong>of</strong> the top twenty lines <strong>of</strong> the actual printed test<br />
patterns <strong>of</strong> printers LP1 top and LP2 bottom. Rulers show ideal line locations.<br />
LINE REGISTRATION QUALITY MEASUREMENT<br />
AND METRICS<br />
An example <strong>of</strong> a Moiré pattern used to extract parameters<br />
<strong>for</strong> quality metrics is shown in Fig. 2. The pattern was created<br />
with a five inch square printed test ruling and a lenticular<br />
grating at a superposition angle indicated by the line at<br />
the top <strong>of</strong> the figure. The Moiré patterns result in wavy lines<br />
along an angled path. Their deviation from a straight path<br />
indicates a faulty printed test pattern that is due to the x<br />
perturbations <strong>of</strong> Eq. (3). For example, the lines in the top<br />
half <strong>of</strong> the figure deviate to the left <strong>of</strong> the expected straight<br />
path and then back again. This corresponds to an increase,<br />
and then a decrease <strong>of</strong> angle x, which corresponds to<br />
changes in x according to Eq. (4). This section describes<br />
how these changes can be extracted, characterized, and used<br />
to <strong>for</strong>m line quality metrics. The underlying line imperfections<br />
<strong>for</strong> two laser printers are illustrated in Figure 4. This<br />
figure compares two portions <strong>of</strong> the printed Ronchi ruling<br />
test pattern. The slices are 8.5 mm long and contain the top<br />
20 printed lines. The figure contains regular tick marks to<br />
indicate line numbers and their expected locations. Observe<br />
at the junction <strong>of</strong> the line sets that the line spacing is not<br />
consistent between the two printers. Line 1 is aligned <strong>for</strong><br />
both prints; however, lines between 6 and 19 do not align,<br />
and lines from printer LP1 (top line set) deviate from the<br />
ideal locations from line 6 onward. The printed lines from<br />
LP2 (bottom line set) also deviate from the ideal locations,<br />
but the deviations are less pronounced and only start to<br />
become large from line 14 onward, indicating the line registration<br />
quality <strong>of</strong> printer LP2 is higher than that <strong>of</strong> printer<br />
LP1. The metrics described in this section will correctly assess<br />
this difference from values extracted over the whole<br />
printed line pattern.<br />
The printed test pattern used <strong>for</strong> the results presented in<br />
this paper is a 55 in. square Ronchi ruling. A lenticular<br />
lens sheet is superimposed at an angle <strong>of</strong> around 4° and the<br />
resulting pattern is scanned at 600 dpi on an HP ScanJet<br />
C7710A flatbed scanner. The scanned image results in a<br />
3000 by 3000 pixel image, which yields 10 pixels per blackwhite<br />
line pair (corresponds to a density <strong>of</strong> 60 line pairs per<br />
inch or a line spacing <strong>of</strong> 0.4233 mm). The targets are<br />
scanned in a monochrome setting and cropped to 2048 by<br />
2048 pixels to limit scan edge effects. The lenticular lens<br />
sheet is a Pacur LENSTAR large <strong>for</strong>mat polyester sheet with<br />
40 lenticules per inch. The sheet is 0.033 in. thick, which<br />
also corresponds to the lenticular focal length. The lenticules<br />
have a radius <strong>of</strong> 0.0146, and a width <strong>of</strong> 0.0251 inches. For<br />
the analysis, the scan is low-pass filtered to emphasize the<br />
lower frequency Moiré patterns <strong>of</strong> interest, using a rotationally<br />
symmetric two-dimensional Gaussian correlation kernel<br />
<strong>of</strong> size 8 and standard deviation parameter <strong>of</strong> 8. Luminance<br />
variability from the scanner, which <strong>of</strong>ten affects banding<br />
metrics, <strong>for</strong> example, is mitigated using this approach because<br />
only the shape <strong>of</strong> the Moiré lines are used in the<br />
metric and not their intensity. The angle between the test<br />
pattern and lenticular grating was determined (near 4°) by<br />
eye to produce Moiré patterns <strong>of</strong> good visibility and measurability<br />
after scanning <strong>for</strong> analysis purposes.<br />
The analysis program extracts the contiguous pixel locations<br />
<strong>of</strong> the local minima (or constant gray-level) <strong>for</strong>ming<br />
a pattern vertically oriented over the page. The groups <strong>of</strong><br />
pixel locations associated with the Moiré patterns are characterized<br />
by a best-fit (least squares) line to pixel minima to<br />
obtain an estimate <strong>of</strong> the Moiré line corresponding to a<br />
perfect line pattern. The groups <strong>of</strong> pixels near the line corresponding<br />
to the actual patterns are identified and<br />
smoothed using a higher-order polynomial (order 32). Since<br />
multiple lines exist over the page, a search <strong>for</strong> local minima<br />
is per<strong>for</strong>med with a best-fit line to identify each Moiré pattern.<br />
To describe this process, denote the scanned Moiré<br />
pattern image as Ix n ,y m ,wherex n and y m respectively represent<br />
the discrete row and column positions <strong>of</strong> the image<br />
matrix. As illustrated in Fig. 2, the origin <strong>of</strong> this coordinate<br />
system is located at the top left pixel. The algorithm searches<br />
<strong>for</strong> Moiré lines by assuming the <strong>for</strong>m<br />
Rx n ;m,b = mx n + b,<br />
where Rx is the y coordinate <strong>of</strong> Moiré line, and m and b<br />
are the slope and y intercept, respectively. The line parameters<br />
are found through an exhaustive search over a range <strong>of</strong><br />
b and m values in order to minimize the following cost<br />
function:<br />
8<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 313
Rawashdeh et al.: Moiré analysis <strong>for</strong> assessment <strong>of</strong> line registration quality<br />
N<br />
Cm,b = Ix n ,Rx n ;m,b,<br />
n=1<br />
where N is the total number <strong>of</strong> pixel rows in the scanned<br />
image. The parameters b and m associated with the best-fit<br />
line can be determined by<br />
m 0 ,b 0 = arg minCm,b.<br />
m,b<br />
9<br />
10<br />
Once the best fit is found over the image, the slope m 0 is<br />
fixed and b is incremented over the column pixels <strong>of</strong> Ix,y<br />
and local minima detected to find the other patterns. Since<br />
the images are relatively simple, the minima appear with<br />
good distinction, and a threshold can be set to ignore insignificant<br />
minimum peaks and collect the set <strong>of</strong> b values corresponding<br />
to local minima denoted by<br />
B = b 1 ,b 2 ,b 3 ,b 4 ,b 5 , ...,b Y ,<br />
11<br />
such that B is a vector, <strong>of</strong> length Y, containing the intercept<br />
values <strong>of</strong> the lines that are the best linear fits to the actual<br />
Moiré lines. The average distance between the minima is<br />
taken as an estimate <strong>of</strong> the Moiré line spacing given by<br />
Pˆ = 1 Y−1<br />
b i+1 − b i .<br />
Y −1i=1<br />
12<br />
The actual curved Moiré pattern can be found by locating<br />
the local minimum <strong>for</strong> each x coordinate in the neighborhood<br />
<strong>of</strong> each fitted line. For some lenticular grids; however,<br />
the locally dark image points appear at the lens intersections,<br />
creating a regular discontinuity over the pattern. To improve<br />
the detection <strong>of</strong> the Moiré pattern pixel, a midluminance<br />
gray level was used. There<strong>for</strong>e, the y coordinates <strong>of</strong> the actual<br />
Moiré patterns were determined by the pixels closest to<br />
the Moiré pattern gray level I m in the neighborhood <strong>of</strong> the<br />
fitted line. The collection <strong>of</strong> points <strong>for</strong> the ith Moiré pattern<br />
is denoted as<br />
Pˆ<br />
2<br />
S i x =arg minIx,y − I m Rx;m 0 ,b i −<br />
y<br />
y Rx;m 0 ,b i + Pˆ<br />
2,<br />
13<br />
where I m is the mean luminance <strong>of</strong> the Moiré patterns. Figure<br />
5 illustrates the results <strong>of</strong> this extraction process <strong>for</strong> two<br />
sample laser printer outputs. A 32-order polynomial was fitted<br />
to the locus <strong>of</strong> points from Eq. (13) in order to smooth<br />
and overlay the Moiré patterns, along with the best fit lines<br />
<strong>for</strong> visual inspection, on the actual scanned image. With the<br />
approach described above the need <strong>for</strong> smoothing is important<br />
because <strong>of</strong> the periodic dark bands <strong>of</strong> the lens intersections<br />
cause regular glitches in the points. While other methods<br />
can be used <strong>for</strong> smoothing, such as the median filter, this<br />
work uses the 32-order polynomial fitted to the points identified<br />
by Eq. (13). The results observed in Fig. 5 demonstrate<br />
that the extraction procedure is indeed capturing the basic<br />
Figure 5. Moiré analysis comparison between two laser printers. Straight<br />
lines indicate the ideal Moiré patterns, and curved lines are best-fit polynomials<br />
to actual Moiré patterns due to printer errors. a Laser printer<br />
LP1. b Laser printer LP2.<br />
elements <strong>of</strong> the Moiré patterns.<br />
The deviation from all lines is characterized by a mean<br />
deviateateachrow,givenby<br />
Y<br />
Lx n = 1 S¯ix n − Rx n ;m 0 ,b i ,<br />
Y i=1<br />
14<br />
where S¯i is the resulting polynomial fit to the points <strong>of</strong> S i in<br />
Eq. (13). The derivative <strong>of</strong> Lx is equal to the tangent <strong>of</strong><br />
angle x; and <strong>for</strong> small values <strong>of</strong> x, can be estimated<br />
with a numerical gradient as follows:<br />
314 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Rawashdeh et al.: Moiré analysis <strong>for</strong> assessment <strong>of</strong> line registration quality<br />
Table II. Statistical measures <strong>of</strong> Moiré pattern deviation quality metrics <strong>for</strong> two laser<br />
Table I. Line registration quality metrics <strong>for</strong> LP1 as a function <strong>of</strong> varying in<br />
5 2.57 2.28 10 −9 13.7 10 −5<br />
degrees.<br />
and one inkjet printers.<br />
Pˆ NAV NMD<br />
Printer Pˆ NAV NMD<br />
2 6.14 2.11 10 −9 13.1 10 −5<br />
LP1 3 4.26 2.08 10 −9 12.7 10 −5<br />
3 4.26 2.08 10 −9 12.7 10 −5<br />
LP2 3.8 3.73 1.79 10 −9 11.7 10 −5<br />
4 3.12 2.29 10 −9 12.9 10 −5<br />
IP 3.7 3.66 2.37 10 −9 11.4 10 −5<br />
x n tanx n 1<br />
2 Lx n+1 − Lx n−1 .<br />
15<br />
The estimate <strong>of</strong> the Moiré angle can be used to compute the<br />
macro quality metric referred to as the normalized average<br />
variance (NAV), given by<br />
1<br />
2 L = x n 2 .<br />
Pˆ 2 N −1 n=1<br />
N<br />
16<br />
Note this metric reflects the average line registration error<br />
over the whole test pattern. A micro quality line metric can<br />
be taken over local portions <strong>of</strong> the test pattern and involve<br />
the row corresponding to the worst deviation. This metric is<br />
called the normalized maximum deviation (NMD), and is<br />
given by<br />
¯ L = 1 Pˆ max x n .<br />
n<br />
17<br />
RESULTS AND ANALYSIS<br />
To demonstrate the robustness <strong>of</strong> the metrics to superposition<br />
angle variation, a pattern from laser printer LP1 was<br />
scanned <strong>for</strong> four different values, and the resulting NAV<br />
and NMD quality metrics are presented in Table I. It can be<br />
seen that the spacing decreases with increasing angle, while<br />
the NAV and NMD measures stay relatively constant. Quantitatively,<br />
the coefficients <strong>of</strong> variances (CV) <strong>for</strong> the metrics<br />
over the variations are 4.9% and 3.3% <strong>for</strong> the NAV and<br />
NMD, respectively. The CV is the ratio <strong>of</strong> the standard deviation<br />
to the mean <strong>of</strong> a data set, and it provides a quantity<br />
related to the measurement resolution, which is affected by<br />
factors such as printer and scanner settings, as well as, properties<br />
<strong>of</strong> the lenticular lenses used, such as lens spacing and<br />
precision.<br />
As an example <strong>of</strong> how the quality metrics respond to<br />
different printers, the NAV and NMD metrics were computed<br />
using the outputs <strong>of</strong> two laser printers (LP1 and LP2,<br />
used in Fig. 5), and an inkjet printer denoted as IP. Table II<br />
shows a numerical comparison between these printer outputs,<br />
as well as the measurement parameters. The CV values<br />
computed from Table I can be used to examine the relative<br />
comparison <strong>of</strong> line registration quality between printers. For<br />
example, the difference between the NAV values as a percentage<br />
<strong>of</strong> their mean is 15% <strong>for</strong> printers LP1 and LP2 in<br />
Table II. This value is greater than the 4.9% variation expected<br />
from the measurement variability and thus indicates<br />
that the large scale (macro) line registration quality <strong>of</strong> LP2 is<br />
better than that <strong>of</strong> LP1 (consistent with observations in Figs.<br />
4 and 5). In addition, the NMD measurements differ by<br />
8.3%, which is greater than the 3.3% CV <strong>for</strong> the NMD measure.<br />
Comparing the inkjet printer IP with LP1, it is evident<br />
from the NAV values that IP has poorer quality (consistent<br />
with examinations <strong>of</strong> scaled-up observation <strong>of</strong> the line quality);<br />
however, the NMD values differ by 10.8% <strong>of</strong> their<br />
mean, which is grater than the 3.3% CV value. This indicates<br />
that even though LP1 has better line registration on a<br />
macro scale (on average across the page), it has greater isolated<br />
deviations than IP.<br />
These results suggest that the above measures can serve<br />
as a quality metric <strong>for</strong> printed line registration. The NAV<br />
measure reflects the average printed line spacing P 0 constancy<br />
over the length <strong>of</strong> the test page. A quasiperiodic pattern<br />
in Lx n reflects banding like intensity variations across<br />
the test page as observed in Fig. 4 <strong>for</strong> LP1. These shape<br />
variations reflected periodic fluctuations in the printed line<br />
spacing, which are likely due to the same problems causing<br />
banding, such as imperfect rollers in the print process direction<br />
or gear noise. Moreover, Lx n can isolate process<br />
motion-related banding causes from other ones that affect<br />
reflectance, such as toner or ink deposition inconsistencies.<br />
CONCLUSION<br />
This work outlines the use <strong>of</strong> Moiré analysis <strong>for</strong> the quantification<br />
<strong>of</strong> line registration. The line registration metrics developed<br />
here are based on modeling the interference between<br />
a lenticular lens sheet and a hardcopy test target containing a<br />
Ronchi ruling or linear grating, and they provide examples<br />
<strong>of</strong> how the resulting Moiré patterns can be used to measure<br />
line registration quality. There are clearly other metrics that<br />
can be derived from the extracted Moiré patterns that can<br />
emphasize other issues depending on the application. For<br />
instance, if the Moiré line deviations are quasiperiodic, it is<br />
likely that these deviations indicate the root cause <strong>of</strong> banding.<br />
There<strong>for</strong>e, metrics based on the periodicity <strong>of</strong> these deviations<br />
over macro regions can be used <strong>for</strong> banding characterization.<br />
The work derived general equations to help in<br />
designing <strong>of</strong> metrics that have good repeatability.<br />
The experimental setup presented in this work suggests<br />
methods <strong>for</strong> volume processing <strong>of</strong> hardcopy samples. This<br />
would require a scanner with an automatic document feeder.<br />
A lenticular lens sheet could then be embedded into the<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 315
Rawashdeh et al.: Moiré analysis <strong>for</strong> assessment <strong>of</strong> line registration quality<br />
scanner glass or fixed on top <strong>of</strong> the glass where the test sheet<br />
could slide over it. The resulting patterns would then be<br />
scanned, and s<strong>of</strong>tware applied to compute the per<strong>for</strong>mance<br />
analyses as described in this paper. The monochrome and<br />
low-resolution scans are relatively easy to produce and analyze<br />
on a computer. A potential problem resulting from an<br />
automatic document feeder is maintaining a constant superposition<br />
angle between test page and lens sheet; however, it<br />
has been shown that the proposed metrics are robust to<br />
small changes in the angle. A more significant problem<br />
would arise from variations in the distance between the test<br />
pattern and lenticular sheet, such as might result from<br />
trapped air or irregular pressure on the test pattern. In this<br />
case the Moiré line will be artificially skewed causing variations<br />
from the distance rather than line mis-registration. It<br />
would be important in such a system to ensure the automatic<br />
feed (or any other system) minimizes this variation <strong>for</strong><br />
accurate metrics.<br />
REFERENCES<br />
1 ISO/IEC 13660:2001 In<strong>for</strong>mation technology - Office equipment:<br />
Measurement <strong>of</strong> image quality attributes <strong>for</strong> hardcopy output, Binary<br />
monochrome text and graphic images (ISO, Geneva), www.iso.org.<br />
2 E. N. Dalal, A. Haley, M. Robb, D. Mashtare, J. Briggs, P. L. Jeran, T. F.<br />
Bouk, and J. Deubert, “INCITS W1.1 Standards <strong>for</strong> Perceptual<br />
Evaluation <strong>of</strong> Text and Line Quality”, Proc. IS&T PICS Conference<br />
(IS&T, Springfield, VA, 2003) pp. 102–103.<br />
3 Y. Kipman, “Image quality metrics <strong>for</strong> printers and media”, Proc. IS&T<br />
PICS Conference (IS&T, Springfield, VA, 1998) pp. 183–187.<br />
4 G. J. Indebetouw and R. Czarnek, Selected Papers on Optical Moiré and<br />
Applications (SPIE, Bellingham, WA, 1992).<br />
5 B. Han, D. Post, and P. Ifju, “Moiré interferometry <strong>for</strong> engineering<br />
mechanics: current practices and future developments”, J. Strain Anal.<br />
Eng. Des. 36, 101–117 (2001).<br />
6 F. Zandman, G. S. Holister, and V. Brcic, “The influence <strong>of</strong> grid<br />
geometry on moire fringe properties”, J. Strain Anal. 1, 1-10 (1965).<br />
7 A. Livnat and O. Kafri, “Moire pattern <strong>of</strong> a linear grid with a lenticular<br />
grating”, Opt. Lett. 7, 253 (1982).<br />
316 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
<strong>Journal</strong> <strong>of</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology® 51(4): 317–327, 2007.<br />
© <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology 2007<br />
Analysis <strong>of</strong> the Influence <strong>of</strong> Vertical Disparities Arising<br />
in Toed-in Stereoscopic Cameras<br />
Robert S. Allison<br />
Department <strong>of</strong> Computer <strong>Science</strong> and Centre <strong>for</strong> Vision Research, York University,<br />
4700 Keele St., Toronto, Ontario M3J 1P3, Canada<br />
E-mail: allison@cs.yorku.ca<br />
Abstract. A basic task in the construction and use <strong>of</strong> a stereoscopic<br />
camera and display system is the alignment <strong>of</strong> the left and<br />
right images appropriately—a task generally referred to as camera<br />
convergence. Convergence <strong>of</strong> the real or virtual stereoscopic cameras<br />
can shift the range <strong>of</strong> portrayed depth to improve visual com<strong>for</strong>t,<br />
can adjust the disparity <strong>of</strong> targets to bring them nearer to the<br />
screen and reduce accommodation-vergence conflict, or can bring<br />
objects <strong>of</strong> interest into the binocular field <strong>of</strong> view. Although camera<br />
convergence is acknowledged as a useful function, there has been<br />
considerable debate over the trans<strong>for</strong>mation required. It is well<br />
known that rotational camera convergence or “toe-in” distorts the<br />
images in the two cameras producing patterns <strong>of</strong> horizontal and<br />
vertical disparities that can cause problems with fusion <strong>of</strong> the stereoscopic<br />
imagery. Behaviorally, similar retinal vertical disparity patterns<br />
are known to correlate with viewing distance and strongly affect<br />
perception <strong>of</strong> stereoscopic shape and depth. There has been<br />
little analysis <strong>of</strong> the implications <strong>of</strong> recent findings on vertical disparity<br />
processing <strong>for</strong> the design <strong>of</strong> stereoscopic camera and display<br />
systems. I ask how such distortions caused by camera convergence<br />
affect the ability to fuse and perceive stereoscopic images. © 2007<br />
<strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology.<br />
DOI: 10.2352/J.<strong>Imaging</strong>Sci.Technol.200751:4317<br />
INTRODUCTION<br />
In many stereoscopic viewing situations it is necessary to<br />
adjust the screen disparity <strong>of</strong> the displayed images <strong>for</strong> viewer<br />
com<strong>for</strong>t, to optimize depth perception or to otherwise enhance<br />
the stereoscopic experience. Convergence <strong>of</strong> the real<br />
or virtual cameras is an effective means <strong>of</strong> adjusting portrayed<br />
disparities. A long-standing question in the stereoscopic<br />
imaging and display literature is what is the best<br />
method to converge the cameras? Humans use rotational<br />
movements to binocularly align the visual axes <strong>of</strong> their eyes<br />
on targets <strong>of</strong> interest. Similarly, one <strong>of</strong> the easiest ways to<br />
converge the cameras is to pan them in opposite directions<br />
to “toe-in” the cameras. However, convergence through<br />
camera toe-in has side effects that can lead to undesirable<br />
distortions <strong>of</strong> stereoscopic depth. 1,2 In this paper we reanalyze<br />
these geometric distortions <strong>of</strong> stereoscopic space in the<br />
context <strong>of</strong> recent findings on the role <strong>of</strong> vertical disparities in<br />
stereoscopic space perception. We focus on a number <strong>of</strong> issues<br />
related to converged cameras and the mode <strong>of</strong> convergence:<br />
The effect <strong>of</strong> rectification; relation between the geometry<br />
<strong>of</strong> the imaging device and the display device; fused and<br />
Received Dec. 5, 2006; accepted <strong>for</strong> publication Mar. 7, 2007.<br />
1062-3701/2007/514/317/11/$20.00.<br />
augmented displays; orthostereoscopy; the relation between<br />
parallax distortions in the display and the resulting retinal<br />
disparity; and the effect <strong>of</strong> these toe-in induced retinal disparities<br />
on depth perception and binocular fusion.<br />
Our interests lie in augmented-reality applications and<br />
stereoscopic heads <strong>for</strong> tele-operation applications. In these<br />
systems a focus is on the match and registration between the<br />
stereoscopic imagery and the “real world” so we will concentrate<br />
on orthostereoscopic or near orthostereoscopic configurations.<br />
These configurations have well known limitations<br />
<strong>for</strong> applications such as visualization and cinema, and<br />
other configurations may result in displays that are more<br />
pleasing and easier to fuse. However, it is important to note<br />
that our basic analysis generalizes to other configurations,<br />
and we will discuss other viewing arrangements when<br />
appropriate. 3,4 In a projector-based display system with separate<br />
right and left projectors, or in binocular head mounted<br />
display (HMD) with independent left and right displays, the<br />
displays/projectors can also be converged mechanically or<br />
optically. In this paper we will also assume a single flat,<br />
fronto-parallel display (i.e., a monitor or projector display)<br />
so that the convergence <strong>of</strong> the projectors is not an issue.<br />
Since the left and right images are projected or displayed<br />
into the same plane we will refer to these configurations as a<br />
“parallel display.” In most cases similar considerations will<br />
apply <strong>for</strong> a HMD with parallel left and right displays.<br />
OPTIONS FOR CAMERA CONVERGENCE<br />
We use the term convergence here to refer to a variety <strong>of</strong><br />
means <strong>of</strong> realigning one stereoscopic half-image with respect<br />
to the other, including toe-in (or rotational) convergence<br />
and translational image shift.<br />
Convergence can shift the range <strong>of</strong> portrayed depth to<br />
improve visual com<strong>for</strong>t and composition. Looking at objects<br />
presented stereoscopically further or nearer than the screen<br />
causes a disruption <strong>of</strong> the normal synergy between vergence<br />
and accommodation in most displays. Normally accommodation<br />
and vergence covary but, in a stereoscopic display, the<br />
eyes should remain focused at the screen regardless <strong>of</strong> disparity.<br />
The accommodation-vergence conflict can cause visual<br />
stress and disrupt binocular vision. 5 Convergence <strong>of</strong> the<br />
cameras can be used to adjust the disparity <strong>of</strong> targets <strong>of</strong><br />
interest to bring them nearer to the screen and reduce this<br />
conflict.<br />
317
Allison: Analysis <strong>of</strong> the influence <strong>of</strong> vertical disparities arising in toed-in stereoscopic cameras<br />
Table I. Typical convergence <strong>for</strong> stereoscopic sensors and displays. “Natural” modes <strong>of</strong><br />
convergence are shown in bold.<br />
DISPLAY/SENSOR<br />
GEOMETRY<br />
Translation<br />
REAL OR VIRTUAL CAMERA CONVERGENCE<br />
Rotation<br />
Flat Horizontal Image Translation Toed-in camera, toed-in<br />
projector combination<br />
Spherical<br />
Differential translation <strong>of</strong><br />
computer graphics images<br />
Image sensor shift<br />
Variable baseline camera<br />
Human viewing <strong>of</strong> planar<br />
stereoscopic displays?<br />
Toed-in stereoscopic camera<br />
or robot head<br />
Haploscope<br />
Human physiological<br />
vergence<br />
Convergence can also be used to shift the range <strong>of</strong> portrayed<br />
depth. For example, it is <strong>of</strong>ten preferable to portray<br />
stereoscopic imagery in the space behind rather than in front<br />
<strong>of</strong> the display. With convergence a user can shift stereoscopic<br />
imagery to appear “inside” the display and reduce interposition<br />
errors between the stereoscopic imagery and the edges<br />
<strong>of</strong> the displays.<br />
Cameras used in stereoscopic imagers have limited field<br />
<strong>of</strong> view and convergence can be used to bring objects <strong>of</strong><br />
interest into the binocular field <strong>of</strong> view.<br />
Finally, convergence or more appropriately translation<br />
<strong>of</strong> the stereoscopic cameras can also be used to adjust <strong>for</strong><br />
differences in a user’s interpupillary distance. The latter<br />
trans<strong>for</strong>mation is not typically called convergence since the<br />
stereoscopic baseline is not maintained.<br />
In choosing a method <strong>of</strong> convergence there are several<br />
issues one needs to consider. What type <strong>of</strong> 2D image trans<strong>for</strong>mation<br />
is most natural <strong>for</strong> the imaging geometry? Can a<br />
3D movement <strong>of</strong> the imaging device accomplish this trans<strong>for</strong>mation?<br />
In a system consisting <strong>of</strong> separate acquisition and<br />
display systems is convergence best achieved by changing the<br />
imaging configuration and/or by trans<strong>for</strong>ming the images<br />
(or projector configuration) prior to display? If an unnatural<br />
convergence technique must be used, what is the impact on<br />
stereoscopic depth perception?<br />
Although camera convergence is acknowledged as a useful<br />
function, there has been considerable debate over the<br />
correct trans<strong>for</strong>mation required. Since the eyes (and the<br />
cameras in imaging applications) are separated laterally, convergence<br />
needs to be an opposite horizontal shift <strong>of</strong> left and<br />
right eyes images on the sensor surface or, equivalently, on<br />
the display. The most appropriate type <strong>of</strong> trans<strong>for</strong>mation to<br />
accomplish this 2D shift—rotation or translation—depends<br />
on the geometry <strong>of</strong> the imaging and display devices. We<br />
agree with the view that the trans<strong>for</strong>mation should reflect<br />
the geometry <strong>of</strong> the display and imaging devices in order to<br />
minimize distortion (see Table I). One could argue that a<br />
“pure” vergence movement should affect the disparity <strong>of</strong> all<br />
objects equally, resulting in a change in mean disparity over<br />
the entire image without any change in relative disparity<br />
between points.<br />
For example, consider a spherical imaging device such<br />
as the human eye where expressing disparity in terms <strong>of</strong><br />
visual angle is a natural coding scheme. A rotational movement<br />
about the optical centre <strong>of</strong> the eye would scan an<br />
image over the retina without distorting the angular relationships<br />
within the image. Thus the natural convergence<br />
movement with such an imaging device is a differential rotation<br />
<strong>of</strong> the two eyes, as occurs in physiological convergence<br />
(although freedom to choose various spherical coordinate<br />
systems complicates the definition <strong>of</strong> disparity 6 ).<br />
A flat sensor is the limiting <strong>for</strong>m <strong>of</strong> spherical sensor<br />
with an infinite radius <strong>of</strong> curvature, and thus the rotation <strong>of</strong><br />
the sensor becomes a translation parallel to the sensor plane.<br />
For displays that rely on projection onto a single flat, frontoparallel<br />
display surface (many stereoscopic displays with the<br />
notable exception <strong>of</strong> some head-mounted displays and haploscopic<br />
systems) depth differences should be represented as<br />
linear horizontal disparities in the image plane. The natural<br />
convergence movement is a differential horizontal shift <strong>of</strong><br />
the images in the plane <strong>of</strong> the display. Acquisition systems<br />
with parallel cameras are well-matched to such display geometry<br />
since a translation on the display corresponds to a<br />
translation in the sensor plane. This model <strong>of</strong> parallel cameras<br />
is typically used <strong>for</strong> the virtual cameras in stereoscopic<br />
computer graphics 7 and the real cameras in many stereoscopic<br />
camera setups.<br />
Thus horizontal image translation <strong>of</strong> the images on the<br />
display is the preferred minimal distortion method to shift<br />
convergence in a stereoscopic rig with parallel cameras when<br />
presented on a parallel display. This analysis corresponds to<br />
current conventional wisdom. If the stereo baseline is to be<br />
maintained then this vergence movement is a horizontal<br />
translation <strong>of</strong> the images obtained from the parallel cameras<br />
rather than a translation <strong>of</strong> the cameras themselves. For example,<br />
in computer-generated displays, the left and right half<br />
images can be shifted in opposite directions on the display<br />
surface to shift portrayed depth with respect to the screen.<br />
With real camera images, a problem with shifting the displayed<br />
images to accomplish convergence is that in doing so,<br />
part <strong>of</strong> each half-image is shifted <strong>of</strong>f <strong>of</strong> the display resulting<br />
in a smaller stereoscopic image.<br />
An alternative is to shift the imaging device (e.g., CCD<br />
array) behind the camera lens, with opposite sign <strong>of</strong> shift in<br />
the two cameras <strong>for</strong>ming the stereo rig. This avoids some <strong>of</strong><br />
the problems associated with rotational convergence discussed<br />
below. Implementing a large, variable range <strong>of</strong> convergence<br />
with mechanical movements or selection <strong>of</strong> subarrays<br />
from a large CCD can be complicated. Furthermore,<br />
many lenses have significant radial distortion and translating<br />
the center <strong>of</strong> the imaging device away from the optical axis<br />
increases the amount <strong>of</strong> radial distortion. Worse, <strong>for</strong><br />
matched lenses the distortions introduced in each sensor<br />
image will be opposite if the sensors are shifted in opposite<br />
directions. This leads to increased disparity distortion.<br />
Toed-in cameras can center the image on the optical axis<br />
and reduce this particular problem.<br />
If we converge nearer than infinity using horizontal im-<br />
318 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Allison: Analysis <strong>of</strong> the influence <strong>of</strong> vertical disparities arising in toed-in stereoscopic cameras<br />
Figure 1. A plan view <strong>of</strong> an array <strong>of</strong> points located in the X-Z plane at<br />
eye level. The solid dots show the true position <strong>of</strong> the points and also their<br />
reconstruction based on images from a parallel camera orthostereoscopic<br />
rig presented at a 0.7 m viewing distance. The open diamond shaped<br />
markers show the reconstructed position <strong>of</strong> the points in the array when<br />
the cameras are converged using horizontal image translation HIT. As<br />
predicted the points that are truly at 1.1 m move in to appear near the<br />
screen distance <strong>of</strong> 0.7 m. Also depth and size should appear scaled<br />
appropriately <strong>for</strong> the nearer distance. But notice that depth ordering and<br />
planarity are maintained. Circles at a distance <strong>of</strong> zero denote the positions<br />
<strong>of</strong> the eyes.<br />
age shift, then far objects should be brought toward the<br />
plane <strong>of</strong> the screen. With convergence via horizontal image<br />
shift, a frontal plane at the camera convergence distance<br />
should appear flat and at the screen distance. However,<br />
depth <strong>for</strong> a given retinal disparity increases approximately<br />
with the square <strong>of</strong> distance. Thus if the cameras are converged<br />
at a distance other than the screen distance to bring a<br />
farther (or nearer) target toward the screen, then the depth<br />
in the scene should be distorted nonlinearly but depth ordering<br />
and planarity are maintained (Figure 1). This apparent<br />
depth distortion is predicted <strong>for</strong> both the parallel and<br />
toed-in configurations. In the toed-in case it would be added<br />
to the curvature effects discussed below. Similar arguments<br />
can be made <strong>for</strong> size distortions in the image (or equivalently<br />
the apparent spacing <strong>of</strong> the dots in Fig. 1). See Woods 1<br />
and Diner and Fender 2 <strong>for</strong> an extended discussion <strong>of</strong> these<br />
distortions.<br />
It is important to note that these effects are predicted<br />
from the geometry and do not always correspond to human<br />
perception. Percepts <strong>of</strong> stereoscopic space tend to deviate<br />
from the geometric predictions based on the Keplerian projections<br />
and Euclidean geometry 6 ). Vergence on its own is<br />
not a strong cue to distance and other depth cues in the<br />
display besides horizontal disparity can affect the interpretation<br />
<strong>of</strong> stereoscopic displays. For example, it has been<br />
known <strong>for</strong> over 100 years that observers can use vertical<br />
disparities in the stereoscopic images to obtain more veritical<br />
estimates <strong>of</strong> stereoscopic <strong>for</strong>m. 8 In recent years, a role <strong>for</strong><br />
vertical disparities in human stereoscopic depth perception<br />
has been confirmed. 9,10<br />
Translation <strong>of</strong> the images on the display or <strong>of</strong> the sensors<br />
behind the lenses maintains the stereoscopic camera<br />
baseline and hence the relative disparities in the acquired or<br />
simulated image. Shifting <strong>of</strong> the images can be used to shift<br />
this disparity range to be centered on the display to ease<br />
viewing com<strong>for</strong>t. However, in many applications this disparity<br />
range is excessive and other techniques may be more<br />
suitable. Laterally shifting the cameras toward or away from<br />
each other increases or decreases the range <strong>of</strong> disparities<br />
corresponding to a given scene. Control <strong>of</strong> the stereo rig<br />
baseline serves a complementary function to convergence by<br />
adjusting the “gain” <strong>of</strong> stereopsis instead <strong>of</strong> simply the mean<br />
disparity. This function is <strong>of</strong>ten very useful <strong>for</strong> mapping a<br />
depth range to a useful or com<strong>for</strong>table disparity range in<br />
applications such as computer graphics, 4,11 photogrammetry,<br />
etc.<br />
In augmented reality or other enhanced vision systems<br />
that fuse stereoscopic imagery with direct views <strong>of</strong> the world<br />
(or with displays from other stereoscopic image sources),<br />
orthostereoscopic configurations (or at least consistent<br />
views) are important. In these systems, proper convergence<br />
<strong>of</strong> the camera systems and calibration <strong>of</strong> image geometry is<br />
required so that objects in the display have appropriate disparity<br />
relative to their real world counterparts. A parallel<br />
camera orthostereoscopic configuration presents true disparities<br />
to the user if presented on a parallel display. Thus,<br />
geometrically at least, we should expect to see true depth. In<br />
practice this seldom occurs because <strong>of</strong> the influence <strong>of</strong> other<br />
depth cues (accommodation-vergence conflict, changes in<br />
effective interpupillary distance with eye movements, flatness<br />
cues corresponding to viewing a flat display, etc.).<br />
In summary, an orthostereoscopic parallel-camera/<br />
parallel-display configuration can present accurate disparities<br />
to the user. 1,7 On parallel displays, convergence by horizontal<br />
shift <strong>of</strong> the images obtained from parallel cameras<br />
introduces no distortion <strong>of</strong> horizontal or vertical screen disparity<br />
(parallax). Essentially, convergence by this method<br />
brings the two half images into register with out changing<br />
relative disparity. This can reduce vergence-accommodation<br />
conflict and improve the ability to fuse the imagery. Geometrically,<br />
one would predict effects on perceived depth—<br />
the apparent depth <strong>of</strong> imagery with respect to the screen and<br />
the depth scaling in the image are affected by the simulated<br />
vergence. 1,13 However, this amounts to a relief trans<strong>for</strong>mation<br />
implying that depth ordering and coplanarity should be<br />
maintained. 2,10<br />
CAMERA TOE-IN<br />
While horizontal image translation is attractive theoretically,<br />
there are <strong>of</strong>ten practical considerations that limit use <strong>of</strong> the<br />
method and make rotational convergence attractive. For example,<br />
with a limited camera field <strong>of</strong> view and a nonzero<br />
stereo baseline there exists a region <strong>of</strong> space near to the<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 319
Allison: Analysis <strong>of</strong> the influence <strong>of</strong> vertical disparities arising in toed-in stereoscopic cameras<br />
Figure 2. a The Toronto IRIS Stereoscopic Head 2 TRISH II, an example <strong>of</strong> a robot head built <strong>for</strong> a wide<br />
range <strong>of</strong> working distances. With such a system, a wide range <strong>of</strong> camera convergence is required to bring<br />
objects <strong>of</strong> interest into view <strong>of</strong> the cameras. With <strong>of</strong>f-the shelf cameras this can be most conveniently achieved<br />
with camera toe-in. b A hypothetical stereo rig with camera field <strong>of</strong> view . Objects in near working space<br />
are out <strong>of</strong> the binocular field <strong>of</strong> view which is indicated by the cross hatch pattern.<br />
cameras that cannot be seen by one or both cameras. In<br />
some applications such as landscape photography this region<br />
<strong>of</strong> space may be irrelevant; in other applications such as<br />
augmented reality or stereoscopic robot heads this may correspond<br />
to a crucial part <strong>of</strong> the normal working range (see<br />
Figure 2). Rotational convergence <strong>of</strong> the cameras can increase<br />
the near working space <strong>of</strong> the system and center the<br />
target in the camera images. 14 Other motivations <strong>for</strong> rotational<br />
convergence include the desire to center the target on<br />
the camera optics (e.g., to minimize camera distortion) and<br />
the relative simplicity and large range <strong>of</strong> motion possible<br />
with rotational mechanisms. Given that rotational convergence<br />
<strong>of</strong> stereo cameras is <strong>of</strong>ten implemented in practice, we<br />
ask what effects the distortions produced by these movements<br />
have on the perception <strong>of</strong> stereoscopic displays?<br />
It is well known that the toed-in configuration distorts<br />
the images in the two cameras producing patterns <strong>of</strong> horizontal<br />
and vertical screen disparities (parallax). Geometrically,<br />
deviations from the parallel-camera configuration may<br />
result in spatial distortion unless compensating trans<strong>for</strong>mations<br />
are introduced mechanically, optically or electronically<br />
in the displayed images, 2,12 <strong>for</strong> example unless a pair <strong>of</strong> projectors<br />
(or HMD with separate left and right displays) with<br />
matched convergence or a parallel display with special distortion<br />
correction techniques are used. 15,16 For the rest <strong>of</strong><br />
this paper we will assume a single projector or display system<br />
(parallel display) and a dual sensor system with parallel<br />
or toed-in cameras.<br />
The effects <strong>of</strong> the horizontal disparities have been well<br />
described in the literature and we review them be<strong>for</strong>e turning<br />
to the vertical disparities in the next section. The depth<br />
distortions due to the horizontal disparities introduced can<br />
be estimated geometrically. 1 The geometry <strong>of</strong> the situation is<br />
illustrated in Figure 3. The imaging space world coordinate<br />
system is centered between the cameras, a is the intercamera<br />
distance and the angle <strong>of</strong> convergence is (using the conventional<br />
stereoscopic camera measure <strong>of</strong> convergence rather<br />
than the physiological one).<br />
Let us assume the cameras converge symmetrically at<br />
point C located at distance F. A local coordinate system is<br />
attached to each camera and rotated ± about the y axis<br />
with respect to the imaging space world coordinate system.<br />
The coordinates <strong>of</strong> a point P=XYZ T in the left and right<br />
cameras is<br />
X<br />
l= X + a − Z sin<br />
l<br />
2cos<br />
Y l<br />
Y<br />
Z<br />
Z cos +X +<br />
2sin,<br />
a<br />
X<br />
r= X − a + Z sin<br />
r<br />
2cos<br />
Y r<br />
Y<br />
Z<br />
Z cos −X −<br />
2sin.<br />
a<br />
After perspective projection onto the converged CCD array<br />
(coordinate frame u-v centered on the optic axis and letting<br />
f=1.0) we get the following image coordinates <strong>for</strong> the left,<br />
u l ,v l T , and right, u r ,v r T ,arrays:<br />
= X + a − Z sin<br />
2cos<br />
u l<br />
v l= X l/Z l Z cos +X +<br />
Y l /Z l 2sin<br />
a<br />
Y<br />
Z cos +X +<br />
2sin,<br />
a<br />
1<br />
2<br />
320 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Allison: Analysis <strong>of</strong> the influence <strong>of</strong> vertical disparities arising in toed-in stereoscopic cameras<br />
Figure 3. <strong>Imaging</strong> and display geometry <strong>for</strong> symmetric toe-in convergence at point C and viewing at distance<br />
D plan view.<br />
Figure 4. Keystone distortion due to toe-in. a Left + and right images <strong>for</strong> a regularly spaced grid <strong>of</strong><br />
points with the stereo camera converged toed-in on the grid. b Corresponding disparity vectors comparing<br />
left eye with right eye views demonstrate both horizontal and vertical components <strong>of</strong> the keystone distortion.<br />
= X − a + Z sin<br />
2cos<br />
u r<br />
v r= X r/Z r Z cos −X −<br />
Y r /Z r 2sin<br />
a<br />
Y<br />
Z cos −X −<br />
2sin.<br />
a<br />
The CCD image is then reprojected onto the display screen.<br />
We assume a single display/projector model with central<br />
projection and a magnification <strong>of</strong> M with respect to the<br />
CCD sensor image resulting in the following screen coordinates<br />
<strong>for</strong> the point in the left, U l ,V l T , and right, U r ,V r T ,<br />
eye images:<br />
U l<br />
V l = M u l<br />
v l,<br />
U r<br />
V r = M u r<br />
v r.<br />
Toeing-in the stereoscopic rig to converge on a surface centers<br />
the images <strong>of</strong> the target in the two cameras but also<br />
introduces a keystone distortion due to the differential perspective<br />
(Figure 4). In contrast convergence by shifting the<br />
CCD sensor behind the camera lens (or shifting the half<br />
images on the display) changes the mean horizontal disparity<br />
but does not entail keystone distortion. For a given focal<br />
length and camera separation, the extent <strong>of</strong> the keystone<br />
distortion is a function <strong>of</strong> the convergence distance and not<br />
the distance <strong>of</strong> the target.<br />
To see how the keystoning affects depth perception, assume<br />
the images are projected onto a screen at distance D<br />
and viewed by a viewer with interocular distance <strong>of</strong> e. If the<br />
magnification from the CCD sensor array to screen image is<br />
3<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 321
Allison: Analysis <strong>of</strong> the influence <strong>of</strong> vertical disparities arising in toed-in stereoscopic cameras<br />
Figure 5. Geometrically predicted perception curved grid <strong>of</strong> displayed<br />
images taken from a toed-in stereoscopic camera rig converged on a<br />
fronto-parallel grid made with 10 cm spacing asterisks based on horizontal<br />
disparities associated size distortion not shown. Camera convergence<br />
distance F and display viewing distance D are 0.70 cm<br />
e=a=62.5 mm; f=6.5 mm; see Fig. 3 and text <strong>for</strong> definitions. The<br />
icon at the bottom <strong>of</strong> the figure indicates the position <strong>of</strong> the world coordinate<br />
frame and the eyeballs.<br />
M and both images are centered on the display then geometrically<br />
predicted coordinates <strong>of</strong> the point in display space<br />
is (after Ref. 1)<br />
d<br />
P d =X d=<br />
eU l + U r <br />
2e − U r − U l <br />
eV l + V r<br />
<br />
<br />
Y d<br />
4<br />
2e − U r − U l <br />
Z<br />
eD<br />
e − U r − U l <br />
where U r −U l is the horizontal screen parallax <strong>of</strong> the point.<br />
If we ignore vertical disparities <strong>for</strong> the moment, converging<br />
the camera causes changes in the geometrically predicted<br />
depth. For instance, if the cameras toe-in to converge<br />
on a frontoparallel surface (parallel to the stereobaseline),<br />
then from geometric considerations the center <strong>of</strong> the object<br />
should appear at the screen distance but the surface should<br />
appear curved (Figure 5). This curvature should be especially<br />
apparent in the presence <strong>of</strong> undistorted stereoscopic<br />
reference imagery as would occur in augmented reality<br />
applications. 16 In contrast, if convergence is accomplished<br />
via horizontal image translation then a frontal plane at the<br />
camera convergence distance should appear flat and at the<br />
screen distance although depth and size will be scaled as<br />
discussed in the previous section.<br />
USE OF VERTICAL DISPARITY IN STEREOPSIS<br />
The pattern <strong>of</strong> vertical disparities in a stereoscopic image<br />
depends on the geometry <strong>of</strong> the stereoscopic rig. With our<br />
spherical retinas disparity is best defined in terms <strong>of</strong> visual<br />
angle. An object that is located eccentric to the median plane<br />
<strong>of</strong> the head is closer to one eye than the other (Figure 6).<br />
Hence, it subtends a larger angle at the nearer eye than at the<br />
further. The vertical size ratio (VSR) between the images <strong>of</strong><br />
an object in the two eyes varies as a function <strong>of</strong> the object’s<br />
eccentricity with respect to the head. Figure 6 also shows the<br />
variation <strong>of</strong> the vertical size ratio <strong>of</strong> the right eye image to<br />
the left eye image <strong>for</strong> a range <strong>of</strong> eccentricities and<br />
distances.<br />
It is evident that, <strong>for</strong> centrally located targets, the gradient<br />
<strong>of</strong> vertical size ratios varies with distance <strong>of</strong> the surface<br />
from the head. This is relatively independent <strong>of</strong> the vergence<br />
state <strong>of</strong> the eyes and the local depth structure. 17 Howard 18<br />
turned this relationship around and suggested that people<br />
could judge the distance <strong>of</strong> surfaces from the gradient <strong>of</strong> the<br />
VSR. Gillam and Lawergren 19 proposed a computational<br />
model <strong>for</strong> the recovery <strong>of</strong> surface distance and eccentricity<br />
based upon processing <strong>of</strong> VSR and VSR gradients. An alternative<br />
computational framework 10,20 uses vertical disparities<br />
to calculate the convergence posture and gaze eccentricity <strong>of</strong><br />
the eyes rather than the distance and eccentricity <strong>of</strong> a target<br />
surface. For our purposes, these models make the same predictions<br />
about the effects <strong>of</strong> camera toe-in. However, the<br />
latter model uses projections onto flat projection surfaces<br />
(hypothetical flat retinae) which is easier <strong>for</strong> visualization<br />
and matches well with our previous discussion <strong>of</strong> camera<br />
toe-in.<br />
With flat imaging planes, disparities are usually measured<br />
in terms <strong>of</strong> linear displacement in the image plane. If<br />
the cameras in a stereoscopic rig are toed in (or if eyes with<br />
flat retinae are converged), then the left and right camera<br />
images have opposite keystone distortion. It is interesting to<br />
note that in contrast to the angular disparity case the gradients<br />
<strong>of</strong> vertical disparities are a function <strong>of</strong> camera convergence<br />
but are affected little by the distance <strong>of</strong> the surface.<br />
These vertical disparity gradients on flat cameras/retinae<br />
provide an indication <strong>of</strong> the convergence angle <strong>of</strong> the cameras<br />
and hence the distance <strong>of</strong> the fixation point.<br />
For a pair <strong>of</strong> objects or <strong>for</strong> depth within an object, the<br />
relationship between relative depth and relative disparity is a<br />
function <strong>of</strong> distance from the observer. To an extent, the<br />
visual system is able to maintain an accurate perception <strong>of</strong><br />
depth <strong>of</strong> an object at various distances despite disparity<br />
varying inversely with the square <strong>of</strong> the distance between the<br />
object and the observer. This “depth constancy” demonstrates<br />
an ability to account <strong>for</strong> the effects <strong>of</strong> viewing distance<br />
on stereoscopic depth. The relationship between the<br />
retinal image size <strong>of</strong> an object and its linear size in the world<br />
is also a function <strong>of</strong> distance. To the degree that vertical<br />
disparity gradients are used as an indicator <strong>of</strong> the distance <strong>of</strong><br />
a fixated surface <strong>for</strong> three-dimensional reconstruction, toe-in<br />
produced vertical disparity gradients would be expected to<br />
indirectly affect depth and size perception. Psychophysical<br />
experiments have demonstrated that vertical disparity gradients<br />
strongly affect perception <strong>of</strong> stereoscopic shape, size and<br />
depth 9,10,21 and implicate vertical disparity processing in human<br />
size and depth constancy.<br />
322 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Allison: Analysis <strong>of</strong> the influence <strong>of</strong> vertical disparities arising in toed-in stereoscopic cameras<br />
Figure 6. a A vertical line located eccentric to the midline <strong>of</strong> the head is nearer to one eye than the other.<br />
Thus it subtends a larger angle in the nearer eye than the further adapted from Howard and Rogers 6 . b The<br />
gradient <strong>of</strong> vertical size ratio <strong>of</strong> the image <strong>of</strong> a surface element in the left eye to that in the right eye varies as<br />
a function <strong>of</strong> distance <strong>of</strong> the surface shown as a series <strong>of</strong> lines: distances <strong>of</strong> 70, 60, 50, 40, and 30 cm in<br />
order <strong>of</strong> steepness.<br />
VERTICAL DISPARITY IN TOED-IN STEREOSCOPIC<br />
CAMERAS<br />
First, consider a stereoscopic camera and parallel display system<br />
that intends to portray realistic depth and that has camera<br />
separation equal to the eye separation. If the camera is<br />
converged using the toe-in method at a fronto-parallel surface<br />
at the distance <strong>of</strong> the screen, then the center <strong>of</strong> the<br />
target will have zero horizontal screen disparity. However,<br />
the camera toe-in will introduce keystone distortion into the<br />
two images with the pattern <strong>of</strong> horizontal disparities predicting<br />
curvature as discussed above. What about the pattern <strong>of</strong><br />
vertical disparities? The pattern <strong>of</strong> vertical disparities produced<br />
by a toed-in camera configuration resembles the gradient<br />
<strong>of</strong> vertical size disparities on the retinae that can arise<br />
due to differential perspective <strong>of</strong> the two eyes. As discussed<br />
in the previous section, this differential perspective <strong>for</strong>ms a<br />
natural and rich source <strong>of</strong> in<strong>for</strong>mative parameters contributing<br />
to human stereoscopic depth perception.<br />
Given that camera toe-in generates such gradients <strong>of</strong><br />
vertical disparity in stereoscopic imagery, is it beneficial to<br />
use camera toe-in to provide distance in<strong>for</strong>mation in a stereoscopic<br />
display? In other words, should the toed-in configuration<br />
be used to converge the cameras and preserve the<br />
sense <strong>of</strong> absolute distance and size, shape and depth constancy?<br />
Perez-Bayas 22 argued that toed-in camera configurations<br />
are more natural since they present these vertical disparities.<br />
The principal problem with this claim is that it<br />
considers the screen parallax <strong>of</strong> stereoscopic images rather<br />
than their retinal disparities. These keystone distortions are<br />
in addition to the natural retinal vertical disparities present<br />
when viewing a scene at the distance <strong>of</strong> the screen.<br />
In order to estimate the effect on depth perception we<br />
need to consider the retinal disparities generated by the stereoscopic<br />
image. The keystone distortion occurs in addition<br />
to the retinal vertical disparity pattern inherent in the image<br />
because it is portrayed on the flat screen. Consider a frontoparallel<br />
surface located at the distance <strong>of</strong> the screen away<br />
from the camera and that we intend to display the surface at<br />
the screen. Projections onto spherical retinas are hard to<br />
visualize so let us consider flat retinae converged (toed-in) at<br />
the screen distance. Alternatively one could imagine another<br />
pair <strong>of</strong> converged cameras viewing the display, one centered<br />
at the center <strong>of</strong> each eye. The images on these converged flat<br />
retinae would <strong>of</strong> course have differential keystone distortion<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 323
Allison: Analysis <strong>of</strong> the influence <strong>of</strong> vertical disparities arising in toed-in stereoscopic cameras<br />
Figure 7. a Simulation <strong>of</strong> the keystone distortion and gradient <strong>of</strong> VSR<br />
present in a stereo half image <strong>for</strong> a toed-in configuration. The plus symbols<br />
show the keystone distortion in the displayed image <strong>of</strong> a grid <strong>for</strong> a<br />
camera converged at 70 cm and the circle symbols indicated the exaggerated<br />
VSR distortion present in the retinal half image <strong>for</strong> an observer<br />
viewing the display at 70 cm flat retina. b Predicted distorted appearance<br />
circles in a set <strong>of</strong> frontal plane surfaces asterisks if depth from<br />
disparity is scaled according to the distance indicated by an exaggerated<br />
VSR. Typically the surface is not mislocalized in depth but curvature is<br />
induced. The predicted curvature based on the on the equations provided<br />
by Duke and Wilcox 28 is also shown diamonds. The simulated positions<br />
<strong>of</strong> the eyes are denoted by circles at zero distance and the screen by a<br />
line at 70 cm.<br />
when viewing a frontal surface such as the screen. When<br />
displaying images from the toed-in stereoscopic camera,<br />
which already have keystone distortion, the result is an exaggerated<br />
gradient <strong>of</strong> vertical disparity in the retinal images<br />
appropriate <strong>for</strong> a much nearer surface. For a spherical retina<br />
the important measure is the gradient <strong>of</strong> vertical size ratios<br />
in the image. The vertical size ratios in the displayed images<br />
imposed by the keystone distortion are in addition to the<br />
natural VSR <strong>for</strong> a frontal surface at the distance <strong>of</strong> the<br />
screen. Clearly, the additional keystone distortion indicates a<br />
nearer surface in this case as well [Figure 7(a)].<br />
From either the flat camera or spherical retina model we<br />
predict spatial distortion if disparities are scaled according to<br />
the vertical disparities, which indicate a closer target. Such a<br />
misjudgement <strong>of</strong> perceived distance would be predicted to<br />
have effects on perceived depth and size [open circles in Fig.<br />
7(b)]. There is little evidence that observers actually<br />
mislocalize surfaces at a nearer distance when a vertical disparity<br />
gradient is imposed. However, there is strong evidence<br />
<strong>for</strong> effects <strong>of</strong> VSR gradients on depth constancy processes.<br />
If a viewer fixates a point on a fronto-parallel screen,<br />
then at all screen distances nearer than infinity the images <strong>of</strong><br />
other points on the screen have horizontal disparity (retinal<br />
but not screen disparity). This is because the theoretical locus<br />
<strong>of</strong> points in three-dimensional space with zero retinal<br />
disparity, which is known as the horopter (the Vieth-Muller<br />
circle), curves inward toward the viewer and away from the<br />
frontal plane. The curvature <strong>of</strong> the horopter increases at<br />
nearer distances (Figure 8). 23 Thus a frontal plane presents a<br />
pattern <strong>of</strong> horizontal disparities that varies with distance. If<br />
depth constancy is to be maintained <strong>for</strong> fronto-parallel<br />
planes then the distance <strong>of</strong> the surface needs to be taken into<br />
account. Rogers and Bradshaw 21 showed that vertical disparity<br />
patterns can have a strong influence on frontal plane<br />
judgements, particularly <strong>for</strong> large field <strong>of</strong> view displays. Specifically,<br />
“flat”—or zero horizontal screen disparity—planes<br />
are perceived as curved if vertical disparity gradients indicate<br />
a distance other than the screen distance.<br />
In our case, the toe-in induced vertical disparity introduces<br />
a cue that the surface is nearer than specified by the<br />
horizontal screen disparity. Thus a zero horizontal screen<br />
disparity pattern <strong>for</strong> a frontal surface at the true distance<br />
would be interpreted as at nearer distance. The disparities<br />
would be less than expected from a frontal plane at the<br />
nearer distance. As a result, surfaces in a scene should appear<br />
curved more concavely than they are in the real scene. Notice<br />
that the distortion is in the opposite direction than the<br />
distortion created by horizontal disparities due to the<br />
keystoning.<br />
Thus the effect <strong>of</strong> vertical disparity introduced by the<br />
keystone distortion is complicated. The vertical disparity introduces<br />
a cue that the surface is nearer than specified by the<br />
horizontal screen disparity. Thus, from vertical disparities,<br />
we would expect a bias in depth perception and concave<br />
distortion <strong>of</strong> stereoscopic space. This may counter the convex<br />
distortions introduced by the horizontal disparities discussed<br />
above. So the surface may appear flatter than expected<br />
from the distorted horizontal disparities. But the<br />
percept is not more “natural” than the parallel configura-<br />
324 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Allison: Analysis <strong>of</strong> the influence <strong>of</strong> vertical disparities arising in toed-in stereoscopic cameras<br />
Figure 8. Disparity <strong>of</strong> a point on a fronto-parallel surface as a function <strong>of</strong> distance. Horizontal disparity <strong>for</strong> a<br />
given eccentricity increases with nearness due to the increasing curvature <strong>of</strong> the Veith-Muller circle see text.<br />
tion. Rather two distortions due to camera toe-in act to<br />
cancel each other out.<br />
Do toed-in configurations provide useful distance<br />
in<strong>for</strong>mation <strong>for</strong> objects at other distances or<br />
<strong>for</strong> nonorthostereoscopic configurations?<br />
Since the toe-in induced vertical disparity gradients are superimposed<br />
upon the natural vertical disparity at the retinae<br />
they do not provide natural distance cues <strong>for</strong> targets near the<br />
display under orthostereoscopic configurations.<br />
Nonorthostereoscopic configurations are more common<br />
than orthostereoscopic and we should consider the effects <strong>of</strong><br />
toe-in on these configurations. Magnification and minification<br />
<strong>of</strong> the images will scale the disparities in the images as<br />
well so that the vertical gradient <strong>of</strong> vertical size ratio will be<br />
relatively unchanged under uni<strong>for</strong>m magnification. Hence<br />
we expect a similar curvature distortion under magnification<br />
or minification.<br />
Hyperstereoscopic and hypostereoscopic configurations<br />
exaggerate and attenuate, respectively, the horizontal and<br />
vertical disparities due to camera toe-in and the magnitude<br />
<strong>of</strong> the stereoscopic distortions will be scaled. However, <strong>for</strong><br />
both configurations the sign <strong>of</strong> the distortion is the same<br />
and vertical disparities from camera toe-in predict concave<br />
curvature <strong>of</strong> stereoscopic space with increased distortion<br />
with an increased stereobaseline.<br />
For surfaces outside the plane <strong>of</strong> the screen, vertical<br />
keystone distortion from toe-in still introduces spatial distortion.<br />
A surface located at a distance beyond the screen in<br />
a parallel camera, orthostereoscopic configuration will have<br />
VSR gradients on spherical retinae appropriate to its distance<br />
due to the imaging geometry. For a toed-in camera<br />
system, all surfaces in the scene will have additional vertical<br />
disparity gradients due to the keystoning. These increased<br />
vertical disparity gradients would indicate a nearer convergence<br />
distance or a nearer surface thus the distance <strong>of</strong> the far<br />
surface should be underestimated and concave curvature introduced.<br />
The distance underestimation would be compounded<br />
by rescaling <strong>of</strong> disparity <strong>for</strong> the near distance<br />
which should compress the depth range in the scene.<br />
What about partial toe-in? For example, let us say we<br />
toed in on a target at 3mand displayed it at 1.0 m with the<br />
centers <strong>of</strong> the image aligned? Would the vertical disparities<br />
in the image indicate a more distant surface, perhaps even<br />
one at 3m (this would be the case if viewed in a haploscope)?<br />
A look at the pattern <strong>of</strong> vertical screen disparities in<br />
this case, however, shows that they are appropriate <strong>for</strong> a<br />
surface that is nearer than the 3msurface, and in fact nearer<br />
than the screen if the half images are aligned on the screen.<br />
Thus when the vertical screen disparities are compounded<br />
by the inherent vertical retinal disparities introduced by<br />
viewing the screen, the toe-in induced distortion actually<br />
indicates a nearer surface rather than the further surface<br />
desired. We will see below that vertical disparity manipulations<br />
can produce the impression <strong>of</strong> a further surface but the<br />
required trans<strong>for</strong>mation is opposite to the one introduced by<br />
camera toe-in.<br />
Do the toed-in configurations improve depth and size<br />
scaling?<br />
Vertical disparities have been shown to be effective in the<br />
scaling <strong>of</strong> depth, shape and size from disparity. 9,21 When the<br />
cameras are toed-in the vertical disparities indicate a nearer<br />
surface. There<strong>for</strong>e, camera toe-in should cause micropsia (or<br />
apparent shrinking <strong>of</strong> linear size) appropriate <strong>for</strong> the nearer<br />
distance. Similarly, depth from disparity should be scaled<br />
appropriate to a nearer surface and depth range should be<br />
compressed. Thus, if toe-in is used to converge an otherwise<br />
orthostereoscopic rig, then image size and depth should be<br />
compressed. Vertical disparity cues to distance are most effective<br />
in a large field <strong>of</strong> view display and the curvature, size<br />
and depth effects are most pronounced in these types <strong>of</strong><br />
displays. 9,21<br />
In the orthostereoscopic case with parallel cameras,<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 325
Allison: Analysis <strong>of</strong> the influence <strong>of</strong> vertical disparities arising in toed-in stereoscopic cameras<br />
there are no vertical screen disparities and the vertical disparities<br />
in the retinal images are appropriate <strong>for</strong> the screen<br />
distance and no size or depth distortions due to vertical<br />
disparity are predicted. Vertical disparities in the retinal (but<br />
not display) images can thus help obtain veridical stereoscopic<br />
perception.<br />
I use computer graphics or image processing to render<br />
stereoscopic images. Can I use VSR to give an<br />
impression <strong>of</strong> different distances? If so how?<br />
Incorporating elements that carry vertical disparity in<strong>for</strong>mation<br />
(<strong>for</strong> example with horizontal edges) can lead to more<br />
veridical depth perception 8 and in this simple sense vertical<br />
disparity cues can assist in the development <strong>of</strong> effective stereoscopic<br />
displays. It is not certain that manipulating vertical<br />
disparity independent <strong>of</strong> vergence would be <strong>of</strong> use to content<br />
creators, but it is possible. In the lab we do this to look<br />
at the effects <strong>of</strong> vertical disparity gradients and to manipulate<br />
the effects <strong>of</strong> vertical disparities with vergence held constant.<br />
We have seen that toe-in convergence introduces a vertical<br />
disparity cue that indicates that a surface is nearer than<br />
other cues indicate. This will scale stereoscopic depth, shape<br />
and size appropriately, particularly <strong>for</strong> large displays. To<br />
make the surface appear further away the opposite trans<strong>for</strong>mation<br />
is required to reduce the vertical disparity gradients<br />
in the retinal image—this essentially entails “toe-out” <strong>of</strong> the<br />
cameras. VSR manipulations, intentional or due to camera<br />
toe-in, exacerbate cue conflict in the display as the distance<br />
estimate obtained from the vertical disparities will conflict<br />
with accommodation, vergence, and other cues to distance.<br />
FUSION OF VERTICAL DISPARITY<br />
In many treatments <strong>of</strong> the camera convergence problem it is<br />
noted that the vertical disparities introduced by toed-in<br />
camera convergence may interfere with the ability to fuse the<br />
images and cause visual discom<strong>for</strong>t. 24 Certainly, vertical fusional<br />
range is known to be less than horizontal fusional<br />
range 23 making it likely that vertical disparities could be<br />
problematic. Tolerance to vertical disparities depends on<br />
several factors including size <strong>of</strong> the display, and the presence<br />
<strong>of</strong> reference surfaces.<br />
When a stereoscopic image pair has an overall vertical<br />
misalignment, such as arises with vertical camera misalignment,<br />
viewers can compensate with vertical vergence and<br />
sensory fusional mechanisms. Vertical vergence is a disjunctive<br />
eye movement where the left and right eyes move in<br />
opposite directions vertically (vertical misalignment can also<br />
<strong>of</strong>ten be partially compensated by tilting the head with respect<br />
to the display). Vertical disparities are integrated over a<br />
fairly large region <strong>of</strong> space to <strong>for</strong>m the stimulus to vertical<br />
vergence. 25 Larger displays increase the vertical vergence response<br />
and the vertical fusional range. Thus we predict that<br />
vertical disparities will be better tolerated in large displays.<br />
In agreement with this Speranza and Wilcox 26 found up to<br />
30 minutes <strong>of</strong> arc <strong>of</strong> vertical disparity could be tolerated in a<br />
stereoscopic IMAX film without significant viewer discom<strong>for</strong>t.<br />
However, convergence via camera toe-in gives local<br />
variations in vertical disparity and thus images <strong>of</strong> objects in<br />
the display have spatially varying vertical disparities. Thus,<br />
averaging retinal vertical disparities over a region <strong>of</strong> space<br />
should be less effective in compensating <strong>for</strong> vertical disparity<br />
due to camera toe-in compared to overall vertical camera<br />
misalignment. Furthermore, any vertical vergence to fuse<br />
one portion <strong>of</strong> the display will increase vertical disparity in<br />
other parts <strong>of</strong> the display.<br />
The ability to fuse a vertically disparate image is reduced<br />
when nearby stimuli have different vertical disparities, particularly<br />
if the target and background are similar in depth. 27<br />
In many display applications the frame <strong>of</strong> the display is visible<br />
and serves as a frame <strong>of</strong> reference. In other applications<br />
such as augmented reality and enhanced vision displays the<br />
stereoscopic imagery may be imposed upon other imagery.<br />
Presence <strong>of</strong> these competing stereoscopic images will be expected<br />
to reduce the tolerance to vertical disparity due to<br />
camera convergence. 27 This indicates that vertical disparity<br />
distortions should be particularly disruptive in augmented<br />
reality displays where the stereoscopic image is superimposed<br />
on other real or synthetic imagery and parallel cameras<br />
or image rectification should be used.<br />
ADAPTATION AND SENSORY INTEGRATION OF<br />
TOE-IN INDUCED VERTICAL DISPARITY<br />
The human visual system relies on a variety <strong>of</strong> monocular<br />
and binocular cues to judge distance and relative depth in a<br />
scene. The effects <strong>of</strong> toe-in induced horizontal and vertical<br />
disparities on depth and distance perception discussed above<br />
will be reduced when viewing a scene rich in these cues. The<br />
extent <strong>of</strong> the perceptual distortion depends on perceptual<br />
biases and the relative effectiveness <strong>of</strong> the various cues. For<br />
example, Bradshaw and Rogers 21 per<strong>for</strong>med an experiment<br />
using dot displays to study size and depth scaling as a function<br />
<strong>of</strong> distance indicated by vertical disparities and vergence.<br />
They argued that use <strong>of</strong> vertical disparity in<strong>for</strong>mation<br />
to drive size and depth constancy requires measuring the<br />
relevant disparity gradients over a fairly large retinal area<br />
whereas vergence signals, correlated with egocentric distance,<br />
could be obtained during binocular viewing <strong>of</strong> a point<br />
source <strong>of</strong> light. Accordingly, when displays were small, subjects<br />
responded as if they were scaling the stimulus appropriate<br />
<strong>for</strong> the distance indicated by vergence; when displays<br />
were large subjects responded as if they were scaling the<br />
stimulus appropriate <strong>for</strong> the distance indicated by vertical<br />
disparity. When other cues reliably indicate a different distance<br />
than toe-in induced vertical disparities the effect <strong>of</strong> the<br />
latter on depth and size perception may be small. However,<br />
latent, even imperceptible, cue conflicts are believed to be a<br />
causal factor in simulator sickness symptoms such as eye<br />
strain and nausea. 5<br />
When sensory conflict is persistent, the visual system<br />
shows remarkable ability to adapt or recalibrate. Following<br />
prolonged viewing <strong>of</strong> a test stimulus that appears curved due<br />
to keystone-type vertical disparity trans<strong>for</strong>mations a nominally<br />
flat stimulus appears curved in the opposite direction.<br />
Duke and Wilcox 28 have claimed this adaptation is driven by<br />
326 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Allison: Analysis <strong>of</strong> the influence <strong>of</strong> vertical disparities arising in toed-in stereoscopic cameras<br />
the curvature in depth induced rather than by the vertical<br />
disparities directly. In general, such an aftereffect can reflect<br />
“habituation” or “fatigue” <strong>of</strong> mechanisms sensitive to the<br />
adapting pattern, or from a recalibration <strong>of</strong> the vertical disparity<br />
signal, or a change in the relative weighting <strong>of</strong> cues<br />
driving depth constancy. At the present time it is unclear<br />
which <strong>of</strong> these adaptive changes can be produced by prolonged<br />
exposure to keystone patterns <strong>of</strong> vertical disparity.<br />
The effects <strong>of</strong> vertical disparities induced by toe-in convergence<br />
also depends on context and may differ depending<br />
on the type <strong>of</strong> task being per<strong>for</strong>med by the subject and<br />
whether they involve size constancy, depth constancy, absolute<br />
distance judgements or other spatial judgements. For<br />
example, Wei et al. 29 reported that full-field vertical disparities<br />
are not used to derive the distance dependent gain term<br />
<strong>for</strong> the linear vestibulo-ocular reflex, a reflexive eye movement<br />
that compensates <strong>for</strong> head movements, under conditions<br />
where vertical disparities drive depth constancy.<br />
CONCLUSIONS<br />
In conclusion, we concur with conventional wisdom that<br />
horizontal image translation is theoretically preferred to<br />
toe-in convergence with parallel stereoscopic displays.<br />
Toed-in camera convergence is a convenient and <strong>of</strong>ten used<br />
technique that is <strong>of</strong>ten well-tolerated 24 despite the fact that it<br />
theoretically and empirically results in geometric distortion<br />
<strong>of</strong> stereoscopic space. The distortion <strong>of</strong> stereoscopic space<br />
should be more apparent in fused or augmented reality displays<br />
where the real world serves as a reference to judge the<br />
disparity distortion introduced by the toe-in technique. In<br />
these cases, and <strong>for</strong> near viewing when the distortions are<br />
large, the distortions may be ameliorated through camera<br />
rectification techniques 15,30 if resampling <strong>of</strong> the images is<br />
practical.<br />
It has been asserted by others that, since camera convergence<br />
through toe-in introduces vertical disparities into<br />
the stereoscopic imagery it should give rise to more natural<br />
or accurate distance perception than the parallel camera<br />
configuration. We have argued in this paper that these assertions<br />
are theoretically unfounded although vertical disparity<br />
gradients are an effective cue <strong>for</strong> depth and size constancy<br />
that could be used by creators <strong>of</strong> stereoscopic content. The<br />
geometrical distortions predicted from the artifactual horizontal<br />
disparities created by camera toe-in may be countered<br />
by opposite distortions created from the vertical disparities.<br />
However, when displayed on a single projector or monitor<br />
display the vertical disparity gradients introduced by<br />
unrectified, toed-in cameras do not correspond to the gradients<br />
experienced by a real user viewing a scene at the<br />
camera convergence distance. This is because the keystoning<br />
due to the camera toe-in is superimposed upon the natural<br />
vertical disparity pattern at the eyes.<br />
Our analysis and data 27 implies that stereoscopic<br />
display/camera systems that fuse or superimpose multiple<br />
stereoscopic images from a number <strong>of</strong> sensors should be<br />
more susceptible to toe-in induced fusion and depthdistortion<br />
problems than displays that present a single stereoscopic<br />
image stream. Analysis <strong>of</strong> toe-in induced vertical<br />
disparity rein<strong>for</strong>ces the recommendation that rectification <strong>of</strong><br />
the stereoscopic imagery should be considered <strong>for</strong> fused stereoscopic<br />
systems such as augmented reality displays or enhanced<br />
vision systems that require toed-in cameras to view<br />
targets at short distances.<br />
ACKNOWLEDGMENTS<br />
The support <strong>of</strong> the Ontario Centres <strong>of</strong> Excellence and<br />
NSERC Canada is gratefully acknowledged. An abbreviated<br />
version <strong>of</strong> this paper was presented at IST/SPIE Electronic<br />
<strong>Imaging</strong> 2004 [R. Allison, Proc. SPIE 5291, 167–178 (2004)].<br />
REFERENCES<br />
1 A. Woods, T. Docherty, and R. Koch, Proc. SPIE 1915, 36 (1993).<br />
2 D. B. Diner and D. H. Fender, Human Engineering in Stereoscopic<br />
Viewing Devices (Plenum Press, New York and London, 1993).<br />
3 L. Lipton, Foundations <strong>of</strong> the Stereoscopic Cinema (Van<br />
Nostrand–Reinhold, New York, 1982).<br />
4 Z. Wartell, L. F. Hodges, and W. Ribarsky, IEEE Trans. Vis. Comput.<br />
Graph. 8(2), 129 (2002).<br />
5 J. P. Wann, S. Rushton, and M. Monwilliams, Vision Res. 35(19), 2731<br />
(1995).<br />
6 I.P.HowardandB.J.Rogers,Depth Perception (I. Porteous, Toronto,<br />
2002).<br />
7 L. Lipton, The Stereographics Developers Handbook (Stereographics<br />
Corp., San Rafael, CA, 1997).<br />
8 H. von Helmholtz, Physiological Optics, English translation by J. P. C.<br />
Southall from the 3rd German edition <strong>of</strong> Handbuch der Physiologischen<br />
Optik, Vos, Hamburg (Dover, New York, 1962).<br />
9 B. J. Rogers and M. F. Bradshaw, Nature (London) 361, 253 (1993).<br />
10 J. Garding, J. Porrill, J. E. W. Mayhew, and J. P. Frisby, Vision Res. 35(5),<br />
703 (1995).<br />
11 M. Siegel and S. Nagata, IEEE Trans. Circuits Syst. Video Technol. 10(3),<br />
387 (2000).<br />
12 A. State, J. Ackerman, G. Hirota, J. Lee, and H. Fuchs, Proc. International<br />
Symposium on Augmented Reality (ISAR) 2001 (IEEE, Piscataway, NJ,<br />
2001) pp. 137–146.<br />
13 V. S. Grinberg, G. Podnar, and M. W. Siegel, Proc. SPIE 2177, 56 (1994).<br />
14 A. State, K. Keller, and H. Fuchs, Proc. International Symposium on<br />
Mixed and Augmented Reality (ISMAR) 2005 (IEEE, Piscataway, NJ,<br />
2005) pp. 28–31.<br />
15 N. Dodgson, Proc. SPIE 3295, 100 (1998).<br />
16 S. Takagi, S. Yamazaki, Y. Saito, and N. Taniguchi, Proc IEEE & ACM<br />
ISAR 2000 (IEEE, Piscataway, NJ, 2000) pp. 68–77.<br />
17 B. Gillam and B. Lawergren, Percept. Psychophys. 34(2), 121 (1983).<br />
18 I. P. Howard, Psychonomic Monograph Supplements 3, 201 (1970).<br />
19 B. Gillam, D. Chambers, and B. Lawergren, Percept. Psychophys. 44, 473<br />
(1988).<br />
20 J. E. W. Mayhew and H. C. Longuet-Higgins, Nature (London) 297, 376<br />
(1982).<br />
21 M. F. Bradshaw, A. Glennerster, and B. J. Rogers, Vision Res. 36(9), 1255<br />
(1996).<br />
22 L. Perez-Bayas, Proc. SPIE 4297, 251 (2001).<br />
23 K. N. Ogle, Researches in Binocular Vision (Hafner, New York, 1964).<br />
24 L. B. Stelmach, W. J. Tam, F. Speranza, R. Renaud, and T. Martin, Proc.<br />
SPIE 5006, 269 (2003).<br />
25 I. P. Howard, X. Fang, R. S. Allison, and J. E. Zacher, Exp. Brain Res.<br />
130(2), 124 (2000).<br />
26 F. Speranza and L. Wilcox, Proc. SPIE 4660, 18 (2002).<br />
27 R. S. Allison, I. P. Howard, and X. Fang, Vision Res. 40(21), 2985 (2000).<br />
28 P. A. Duke and L. M. Wilcox, Vision Res. 43(2), 135 (2003).<br />
29 M. Wei, G. C. DeAngelis, and D. E. Angelaki, J. Neurosci. 23, 8340<br />
(2003).<br />
30 O. Faugeras and Q. Luong, The Geometry <strong>of</strong> Multiple Images (MIT<br />
Press, Cambridge, MA, 2001).<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 327
<strong>Journal</strong> <strong>of</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology® 51(4): 328–336, 2007.<br />
© <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology 2007<br />
Improved B-Spline Contour Fitting Using Genetic<br />
Algorithm <strong>for</strong> the Segmentation <strong>of</strong> Dental Computerized<br />
Tomography Image Sequences<br />
Xiaoling Wu, Hui Gao, Hoon Heo, Oksam Chae, Jinsung Cho, Sungyoung Lee and Young-Koo Lee<br />
Department <strong>of</strong> Computer Engineering, Kyunghee University, 1 Seochun-ri, Kiheung-eup, Yongin-si,<br />
Kyunggi-do, 449-701, South Korea<br />
E-mail: yklee@khu.ac.kr<br />
Abstract. In the dental field, 3D tooth modeling, in which each tooth<br />
can be manipulated individually, is an essential component <strong>of</strong> the<br />
simulation <strong>of</strong> orthodontic surgery and treatment. However, in dental<br />
computerized tomography slices teeth are located closely together<br />
or inside alveolar bone having an intensity similar to that <strong>of</strong> teeth.<br />
This makes it difficult to individually segment a tooth be<strong>for</strong>e building<br />
its 3D model. Conventional methods such as the global threshold<br />
and snake algorithms fail to accurately extract the boundary <strong>of</strong> each<br />
tooth. In this paper, we present an improved contour extraction algorithm<br />
based on B-spline contour fitting using genetic algorithm.<br />
We propose a new fitting function incorporating the gradient direction<br />
in<strong>for</strong>mation on the fitting contour to prevent it from invading the<br />
areas <strong>of</strong> other teeth or alveolar bone. Furthermore, to speed up the<br />
convergence to the best solution we use a novel adaptive probability<br />
<strong>for</strong> crossover and mutation in the evolutionary program <strong>of</strong> the genetic<br />
algorithm. Segmentation results <strong>for</strong> real dental images demonstrate<br />
that our method can accurately determine the boundary <strong>for</strong><br />
individual teeth as well as its 3D model while other methods fail.<br />
Independent manipulation <strong>of</strong> each tooth model demonstrates the<br />
practical usage <strong>of</strong> our method. © 2007 <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong><br />
and Technology.<br />
DOI: 10.2352/J.<strong>Imaging</strong>Sci.Technol.200751:4328<br />
INTRODUCTION<br />
The accurate 3D modeling <strong>of</strong> the mandible and the simulation<br />
<strong>of</strong> tooth movement play an important role in preoperative<br />
planning <strong>for</strong> dental and maxill<strong>of</strong>acial surgery. The 3D<br />
reconstruction <strong>of</strong> the teeth can be used in virtual reality<br />
based training <strong>for</strong> orthodontics students and <strong>for</strong> preoperatory<br />
assessment by dental surgeons. For 3D modeling tooth<br />
segmentation to extract the individual contour <strong>of</strong> a tooth is<br />
<strong>of</strong> critical importance. Automated tooth segmentation methods<br />
from 3D digitized images have been researched <strong>for</strong> the<br />
measurement and simulation <strong>of</strong> orthodontic procedures. 1<br />
These methods provide interstices along with their locations<br />
and orientations between the teeth <strong>for</strong> segmentation result.<br />
However, it does not give individual tooth contour in<strong>for</strong>mation<br />
which manifests more details that are helpful in dental<br />
study. A thresholding method, used in the existing segmentation<br />
and reconstruction systems, is known to be efficient<br />
<strong>for</strong> automatic hard tissue segmentation. 2,3 Some morphological<br />
filtering methods are used <strong>for</strong> creating intermediary<br />
Received Oct. 28, 2006; accepted <strong>for</strong> publication Mar. 30, 2007.<br />
1062-3701/2007/514/328/9/$20.00.<br />
slices by interpolation <strong>for</strong> modeling teeth in 3D. 4 The morphological<br />
operations are also combined with the thresholding<br />
method <strong>for</strong> dental segmentation in x-ray films. 2 However,<br />
neither the thresholding method nor the morphological<br />
filtering method is suitable <strong>for</strong> separating individual tooth<br />
regions using tooth computerized tomography (CT) slices,<br />
because some teeth touch each other and some are located<br />
inside <strong>of</strong> alveolar bone with a CT slice intensity pr<strong>of</strong>ile similar<br />
to teeth. 5 A modified watershed algorithm was suggested<br />
to create closed-loop contours <strong>of</strong> teeth while alleviating the<br />
over-segmentation problem <strong>of</strong> the watershed algorithm. 5 Although<br />
this reduces the number <strong>of</strong> regions significantly, it<br />
still produces many irrelevant basins that make it difficult to<br />
define an accurate tooth contour. A seed-growing segmentation<br />
algorithm 6 was suggested based on B-spline fitting <strong>for</strong><br />
arbitrary shape segmentation in sequential images. The best<br />
contour <strong>of</strong> an object is determined by fitting the initial contour<br />
passed by previous frame to the edges detected in the<br />
current frame. For the fitting operation, the objective function<br />
defined by the sum <strong>of</strong> distances between the initial contour<br />
and the object edges is used. For this algorithm to work<br />
properly, the complete object boundary should be extracted<br />
by global thresholding and the object should be located<br />
apart from other objects. If other objects are located nearby<br />
as in the case <strong>of</strong> the tooth CT image, the shape <strong>of</strong> the initial<br />
contour should be very close to the actual object contour to<br />
prevent being fitted to the boundaries <strong>of</strong> the nearby objects.<br />
Many snake algorithms have been proposed <strong>for</strong> medical<br />
image analysis applications. 7–10 However, in the CT image<br />
sequence where objects are closely located, the classical snake<br />
algorithms have not yet been successful due to difficulties in<br />
initialization and the existence <strong>of</strong> multiple extrema. It is only<br />
successful when it is initialized close to the structure <strong>of</strong> interest<br />
and there is no object which has similar intensity values<br />
to those <strong>of</strong> interest. 7 The snake models <strong>for</strong> object<br />
boundary detection search <strong>for</strong> an optimal contour that minimizes<br />
(or maximizes) an objective function. The objective<br />
function generally consists <strong>of</strong> the internal energy representing<br />
the properties <strong>of</strong> a contour shape and the external potential<br />
energy depending on the image <strong>for</strong>ce. The final shape<br />
<strong>of</strong> the contour is influenced by how these two energy terms<br />
are represented. However, many snakes tend to shrink when<br />
328
Wu et al.: Improved B-spline contour fitting using genetic algorithm <strong>for</strong> the segmentation <strong>of</strong> dental...<br />
its external energy is relatively small due to the lack <strong>of</strong> image<br />
<strong>for</strong>ces. 7 Some snakes also suffer from the limited flexibility <strong>of</strong><br />
representing the contour shape and a large number <strong>of</strong> derivative<br />
terms in their internal energy representation. A<br />
B-spline based snake has been developed as a B-spline snake<br />
and B-snake to enhance the geometric flexibility and optimization<br />
speed by means <strong>of</strong> a small number <strong>of</strong> control<br />
points instead <strong>of</strong> snaxels. 11,12 B-spline snake controls contour<br />
shapes by a stiffening parameter as well as its control<br />
points, and detects object boundaries in noisy environments<br />
by using gradient magnitude in<strong>for</strong>mation instead <strong>of</strong> edge<br />
in<strong>for</strong>mation. This algorithm introduces a stiffening factor to<br />
the B-spline function 13 that varies the spacing between the<br />
spline knots and the number <strong>of</strong> sampled points used during<br />
the evaluation <strong>of</strong> the objective function. In addition, the<br />
factor controls the smoothness <strong>of</strong> curve and reduces the<br />
computation <strong>of</strong> the cost function. Although the algorithm<br />
was proposed to extract the contour <strong>of</strong> a de<strong>for</strong>mable object<br />
in a single image, it can be applied to the tooth segmentation<br />
in CT slices. However, in tooth CT data, the algorithm may<br />
cause the contour <strong>of</strong> a tooth to expand to include contours<br />
<strong>of</strong> nearby teeth and alveolar bone, or it may cause the contour<br />
to be contracted to a small region.<br />
A B-spline fitting algorithm employing a genetic algorithm<br />
(GA) was used to overcome local extrema indwelling<br />
in the vicinity <strong>of</strong> an object <strong>of</strong> interest. 14–17 In this case, it was<br />
shown that the GA does not require exhaustive search while<br />
avoiding high-order derivatives <strong>for</strong> curve fitting or matching<br />
problems. 18,19 However, the conventional GA-based B-spline<br />
fitting still suffers from the influence <strong>of</strong> other objects and<br />
<strong>of</strong>ten fails to extract the object boundary from the image<br />
sequences when similar objects are adjacent to each other.<br />
In this paper, we propose an improved B-spline contour<br />
fitting algorithm using a GA to generate a smooth and accurate<br />
tooth boundary <strong>for</strong> the 3D reconstruction <strong>of</strong> a tooth<br />
model. We devise a new B-spline fitting function by incorporating<br />
the gradient direction in<strong>for</strong>mation on the fitting<br />
contours to search the tooth boundary while preventing it<br />
from being fitted to neighboring spurious edges. We also<br />
present an evolution method to accelerate the search speed<br />
by means <strong>of</strong> automatic and dynamic determination <strong>of</strong> GA<br />
probabilities <strong>for</strong> crossover and mutation. Experimental results<br />
show that our method can successfully extract the individual<br />
tooth boundary, compared with other methods<br />
which fail to do so.<br />
BACKGROUND<br />
Dental CT images have the following two distinct characteristics:<br />
(1) An individual tooth <strong>of</strong>ten appears with neighboring<br />
hard tissues such as other teeth and alveolar bone, and<br />
(2) these neighboring hard tissues have the same or similar<br />
intensity values to the tooth <strong>of</strong> interest. Thus, the fixed<br />
threshold value <strong>for</strong> each tooth in each slice is not effective as<br />
shown in Figure 1. When we try to obtain a tooth region by<br />
thresholding method, the lower and upper limits <strong>of</strong> a threshold<br />
value can be displayed at each slice <strong>for</strong> a given tooth by<br />
the two curves in Fig. 1. Any threshold value within the limit<br />
Figure 1. Threshold values <strong>for</strong> a certain tooth computed at different slices<br />
by manual.<br />
produces the tooth region with the accuracy better than<br />
90%. It shows us that individual segmentation method is<br />
required <strong>for</strong> each tooth in each slice.<br />
There are many segmentation methods, each <strong>of</strong> which<br />
have their own limitations in separating individual tooth<br />
regions on CT images. 3–6 An optimal thresholding scheme 20<br />
can be attempted by taking advantage <strong>of</strong> the fact that the<br />
shape and intensity <strong>of</strong> each tooth changes gradually through<br />
the CT image slices.<br />
However, even if an optimal threshold is determined <strong>for</strong><br />
every slice, the result <strong>of</strong> the segmentation is found unsatisfactory<br />
because <strong>of</strong> neighboring hard tissue. For the 3D reconstruction<br />
<strong>of</strong> an individual tooth model, the tooth boundary<br />
needs to be defined more precisely.<br />
B-Spline Contour Fitting<br />
The B-spline curve has attractive properties <strong>for</strong> the representation<br />
<strong>of</strong> an object contour with arbitrary shape. They are<br />
also suitable <strong>for</strong> the curve fitting process and are summarized<br />
as follows.<br />
• An object <strong>of</strong> any shape, including those subsuming angular<br />
points, can be represented by a set <strong>of</strong> control<br />
points, a knot sequence, and a basis function. The shape<br />
<strong>of</strong> the contour can be adjusted by simply repositioning<br />
the control points in many fitting problems where the<br />
knot sequence and basis function can be fixed.<br />
• Little else remains to be different in the shape <strong>of</strong> the<br />
contour by deducting the number <strong>of</strong> control points<br />
within some tolerable limit <strong>for</strong> the purpose <strong>of</strong> reducing<br />
in<strong>for</strong>mation needed <strong>for</strong> fitting process. This allows the<br />
fitting process to be faster with fewer variables over<br />
which to optimize.<br />
We choose the uni<strong>for</strong>m cubic closed B-spline curve,<br />
shown as follows in Eqs. (1) and (2), to describe the object<br />
contours in the image.<br />
rs = r xs<br />
r y s =<br />
n−1<br />
x i B 0 s − i<br />
i=0<br />
n−1<br />
y i B 0 s − i, 1<br />
<br />
i=0<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 329
Wu et al.: Improved B-spline contour fitting using genetic algorithm <strong>for</strong> the segmentation <strong>of</strong> dental...<br />
B 0 s =s 3 /2 − s 2 + 2/3 if t 0 s t 1 ,<br />
2−s 3 /6 if t 1 s t 2 ,<br />
0 otherwise<br />
In the equations, rs represents the coordinate <strong>of</strong> a contour<br />
pixel at a specific value <strong>of</strong> parameter s and x i ,y i represents<br />
coordinates <strong>of</strong> ith control point. The B-spline basis functions<br />
are translated copies <strong>of</strong> B 0 s. In the case <strong>of</strong> tooth<br />
segmentation we use a closed uni<strong>for</strong>m knot sequence, as<br />
t 0 ,t 1 ,...,t n =0,1,...,n and t 0 =t n where n is the total<br />
number <strong>of</strong> the control points.<br />
The B-spline fitting function f is represented in Eq. (3)<br />
(Ref. 11) as follows:<br />
M−1<br />
f = Irs k ,<br />
k=0<br />
where M is the total number <strong>of</strong> contour points. The fitting<br />
function is maximized when the contour con<strong>for</strong>ms to the<br />
object boundary. The B-spline fitting function makes use <strong>of</strong><br />
only external <strong>for</strong>ce computed based on the gradient magnitude<br />
on the contour. The smoothness constraint is implicitly<br />
represented by the B-spline itself.<br />
B-spline Contour Fitting using Genetic Algorithm<br />
The genetic algorithm is a probabilistic technique <strong>for</strong> searching<br />
<strong>for</strong> an optimal solution. The optimal solution is described<br />
by a vector, called a “chromosome,” which can be<br />
obtained by maximizing a fitting function. Hence the definition<br />
<strong>of</strong> the fitting function significantly affects the solution<br />
state. A sequence <strong>of</strong> evolutionary operations is repeated <strong>for</strong> a<br />
chromosome to evolve to its final state. The end <strong>of</strong> the evolutionary<br />
operation is determined by checking the fitness<br />
values, which represent the goodness <strong>of</strong> each chromosome in<br />
the population.<br />
A chromosome is a collection <strong>of</strong> genes, and a gene represents<br />
the control point <strong>of</strong> B-spline. Since the chromosome<br />
represents a complete contour and a gene uses the actual<br />
location <strong>of</strong> a control point, the search algorithm has neither<br />
ambiguity on the contour location nor potential bias to particular<br />
shapes. To reduce the size <strong>of</strong> a gene, we use the index<br />
value as a gene, instead <strong>of</strong> two coordinate values. 16,17 Composing<br />
a search area based on the indices provides a search<br />
area with arbitrary shape, where it is confined to search <strong>for</strong><br />
the final position <strong>of</strong> the control point to be found out. This<br />
scheme <strong>of</strong> chromosome guarantees that gene in<strong>for</strong>mation<br />
does not spread over the chromosome, which results in short<br />
length and order <strong>of</strong> schema. 16 Accordingly, there is a high<br />
probability to converge fast. A new generation is made<br />
through the sequence <strong>of</strong> evolutionary operations and, during<br />
the evolutionary processes, crossover and mutation steps affect<br />
the quality and speed <strong>of</strong> final solution significantly.<br />
2<br />
3<br />
IMPROVED B-SPLINE CONTOUR FITTING USING<br />
GENETIC ALGORITHM<br />
Fitting Function Based on Gradient Magnitude and<br />
Direction<br />
The fitting function measures the fitness <strong>of</strong> the possible contour<br />
to the object boundary in the current slice. The fitness<br />
value is the basis <strong>for</strong> determining the termination <strong>of</strong> the<br />
evolutionary process and selecting elite chromosomes <strong>for</strong><br />
mating pool generation. In the existing active contour models,<br />
the fitting function consists <strong>of</strong> the internal <strong>for</strong>ces controlling<br />
the smoothness <strong>of</strong> the contour and the external <strong>for</strong>ce<br />
used <strong>for</strong> representing the object boundary in<strong>for</strong>mation in<br />
the image. 7,12 One drawback <strong>of</strong> this representation is that it<br />
requires the determination <strong>of</strong> the weight values balancing<br />
these two components.<br />
B-spline snake makes use <strong>of</strong> a simple fitting function<br />
with only external <strong>for</strong>ce computed based on the gradient<br />
magnitude on the contour. The internal <strong>for</strong>ce terms are replaced<br />
by using a stiffening parameter and implicit smoothness<br />
constraint <strong>of</strong> the B-spline representation <strong>of</strong> a contour.<br />
However, in the image data such as the tooth CT image<br />
slices, those fitting functions <strong>of</strong>ten generate the contour fitted<br />
to the boundary <strong>of</strong> nearby object. They also generate the<br />
contour contracted to a small region unless the stiffening<br />
parameter is set properly.<br />
Note that the magnitude <strong>of</strong> the intensity difference may<br />
vary between the inside and outside <strong>of</strong> an object contour.<br />
However, if the relative intensity between two sides <strong>of</strong> a contour<br />
is maintained throughout the contour, the sign <strong>of</strong> the<br />
intensity difference made by two sides is inverted when the<br />
contour expands out to the boundary <strong>of</strong> another object.<br />
Hence, when fixing moving direction <strong>of</strong> parameter s along<br />
the curve, we are able to have knowledge <strong>of</strong> which side is<br />
inside (or outside) in advance. This enables us to know<br />
whether the contour is fitted to the object <strong>of</strong> interest or other<br />
adjacent objects. In this paper, the fitting function to be<br />
maximized is designed to take advantage <strong>of</strong> this property <strong>of</strong><br />
the data. This gradient direction in<strong>for</strong>mation allows the fitness<br />
function to penalize the portion <strong>of</strong> a contour fitted to<br />
the neighboring object.<br />
To compute the fitness value <strong>for</strong> a possible solution (or<br />
chromosome), we first generate the contour points from the<br />
B-spline representation <strong>of</strong> the solution and trace the contour<br />
as shown in Figure 2(a). At the kth contour point rs k ,a<br />
unit normal vector ns k is computed. Next, the inner region<br />
i<br />
and outer region pixel location p k and p o k , respectively, are<br />
identified by using ns k computed at the kth point rs k <br />
according to<br />
and<br />
p k o = rs k + ns k <br />
p k i = rs k − ns k .<br />
Then, the fitness value is determined based on gradient<br />
magnitude and direction in<strong>for</strong>mation, k ,ateachcontour<br />
point according to<br />
4<br />
5<br />
330 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Wu et al.: Improved B-spline contour fitting using genetic algorithm <strong>for</strong> the segmentation <strong>of</strong> dental...<br />
Figure 2. a Definition <strong>of</strong> inner and outer regions. b Illustration <strong>for</strong><br />
fitting function—right object is <strong>of</strong> interest, with adjacent left object, and<br />
thick black curve is a fitting curve. c Twisted contour.<br />
where<br />
and<br />
k =<br />
M−1<br />
f = k − k ,<br />
k=0<br />
k =Irs k if Ip k i − Ip k o 0,<br />
− Irs k if Ip k i − Ip k o 0,<br />
C, rs k = rs j <br />
, ∀ j 0,1, ... ,M −1 ∧ j k.<br />
0, rs k rs j <br />
Ip k i and Ip k o are intensity values <strong>of</strong> the inside and outside<br />
<strong>of</strong> the kth contour point, respectively. This equation is further<br />
illustrated by Fig. 2(b), where some portion <strong>of</strong> the contour<br />
attaches to another object and in this portion<br />
Ip k i Ip k o , so we assign the negative gradient magnitude<br />
to penalize the fitness value. The figure also shows that in<br />
other portions the contour correctly con<strong>for</strong>ms to the tooth<br />
boundary and in these portions Ip k i Ip k o , so we assign<br />
the positive gradient magnitude to the fitness value. Note<br />
that when there is no difference <strong>of</strong> gradient direction, which<br />
may happen if inner and outer pixel values are identical,<br />
then Ip k i =Ip k o . This aims at preventing the contour from<br />
being misfitted when the contour lies inside an object region<br />
having uni<strong>for</strong>m intensity values, such as the inside region <strong>of</strong><br />
a tooth.<br />
A constant-valued penalty C is deducted from the fitness<br />
value when the contour is twisted as shown in Fig. 2(c).<br />
Our experimental results showed that setting the penalty too<br />
high hindered searching the contour maximizing the sum <strong>of</strong><br />
gradient magnitudes. The proposed fitting method yields the<br />
best per<strong>for</strong>mance when C is set to around 0.1% <strong>of</strong> the sum<br />
<strong>of</strong> gradient magnitudes.<br />
6<br />
Improved Adaptive Evolutionary Operations<br />
The evolutionary process generates a new population <strong>of</strong> possible<br />
solutions through the following three genetic operators:<br />
reproduction (or selection), crossover, and mutation. The<br />
selection operation constructs the mating pool from the current<br />
population <strong>for</strong> the crossover operation. The results presented<br />
here use a tournament selection scheme. 16 The crossover<br />
operation generates two child chromosomes by<br />
swapping genes between the two parent chromosomes. In<br />
this paper we present one point cutting scheme by improved<br />
adaptive crossover probability. We also use an adaptive mutation<br />
probability scheme <strong>for</strong> our evolutionary process.<br />
The conventional GA generally uses fixed crossover and<br />
mutation probabilities. Adaptive genetic algorithm 21 (AGA)<br />
was proposed by Srinivas et al. that uses variable crossover<br />
and mutation probabilities that are determined automatically<br />
based on fitness values during fitting process <strong>for</strong> fast<br />
convergence rate. The probabilities <strong>for</strong> evolution are, there<strong>for</strong>e,<br />
no longer required to be set to constants. At the beginning<br />
stage <strong>of</strong> the fitting process, we consider all the possibilities<br />
<strong>of</strong> control point locations in the search area. As the<br />
process goes on, we obtain the evolutionary probabilities<br />
such that the possible solution near the optimal solution<br />
quickly converges to the actual solution. In AGA, 21 the crossover<br />
probability is adaptively determined depending on the<br />
fitness value f, according to<br />
f best − f<br />
1 , f f avg ,<br />
p c =k f best − f avg<br />
k 2 , f f avg ,<br />
where f best and f avg are the best and average fitness values in<br />
the mating pool, respectively, and k 1 and k 2 are constants<br />
and set to 1.0. Hence, if f=f best when ff avg , f is preserved,<br />
although the value <strong>of</strong> k 1 ensures high occurrence <strong>of</strong> crossover.<br />
If ff avg , crossover is operated without exceptions,<br />
since its corresponding chromosome has low fitness value.<br />
The mutation operation is also implemented by using<br />
the mutation probability p m as follows:<br />
f best − f<br />
3 , f f avg ,<br />
p m =k f best − f avg<br />
k 4 , f f avg ,<br />
where k 3 and k 4 are constants set to 0.5. As in the case <strong>of</strong><br />
crossover, the mutation operation does not affect the chromosome<br />
with the best fitness value. However if ff avg its<br />
mutation operation takes place with the most ambiguity<br />
since k 3 =0.5.<br />
In this paper we propose an improved adaptive crossover<br />
probability. To maintain the solution with high fitness<br />
value, we generate a random number p r and consider the<br />
relationship <strong>of</strong> p r with p c1 and p c2 ,wherep c1 and p c2 denote<br />
crossover probabilities generated from two parent chromosomes,<br />
father chromosome and mother chromosome respectively.<br />
When two parent chromosomes are selected, two children<br />
are generated as follows.<br />
7<br />
8<br />
(1) Generate a random number p r between0and1to<br />
determine the adaptive crossover probability, generate<br />
a random number p l between0and1todetermine<br />
the crossing site, and generate a random<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 331
Wu et al.: Improved B-spline contour fitting using genetic algorithm <strong>for</strong> the segmentation <strong>of</strong> dental...<br />
number p s between 0 and 1 to determine which<br />
side <strong>of</strong> the crossing site the portion <strong>of</strong> the chromosome<br />
should exchange with the corresponding portion<br />
<strong>of</strong> its mate.<br />
(2) Replace f in Eq. (7) by the fitness value <strong>of</strong> each<br />
parent <strong>for</strong> computing the crossover probabilities,<br />
p c1 and p c2 .<br />
(3) If p r p c1 and p r p c2 , put the two parents to the<br />
next generation without change.<br />
(4) If p r is between p c1 and p c2 , thus p c1 p c2 and<br />
p s 0.5 then the left portion <strong>of</strong> the father chromosome<br />
should be exchanged with the corresponding<br />
portion <strong>of</strong> the mother chromosome to generate one<br />
child and put mother chromosome directly to the<br />
generation as another child. If p s 0.5 then the<br />
right portion from the father chromosome should<br />
be exchanged to generate one child and another<br />
child is a copy <strong>of</strong> the mother chromosome. Similarly<br />
if p c1 p c2 then the mother chromosome<br />
should be changed and put to the next generation<br />
while the father chromosome is put to the next<br />
generation without any change. In addition, the<br />
crossover scheme is determined by the value <strong>of</strong> p s .<br />
(5) If p r is less than both p c1 and p c2 , generate two child<br />
chromosomes as the normal crossover method<br />
does.<br />
In the proposed operation, the chromosomes with high<br />
fitness values can survive until a new chromosome with<br />
higher fitness is created. It supports rapid searching <strong>for</strong> an<br />
optimal solution by taking advantage <strong>of</strong> the crossover<br />
scheme swapping either side to the crossing site.<br />
EXPERIMENTAL EVALUATION<br />
We tested the proposed contour segmentation with two<br />
kinds <strong>of</strong> sets <strong>of</strong> data: synthetic images and two sets <strong>of</strong> real<br />
dental CT image sequences with a slice thickness <strong>of</strong> 0.67mm<br />
and 1mm and x-y resolution <strong>of</strong> 512512. Visual C++ with<br />
DICOM libraries 22 <strong>for</strong> reading 16-bit CT images and the 3D<br />
graphics library OpenGL were used as tools to implement<br />
the proposed algorithm. CT images are saved in DICOM<br />
<strong>for</strong>mat, an international standard <strong>for</strong> medical images, after<br />
acquisition through the commercially available Shimadzu<br />
Ltd. SCT-7800 CT scanner. The test data were prepared to<br />
reveal the capability <strong>of</strong> the proposed algorithm in finding an<br />
accurate boundary among many similar objects nearby. We<br />
compared the proposed algorithm with the existing B-spline<br />
snake algorithm that uses the gradient magnitude based external<br />
<strong>for</strong>ce in the fitting function. 11<br />
First, we applied these algorithms to a synthetic image<br />
similar to a tooth surrounded by alveolar bone. To generate<br />
the results, we constructed a B-spline contour with 8 control<br />
points and selected 20 initial chromosomes <strong>for</strong> each<br />
4040 window. For the following examples <strong>of</strong> B-spline<br />
snake the stiffening parameter is set to 2. As shown in Figure<br />
3, the proposed algorithm extracts an accurate object<br />
boundary while the existing B-spline snake fails.<br />
We also applied the two algorithms to real CT image<br />
Figure 3. Contours extracted from the synthetic data number <strong>of</strong> control<br />
points CP=8. a By B-spline snake method. b By the proposed<br />
method.<br />
sequences where an individual tooth <strong>of</strong>ten appears with<br />
neighboring hard tissues such as other teeth and alveolar<br />
bone. If too many control points are used <strong>for</strong> a contour, it<br />
reduces the smoothing effect on the curve and consequently<br />
generates twisted parts <strong>of</strong> contour as shown in Figure 4.<br />
Figure 5 shows part <strong>of</strong> test results using different set <strong>of</strong> slices,<br />
which have lower resolution. Since the test image is small, a<br />
1010 search area suffices <strong>for</strong> a control point.<br />
As shown in Fig. 5, an individual tooth <strong>of</strong>ten appears<br />
with neighboring hard tissues such as other teeth and alveolar<br />
bone, and the proposed algorithm produces better results<br />
than B-spline snake. The difference in the results stems from<br />
the fitting function.<br />
Part <strong>of</strong> the segmentation results <strong>of</strong> slice sequences is<br />
shown in Figure 6 and those <strong>of</strong> a molar having a more<br />
complicated shape are shown in Figure 7. In Fig. 6, the figures<br />
at the far left side show the results <strong>of</strong> teeth initialization<br />
<strong>for</strong> the first slice by applying a proper threshold to each<br />
tooth interactively. As the segmentation is per<strong>for</strong>med slice by<br />
slice, in contrast with the results <strong>of</strong> proposed method, malfitting<br />
error contained in the results <strong>of</strong> the existing method<br />
increases.<br />
332 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Wu et al.: Improved B-spline contour fitting using genetic algorithm <strong>for</strong> the segmentation <strong>of</strong> dental...<br />
Figure 4. Tooth contours extracted from CT image CP=16. a By the<br />
proposed method. b By B-spline snake.<br />
Table I lists part <strong>of</strong> the numerical results <strong>of</strong> the segmentation.<br />
N is the number <strong>of</strong> slices over which each tooth<br />
spans. FPE (false positive error) is the percent <strong>of</strong> area reported<br />
as a tooth by the algorithm, but not by manual segmentation.<br />
FNE (false negative error) is the percent <strong>of</strong> area<br />
reported by manual segmentation, but not by the algorithm.<br />
Similarity and dissimilarity indices, 23,10 which show the<br />
amount <strong>of</strong> agreement and disagreement, S agr and S dis ,respectively,<br />
between the algorithm area A alg and the manual<br />
segmentation area A man , are computed according to<br />
S agr =2 A man A alg<br />
A man + A alg<br />
,<br />
9<br />
S dis =2 A man A alg − A man A alg<br />
A man + A alg<br />
. 10<br />
Figure 5. Tooth contours extracted from CT image sequence CP=8.<br />
a By the proposed method. b By B-spline snake.<br />
These indices are calculated <strong>for</strong> validation on N slices <strong>of</strong><br />
each tooth. Averaged values <strong>of</strong> S agr as well as its minimum<br />
and maximum values are shown in Table I, and we conclude<br />
that the proposed method <strong>for</strong> segmentation isolates individual<br />
region <strong>of</strong> tooth successfully, in contrast with the results<br />
<strong>of</strong> B-spline snake shown in Table II.<br />
The proposed fitting method is designed <strong>for</strong> the fast<br />
contour extraction by the improved crossover method which<br />
uses a random number <strong>for</strong> copying genes <strong>of</strong> a superior chromosome<br />
to an inferior one when the random number falls<br />
into the range <strong>of</strong> crossover probabilities <strong>of</strong> its parents, p c1<br />
and p c2 . Furthermore, the proposed crossover method decides<br />
which part <strong>of</strong> crossing site will be exchanged between<br />
parent chromosomes. The decided part fosters chromo-<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 333
Wu et al.: Improved B-spline contour fitting using genetic algorithm <strong>for</strong> the segmentation <strong>of</strong> dental...<br />
Table I. Segmentation results <strong>for</strong> 8 teeth <strong>of</strong> the proposed method from the same scans<br />
<strong>of</strong> CT set.<br />
Tooth N FPE% FNE% S agr S min S max S dis<br />
Figure 6. Tooth contours extracted from CT image sequence CP=16.<br />
a By the proposed method. b By B-spline snake.<br />
1 20 4.43 8.37 0.935 0.915 0.977 0.131<br />
2 22 7.88 3.45 0.945 0.916 0.973 0.111<br />
3 25 8.96 4.48 0.935 0.901 0.968 0.131<br />
4 24 8.46 6.47 0.926 0.905 0.970 0.148<br />
5 27 5.81 8.29 0.929 0.917 0.967 0.143<br />
6 26 2.07 7.05 0.953 0.923 0.971 0.094<br />
7 25 5.21 3.79 0.955 0.927 0.976 0.089<br />
8 23 5.69 1.42 0.965 0.932 0.983 0.069<br />
Table II. Segmentation results <strong>for</strong> 8 teeth <strong>of</strong> B-spline snake from the same scans <strong>of</strong> CT<br />
set.<br />
Tooth N FPE% FNE% S agr S min S max S dis<br />
1 20 6.12 27.21 0.814 0.574 0.952 0.373<br />
2 22 26.01 1.16 0.879 0.628 0.956 0.241<br />
3 25 45.86 11.28 0.756 0.316 0.897 0.487<br />
4 24 29.89 4.59 0.842 0.764 0.941 0.313<br />
5 27 28.06 8.06 0.836 0.726 0.933 0.328<br />
6 26 15.09 8.81 0.884 0.818 0.948 0.232<br />
7 25 27.98 5.03 0.852 0.755 0.936 0.296<br />
8 23 10.12 3.89 0.932 0.771 0.972 0.136<br />
Figure 8. Comparison <strong>of</strong> convergence rates.<br />
Figure 7. Extracted contours <strong>of</strong> molar CP=32. a By the proposed<br />
method. b By B-spline snake.<br />
somes to be competent with a high fitness value. We implement<br />
two genetic B-spline fittings with existing crossover<br />
methods to analyze the per<strong>for</strong>mance <strong>of</strong> the proposed crossover.<br />
Both existing methods generate the initial population<br />
randomly, with uni<strong>for</strong>m distribution, while using different<br />
crossover methods. “Method A” uses a fixed p c <strong>of</strong> 0.75 and<br />
“Method B” uses AGA, which determines p c adaptively. Figure<br />
8 compares the convergence rate <strong>of</strong> the proposed crossover<br />
method with those <strong>of</strong> the existing methods in terms <strong>of</strong><br />
the fitness value along chromosome generation. The figure<br />
shows that the proposed crossover method results in a better<br />
334 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Wu et al.: Improved B-spline contour fitting using genetic algorithm <strong>for</strong> the segmentation <strong>of</strong> dental...<br />
Figure 10. Manipulation <strong>of</strong> tooth. a Every tooth can be manipulated.<br />
b Simulation <strong>of</strong> having tooth out.<br />
Figure 9. Wireframe models <strong>of</strong> tooth and mandible. a 3D reconstruction<br />
<strong>of</strong> tooth. b 3D reconstruction <strong>of</strong> mandible.<br />
convergence rate than either method A or B. The proposed<br />
crossover method preserves the chromosomes with high fitness<br />
<strong>for</strong> fast convergence and the results shows it is effective<br />
to randomly select either side to crossing site <strong>for</strong> improved<br />
crossover operation.<br />
Individual segmentation <strong>of</strong> all teeth can be used to reconstruct<br />
a model <strong>of</strong> the mandible, as shown in Figures 9<br />
and 10. Every tooth can be separated from the jaw <strong>for</strong> simulation<br />
<strong>of</strong> dental treatments.<br />
CONCLUSIONS<br />
In this paper, we presented the improved genetic B-spline<br />
curve fitting algorithm <strong>for</strong> extracting individual smooth<br />
tooth contours from CT slices while preventing the contour<br />
from being twisted. This enables us to obtain individual accurate<br />
contours <strong>of</strong> teeth by overcoming the problem <strong>of</strong> the<br />
contour <strong>of</strong> a tooth expanding out to other teeth boundaries<br />
in the fitting process. Furthermore, we also devised the<br />
crossover method which accelerates convergence rate by<br />
means <strong>of</strong> both conserving chromosomes with high fitness<br />
value and allowing exchange <strong>of</strong> either side <strong>of</strong> cross site. The<br />
test results show that the proposed segmentation algorithm<br />
successfully extracts a smooth tooth contour under specific<br />
conditions such as the existence <strong>of</strong> objects in close vicinity.<br />
This paper also demonstrated the possibility <strong>of</strong> reconstruction<br />
<strong>of</strong> a 3D model in which each tooth was modeled<br />
and manipulated separately <strong>for</strong> the simulation <strong>of</strong> dental surgery.<br />
These anatomical 3D models can be used <strong>for</strong> facilitating<br />
diagnoses, pre-operative planning and prosthesis design.<br />
They will provide radiography <strong>of</strong> the mandible, an accurate<br />
mechanical model <strong>of</strong> the individual tooth and that <strong>of</strong> its root<br />
<strong>for</strong> endodontics and orthodontic operations. Hence the 3D<br />
reconstruction <strong>of</strong> the teeth can be used in virtual reality<br />
based training <strong>for</strong> orthodontics students and <strong>for</strong> preoperatory<br />
assessment by dental surgeons.<br />
ACKNOWLEDGMENTS<br />
This research was supported by the MIC (Ministry <strong>of</strong> In<strong>for</strong>mation<br />
and Communication), Korea, under the ITRC (In<strong>for</strong>mation<br />
Technology Research Center) support program<br />
supervised by the IITA (Institute <strong>of</strong> In<strong>for</strong>mation Technology<br />
Advancement) (IITA-2006–C1090–0602–0002). The authors<br />
are grateful to K. Blankenship and Y. Blankenship <strong>for</strong> their<br />
ef<strong>for</strong>t in pro<strong>of</strong>reading this paper.<br />
REFERENCES<br />
1 T. Kondo, S. H. Ong, and K. W. C. Foong, “Tooth segmentation <strong>of</strong><br />
dental study models using range images”, IEEE Trans. Med. <strong>Imaging</strong> 23,<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 335
Wu et al.: Improved B-spline contour fitting using genetic algorithm <strong>for</strong> the segmentation <strong>of</strong> dental...<br />
350–362 (2004).<br />
2 E. H. Said, D. E. M. Nassar, G. Fahmy, and H. H. Ammar, “Teeth<br />
segmentation in digitized dental X-ray films using mathematical<br />
morphology”, IEEE Trans. Inf. Forensics Security 1, 178–189 (2006).<br />
3 J. H. Ryu, H. S. Kim, and K. H. Lee, “Contour based algorithms <strong>for</strong><br />
generating 3D CAD models from medical images”, Int. J. Adv. Manuf.<br />
Technol. 24, 112–119 (2004).<br />
4 A. G. Bors, L. Kechagias, and I. Pitas, “Binary Morphological Shape-<br />
Based Interpolation Applied to 3-D Tooth Reconstruction”, IEEE Trans.<br />
Med. <strong>Imaging</strong> 21, 100–108 (2002).<br />
5 G. Bohm, C. Knoll, V. G. Colomer, M. Alcaniz-Raya, and S. Albalat,<br />
“Three-dimensional segmentation <strong>of</strong> bone structures in CT images”,<br />
Proc. SPIE 3661, 277–286 (1999).<br />
6 S. Liu and W. Ma, “Seed-growing segmentation <strong>of</strong> 3D surfaces from<br />
CT-contour data”, Computer-Aided Design 31, 517–536 (1999).<br />
7 M. Kass, A. Witkin, and D. Terzopoulos, “Snakes: Active contour<br />
models”, Int. J. Comput. Vis. 1, 321–331 (1988).<br />
8 C. Han and W. S. Kerwin, “Detecting objects in image sequences using<br />
rule-based control in an active contour model”, IEEE Trans. Biomed.<br />
Eng. 50, 705–710 (2003).<br />
9 A. Klemencic, S. Kovacic, and F. Pernus, “Automated segmentation <strong>of</strong><br />
muscle fiber images using active contour models”, Cytometry 32,<br />
317–326 (1998).<br />
10 J. Klemencic, V. Valencic, and N. Pecaric, “De<strong>for</strong>mable contour based<br />
algorithm <strong>for</strong> segmentation <strong>of</strong> the hippocampus from MRI”, in Proc. 9th<br />
International Conference on Computer Analysis <strong>of</strong> Images and Patterns,<br />
Lect. Notes Comput. Sci. 2124, 298–308 (2001).<br />
11 P. Brigger, J. Hoeg, and M. Unser, “B-Spline snakes: A flexible tool <strong>for</strong><br />
parametric contour detection”, IEEE Trans. Image Process. 9, 1484–1496<br />
(2000).<br />
12 S. Menet, P. Saint-Marc, and G. Medioni, “Active contour models:<br />
overview, implementation and applications. Systems”, in Proc. Man and<br />
Cybernetics (IEEE Press, Piscataway, NJ, 1990), pp. 194–199.<br />
13 G. Farin, Curves and surfaces <strong>for</strong> CAGD, 4th ed. (Academic Press, New<br />
York, 1996), pp. 141–168.<br />
14 M. Cerrolaza, W. Annicchiarico, and M. Martinez, “Optimization <strong>of</strong> 2D<br />
boundary element models using B-splines and genetic algorithms”, in<br />
Engineering Analysis with Boundary Elements (Elsevier, Ox<strong>for</strong>d, 2000),<br />
Vol. 24, pp. 427–440.<br />
15 W. Annicchiarico and M. Cerrolaza, “Finite elements, genetic algorithms<br />
and B-splines: a combined technique <strong>for</strong> shape optimization”, Finite<br />
Elem. Anal. Design 33, 125–141 (1999).<br />
16 C. Ooi and P. Liatsis, “Co-evolutionary-based active contour models in<br />
tracking <strong>of</strong> moving obstacles”, Proc. International Conference on<br />
Advanced Driver Assistance Systems (IEEE, London, 2001), pp. 58–62.<br />
17 L. A. MacEachern and T. Manku, “Genetic algorithms <strong>for</strong> active contour<br />
optimization”, Proc. IEEE Int. Sym. <strong>for</strong> Circuits and Systems (IEEE Press,<br />
Piscataway, NJ, 1998), pp. 229–232.<br />
18 M.-S. Dao, F. G. B. De Natale, and A. Massa, “Edge Potential Functions<br />
(EPF) and Genetic Algorithms (GA) <strong>for</strong> Edge-Based Matching <strong>of</strong> Visual<br />
Objects”, IEEE Trans. Multimedia 9, 120–135 (2007).<br />
19 L. Ballerni and L. Bocchi, “Multiple Genetic Snakes <strong>for</strong> Bone<br />
Segmentation”, in Proc. EvoWorkshops, Lect. Notes Comput. Sci. 2611,<br />
346–356 (2003).<br />
20 R. C. Gonzalez and R. E. Woods, Digital Image Processing (Addison<br />
Wesley, Reading, MA, 1993), pp. 447–455.<br />
21 M. Srinivas and L. M. Patnaik, “Adaptive probabilities <strong>of</strong> crossover and<br />
mutation in genetic algorithms”, IEEE Trans. Syst. Man Cybern. 24,<br />
656–667 (1994).<br />
22 OFFIS Institute <strong>for</strong> In<strong>for</strong>mation Technology website, http://www.<strong>of</strong>fis.de/<br />
index-e.php/, accessed October 2006.<br />
23 A. P. Zijdenbos, B. M. Dawant, R. A. Margolin, and A. C. Palmer,<br />
“Morphometric analysis <strong>of</strong> white matter lesions in MR images: Method<br />
and validation”, IEEE Trans. Med. <strong>Imaging</strong> 13, 716–724 (1994).<br />
336 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
<strong>Journal</strong> <strong>of</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology® 51(4): 337–347, 2007.<br />
© <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology 2007<br />
Colorimetric Characterization Model <strong>for</strong> Plasma Display<br />
Panel<br />
Seo Young Choi, Ming Ronnier Luo and Peter Andrew Rhodes<br />
Department <strong>of</strong> Color & Polymer Chemistry, University <strong>of</strong> Leeds, Leeds, United Kingdom LS2 9JT<br />
E-mail: seoyoung228@googlemail.com<br />
EunGiHeoandImSuChoi<br />
PDP Division, Samsung SDI, 508 Sungsung-Dong, Chonan City, Chungchongnam-Do 330–300, South Korea<br />
Abstract. This paper describes a new device characterization<br />
model applicable to plasma display panels (PDP). PDPs are inherently<br />
dissimilar to cathode ray tube and liquid crystal display devices,<br />
and so new techniques are needed to model their color characteristics.<br />
The intrinsic properties and distinct colorimetric<br />
characteristics are first introduced followed by model development.<br />
It was found that there was a large deviation in colorimetric additivity<br />
and a variation in color due to differences in the number <strong>of</strong> pixels in<br />
a color patch (pattern size). Three colorimetric characterization<br />
models, which define the relationship between the number <strong>of</strong> sustain<br />
pulses and CIE XYZ values, were successfully derived <strong>for</strong> full<br />
pattern size: A three-dimensional lookup table (3D-LUT) model, a<br />
single-step polynomial model and a two-step polynomial model including<br />
three 1D LUTs with a trans<strong>for</strong>mation matrix. The single-step<br />
and two-step polynomial models having more than 8 terms and the<br />
3D LUT model produced the most accurate results. However, the<br />
single-step polynomial model was selected and extended to other<br />
pattern sizes because <strong>of</strong> its simplicity and good per<strong>for</strong>mance. Finally,<br />
a comprehensive model was proposed which can predict CIE<br />
XYZ at sizes different to that used <strong>for</strong> the training set. It was found<br />
that one combined training set <strong>for</strong>med using two different pattern<br />
sizes could give better results than a single-size training set.<br />
© 2007 <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology.<br />
DOI: 10.2352/J.<strong>Imaging</strong>Sci.Technol.200751:4337<br />
INTRODUCTION<br />
Large size displays such as plasma display panel (PDP), liquid<br />
crystal display (LCD), and digital light processing (DLP)<br />
TV are promising candidates <strong>for</strong> replacing the cathode ray<br />
tube (CRT) displays that are currently in widespread use.<br />
Plasma display technology has the following characteristics:<br />
it is thin and light, has superior video image per<strong>for</strong>mance<br />
and uses a large screen size from 42 to 100 inches. These<br />
enable PDPs to be used inside stores <strong>for</strong> product promotion<br />
and, increasingly, <strong>for</strong> home theater. One <strong>of</strong> the desirable<br />
properties <strong>for</strong> large displays is to achieve a “lifelike” appearance<br />
under a range <strong>of</strong> practical viewing conditions as judged<br />
in terms <strong>of</strong> color and image quality. It is there<strong>for</strong>e important<br />
to investigate its colorimetric behavior and to make a characterization<br />
model based on the intrinsic physical properties<br />
<strong>of</strong> a PDP. Already, much research has been conducted to<br />
investigate the use <strong>of</strong> LCD and CRT technology, however<br />
relatively little work has been per<strong>for</strong>med <strong>for</strong> plasma<br />
displays. 1–5 Only one paper deals with PDP characterization<br />
based on the gain-<strong>of</strong>fset-gamma (GOG) model; (previously<br />
developed <strong>for</strong> CRT) at one pattern size, however the physical<br />
properties <strong>of</strong> PDPs and the pattern-size effect were not considered<br />
in this model. 6 In other words, the model could not<br />
be successfully extended to different pattern sizes. The International<br />
Electrotechnical Commission (IEC) has issued a<br />
standard IEC 61966-5 which includes methods and parameters<br />
<strong>for</strong> investigating the use <strong>of</strong> PDPs to display colored<br />
images in multimedia applications. 7 It does include the pattern<br />
size effect as a display area ratio characteristic, but<br />
changes in other characteristics such as color gamut due to<br />
pattern size were not considered. It assumes that a PDP’s<br />
RGB channels are independent. Un<strong>for</strong>tunately, PDPs typically<br />
exhibit significant additivity failure when compared to<br />
other display technologies. It is there<strong>for</strong>e essential that this<br />
additivity failure should be taken into account during the<br />
development <strong>of</strong> any device characterization model.<br />
The simplified structure <strong>of</strong> a PDP RGB cell is shown in<br />
Figure 1. A PDP is composed <strong>of</strong> two glass plates having a<br />
100 m gap and filled with a rare gas mixture which can<br />
<br />
IS&T member.<br />
Received Jul. 27, 2006; accepted <strong>for</strong> publication Mar. 1, 2007.<br />
1062-3701/2007/514/337/11/$20.00.<br />
Figure 1. The structure <strong>of</strong> a PDP’s RGB cells.<br />
337
Choi et al.: Colorimetric characterization model <strong>for</strong> plasma display panel<br />
and resultant CIE XYZ values at one particular pattern size.<br />
In addition, one <strong>of</strong> these models was extended to take into<br />
account other pattern sizes.<br />
Figure 2. A 16.7 ms frame includes 8 subfields. The black boxes are the<br />
durations <strong>of</strong> the sustain periods proportional to 1,2,4,…,128.<br />
include a 500 torr, Xe–Ne or Xe–Ne–He mixture which,<br />
when excited, results in the Xe atoms emitting vacuum ultraviolet<br />
(vuv) radiation at 147 and 173 nm, respectively.<br />
This vuv radiation then excites the red, green, and blue<br />
phosphors located on the rear glass plate. The discharge also<br />
emits red-orange visible light due to neon, which causes a<br />
subsequent reduction in color purity (see the Colorimetric<br />
characteristics <strong>of</strong> a PDP section). Each pixel has three individual<br />
RGB microdischarge cells. An alternating current<br />
(ac) is generated by dielectric barrier discharge operating in<br />
a glow regime to generate plasma in each cell. The ac voltage<br />
is approximately rectangular with a frequency in the order <strong>of</strong><br />
100 kHz and a rise time <strong>of</strong> about 200–300 ns. 8 Different<br />
intensity levels are obtained via the modulation <strong>of</strong> the number<br />
<strong>of</strong> ac pulses (sustain pulses) in a discharge cell. For CRT<br />
and LCD, intensity levels are controlled differently from a<br />
PDP, i.e., according to voltage level. In addition, the luminance<br />
output <strong>of</strong> a PDP is dependent on the pattern size<br />
displayed, even when the RGB input values remain the same.<br />
The average level <strong>of</strong> input video signal—a product <strong>of</strong> RGB<br />
input values and pattern size—increases in proportion to the<br />
increase in pattern size. This is also accompanied by an increase<br />
in power consumption. As a result, there is a need to<br />
regulate the power consumed <strong>for</strong> large area patterns by<br />
means <strong>of</strong> the automatic power control (APC) function. Specifically,<br />
this regulates power consumption to within a certain<br />
upper limit. Moreover, luminance output is affected by<br />
the APC function and generates different values dependent<br />
on pattern size.<br />
This paper describes an investigation into the colorimetric<br />
characteristics <strong>of</strong> a PDP and the development <strong>of</strong> three<br />
device characterization models which describe the relationship<br />
between the number <strong>of</strong> sustain pulses <strong>of</strong> RGB input<br />
PHYSICAL PROPERTIES OF A PDP<br />
Overall Transfer Process <strong>of</strong> the Input Video Signal<br />
As mentioned earlier, one unique feature <strong>of</strong> PDP is that, <strong>for</strong><br />
afixedRGB input, its luminance output varies according to<br />
the pattern size displayed. The average level <strong>of</strong> an input<br />
video signal increases not only in proportion to the RGB<br />
input but also to the increase in pattern size. Furthermore,<br />
power demand also grows, because a larger input video signal<br />
leads to a bigger discharge current in the PDP. To protect<br />
the electronic components from damage, it is necessary to<br />
impose a limit on power consumption. This is accomplished<br />
by controlling the discharge current. There are two methods<br />
<strong>for</strong> controlling this: adjusting the number <strong>of</strong> RGB sustain<br />
pulses and modifying the input level <strong>of</strong> the video signal. The<br />
PDP used in this study adopts the first method. The number<br />
<strong>of</strong> RGB sustain pulses is adjusted through the APC function,<br />
which is determined by the average intensity level <strong>of</strong> the<br />
input video signal. To explain the role <strong>of</strong> the APC function<br />
here, it is assumed that each discharge cell can display 256<br />
gray levels. Unlike a CRT, each cell is only capable <strong>of</strong> being<br />
turned on or <strong>of</strong>f (binary). Each gray level is obtained by<br />
modulating the number <strong>of</strong> sustain pulses during one frame<br />
(16.7 ms, 60 frames per second=60 Hz). A frame is divided<br />
into eight subfields, having weight ratios <strong>of</strong> 1, 2, 4…,128<br />
(Figure 2). The function <strong>of</strong> a subfield is to modulate light<br />
output over time. This is accomplished by dividing each<br />
video frame into shorter time periods where each cell is<br />
either turned on or <strong>of</strong>f. Each subfield has a sustain period<br />
(see black box in Fig. 2) whose duration is proportional to<br />
weight ratios, and an address period (see white box in Fig. 2)<br />
whose duration is the same <strong>for</strong> 8 subfields. The address periodisusedtoswitchonor<strong>of</strong>fagivencell.An8-bit<br />
binary<br />
coding is used to obtain 256 gray levels since there are 256<br />
possible levels that can be achieved by assigning on/<strong>of</strong>f to<br />
any combination <strong>of</strong> the eight subfields. In practice, the number<br />
<strong>of</strong> sustain pulses is determined by the sum <strong>of</strong> the product<br />
<strong>of</strong> the sustain pulse limit and the “weight ratios” which<br />
correspond to the proportion <strong>of</strong> subfields turned on. This<br />
calculation is shown in Table I. The sustain pulse limit, as<br />
mentioned previously, safeguards the display’s electronics. It<br />
Table I. One example <strong>of</strong> subfield configuration and the calculation process <strong>for</strong> the number <strong>of</strong> sustain pulses used <strong>for</strong> a color patch with input value <strong>of</strong> 5.<br />
Subfield 1 2 3 4 5 6 7 8<br />
Weight ratios 0.004<br />
1/255<br />
0.008 0.016 0.031 0.063 0.125 0.251 0.502<br />
128/255<br />
Configuration a 1 0 1 0 0 0 0 0 8 bits<br />
The sustain pulse limit, 2600, is given in the APC table defined by the manufacturer<br />
Calculation<br />
26000.004+26000.016=52<br />
the practical number <strong>of</strong> sustain pulses assigned to RGB cells<br />
a Binary coding: 0 is <strong>of</strong>f and 1 is on. This configuration corresponds to input value 5.<br />
Weight<br />
ratios’<br />
sum=1<br />
338 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Choi et al.: Colorimetric characterization model <strong>for</strong> plasma display panel<br />
Figure 3. Flowchart explaining the trans<strong>for</strong>mation <strong>of</strong> input video signal to<br />
the emission <strong>of</strong> light in a PDP.<br />
Figure 4. Plot <strong>of</strong> CIE X values versus the number <strong>of</strong> red sustain pulses at 4<br />
pattern sizes. Points 1, 2, and3 correspond to the maximum number<br />
<strong>of</strong> sustain pulses at 100%, 60%, and 30% pattern size, respectively.<br />
Table II. Maximum CIE XYZ <strong>for</strong> RGB and the range <strong>of</strong> the number <strong>of</strong> sustain pulses at 4, 30, 60, and 100% pattern size.<br />
Color CIE XYZ 4% 30% 60% 100%<br />
Red Max X 498.1 454.5 330.5 230.4<br />
Green Max Y 593.9 546.2 395.7 284.7<br />
Blue Max Z 1311.6 1175.6 809.9 551.7<br />
Sustain pulse range 0–2594 0–2594 0–1826 0–1260<br />
is determined from a manufacturer-defined APC table according<br />
to the average input video signal. Eight subfield<br />
combinations yield 256 gray levels corresponding to the supplied<br />
input values; however, the actual number <strong>of</strong> sustain<br />
pulses (and hence the brightness <strong>of</strong> each level) is controlled<br />
by the sustain pulse limit.<br />
The overall transfer process <strong>of</strong> input video signal to<br />
output stimulus can be expressed in terms <strong>of</strong> the steps<br />
shown in Figure 3, which includes an example <strong>for</strong> a full<br />
white pattern. RGB input values <strong>for</strong> the video signal are<br />
trans<strong>for</strong>med into the number <strong>of</strong> RGB sustain pulses via the<br />
PDP’s logic board. These are calculated from the sub-field<br />
configuration corresponding to the RGB input value and the<br />
sustain pulse limit assigned by the APC table. The number<br />
<strong>of</strong> sustain pulses are the same as the number <strong>of</strong> plasma<br />
discharges occurring in each cell. A succession <strong>of</strong> discharge<br />
pulses occurs between two sustain electrodes inside the front<br />
plate <strong>of</strong> the RGB cells according to the number <strong>of</strong> RGB<br />
sustain pulses assigned. The rare gas mixture (Xe–Ne or<br />
Xe–Ne–He) emits vacuum ultraviolet (vuv) photons at 147<br />
and 173 nm during discharge in RGB cells and those intensities<br />
are governed by the number <strong>of</strong> RGB sustain pulses.<br />
Xenon is used as a vuv emitter and neon acts as a buffer gas<br />
which lowers the breakdown voltage, i.e., the minimum voltage<br />
to initiate plasma. The vuv photons are converted into<br />
visible photons by the phosphor materials deposited on the<br />
inner walls <strong>of</strong> RGB discharge cells. Based on an understanding<br />
<strong>of</strong> this process, the final characterization model was developed<br />
between CIE XYZ values and the number <strong>of</strong> RGB<br />
sustain pulses (rather than simply RGB input values).<br />
Pattern Size Effect<br />
As mentioned in the Overall Transfer Process <strong>of</strong> the Input<br />
Video Signal section brightness varies according to pattern<br />
size. Figure 4 depicts CIE X values plotted against the normalized<br />
number <strong>of</strong> R sustain pulses. In the figure, there are<br />
four sets <strong>of</strong> data points corresponding to pattern sizes <strong>of</strong> 4%,<br />
30%, 60%, and 100% respectively. Each set includes 26 equal<br />
steps <strong>of</strong> the red channel. The decrease <strong>of</strong> maximum X value<br />
with increasing size shown in Fig. 4 can be explained due to<br />
an increased power demand by larger pattern sizes. To limit<br />
the total power consumption <strong>of</strong> a PDP, the number <strong>of</strong> sustain<br />
pulses must be lowered <strong>for</strong> larger color patches by the<br />
APC function. This results in the decrease in maximum X<br />
value with increasing size shown in Fig. 4. Table II illustrates<br />
also the size effect on color patches at 4% and 30% <strong>of</strong> the<br />
total screen area. It can be seen that they have a higher<br />
maximum range <strong>of</strong> the number <strong>of</strong> sustain pulses which result<br />
in larger CIE XYZ values than those <strong>for</strong> the 60% and<br />
100% pattern sizes. The highest number <strong>of</strong> sustain pulses <strong>for</strong><br />
the R primary color at 100% pattern size is 1260 while the<br />
4% pattern size is 2594, even though the input RGB values<br />
are the same (0,255,0). Hence, the resultant maximum Y<br />
value <strong>for</strong> the 4% pattern size is higher than <strong>for</strong> the 100%<br />
pattern size. This implies that the number <strong>of</strong> sustain pulses<br />
should be used as an input color specification, contrary to<br />
conventional approaches to display characterization which<br />
only consider the input digital RGB values.<br />
EXPERIMENTAL METHODS<br />
The colorimetric characterization <strong>of</strong> a 42-inch Samsung<br />
high-definition PDP (model PPM42H3) was evaluated. Its<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 339
Choi et al.: Colorimetric characterization model <strong>for</strong> plasma display panel<br />
Table III. a Color patches <strong>for</strong> three characterization models at 100% pattern size.<br />
b Color patches used <strong>for</strong> developing single-step polynomial models at other pattern<br />
sizes.<br />
a<br />
Models<br />
at 100%<br />
pattern<br />
3D<br />
LUT<br />
Single-step<br />
polynomial<br />
model<br />
Three<br />
1D LUT<br />
Two-step<br />
polynomial model<br />
Trans<strong>for</strong>mation<br />
matrix<br />
Training<br />
sets<br />
Testing<br />
set<br />
6-<br />
level<br />
3D LUT<br />
6-, 5-, 4-<br />
& 3-level<br />
3D LUT<br />
52 steps<br />
<strong>of</strong> RGB<br />
115-color test set<br />
6-, 5-, 4-<br />
& 3-level<br />
3D LUT<br />
b<br />
4%, 30%, and 60%<br />
Pattern<br />
size<br />
Training<br />
sets<br />
Test sets<br />
5-, 4-, and 3-level 3D LUT<br />
at each pattern size<br />
Three 27-color test sets<br />
at three pattern sizes<br />
pixel resolution is 1024768 with an aspect ratio <strong>of</strong> 16:9<br />
and it is capable <strong>of</strong> addressing 512 intensity levels per channel,<br />
although only 256 were used here. Its pixel pitch is<br />
0.912 mm H0.693 mm V so that the total display area<br />
is 933.89532.22 mm 2 . Color patches were displayed in the<br />
middle <strong>of</strong> the screen and generated by a computercontrolled<br />
graphic card equipped with digital visual interface<br />
(DVI) output. This allows the PDP’s logic board to receive a<br />
digital signal directly from the computer.<br />
Contrary to typical display characterization, the number<br />
<strong>of</strong> sustain pulses was used as input color specification in this<br />
study, as explained previously in the Pattern Size Effect section.<br />
Color measurements were taken using a Minolta<br />
CS-1000 tele-spectroradiometer (TSR) in a dark room. The<br />
repeatability <strong>of</strong> both the TSR and PDP was evaluated using<br />
15 colors measured twice over a two-month period. The<br />
*<br />
median and maximum E ab were 0.38 and 1.18 during this<br />
time. This per<strong>for</strong>mance is considered satisfactory. Measurement<br />
patches consisted <strong>of</strong> rectangles which were 100%, 80%,<br />
60%, 45%, 30%, or 4% <strong>of</strong> the display area. The background<br />
was set to black (except <strong>for</strong> the 100% case).<br />
Three characterization models were first developed <strong>for</strong><br />
the 100% pattern size. Table III(a) describes the data set<br />
generated <strong>for</strong> the three types <strong>of</strong> characterization models: 3D<br />
LUT, single-step polynomial, and two-step polynomial<br />
model. For the 3D LUT model, 6 levels were used. For the<br />
other two models, 6-, 5-, 4-, and 3-level 3D LUTs were compared<br />
in order to determine which training set gave the optimum<br />
per<strong>for</strong>mance with the fewest number <strong>of</strong> measurements.<br />
The 6-level 3D LUT was first generated using 1, 15,<br />
43, 66, 105, and 255 digital input values <strong>for</strong> each <strong>of</strong> the RGB<br />
channels. These values were empirically determined to have<br />
approximately uni<strong>for</strong>m coverage <strong>of</strong> the CIE XYZ destination<br />
color space. The distribution <strong>of</strong> the 6-level training set is<br />
Figure 5. Plot <strong>of</strong> 216 colors <strong>of</strong> the 6-level 3D LUT in a XY, b YZ, and<br />
c a * b * plane, respectively.<br />
shown as an XY and YZ projection in Figures 5(a) and 5(b),<br />
respectively. In addition, an a * b * diagram depicting the<br />
whole <strong>of</strong> the 6-level training set can be seen in Fig. 5(c).<br />
Among the 6 digital input values, 1, 15, 43, 105, and 255<br />
were used to make the 5-level 3D LUT; 1, 15, 66, and 255<br />
340 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Choi et al.: Colorimetric characterization model <strong>for</strong> plasma display panel<br />
were used <strong>for</strong> the 4-level 3D LUT; and 1, 43, and 255 <strong>for</strong> the<br />
3-level 3D LUT.<br />
Another model is called the “two-step polynomial”<br />
model which includes three 1D-LUTs between the normalized<br />
RGB luminance values and the number <strong>of</strong> RGB sustain<br />
pulses. This is followed by a trans<strong>for</strong>mation from normalized<br />
RGB luminance values to XYZ values via trans<strong>for</strong>mation<br />
matrix. Three 1D-LUTs were created <strong>for</strong> each <strong>of</strong> the<br />
RGB channels including 52 equal steps in RGB space. Linear<br />
interpolation was then used to predict the normalized luminance<br />
values between the data points. Six trans<strong>for</strong>mation<br />
matrices were used. One <strong>of</strong> them was the primary matrix<br />
obtained by measurements <strong>of</strong> RGB primary colors. This matrix<br />
can be used <strong>for</strong> the ideal case where there is little interaction<br />
among RGB channels and little unwanted emission.<br />
<strong>Additional</strong>ly, five trans<strong>for</strong>mation matrices <strong>for</strong> nonideal cases,<br />
were derived using polynomial regression between the measured<br />
XYZ values <strong>for</strong> 6-, 5-, 4-, and 3-level 3D LUTs and<br />
their corresponding normalized RGB luminance values. The<br />
cross terms RG, RB, GB, and RGB were included in the<br />
matrix to compensate <strong>for</strong> cross-channel interaction. The<br />
square terms R 2 , G 2 , and B 2 were also included, although<br />
these terms have no particular physical meaning. These<br />
terms were chosen because they were included in some previous<br />
display characterization studies. 1,9<br />
The per<strong>for</strong>mance <strong>of</strong> the three models was evaluated using<br />
three test sets. The first set includes 444 bright<br />
color patches that were chosen to correspond to L * values <strong>of</strong><br />
45, 85, 95, and 99 <strong>for</strong> each <strong>of</strong> the RGB channels. Two additional<br />
sets, including 24 colors L * 40 and 27 colors composed<br />
<strong>of</strong> three L * values (20, 60, and 90), were also added to<br />
verify model per<strong>for</strong>mance. The three sets were merged to<br />
<strong>for</strong>m a combined test set <strong>of</strong> 115 colors. The color difference<br />
E * ab between the predicted and measured values <strong>for</strong> these<br />
test colors was calculated to evaluate the accuracy <strong>of</strong> characterization<br />
models. All measured tristimulus values were corrected<br />
by subtracting those <strong>of</strong> the black level.<br />
A subsequent experiment was carried out to investigate<br />
model per<strong>for</strong>mance <strong>for</strong> different pattern sizes. Only the<br />
single-step polynomial model was further developed in this<br />
experiment. Three training sets at each pattern size were<br />
used to generate the 3D LUTs [see Table III(b)]. Per<strong>for</strong>mance<br />
was then evaluated by measuring 27 test colors consisting<br />
<strong>of</strong> combinations <strong>of</strong> three input levels producing L *<br />
values <strong>of</strong> 20, 60, and 90 <strong>for</strong> each channel.<br />
Table IV gives the terms used to develop the single-step<br />
model and the trans<strong>for</strong>mation matrices <strong>of</strong> the two-step<br />
model. The polynomial coefficients were computed from experimental<br />
data consisting <strong>of</strong> 216, 125, 64, or 27 colors measured<br />
from the 6-, 5-, 4-, and 3-level 3D LUTs, respectively.<br />
Each sample includes a set <strong>of</strong> RGB sustain numbers and<br />
their corresponding XYZ values. In the two-step model’s<br />
trans<strong>for</strong>mation matrix, a polynomial relationship was determined<br />
between the normalized RGB luminances and XYZ<br />
values. All calculations were executed using Matlab.<br />
Table IV. Detailed description <strong>of</strong> the terms used in the single- and two-step polynomial<br />
models.<br />
The matrices consisting <strong>of</strong><br />
various polynomial coefficients<br />
Independent variables<br />
33 R, G, B<br />
34 R, G, B,1<br />
35 R, G, B, RGB,1<br />
38 R, G, B, RG, RB, GB, RGB,1<br />
311 R, G, B, R 2 , G 2 , B 2 , RG, RB, GB, RGB,1<br />
320 311 plus R 3 , G 3 , B 3 , R 2 G, R 2 B, G 2 R, G 2 B,<br />
B 2 R, B 2 G<br />
335 320 plus R 3 G, R 3 B, G 3 R, G 3 B, B 3 R, B 3 G,<br />
R 2 GB, RG 2 B, RGB 2 , R 4 , G 4 , B 4 , R 2 G 2 , R 2 B 2 , G 2 B 2<br />
COLORIMETRIC CHARACTERISTICS OF A PDP<br />
Spectral Characteristics<br />
The spectral power distributions <strong>of</strong> the maximum intensity<br />
RGB primaries are shown in Figure 6(a). Two kinds <strong>of</strong> green<br />
phosphors were used <strong>for</strong> boosting luminance and stabilizing<br />
discharge: Zn 2 SiO 4 :Mn with a broad band at 526 nm and<br />
YBO 3 :Tb with a sharp peak at 544 nm, respectively. To improve<br />
red saturation, two types <strong>of</strong> red phosphors were<br />
mixed: Y,GdBO 3 :Eu with three main peaks at 593, 611,<br />
and 628 nm, together with YV,PO 4 :Eu which has a sharp<br />
peak at 620 nm. The <strong>for</strong>mer red phosphor appears redorange<br />
due to its main 593 nm emission peak, and it also<br />
possesses the highest conversion efficiency <strong>of</strong> vuv radiation<br />
into red visible light. On the other hand, YV,PO 4 :Eu,<br />
having a sharp main peak at 620 nm, <strong>of</strong>fers good red color<br />
purity. BaMaAl 10 O 17 :Eu was employed as the blue phosphor.<br />
It generates high luminance from a vuv excitation<br />
source but is weak under the harsh conditions <strong>of</strong> high energy<br />
vuv radiation. Consequently, it plays a key role in display<br />
longevity. For these reasons, the spectral properties <strong>of</strong> a<br />
PDP appear to be more complex than the other kinds <strong>of</strong><br />
displays. Fig. 6(b) is an enlargement <strong>of</strong> Fig. 6(a) and illustrates<br />
the red-orange emission <strong>of</strong> Ne gas. The intensity <strong>of</strong><br />
red-orange emission due to Ne gas was quite small compared<br />
with the other main peaks caused by phosphor materials.<br />
Hence the maximum radiance values used <strong>for</strong> Figs.<br />
6(a) and 6(b) are different. The fluctuations cause decrease<br />
in color purity as mentioned earlier. The characteristic redorange<br />
Ne gas emission at 585.2 nm results from atomic<br />
electronic transitions from the higher energy 2p quantum<br />
state to the lower lying 1s energy level. 10<br />
Temporal Stability<br />
A PDP needs time to reach a steady state <strong>for</strong> accurate measurement.<br />
As a result, temporal stability was evaluated over<br />
60 minutes using four pattern sizes, each consisting <strong>of</strong> a<br />
white color as shown in Figure 7. The white color at size 4%<br />
and having the highest number <strong>of</strong> sustain pulses shows the<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 341
Choi et al.: Colorimetric characterization model <strong>for</strong> plasma display panel<br />
Figure 6. a Spectral power distributions <strong>of</strong> normalized RGB primaries. b Visible emission <strong>of</strong> neon gas<br />
between 580 and 680 nm.<br />
Figure 7. Plot <strong>of</strong> relative luminance values over time <strong>for</strong> white at 4, 30,<br />
60, and 100% pattern size.<br />
Figure 8. a White color patch at 30% pattern size with a black background.<br />
b White color patch at 30% pattern size with a white<br />
background.<br />
largest gap between its initial state and a steady state. For the<br />
4% and 30% color patches, the luminance decreases significantly<br />
at the beginning. Although the time to reach a steady<br />
state is similar <strong>for</strong> all four pattern sizes (around 40 min), the<br />
decrease in luminance is dependent on pattern size. A decrease<br />
in pattern size leads to an increase the number <strong>of</strong><br />
RGB sustain pulses accompanied by an increase in temperature<br />
<strong>of</strong> the RGB cells. If a small color patch is displayed on<br />
a PDP, the initial temperature is higher than that <strong>for</strong> a larger<br />
sized patch; however this is mitigated by a greater rate <strong>of</strong><br />
temperature change. As a result, the time to reach a steady<br />
state is not dependent on pattern size. Conversely, the decrease<br />
in luminance ratio is inversely proportional to pattern<br />
size due to the shorter time needed to reach a stable RGB<br />
cell temperature.<br />
Spatial Independence<br />
The two color patches shown in Figures 8(a) and 8(b) have<br />
the same center square color and different backgrounds.<br />
Spatial independence defines by how much the central white<br />
color is affected by changes in background color. 11 The central<br />
white colors <strong>of</strong> Figs. 8(a) and 8(b) have quite different<br />
CIE XYZ values (445, 444, 550) and (160, 166, 186), respectively,<br />
because the colorimetric characteristics <strong>of</strong> a PDP are<br />
dictated by the number <strong>of</strong> RGB sustain pulses as determined<br />
by the APC rather directly from the RGB input values. For<br />
example, in order to produce the same light output <strong>for</strong> the<br />
white in Figs. 8(a) and 8(b), there is a need to have different<br />
numbers <strong>of</strong> sustain pulses.<br />
Color Gamut<br />
The definition <strong>of</strong> color gamut is the range <strong>of</strong> colors that is<br />
achievable on a given color reproduction medium under a<br />
specified set <strong>of</strong> viewing conditions. Figure 9 shows the color<br />
gamut under dark viewing conditions defined by the primary<br />
and secondary colors <strong>of</strong> a PDP. The ranges <strong>of</strong><br />
tristimulus values displayed differ according to pattern size<br />
(as explained in Pattern size effect section). The color gamuts<br />
<strong>of</strong> four pattern sizes are plotted in a CIELAB a * b * diagram<br />
[Fig. 9(a)]. In addition, color gamuts <strong>of</strong> four pattern<br />
sizes at a hue angle <strong>of</strong> 137° are compared using a CIELAB<br />
C * L * diagram [Fig. 9(b)]. The CIE L * a * b * color coordinates<br />
were calculated based on the peak white at 4% pattern size<br />
representing an L * <strong>of</strong> 100 848.4 cd/m 2 . There<strong>for</strong>e, the<br />
C * L * coordinates <strong>of</strong> the white color patches <strong>for</strong> 30, 60, and<br />
100% pattern sizes show lower values than those at the 4%<br />
pattern size. As pattern size decreases, the range <strong>of</strong> colors<br />
achievable on a PDP becomes larger.<br />
Tone Reproduction Curve<br />
The tone reproduction curve depicts the relationship between<br />
RGB input values and their resultant luminance val-<br />
342 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Choi et al.: Colorimetric characterization model <strong>for</strong> plasma display panel<br />
Figure 9. a Color gamuts <strong>for</strong> the 4, 30, 60, and 100% pattern sizes plotted in an a * b * diagram. b Color<br />
gamuts at a hue angle <strong>of</strong> 137° <strong>for</strong> the 4, 30, 60, and 100% pattern sizes plotted in a C * L * diagram.<br />
ues. The RGB luminances <strong>of</strong> a CRT are controlled by cathode<br />
voltages. For PDPs, on the other hand, the number <strong>of</strong><br />
RGB sustain pulses controls luminance. Figures 10(a) and<br />
10(b) illustrate the intrinsic properties <strong>of</strong> the PDP studied.<br />
Figures 10(a) and 10(b) contain plots <strong>of</strong> normalized Y values<br />
<strong>for</strong> the 4% size (white luminance is 1) against the normalized<br />
number <strong>of</strong> sustain pulses and input values, respectively.<br />
It can be clearly seen in Fig. 10(a) that an increase in the<br />
number <strong>of</strong> RGB sustain pulses leads to an increase in RGB<br />
luminance. Furthermore, a larger patch size has a smaller<br />
range <strong>of</strong> sustain pulse numbers. This is because a larger<br />
pattern size is constrained to a lower maximum number <strong>of</strong><br />
sustain pulses in order avoid exceeding power consumption<br />
limitations. The points indicated by the arrows correspond<br />
to 95% relative luminance, with respect to the maximum<br />
white luminance value at each pattern size.<br />
The intrinsic TRC <strong>of</strong> a PDP [Figs. 10(a) and 10(b)]<br />
should be modified so that the slope <strong>of</strong> the low luminance<br />
range is much smaller than that at high luminances. Figure<br />
10(c) shows the result after gamma is modified by adjusting<br />
the number <strong>of</strong> sustain pulses. The shape <strong>of</strong> the TRC after<br />
gamma modification is the same regardless <strong>of</strong> pattern size.<br />
The usable range <strong>of</strong> the number <strong>of</strong> sustain pulses, however,<br />
depends on pattern size. For example, to produce a white<br />
patch on this PDP using an input value <strong>of</strong> 255, either 466,<br />
766, 1376, or 2594 RGB sustain pulses are assigned, respectively,<br />
<strong>for</strong> the 100%, 60%, 30%, and 4% pattern sizes. The<br />
number <strong>of</strong> sustain pulses available <strong>for</strong> white at 100% pattern<br />
size, 0 to 466, are quantized to 256 levels to make white<br />
luminance follow a power function <strong>of</strong> approximately 2.2 (the<br />
gamma value).<br />
Additivity<br />
The channel additivity was evaluated <strong>for</strong> white at four pattern<br />
sizes. The results are given in Table V in terms <strong>of</strong> percentage<br />
change in tristimulus values and color difference<br />
E * ab . The tristimulus values <strong>of</strong> the RGB patches having<br />
the same number <strong>of</strong> sustain pulses as the peak white patch<br />
were measured to evaluate additivity <strong>for</strong> each pattern size. A<br />
substantial difference between the tristimulus values <strong>for</strong><br />
white and the sum <strong>of</strong> the red, green and blue channels was<br />
found. The latter is larger than the <strong>for</strong>mer by about 15%<br />
*<br />
which corresponds to approximately 6 E ab units. The available<br />
power to a cell drops as other cells become active, leading<br />
to a reduction in brightness. This means that, <strong>for</strong> example,<br />
in terms <strong>of</strong> luminance, R+G+Bwhite. To<br />
counteract the problem <strong>of</strong> deviation from additivity, several<br />
matrix coefficients—including RGB channel cross terms<br />
(RG, RB, GB, and RGB)—were incorporated into the trans<strong>for</strong>mation<br />
matrices <strong>of</strong> the two-step model. In addition, a 3D<br />
LUT model was implemented in which many measured data<br />
points were included so as to compensate <strong>for</strong> the inherent<br />
additivity failure <strong>of</strong> a PDP.<br />
COLORIMETRIC CHARACTERIZATION MODEL FOR<br />
A PDP<br />
Testing the Models’ Per<strong>for</strong>mance at 100% Pattern Size<br />
As introduced in the third section, three types <strong>of</strong> characterization<br />
models were developed: 3D-LUT, single-step polynomial,<br />
and two-step polynomial. These were tested using the<br />
115-test color set. All <strong>of</strong> the models developed here are based<br />
on a 100% pattern size. The results are summarized in Table<br />
*<br />
VI in terms <strong>of</strong> mean and 95th percentile E ab units.<br />
It can be seen that the 3D-LUT model using tetrahedral<br />
interpolation 12 gave a reasonable prediction to the test data<br />
*<br />
with a mean and a 95th percentile <strong>of</strong> 1.3 and 2.5 E ab units.<br />
The two-step model using the primary matrix in which their<br />
coefficients were based on the measurement data gave quite<br />
poor per<strong>for</strong>mance with a mean difference <strong>of</strong> 4.3.<br />
Comparing different single-step polynomial models,<br />
there is a trend that the higher order polynomial models<br />
per<strong>for</strong>med better than the lower ones. However, this is only<br />
true <strong>for</strong> the models developed using more training samples.<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 343
Choi et al.: Colorimetric characterization model <strong>for</strong> plasma display panel<br />
For those models developed using the 3-level and 4-level<br />
training samples, the higher term polynomial models did<br />
not exhibit more accurate prediction than the lower order<br />
models. This could be caused by over-fitting the measurement<br />
noise when using higher order polynomial models<br />
based on small number <strong>of</strong> training samples. Overall, the<br />
311 polynomial model developed using the 4-level training<br />
data set (which included 64 colors) was found to be<br />
acceptable <strong>for</strong> industrial applications. Using the 320<br />
model can lead to further improvements in the modeling<br />
per<strong>for</strong>mance.<br />
In comparing the single- and two-step models, the<br />
single-step model per<strong>for</strong>med slightly better than the twostep<br />
model except <strong>for</strong> the 33 model. This implies that<br />
single-step polynomial models with a higher order already<br />
consider the cross-talk between different channels in PDPs.<br />
There is needless to include a 1D-LUT normalization.<br />
Figure 11 shows different polynomial per<strong>for</strong>mances <strong>for</strong><br />
*<br />
a single-step model in terms <strong>of</strong> mean E ab using the training<br />
and testing data sets. It can be seen that the models<br />
predicted more accurately when the terms increase until<br />
reaching 311 and 320 polynomial models. For the<br />
335 model, this fits the training data set best, however it<br />
per<strong>for</strong>med poorly <strong>for</strong> the testing data set due to modeling<br />
the noise in the training data set.<br />
In real applications, both the <strong>for</strong>ward and reverse characterization<br />
models are used, i.e., from device signal to XYZ<br />
and vice versa. However, not all models are analytically invertible<br />
and so reverse models were developed having the<br />
same structure as the <strong>for</strong>ward model. The numerical reversibility<br />
<strong>of</strong> the single-step model was also tested. The testing<br />
procedure is shown in Figure 12 and does not require any<br />
color measurement. Here the 115-test color set, defined in<br />
terms <strong>of</strong> XYZ, was again used to first predict RGB sustain<br />
pluses using the reverse model and then further predict the<br />
corresponding XYZ via the <strong>for</strong>ward model. Finally, the color<br />
difference was calculated between the target XYZ and predicted<br />
XYZ values. The results <strong>for</strong> each combination <strong>of</strong> <strong>for</strong>ward<br />
and reverse polynomial model developed by the 4-level<br />
training data are given in Table VII. It can be seen that the<br />
311 polynomial model can give acceptable per<strong>for</strong>mance<br />
(its mean and 95th percentile are 0.3 and 0.8 E * ab , respectively).<br />
This can be further improved by using the 320<br />
model. Both models outper<strong>for</strong>med the other models<br />
studied.<br />
Figure 10. a The relationship between normalized number <strong>of</strong> sustain<br />
pulses and normalized white luminance <strong>for</strong> 4, 30, 60, and 100% pattern<br />
sizes be<strong>for</strong>e the modification <strong>of</strong> gamma. b The relationship between<br />
normalized input values and normalized white luminance <strong>for</strong> four pattern<br />
sizes be<strong>for</strong>e the modification <strong>of</strong> gamma. c The relationship between<br />
normalized input values and normalized white luminance <strong>for</strong> four pattern<br />
sizes after modifying gamma.<br />
Testing the Models’ Per<strong>for</strong>mance at Different Pattern<br />
Sizes<br />
Using the same approach as <strong>for</strong> the 100% pattern size (previous<br />
section), different single-step polynomial models developed<br />
using 3-, 4-, and 5-level 3D-LUT training data <strong>for</strong><br />
each <strong>of</strong> the 4%, 30%, and 60% pattern sizes. Very similar<br />
per<strong>for</strong>mances were found and so only the results from the<br />
30% pattern size are reported in Table VIII in terms <strong>of</strong> mean<br />
*<br />
and 95th percentile E ab values. The results showed that the<br />
311 and 320 polynomial models using 4- or 5-level<br />
training data gave a reasonable prediction. These results are<br />
very similar to those found at 100% pattern size (see Table<br />
VI).<br />
Developing a Single Characterization Model<br />
As mentioned in the Pattern size effect section, light output<br />
is proportional to the number <strong>of</strong> sustain pulses and the<br />
range these are regulated by the APC according to pattern<br />
size. Hence the characterization models developed earlier are<br />
only applicable to a single pattern size. A new method is<br />
developed here which aims to predict the colors displayed at<br />
different pattern sizes. In order to make a single model<br />
which can predict CIE XYZ at pattern sizes other than that<br />
used <strong>for</strong> the training set, it is necessary to select an appropriate<br />
training set covering the whole range <strong>of</strong> sustain pulses<br />
used <strong>for</strong> the test set. For example, it needs to predict CIE<br />
XYZ values <strong>of</strong> several colors in 80% and 45% sizes. There<br />
are two approaches to the selection <strong>of</strong> a training set <strong>for</strong> this<br />
purpose. First, a set having smaller size than 45% can be<br />
used because this can cover a higher range <strong>of</strong> sustain pulses<br />
than those available <strong>for</strong> the 80% and 45% sizes. Second, two<br />
344 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Choi et al.: Colorimetric characterization model <strong>for</strong> plasma display panel<br />
Table V. Tristimulus additivity failure and corresponding color difference <strong>for</strong> white at 4, 30, 60, and 100% pattern size.<br />
Pattern<br />
size<br />
RGB<br />
IVs<br />
Y<br />
cd/m 2 <br />
APC<br />
level<br />
No. <strong>of</strong><br />
sustain X Y Z<br />
*<br />
E ab<br />
100% 255 166.1 255 466 15.1% 13.9% 16.4% 5.6<br />
60% 255 263.2 219 766 16.7% 15.2% 19.2% 6.5<br />
30% 255 443.6 146 1376 17.6% 17.8% 19.9% 6.6<br />
4% 255 848.0 0 2594 16.7% 16.6% 21.1% 6.6<br />
Table VI. Testing the per<strong>for</strong>mance in terms <strong>of</strong> E * ab <strong>of</strong> the characterization models using the 115-color test set. The models were developed based on 6-, 5-, 4- and 3-level training sets.<br />
Training<br />
set<br />
3D<br />
LUT<br />
Single-step polynomial model<br />
Two-step polynomial model<br />
35 38 311 320 335 Primary matrix a 33 34 38 311 320 335<br />
6-level Mean 1.3 3.8 3.1 1.5 1.2 1.2 3.2 3.5 1.7 1.5 1.3 1.4<br />
95th 2.5 9.4 11.7 3.8 2.7 2.8 8.3 8.3 4.5 3.4 2.9 3.0<br />
5-level Mean 3.5 2.8 1.4 1.2 1.1 4.3 3.1 3.2 1.6 1.5 1.3 1.3<br />
95th 8.3 9.4 3.3 2.7 2.3 7.4 7.4 4.1 3.2 2.7 2.5<br />
4-level Mean 3.4 2.5 1.5 1.2 3.9 6.9 3.0 3.0 1.6 1.5 1.3 3.8<br />
95th 7.8 7.8 3.7 2.8 11.7 6.6 6.8 3.6 3.4 3.0 10.3<br />
3-level Mean 3.2 2.3 1.6 7.0 69.5 3.0 3.1 1.5 1.6 3.6 17.9<br />
95th 8.0 7.2 3.5 22.9 305.5 6.8 6.7 3.7 3.6 11.5 55.2<br />
a The primary matrix was obtained from RGB primary colors.<br />
Table VII. Reversibility result <strong>of</strong> polynomials <strong>for</strong> the 4-level training set in terms <strong>of</strong><br />
E ab .<br />
35 38 311 320 335<br />
Mean 2.3 1.6 0.3 0.2 1.7<br />
95th 5.9 4.9 0.8 0.5 4.8<br />
Figure 11. A comparison <strong>of</strong> average E * ab values against the terms used<br />
in the single-step polynomial model <strong>for</strong> the test and training set.<br />
training sets—one set having smaller pattern size than 45%<br />
and another having 100% size—can be combined to make a<br />
new training set. The reasoning behind the second method<br />
is to improve model accuracy. If two color patches at different<br />
pattern sizes but with same RGB sustain pulses are measured,<br />
a subtle difference in their XYZ values could be<br />
found. One practical example is that the luminance values<br />
<strong>for</strong> 30%, 60% and 100% pattern sizes using the same number<br />
<strong>of</strong> sustain pulses (408), are 111, 107, and 104, respectively.<br />
This is due to the available power to a cell being<br />
slightly different because <strong>of</strong> the different number <strong>of</strong> activated<br />
cells at different pattern sizes. Although the first training set<br />
may be sufficient, the polynomial coefficients computed by<br />
the second training set can be expected to take into account<br />
this small color difference due to pattern size. In real applications,<br />
the number <strong>of</strong> sustain pulses used to display complex<br />
images typically corresponds to those associated with<br />
40–50 % pattern sizes. There<strong>for</strong>e, we generated the first<br />
training set as 4-level 3D LUT <strong>for</strong> the 30% pattern size. In<br />
addition, a second training set was produced by combining<br />
two 4-level 3D LUTs <strong>of</strong> 100% and 30% pattern sizes. These<br />
two 4-level 3D LUTs were composed <strong>of</strong> different combinations<br />
<strong>of</strong> RGB sustain pulses. The test data included three<br />
27-color test sets at 80%, 60%, and 45% sizes. Tables IX(a)<br />
and IX(b) summarize the results from the first and second<br />
training sets. The results from the second training set show<br />
that the polynomials with 11, 20, and 35 terms per<strong>for</strong>med<br />
well and gave similar predictive accuracy <strong>for</strong> the three pattern<br />
sizes in Table IX(b). Table IX(a) summarizes the results<br />
<strong>for</strong> the models developed using only the 30% pattern size.<br />
The results in Table IX(b) are worse than those in Table<br />
IX(b) in all cases. This demonstrates that it is better to use<br />
the combined training set <strong>of</strong> 100% and 30% sizes <strong>for</strong> predicting<br />
midsized test colors.<br />
CONCLUSIONS<br />
The physical properties <strong>of</strong> a PDP which affect colorimetric<br />
characterization were examined. Also, colorimetric characteristics<br />
unique to PDP displays were investigated. Among<br />
those, a pattern-size influence and a substantial additivity<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 345
Choi et al.: Colorimetric characterization model <strong>for</strong> plasma display panel<br />
Figure 12. The process <strong>for</strong> testing reversibility.<br />
Table VIII. A comparison <strong>of</strong> the per<strong>for</strong>mances in terms <strong>of</strong> E * ab using the 27-color test set and 5-, 4-, and 3-level training sets at 30% pattern<br />
size.<br />
Training<br />
set<br />
Single-step polynomial model<br />
35 38 311 320 335<br />
5-level Mean 3.8 3.3 1.1 1.3 1.5<br />
95th 8.8 9.1 2.2 3.0 3.4<br />
4-level Mean 3.7 3.0 1.4 1.5 3.2<br />
95th 7.6 6.6 2.4 2.6 8.8<br />
3-level Mean 3.9 3.2 2.2 8.6 26.9<br />
95th 7.7 6.8 4.2 18.2 52.0<br />
Table IX. a Comparing models’ per<strong>for</strong>mance E * ab using each 27-color test set at 80%, 60%, and 45% pattern size. Each model was developed<br />
using the 4-level training set at 30% pattern size. b Comparing models’ per<strong>for</strong>mance E * ab using each 27-color test set at 80%, 60%, and 45%<br />
pattern size. Each model was developed using two 4-level training sets at 30% and 100% pattern sizes.<br />
a<br />
Training<br />
set<br />
Test<br />
set<br />
*<br />
E ab<br />
Single-step polynomial model<br />
35 38 311 320 335<br />
30%<br />
pattern<br />
size<br />
4-level<br />
80% Mean 5.1 5.3 2.1 2.5 2.9<br />
95th 9.5 10.6 4.2 5.2 7.6<br />
60% Mean 4.8 4.9 1.9 2.2 2.4<br />
95th 9.5 10.1 4.3 4.8 6.7<br />
45% Mean 3.9 3.7 1.8 1.7 2.4<br />
95th 7.7 7.6 3.2 3.1 6.3<br />
Training<br />
set<br />
Test<br />
set<br />
b<br />
*<br />
E ab<br />
Single-step polynomial model<br />
35 38 311 320 335<br />
Mixture<br />
<strong>of</strong><br />
100%<br />
4-level<br />
&<br />
30%<br />
4-level<br />
pattern<br />
size<br />
80% Mean 4.4 4.8 1.7 1.5 1.1<br />
95th 9.0 11.0 3.5 3.3 2.5<br />
60% Mean 4.2 4.4 1.7 1.5 1.3<br />
95th 8.8 10.3 3.7 2.8 2.3<br />
45% Mean 3.3 3.1 1.6 1.4 1.6<br />
95th 6.8 6.9 2.7 2.0 3.2<br />
failure were found. These must necessarily be considered<br />
when making an accurate colorimetric characterization<br />
model.<br />
Initially, three characterization methods were derived<br />
between the number <strong>of</strong> sustain pulses and CIE XYZ values<br />
at 100% pattern size in order to determine an appropriate<br />
model <strong>for</strong> a PDP. In the <strong>for</strong>ward direction, single- and twostep<br />
polynomial models, which each have more than 8<br />
terms, and a 3D LUT model showed the best results <strong>for</strong> the<br />
6-, 5-, and 4-level training set. However, the single-step<br />
model was eventually selected because <strong>of</strong> its simplicity. The<br />
required number <strong>of</strong> training set samples needed to obtain<br />
good model per<strong>for</strong>mance and requiring the least measurement,<br />
was 64 <strong>for</strong> the 4-level 3D LUT. Also the reversibility <strong>of</strong><br />
the single-step model was evaluated using a 4-level 3D LUT<br />
and this was shown to produce satisfactory results <strong>for</strong> 11-<br />
and 20-term polynomials. There<strong>for</strong>e, the single-step model<br />
was extended to various other pattern sizes. Their results<br />
346 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Choi et al.: Colorimetric characterization model <strong>for</strong> plasma display panel<br />
validated that the polynomial regression method using the<br />
4-level training set was a good characterization model <strong>for</strong><br />
this PDP display.<br />
Finally, one comprehensive training set consisting <strong>of</strong><br />
two 4-level 3D LUTs corresponding to 100% and 30% pattern<br />
sizes were produced to predict CIE XYZ at intermediate<br />
pattern sizes (i.e., sizes which were not present in the training<br />
set). These outcomes demonstrated that the single-step<br />
model could be successfully applied to estimate colors at<br />
different pattern sizes using just one combined training set.<br />
REFERENCES<br />
1 N. Katoh, T. Deguchi, and R. S. Berns, “An Accurate Characterization <strong>of</strong><br />
CRT Monitor (I) Verification <strong>of</strong> Past Studies and Clarification <strong>of</strong><br />
Gamma”, Opt. Rev. 8, 305 (2001).<br />
2 N. Katoh, T. Deguchi, and R. S. Berns, “An Accurate Characterization <strong>of</strong><br />
CRT Monitor (II) Proposal <strong>for</strong> an Extension to CIE Method and its<br />
Verification”, Opt. Rev. 8, 397 (2001).<br />
3 M. D. Fairchild and J. E. Gibson, “Colorimetric Characterization <strong>of</strong><br />
Three Computer Displays (LCD and CRT)”, Munsell Color <strong>Science</strong><br />
Laboratory Technical Report, http://www.cis.rit.edu/mcsl/research/PDFs/<br />
GibsonFairchild.pdf (2000).<br />
4 Y. S. Kwak and L. MacDonald, “Characterization <strong>of</strong> a Desktop LCD<br />
Projector”, Displays 21, 179 (2000).<br />
5 D. R. Wyble and H. Zhang, “Colorimetric Characterization Model <strong>for</strong><br />
DLP Projectors”, Proc. IS&T/SID 11th Color <strong>Imaging</strong> Conference (IS&T,<br />
Springfield, VA, 2003), pp. 346–350.<br />
6 G. Kutas and P. Bodrogi, “Colorimetric Characterization <strong>of</strong> HD-PDP<br />
Device”, in IS&T’s 2nd European Conference on Color Graphics, <strong>Imaging</strong><br />
and Vision, (IS&T, Springfield, VA, 2004,). pp. 65–69.<br />
7 Multimedia Systems and Equipment—Color Measurement and<br />
Management, Part 5: Equipment using Plasma Display Panels, IEC<br />
61966–5, 2001.<br />
8 J. P. Boeuf, “Plasma Display Panels: Physics, Recent Developments and<br />
Key Issues”, J. Phys. D 36, R53 (2003).<br />
9 P. Bodrogi and J. Schanda, “Testing a Calibration Method <strong>for</strong> Color CRT<br />
Monitors. A Method to Characterize the Extent <strong>of</strong> Spatial<br />
Interdependence and Channel Interdependence”, Displays 16(3), 123<br />
(1995).<br />
10 R. S. Van Dyck, C. E. Johnson, and H. A. Shugart, “Lifetime Lower<br />
Limits <strong>for</strong> the 3p 0 and 3p 2 Metastable States <strong>of</strong> Neon, Argon and<br />
Krypton”, Phys. Rev. A 5, 991 (1972).<br />
11 P. Green and L. MacDonald, Color Engineering: Achieving Device<br />
Independent Color (John Wiley and Sons Ltd, West Sussex, UK, 2002),<br />
p. 158.<br />
12 H. R. Kang, Color Technology <strong>for</strong> Electronic <strong>Imaging</strong> Devices (SPIE<br />
Optical Engineering Press, Bellingham, WA, 1997), p. 64.<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 347
<strong>Journal</strong> <strong>of</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology® 51(4): 348–359, 2007.<br />
© <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology 2007<br />
Real-Time Color Matching Between Camera and LCD<br />
Based on 16-bit Lookup Table Design in Mobile Phone<br />
Chang-Hwan Son<br />
School <strong>of</strong> Electrical Engineering and Computer <strong>Science</strong>, Kyungpook National University, 1370,<br />
Sankyuk-dong, Buk-gu, Daegu 702-701, Korea<br />
Cheol-Hee Lee<br />
Major <strong>of</strong> Computer Engineering, Andong National University, 388, Seongcheon-dong, Andong,<br />
Gyeongsangbuk-Do 760-747, Korea<br />
Kil-Houm Park and Yeong-Ho Ha <br />
School <strong>of</strong> Electrical Engineering and Computer <strong>Science</strong>, Kyungpook National University, 1370,<br />
Sankyuk-dong, Buk-gu, Daegu 702-701, Korea<br />
E-mail: yha@ee.knu.ac.kr<br />
Abstract. Based on the concept <strong>of</strong> multimedia convergence, imaging<br />
devices, such as cameras, liquid crystal displays (LCDs), and<br />
beam projectors, are now built-in to mobile phones. As such, mobile<br />
cameras capture still images or moving pictures, then store them as<br />
digital files, making it possible <strong>for</strong> users to replay moving pictures<br />
and review captured still images. Increasingly, users want LCD in<br />
the mobile phone (we call it mobile LCD hereafter) to reproduce the<br />
same colors as the real scene. Accordingly, this paper proposes a<br />
method <strong>for</strong> color matching between mobile camera and mobile LCD<br />
that includes characterizing the mobile camera and mobile LCD,<br />
gamut mapping, camera noise reduction, and a 16-bit lookup table<br />
(LUT) design. First, to estimate the CIELAB values <strong>for</strong> the objects in<br />
the real scene, mobile camera characterization is achieved through<br />
polynomial regression <strong>of</strong> the optimal order determined by investigating<br />
the relation between captured RGB values and measured<br />
CIELAB values <strong>for</strong> a standard color chart. Thereafter, mobile LCD<br />
characterization is conducted based on 16-bit/pixel processing because<br />
<strong>of</strong> the reduced bit depth <strong>of</strong> the images displayed on a mobile<br />
LCD. In addition, a sigmoid model is used to find the luminance<br />
value corresponding to the RGB control signal, instead <strong>of</strong> using gain<br />
<strong>of</strong>fset gamma and S-curve models due to the adjustment <strong>of</strong> luminance<br />
curve made by a system designer <strong>for</strong> preference color reproduction.<br />
After completing the two types <strong>of</strong> characterization, gamut<br />
mapping is per<strong>for</strong>med to connect the source medium (mobile camera)<br />
with the target medium (mobile LCD), then a combination <strong>of</strong><br />
sigmoid functions with different parameters to control the shape is<br />
applied to the luminance component <strong>of</strong> the gamut-mapped CIELAB<br />
values to reduce camera noise. Finally, a three-dimensional RGB<br />
LUT is constructed using 16-bit/pixel-based data to enable color<br />
matching <strong>for</strong> moving pictures and inserted into the mobile phone.<br />
Experimental results show that moving pictures transmitted by a<br />
mobile camera can be realistically reproduced on a mobile LCD<br />
without any additional computation or memory burden. © 2007 <strong>Society</strong><br />
<strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology.<br />
DOI: 10.2352/J.<strong>Imaging</strong>Sci.Technol.200751:4348<br />
<br />
IS&T Member.<br />
Received Dec. 1, 2006; accepted <strong>for</strong> publication Mar. 30, 2007.<br />
1062-3701/2007/514/348/12/$20.00.<br />
INTRODUCTION<br />
With the appearance <strong>of</strong> multimedia convergence in mobile<br />
phones that can now provide such functions as web browsing,<br />
3D games, television broadcasting, and image capturing,<br />
in addition to communication, manufacturers have invested<br />
heavily in super highway communication network, nextgeneration<br />
memory chips, and encryption technology <strong>for</strong><br />
reliable e-commerce operations. The use <strong>of</strong> color reproduction<br />
technology in mobile phones has also been recently<br />
introduced to support the development <strong>of</strong> mobile cameras,<br />
mobile beam projectors, and mobile liquid crystal displays<br />
(LCDs). In particular, with the rapid increase in mobile<br />
cameras, mobile phones can now capture and store still images<br />
or moving pictures as digital files, making it possible <strong>for</strong><br />
users to replay the moving pictures and review captured still<br />
images anytime and anywhere. However, mobile LCDs are<br />
currently unable to reproduce the original colors captured by<br />
a mobile camera due to a reduced bit-depth, lower backlight<br />
luminance, and weak resolution. 1 In addition, mobile cameras<br />
have a small lens, low dynamic range, and poor modulation<br />
transfer function (MTF), plus each device senses or<br />
displays in a different way, as they have unique<br />
characteristics. 2 As a result, there is a significant difference in<br />
the color appearance when captured images are displayed on<br />
a mobile LCD. There<strong>for</strong>e, real-time colormatching between<br />
mobile camera and mobile LCD in a mobile phone needs to<br />
be considered to ensure a better image quality.<br />
The aim <strong>of</strong> color matching is to achieve color consistency<br />
even when an image moves across various devices and<br />
undergoes many color trans<strong>for</strong>mations. 3 Several color<br />
matching approaches have already been suggested, <strong>for</strong> example,<br />
a simple method is to transmit the RGB digital values<br />
from the original device to the reproducing device, referred<br />
to as device-dependent color matching. Yet, since this<br />
method is no more than physical data transmission, accurate<br />
color matching cannot be achieved across various devices.<br />
Meanwhile, spectral-based approaches match the spectral reflectance<br />
curves <strong>of</strong> the original and reproduced colors, so the<br />
original and reproduction look the same under any il-<br />
348
Son et al.: Real-time color matching between camera and LCD based on 16-bit lookup table design in mobile phone<br />
Figure 1. The block diagram <strong>of</strong> the proposed method.<br />
luminant: i.e., there is no metamerism. However, the computation<br />
<strong>of</strong> reflectance is very complex and time consuming,<br />
making a spectral-based approach inappropriate <strong>for</strong> realtime<br />
color matching. Another method is colorimetric color<br />
matching to reproduce the same CIE chromaticity and relative<br />
luminance compared with the original color. This has<br />
already been widely applied to imaging devices, such as<br />
monitors, printers, and scanners based on the international<br />
color consortium (ICC) pr<strong>of</strong>ile, yet not to mobile phones,<br />
which have only been considered as a means <strong>of</strong> communication<br />
until quite recently. However, with multimedia convergence,<br />
mobile manufacturers have become aware <strong>of</strong> the<br />
importance <strong>of</strong> the ICC pr<strong>of</strong>ile <strong>for</strong> color matching between<br />
mobile cameras and mobile LCDs. Accordingly, this paper<br />
presents a real-time color matching system <strong>for</strong> mobile cameras<br />
and mobile LCDs based on the concept <strong>of</strong> the ICC<br />
pr<strong>of</strong>ile.<br />
The proposed color matching system is composed <strong>of</strong><br />
four steps: Characterization <strong>of</strong> the mobile LCD and mobile<br />
camera, gamut mapping, noise reduction, and a<br />
16-bit-based lookup table (LUT) design. The device characterization<br />
defines the relationship between the tristimulus<br />
values (CIEXYZ or CIELAB) and RGB digital values. In general,<br />
mobile camera characterization is modeled by a polynomial<br />
regression, and the more the polynomial order increases,<br />
the better the per<strong>for</strong>mance. However, <strong>for</strong> a higher<br />
polynomial order, most estimated tristimulus values exceed<br />
the boundary <strong>of</strong> the maximum lightness and chroma, making<br />
the implementation <strong>of</strong> mobile camera characterization<br />
difficult, as the relation between the tristimulus values and<br />
digital RGB values has not been analyzed. Thus a polynomial<br />
order is suggested based on investigating the relation<br />
between RGB digital values trans<strong>for</strong>med using the opponent<br />
color theory and CIELAB values. Meanwhile, <strong>for</strong> the mobile<br />
LCD characterization, a sigmoid function instead <strong>of</strong> a conventional<br />
method, such as the gain <strong>of</strong>fset gamma (GOG) or<br />
S-curve model, is used to estimate the luminance curve<br />
made by the system designer to achieve a preferable color<br />
reproduction or to improve the perceived contrast <strong>of</strong> the<br />
mobile LCD. Furthermore, the characterization is conducted<br />
based on 16-bit data processing, as a mobile LCD is controlled<br />
based on 16-bit data, in contrast to digital TVs or<br />
monitors with 24-bit data. After completing the two types <strong>of</strong><br />
characterization, a gamut-mapping algorithm is applied to<br />
connect the source medium (mobile camera) with the target<br />
medium (mobile LCD).<br />
Although the three processes mentioned above are sufficient<br />
to obtain colorimetric color matching <strong>for</strong> still images,<br />
noise reduction and an LUT design still need to be considered<br />
to achieve real-time color matching <strong>for</strong> moving pictures.<br />
In a mobile camera, various camera noises, such as<br />
CCD noise and thermal noise, are incorporated into moving<br />
pictures and further amplified after color matching, thereby<br />
degrading the image quality, especially in the dark region <strong>of</strong><br />
the achromatic axis. Thus, to solve this problem, a combination<br />
<strong>of</strong> two sigmoid functions with different parameters to<br />
control the shape is applied to the lightness component <strong>of</strong><br />
the gamut-mapped tristimulus values to change the contrast<br />
ratio. As a result, the lightness values <strong>for</strong> the camera noise<br />
are reduced in the dark region <strong>of</strong> the achromatic axis,<br />
thereby reducing the amplified camera noises. In addition, a<br />
three-dimensioinal (3D) RGB LUT is designed based on<br />
16-bit data to reduce the complex computation <strong>of</strong> serialbased<br />
processing and facilitate color matching <strong>for</strong> moving<br />
pictures.<br />
PROPOSED METHOD<br />
Figure 1 shows a block diagram <strong>of</strong> the proposed algorithm<br />
that can achieve real-time color matching between a mobile<br />
camera and a mobile LCD. First, to predict the CIELAB<br />
values <strong>of</strong> arbitrary objects in a real scene, the mobile camera<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 349
Son et al.: Real-time color matching between camera and LCD based on 16-bit lookup table design in mobile phone<br />
characterization is conducted by finding the relation between<br />
the RGB digital values <strong>of</strong> a standard color chart captured<br />
in a lighting booth and CIELAB values measured using<br />
a colorimeter. The CIELAB values estimated from the<br />
mobile camera characterization <strong>of</strong> the input RGB values are<br />
then trans<strong>for</strong>med into an achievable color range that can be<br />
reproduced by the mobile LCD, referred to as gamut mapping.<br />
Next, the lightness values <strong>of</strong> the gamut-mapped<br />
CIELAB values are changed using the parameters <strong>of</strong> a sigmoid<br />
function provided by a visual experiment to reduce the<br />
camera noise incorporated into a moving picture, then combined<br />
with two untouched color signals. Thereafter, the<br />
modified gamut-mapped CIELAB values are converted into<br />
color matched RGB values <strong>for</strong> display on the mobile LCD<br />
based on a sigmoid-based mobile-LCD characterization to<br />
consider the luminance curve adjusted <strong>for</strong> the preferred<br />
color reproduction, along with 16-bit data processing due to<br />
the reduced bit depth in the mobile LCD. Finally, a 3D-RGB<br />
LUT is constructed using the 16-bit/pixel-based data to enable<br />
color matching <strong>for</strong> moving pictures and inserted into<br />
the mobile phone, thereby allowing 24-bit moving pictures<br />
to be reproduced on a mobile LCD with a higher quality<br />
image.<br />
CHARACTERIZATION OF THE MOBILE LCD BASED<br />
ON 16-bit DATA PROCESSING<br />
The display characterization predicts the tristimulus value<br />
<strong>for</strong> the input digital value and may be conducted by a<br />
measurement-based approach or modeling-based<br />
approach. 4–6 A measurement-based approach measures a lot<br />
<strong>of</strong> patches made by combination <strong>of</strong> input digital values using<br />
a colorimeter and estimates the tristimulus value by the<br />
interpolation method or polynomial regression <strong>for</strong> an arbitrary<br />
digital value. There<strong>for</strong>e, this approach improves the<br />
characterization accuracy, yet requires a lot <strong>of</strong> measurement<br />
data and extensive memory and is relatively complex. Meanwhile,<br />
a modeling-based approach finds the relationship between<br />
the digital input data and tristimulus value based on a<br />
mathematical function with a smaller number <strong>of</strong> data measurements.<br />
The GOG and S-curve models have been used as<br />
typical mathematical functions and have been applied to different<br />
types <strong>of</strong> display. In general, the GOG model is appropriate<br />
<strong>for</strong> CRT display because its electro-optical transfer<br />
function, the relationship between the grid voltage and beam<br />
current, follows a power curve shape, while LCD display is a<br />
binary device that switches from an OFF state to an ON state<br />
and follows the S-shaped curve, thereby adapting the<br />
S-curve model <strong>for</strong> LCD characterization. The overall procedure<br />
<strong>of</strong> modeling-based approach is identical except that<br />
electro-optical transfer function is modeled with a different<br />
mathematical function. The first step <strong>of</strong> modeling-based<br />
characterization is to convert the digital value to luminance<br />
value <strong>for</strong> each RGB channel. This can be done by estimating<br />
the coefficient <strong>of</strong> mathematical function with optimization<br />
programming. In the case <strong>of</strong> the GOG model, mathematical<br />
function can be described as<br />
Y ch =k g,ch d ch<br />
2 N −1 + k o,ch,ch<br />
, 1<br />
where ch represents RGB channel, d ch is input digital value,<br />
and N is the bit number; k g,ch ,k o,ch , are the gain, <strong>of</strong>fset,<br />
and gamma parameters <strong>of</strong> the GOG model, respectively. Y ch<br />
is the normalized luminance value corresponding to the normalized<br />
input digital value <strong>for</strong> each channel. To get all parameters<br />
<strong>of</strong> the GOG model, the digital value <strong>of</strong> each channel<br />
is independently sampled by a uni<strong>for</strong>m M interval,<br />
which assumes no channel interaction that the light emitted<br />
from a pixel location is dependent only on R, G, B triplet <strong>for</strong><br />
that pixel and is independent <strong>of</strong> input digital value <strong>for</strong> other<br />
pixels. 7 Then, the CIEXYZ values <strong>for</strong> M-sample digital values<br />
are acquired by measuring the displayed patches created<br />
by M-sampled digital values with a colorimeter. At this time,<br />
even though displayed patches are made by 8-bit data <strong>for</strong><br />
each channel, the 8-bit based M-sampled RGB digital values<br />
corresponding to the measured CIEXYZ values must be<br />
practically converted to 5,6,5-bit data in the mobile LCD,<br />
and thus the digital values <strong>of</strong> the 8-bit based R-channel and<br />
B-channel are divided by 8, while that <strong>of</strong> the G-channel is<br />
dividedby4:<br />
d R = d R<br />
2 R, d G = d G<br />
2 G, d B = d B<br />
, 2<br />
B<br />
2<br />
where d R , d G , and d B is the digital value <strong>of</strong> displayed patch<br />
<strong>for</strong> each channel and R,G,B is the difference <strong>of</strong> bitnumber<br />
between 8-bit/channel <strong>of</strong> patches and bit-number/<br />
channel in mobile LCD. There<strong>for</strong>e, d ch in Eq. (1) is substituted<br />
with d R ,d G ,d B , which is used to find the luminance<br />
curve <strong>of</strong> mobile LCD.<br />
Of the measured CIEXYZ values, the Y values are selected<br />
as Y ch , assuming that the shapes <strong>of</strong> X, Y, and Z are<br />
identical after normalization, which is referred to the<br />
channel-chromaticity constancy that spectrum <strong>of</strong> light from<br />
a channel has the same basic shape and only undergoes a<br />
scaling in amplitude as the digital value <strong>for</strong> each channel is<br />
varied. 7 Finally, the pairs <strong>of</strong> M-sampled digital value<br />
d R ,d G ,d B and Y values are substituted in Eq. (1), yielding<br />
all parameters <strong>of</strong> the GOG model using optimization nonlinear<br />
programming. The second step is to trans<strong>for</strong>m the<br />
luminance value <strong>of</strong> each channel calculated by the GOG<br />
model to the CIEXYZ value. This stage can be simply<br />
achieved by a matrix operation:<br />
X r,max X g,max X R<br />
Y Y r,max Y g,max Y b,max Y G<br />
Z=X Z r,max Z g,max Z b,maxY Y B,<br />
where Y R , Y G , and Y B are the luminance values <strong>of</strong> each channel,<br />
Y ch=R,G,B . In each column, the matrix coefficients are the<br />
CIEXYZ value at the maximum digital value <strong>of</strong> each channel,<br />
and can be directly measured with a colorimeter.<br />
Through the above-mentioned two steps, display characterization<br />
can be accomplished. In the case <strong>of</strong> the S-curve<br />
3<br />
350 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Son et al.: Real-time color matching between camera and LCD based on 16-bit lookup table design in mobile phone<br />
Figure 2. Electro-optical transfer function <strong>for</strong> mobile LCD; a GOG model, b GOG model except saturation<br />
region, c S-curve model, and d sigmoid model.<br />
model, only the power-curved function shown in Eq. (1) is<br />
replaced with the S-shaped mathematical function in the<br />
process <strong>of</strong> display characterization<br />
d ch /2 N −1 ch<br />
Y ch =A ch<br />
d ch /2 N −1 ch + C ch ,<br />
where A ch , ch , ch , and C ch are parameters, respectively.<br />
Equation (4) has various S-shaped curves according to the<br />
parameter values, and if both ch , and C ch is zero, Eq. (4)<br />
follows the gamma curve as Eq. (1). All parameters in Eq.<br />
(4) can be obtained by applying the same process, the first<br />
step explained in the GOG model. Using these parameters,<br />
the input digital value is converted into the luminance value<br />
and then trans<strong>for</strong>med into the CIEXYZ value through a matrix<br />
operation.<br />
To conduct the characterization <strong>of</strong> mobile LCD, we apply<br />
conventional methods to a cellular phone, a Samsung<br />
SCH-500. In a mobile phone, each RGB pixel value is represented<br />
by (5,6,5) bit and image size is fixed at 240320.<br />
Figures 2(a)–2(c) show the electro-optical transfer function<br />
4<br />
resulting from the GOG model, the GOG model without the<br />
saturation region, and the S-curve model. In Figure 2, three<br />
types <strong>of</strong> lines represent the estimated luminance values obtained<br />
by using conventional characterization <strong>for</strong> each channel,<br />
while three types <strong>of</strong> marks indicate the measured luminance<br />
values <strong>for</strong> each channel. In Fig. 2, the shape <strong>of</strong> the<br />
electro-optical transfer function <strong>for</strong> mobile display is different<br />
from a power-curved shape <strong>of</strong> CRT display or S-curved<br />
shape <strong>of</strong> LCD display. As the input digital value moves toward<br />
the middle point, the gradient <strong>of</strong> the luminance curve<br />
rapidly increases and immediately decreases, producing a<br />
saturation region. This is due to the adjustment <strong>of</strong> the luminance<br />
curve by the system designer intended to enhance<br />
the contrast ratio and overcome the low channel-bit number.<br />
As a result, a conventional GOG model or S-curve model<br />
does not follow the luminance curve <strong>of</strong> the saturation region<br />
and is not directly applied to mobile LCD. There<strong>for</strong>e, we<br />
used the sigmoid function to model the electro-optical<br />
transfer function <strong>of</strong> mobile LCD based on visual observation<br />
<strong>of</strong> the luminance curve. The sigmoid function is expressed<br />
as<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 351
Son et al.: Real-time color matching between camera and LCD based on 16-bit lookup table design in mobile phone<br />
Table I. Estimated parameter <strong>of</strong> sigmoid function.<br />
a-parameter<br />
c-parameter<br />
R-channel 11.9149 0.4647<br />
G-channel 11.1892 0.4508<br />
B-channel 11.3273 0.4359<br />
Table II. Per<strong>for</strong>mance <strong>of</strong> mobile LCD characterization with various methods.<br />
*<br />
Average E ab<br />
*<br />
Maximum E ab<br />
GOG model 15.655 32.4424<br />
GOG model except<br />
8.7614 17.5898<br />
saturation region<br />
S-curve model 6.9801 15.2279<br />
Sigmoid model 3.9683 14.6831<br />
1<br />
sigmoidx,a,c =<br />
1 + exp− ax − c . 5<br />
The sigmoid function is a symmetrical function with respect<br />
to c and is a constant value if it is zero. The shape <strong>of</strong> a<br />
sigmoid function depends on the absolute value <strong>of</strong> a, and as<br />
the absolute value <strong>of</strong> a increases, the gradient <strong>of</strong> the sigmoid<br />
function rapidly increases with respect to c. Figure 2(d)<br />
shows the electro-optical transfer function resulting from<br />
the sigmoid model. In Fig. 2(d), the estimated luminance<br />
curve closely follows the measured luminance value and it is<br />
predicted that the estimation error will be reduced. The estimated<br />
coefficients <strong>of</strong> the sigmoid function are shown in<br />
Table I. The estimated curve is nearly symmetric with respect<br />
to 0.45 and the absolute value <strong>of</strong> a to determine the<br />
shape <strong>of</strong> the sigmoid function is independent <strong>of</strong> channel and<br />
is almost the same.<br />
To evaluate the per<strong>for</strong>mance <strong>of</strong> each method, the<br />
CIE1976 color difference E * ab was used to measure the<br />
characterization error, which is the Euclidian distance between<br />
estimated CIELAB value and measured CIELAB<br />
value. Sixty-four patches were tested and Table II shows the<br />
characterization error <strong>of</strong> various model-based methods. The<br />
GOG model had the largest color difference and the characterization<br />
error was still severe although the GOG model was<br />
used except in the saturation region. For the S-curve model,<br />
*<br />
the average E ab was approximately 6.9 and is normal color<br />
difference. However, in the middle region, the estimated luminance<br />
value shows a significant difference compared with<br />
measured luminance value. The sigmoid model has a good<br />
average color difference smaller than 6.0, which is indistinguishable<br />
in human vision.<br />
DECISION OF POLYNOMIAL ORDER FOR THE<br />
CHARACTERIZATION OF MOBILE CAMERA<br />
The camera characterization is to find the relationship between<br />
the tristimulus value and digital RGB value. Through<br />
the accurate camera characterization, we can get in<strong>for</strong>mation<br />
about an object color in real scene and reproduce the object<br />
color on mobile LCD. The general procedure <strong>of</strong> camera<br />
characterization is shown in Figure 3. First, a standard color<br />
chart such as a Macbeth or Gretag Color Chart is placed<br />
with 0/45° geometry in a lighting booth, where an illuminant<br />
is set at D65 to reflect the perceived color corresponding<br />
to a daylight condition. 8 The standard color chart<br />
is then captured by a mobile camera set with aut<strong>of</strong>ocusing to<br />
avoid color clipping. Captured RGB digital values <strong>of</strong> each<br />
patch in the standard color chart are averaged to reduce the<br />
camera noise and nonuni<strong>for</strong>mity <strong>of</strong> illumination. Next, the<br />
tristimulus value <strong>of</strong> the standard color chart is acquired by<br />
measuring each patch <strong>of</strong> the color chart or standard data<br />
provided by the manufacturer. Finally, polynomial regression<br />
with least square fitting is applied to find the relationship<br />
between captured RGB digital values and measured<br />
tristimulus values. 9,10<br />
In general, the per<strong>for</strong>mance <strong>of</strong> camera characterization<br />
becomes better as the polynomial order increases. Practically,<br />
Figure 3. The procedure <strong>for</strong> mobile camera characterization.<br />
352 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Son et al.: Real-time color matching between camera and LCD based on 16-bit lookup table design in mobile phone<br />
Figure 4. The characteristics <strong>of</strong> cellular camera; a L * vs average RGB<br />
value <strong>of</strong> gray sample, b a * vs R−B, andb * vs G−B.<br />
<strong>for</strong> a higher polynomial order, most estimated tristimulus<br />
values exceed the boundary <strong>of</strong> the maximum lightness and<br />
chroma, and there is difficulty in implementing mobile camera<br />
characterization. This is because the characteristic curve<br />
<strong>of</strong> the mobile camera, i.e., the relationship between the<br />
tristimulus value and digital RGB value, is not analyzed to<br />
suggest an appropriate polynomial order. To determine the<br />
polynomial order, the RGB digital value is manipulated<br />
based on opponent color theory and is compared with the<br />
Figure 5. The characteristics <strong>of</strong> PDA camera; a L * vs average RGB<br />
value <strong>of</strong> gray sample, b a * vs R−B, andb * vs G−B.<br />
CIELAB value. The CIELAB space is an opponent color coordinate<br />
composed <strong>of</strong> a lightness signal and two types <strong>of</strong><br />
color signals obtained by the difference <strong>of</strong> three color signals.<br />
Thus the RGB digital value is trans<strong>for</strong>med into a lightness<br />
signal and two color signals, R+G+B/3, R−B, and<br />
G−B, just as in opponent color space. Figures 4 and 5<br />
show the relationship between the manipulated RGB values<br />
and CIELAB values <strong>for</strong> a cellular camera and PDA camera.<br />
From a visual evaluation, the distribution <strong>of</strong> the measure-<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 353
Son et al.: Real-time color matching between camera and LCD based on 16-bit lookup table design in mobile phone<br />
ment data was found to be slightly dispersed due to nonuni<strong>for</strong>m<br />
illumination intensities according to the spatial position<br />
on the color chart, where the lux-meter measurements<br />
<strong>for</strong> the four corners were 1990 lux, 2055 lux, 1955 lux, and<br />
1922 lux, respectively. Although ef<strong>for</strong>ts were made to correct<br />
the nonuni<strong>for</strong>mity <strong>of</strong> the illumination intensity, the modeling<br />
<strong>of</strong> lux-meter measurements according to their distance<br />
from the center <strong>of</strong> the color chart is not trivial work due to<br />
their random distribution, there<strong>for</strong>e, this issue has been carried<br />
over to future work. However, it was still clear that the<br />
manipulated RGB digital values were roughly linear to the<br />
CIELAB values:<br />
R + G + B L * , R − B a * , G − B b * . 6<br />
3<br />
There<strong>for</strong>e, a first-order polynomial is adopted, and mathematical<br />
modeling <strong>of</strong> mobile camera characterization is expressed<br />
as linear equations:<br />
L * =1+ L,R R + L,G G + L,B B,<br />
a * =1+ a,R R + a,G G + a,B B,<br />
b * =1+ b,R R + b,G G + b,B B.<br />
Equation (7) can be equally expressed in vector <strong>for</strong>m,<br />
P = V T ,<br />
7<br />
1 , ...,1 n<br />
L,1 , a,1 , b,1<br />
R 1 , ...R n L,2 , a,2 , b,2<br />
V =1 = 8<br />
G 1 , ...G n L,3 , a,3 , b,3<br />
B 1 , ...B n, L,4 , a,4 , b,4,<br />
1<br />
P =L * , a * *<br />
1 , b 1<br />
· · ·<br />
· · ·<br />
L * n , a * n , b<br />
*,<br />
n<br />
where n is the patch number <strong>of</strong> color chart. The ultimate<br />
goal <strong>of</strong> mobile camera characterization is deriving the coefficients<br />
<strong>of</strong> the linear equation, which can be obtained by<br />
pseudoinverse trans<strong>for</strong>mation <strong>of</strong> Eq. (8):<br />
= VV T −1 VP.<br />
Using the derived coefficients, an arbitrary captured digital<br />
value can be converted into the CIELAB value. However,<br />
some <strong>of</strong> the estimated CIELAB values may exceed the maximum<br />
value <strong>of</strong> CIELAB space caused by the error <strong>of</strong> linear<br />
regression. To solve this problem, the lightness value is subtracted<br />
from the amount <strong>of</strong> excessive lightness, and two<br />
color signals are linearly compressed while preserving their<br />
hue value:<br />
9<br />
Table III. Estimation errors <strong>of</strong> mobile camera characterization.<br />
Cellular camera<br />
Samsung SCH-100<br />
PDA camera<br />
Samsung SPH-M400<br />
*<br />
Average E ab<br />
L * = L * *<br />
− L max − 100,<br />
*<br />
Maximum E ab<br />
4.3605 12.2098<br />
6.2638 16.8828<br />
10<br />
a * = k a * *<br />
/a max , b * = b * /a * a * , 11<br />
where k is the constant value <strong>for</strong> color-signal compression.<br />
L max and a max are the estimated maximum lightness value<br />
and color signal value, respectively. Table III shows the per<strong>for</strong>mance<br />
<strong>of</strong> mobile camera characterization <strong>for</strong> the color<br />
chart; the PDA camera shows poorer per<strong>for</strong>mance than the<br />
cellular camera. When observing the moving picture transmitted<br />
from the mobile camera, the PDA camera is subject<br />
to more noise than the cellular camera, which will produce a<br />
large characterization error.<br />
REAL-TIME COLOR MATCHING BETWEEN MOBILE<br />
CAMERA AND MOBILE LCD BASED ON 16-bit<br />
LUT DESIGN INCLUDING NOISE REDUCTION<br />
The process <strong>of</strong> colorimetric color matching reproduces the<br />
same CIE chromaticity and relative luminance compared<br />
with the original color, and has been widely applied to imaging<br />
devices such as monitors, printers, and scanners based<br />
on the ICC pr<strong>of</strong>ile, but not to mobile phones because mobile<br />
phones have been considered to be primarily communication<br />
devices. However, with multimedia convergence, mobile<br />
manufacturers have become aware <strong>of</strong> the necessity <strong>of</strong> color<br />
matching between mobile camera and mobile LCD. With the<br />
characterization <strong>of</strong> mobile camera and mobile LCD, to<br />
implement the color matching system, achievable ranges <strong>of</strong><br />
colors (gamut) must be considered. Figure 6 shows the<br />
gamut difference between a mobile camera under D65 environment<br />
(point) and a mobile LCD (solid color). As shown<br />
in Fig. 6, the gamut <strong>of</strong> mobile camera is larger than that <strong>of</strong><br />
mobile LCD, and has a regular <strong>for</strong>m resulting from the use<br />
<strong>of</strong> linear equations. Thus, significant parts <strong>of</strong> the mobile<br />
camera gamut can be unachievable by the gamut <strong>of</strong> mobile<br />
LCD, and it is necessary to alter the original colors (mobile<br />
camera) to ones that a given output medium (mobile LCD)<br />
is capable <strong>of</strong> reproducing. This power is frequently referred<br />
to as gamut mapping. In this paper, gamut mapping with<br />
variable and multiple anchor points is used to reduce any<br />
sudden color changes on the gamut region boundary and<br />
increase the lightness range reduced in conventional gamut<br />
mapping toward an anchor point. 11<br />
In general, the per<strong>for</strong>mance <strong>of</strong> colorimetric color<br />
matching between cross media, such as a monitor and<br />
printer, depends on the gamut mapping and device characterization.<br />
However, in the case <strong>of</strong> a mobile camera, various<br />
camera noises, such as CCD noise and thermal noise, can be<br />
354 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Son et al.: Real-time color matching between camera and LCD based on 16-bit lookup table design in mobile phone<br />
Figure 6. Gamut mismatches between mobile camera point and mobile<br />
LCD solid color; a projected to a * ,b * plane and b projected<br />
to L * ,b * plane.<br />
included in moving pictures and become amplified after<br />
color matching, especially in the dark regions <strong>of</strong> the achromatic<br />
axis, although not in the chromatic region due to the<br />
blending <strong>of</strong> the reproduced signal. To solve this problem, the<br />
combination <strong>of</strong> two sigmoid functions with different parameters<br />
is applied to the lightness component <strong>of</strong> gamutmapped<br />
tristimulus values to change the contrast ratio. The<br />
sigmoid function is expressed as<br />
n=i<br />
1<br />
S i = e<br />
n=0 −100x n /m − x 0 2 /2 2 ,<br />
2<br />
i =1,2,. .,m,<br />
S i − minS<br />
S LUT =<br />
maxS − minS L *<br />
*<br />
*<br />
max out − L min out + L min out .<br />
12<br />
13<br />
Equation (12) is a discrete cumulative normal function S,<br />
where x 0 and are the mean and standard deviation <strong>of</strong> the<br />
normal distribution, respectively, and m is the number <strong>of</strong><br />
points used in the discrete lookup table. X n is the gamutmapped<br />
lightness component <strong>of</strong> CIELAB values and this<br />
value is then scaled into dynamic range <strong>of</strong> the mobile LCD,<br />
*<br />
*<br />
as given in Eq. (13), where L min out and L max out are blackpoint<br />
and white-point lightness value <strong>of</strong> the mobile LCD. In<br />
Eq. (12), x 0 controls the centering <strong>of</strong> the sigmoid function,<br />
Figure 7. Modified sigmoid function <strong>for</strong> noise reduction; a sigmoid<br />
functions with different parameters and b the combination <strong>of</strong> two sigmoid<br />
functions.<br />
and controls the shape. To find the parameters to conceal<br />
the camera noise through the lightness remapping, visual<br />
experiments were repeated based on adjustment <strong>of</strong> two parameters,<br />
and thus we found that the combination <strong>of</strong> two<br />
sigmoid functions is needed. In Figure 7(a), the solid line<br />
with x 0 =30 and =11.025 is the optimal curve to reduce<br />
the camera noise in the dark region, yet remapped lightness<br />
values in the bright region are significantly increased, <strong>for</strong>ming<br />
a saturation region. Thus another sigmoid function with<br />
x 0 =40 and =27.35, represented by dotted line in Fig. 7(a),<br />
is applied to the gamut-mapped light values larger than the<br />
input lightness value <strong>of</strong> 20 in order to make the lightness<br />
value <strong>of</strong> the reproduced image similar to that <strong>of</strong> the original<br />
image. Ultimately, the combination <strong>of</strong> two sigmoid functions,<br />
as shown in Fig. 7(b) expressed by the solid line, decreases<br />
the lightness value <strong>of</strong> camera noise in the dark region<br />
<strong>of</strong> achromatic axis, and from this result, amplified camera<br />
noise is hardly observed by a human eye.<br />
This kind <strong>of</strong> serial processing mentioned above, including<br />
the characterization, gamut mapping, and noise reduc-<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 355
Son et al.: Real-time color matching between camera and LCD based on 16-bit lookup table design in mobile phone<br />
Table IV. Example <strong>of</strong> bit quantization.<br />
Table V. The example <strong>of</strong> proposed 3D-RGB LUT.<br />
6-bit quantization<br />
8-bit input data && 8-bit masking data<br />
R G B R G B<br />
R-channel: 00110<br />
G-channel: 011001<br />
B-channel: 10011<br />
00110011 && 11111000<br />
01100110 && 11111100<br />
10011001 && 11111000<br />
tion, has computational complexity and is not appropriate<br />
<strong>for</strong> real-time processing. There<strong>for</strong>e, a 3D-RGB LUT is constructed<br />
based on N-grid points <strong>for</strong> each channel. The input<br />
RGB digital values are uni<strong>for</strong>mly sampled by nnn grid<br />
points, which are processed by serial color matching, resulting<br />
in new corresponding output RGB values. The input<br />
RGB digital value and output RGB digital values are stored<br />
in the 3D-LUT and arbitrary input RGB values are calculated<br />
by interpolation. This 3D-LUT can be inserted into the<br />
mobile LCD without any difficulties associated with memory<br />
and computation. In actuality, in a mobile phone, a moving<br />
picture has 8-bits per channel, while the displayed RGB image<br />
on the LCD screen is represented by 5,6,5 bits per<br />
channel. Thus, be<strong>for</strong>e displaying an image on the LCD<br />
screen, 24-bit moving picture data is quantized into 16-bit<br />
data through a bit operation used in program language. For<br />
26 64 0 24 64 17<br />
32 64 0 25 64 18<br />
0 0 6 3 3 4<br />
6 0 6 10 3 2<br />
13 0 6 14 0 1<br />
19 0 6 17 0 1<br />
26 0 6 21 0 3<br />
32 0 6 23 0 4<br />
0 13 6 0 19 4<br />
example, suppose that the moving picture data to be displayed<br />
is (51,102,153). The final data are calculated by applying<br />
the AND operation (&&) with 8-bit masking data.<br />
Table IV shows an example <strong>of</strong> the AND operation.<br />
EXPERIMENTS AND RESULTS<br />
To conduct a subjective experiment <strong>of</strong> colorimetric color<br />
matching, test images were captured using a mobile camera;<br />
these included both face image and color chart images cap-<br />
Figure 8. The experimental results with the cellular phone; a and b device-dependent color matching, c<br />
and d proposed color matching.<br />
356 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Son et al.: Real-time color matching between camera and LCD based on 16-bit lookup table design in mobile phone<br />
Figure 9. The experimental results with the PDA camera; a and b device-dependent color matching, c<br />
and d proposed color matching.<br />
tured in a lighting booth with D65 illumination. Statistically,<br />
the face image is one <strong>of</strong> the most frequently captured images,<br />
and people are very sensitive to their skin color displayed<br />
on mobile LCD. For this reason, a face image representing<br />
the skin color was selected as a test image. Similarly,<br />
the reason why the color chart image captured under D65<br />
illumination was used as a test image was that the characterization<br />
<strong>of</strong> the mobile camera was conducted under D65<br />
illumination, and subjective per<strong>for</strong>mance <strong>of</strong> color matching<br />
can be easily evaluated by comparing the displayed image<br />
with the real object as seen in the lighting booth. In addition,<br />
device-dependent color matching was compared to<br />
evaluate the per<strong>for</strong>mance <strong>of</strong> colorimetric color matching.<br />
Device-dependent color matching directly transmits the captured<br />
image to mobile LCD, while colorimetric color matching<br />
sends the captured image through the 3D-RGB LUT,<br />
which is quantized and transmitted to the mobile LCD.<br />
Table V shows a part <strong>of</strong> the data set stored in the 3D-RGB<br />
LUT designed to the 16-bit system. In the R channel and B<br />
channel, the maximum digital value is 2 5 , whereas<br />
G-channel’s maximum digital value is 2 6 . Based on the<br />
16-bit LUT, colorimetric color matching between mobile<br />
camera and mobile LCD can be processed in real time.<br />
SUBJECTIVE EXPERIMENT OF DEVELOPED COLOR<br />
MATCHING BASED ON 16-bit LUT DESIGN<br />
Figure 8 shows the captured images that are displayed on the<br />
cellular phone. Figures 8(a) and 8(b) show the images resulting<br />
from device-dependent color matching, while Figs.<br />
8(c) and 8(d) show the images resulting from LUT-based<br />
colorimetric color matching. In Fig. 8(a), even though the<br />
picture is taken against the light, the face region is very<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 357
Son et al.: Real-time color matching between camera and LCD based on 16-bit lookup table design in mobile phone<br />
Figure 11. Quality evaluations <strong>of</strong> device-dependent color matching and<br />
proposed color matching.<br />
seen in the D65 daylight, especially the red and green hues.<br />
Figure 9 shows the results <strong>of</strong> color matching <strong>for</strong> a PDA<br />
phone; the same effect is shown. Figure 10 shows the resulting<br />
images <strong>of</strong> colorimetric color matching considering the<br />
noise reduction. Figure 10(a) is the resulting image obtained<br />
by conventional colorimetric color matching, and its image<br />
quality is significantly degraded by the camera noise. By applying<br />
the combination <strong>of</strong> two sigmoid functions in Fig.<br />
7(b) to conventional colorimetric matching, the contrast ratio<br />
<strong>of</strong> reproduced image is changed and from this result,<br />
camera noise is not observable to the human eye, as shown<br />
in Fig. 10(b). Consequently, color matching based on 3D-<br />
LUT accurately reproduces the object color seen in the real<br />
scene and thus improves the color fidelity <strong>of</strong> the mobile<br />
display. For moving pictures, the same results decreasing the<br />
camera noise can be achieved with no problems <strong>of</strong> computation<br />
and memory.<br />
Figure 10. The results <strong>of</strong> noise reduction; a be<strong>for</strong>e lightness remapping<br />
and b after lightness remapping.<br />
bright due to the tendency <strong>of</strong> the electro-optical transfer<br />
function <strong>of</strong> mobile LCD to saturate the bright region as<br />
shown. In addition, the colorfulness <strong>of</strong> “table” and “cloth”<br />
region is more decreased than the original color, and the<br />
image quality is degraded. As shown in Fig. 8(c), the skin<br />
color in the “face” region is more natural and realistic than<br />
in Fig. 8(a), and the object colors such as “cloth” and “table”<br />
arewellreproducedonLCD.FortheMacbethColorchart<br />
seen in Fig. 8(b), colors <strong>of</strong> each patch are washed out and<br />
exhibit major differences in appearance, compared with<br />
original color seen in the D65 lighting booth, because<br />
device-dependent color matching is only physical data transmission.<br />
On the other hand, the result shown in Fig. 8(d)<br />
adquately represents the colorfulness <strong>of</strong> the original color<br />
QUANTITATIVE EVALUATION OF THE DEVELOPED<br />
COLOR MATCHING<br />
To evaluate colorimetric color matching based on 16-bit<br />
RGB LUT, a Macbeth Color Chart composed <strong>of</strong> 24 patches<br />
was used as a test image. For quantitative evaluation <strong>of</strong> the<br />
device dependent color matching, the Macbeth Color Chart<br />
is previously captured in the D65 lighting booth, and is displayed<br />
on mobile LCD. Then, the CIELAB value <strong>of</strong> each<br />
patch is measured using a colorimeter and compared with<br />
the CIELAB data <strong>of</strong> the Macbeth Color Chart measured in<br />
the D65 lighting booth, thus calculating the CIE 1976 color<br />
difference. For the proposed color matching, the Macbeth<br />
Color Chart is again captured in the D65 lighting booth, and<br />
is displayed on mobile LCD through use <strong>of</strong> the 16-bit RGB<br />
LUT. Then, the CIELAB value <strong>of</strong> each patch is measured<br />
using a colorimeter and compared with the CIELAB data <strong>of</strong><br />
Macbeth Color Chart measured in the D65 lighting booth.<br />
Figure 11 shows the result <strong>of</strong> quantitative evaluation using<br />
1976 Color difference. In Fig. 11, several patches corresponding<br />
to the proposed color matching have a larger color<br />
difference than <strong>for</strong> the conventional method, due to the<br />
characterization errors <strong>of</strong> the mobile camera and mobile<br />
LCD. However, the proposed color matching has a lower<br />
average color difference <strong>of</strong> 15.56, whereas device-dependent<br />
color matching has the average color difference <strong>of</strong> 24.395.<br />
358 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Son et al.: Real-time color matching between camera and LCD based on 16-bit lookup table design in mobile phone<br />
There<strong>for</strong>e, the proposed color matching achieves better colorimetric<br />
reproduction than the conventional method, and it<br />
is concluded that object color transmitted by mobile camera<br />
in real time can be accurately and realistically reproduced on<br />
a mobile LCD.<br />
CONCLUSIONS<br />
This paper presented a method <strong>for</strong> real-time color matching<br />
between mobile camera and mobile LCD, involving characterization<br />
<strong>of</strong> the mobile camera and mobile LCD, gamut<br />
mapping, noise reduction, and a LUT design. The characterization<br />
<strong>of</strong> the mobile LCD is conducted based on 16-bit<br />
processing, plus a sigmoid function is used to estimate the<br />
electro-optical transfer function. Meanwhile, <strong>for</strong> the characterization<br />
<strong>of</strong> the mobile camera, the optimal polynomial order<br />
is determined by trans<strong>for</strong>ming the captured RGB data<br />
into opponent color space and finding the relationship between<br />
the trans<strong>for</strong>med RGB values and the measured<br />
CIELAB values. Following the two types <strong>of</strong> characterization,<br />
gamut mapping is executed to overcome the gamut difference<br />
between the mobile camera and the mobile LCD, then<br />
noise reduction processing is applied to the lightness component<br />
<strong>of</strong> the gamut-mapped CIELAB values. Finally, to reduce<br />
the complex computation <strong>of</strong> serial-based color matching,<br />
a 3D RGB LUT is designed based on 16-bit data and<br />
inserted into the mobile phone. Experiments demonstrated<br />
that the proposed color matching realistically reproduced<br />
object colors from a real scene on a mobile LCD and improved<br />
the fidelity color <strong>of</strong> the mobile display. The LUT was<br />
also designed without any further computation or memory<br />
burden, making real-time processing possible.<br />
Acknowledgments<br />
This work is financially supported by the Ministry <strong>of</strong> Education<br />
and Human Resources Development (MOE), the<br />
Ministry <strong>of</strong> Commerce, Industry and Energy (MOCIE), and<br />
the Ministry <strong>of</strong> Labor (MOLAB) through the fostering<br />
project <strong>of</strong> the Lab <strong>of</strong> Excellency.<br />
REFERENCES<br />
1 J. Luo, “Displaying images on mobile device: capabilities, issues, and<br />
solutions”, Wirel. Commun. Mob. Comput. 2, 585–594 (2002).<br />
2 J. Y. Hardeberg, Acquisition and reproduction <strong>of</strong> color images:<br />
colorimetric and multispectral approaches, Universal Publishers,<br />
Dissertation.com, 2001.<br />
3 H. R. Kang, Color Technology <strong>for</strong> Electronic Image Device (SPIE Optical<br />
Engineering Press, Bellingham, WA, 1996).<br />
4 R. S. Berns, “Methods <strong>for</strong> characterizing CRT displays”, Displays 16,<br />
173–182 (1996).<br />
5 Y. S. Kwak and L. W. MacDonald, “Characterisation <strong>of</strong> a desktop LCD<br />
projector”, Displays 21, 179–194 (2000).<br />
6 N. Tamura, N. Tusmura, and Y. Miyake, “Masking model <strong>for</strong> accurate<br />
colorimetric characterization <strong>of</strong> LCD”, Proc. IS&T/SID Tenth Color<br />
<strong>Imaging</strong> Conference (IS&T, Jmigtiel, VA, 2002), 312–316.<br />
7 G. Sharma, “LCD versus CRTs color-calibration and gamut<br />
consideration”, Proc. IEEE 90, 605–622 (2002).<br />
8 M. D. Fairchild, Color Appearance Models (Addison-Wesley, Reading,<br />
MA, 1998).<br />
9 G. Hong, M. R. Luo, and P. A. Ronnier, “A study <strong>of</strong> digital camera<br />
colorimetric characterization based on polynomial modeling”, Color<br />
Res. Appl. 26, 76–84 (2001).<br />
10 M. R. Pointer, G. G. Attridge, and R. E. Jacobson, “Practical camera<br />
characterization <strong>for</strong> colour measurement”, <strong>Imaging</strong> Sci. J., 49, 63–80<br />
(2001).<br />
11 C. S. Lee, Y. W. Park, S. J. Cho, and Y. H. Ha, “Gamut mapping<br />
algorithm using lightness mapping and multiple anchor points <strong>for</strong> linear<br />
tone and maximum chroma reproduction”, J. <strong>Imaging</strong> Sci. Technol. 45,<br />
209–223 (2001).<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 359
<strong>Journal</strong> <strong>of</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology® 51(4): 360–367, 2007.<br />
© <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology 2007<br />
Solving Under-Determined Models in Linear Spectral<br />
Unmixing <strong>of</strong> Satellite Images: Mix-Unmix Concept<br />
(Advance Report)<br />
Thomas G. Ngigi and Ryutaro Tateishi<br />
Center <strong>for</strong> Environmental Remote Sensing, Chiba University, 1-33 Yayoi, Inage, Chiba, 263-8522, Japan<br />
E-mail: tgngigi@hotmail.com<br />
Abstract. This paper reports on a simple novel concept <strong>of</strong> addressing<br />
the problem <strong>of</strong> underdetermination in linear spectral unmixing.<br />
Most conventional unmixing techniques fix the number <strong>of</strong> endmembers<br />
on the dimensionality <strong>of</strong> the data, and none <strong>of</strong> them can<br />
derive multiple 2 + end-members from a single band. The concept<br />
overcomes the two limitations. Further, the concept creates a processing<br />
environment that allows any pixel to be unmixed without any<br />
sort <strong>of</strong> restrictions (e.g., minimum determinable fraction), impracticalities<br />
(e.g., negative fractions), or trade-<strong>of</strong>fs (e.g., either positivity<br />
or unity sum) that may be associated with conventional unmixing<br />
techniques. The proposed mix-unmix concept is used to generate<br />
fraction images <strong>of</strong> four spectral classes from Landsat 7 ETM+data<br />
(aggregately resampled to 240 m) first principal component only.<br />
The correlation coefficients <strong>of</strong> the mix-unmix image fractions versus<br />
reference image fractions <strong>of</strong> the four end-members are 0.88, 0.80,<br />
0.67, and 0.78. © 2007 <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and<br />
Technology.<br />
DOI: 10.2352/J.<strong>Imaging</strong>Sci.Technol.200751:4360<br />
PROBLEM STATEMENT / INTRODUCTION, AND<br />
OBJECTIVE<br />
“… the number <strong>of</strong> bands must be more than the number <strong>of</strong><br />
end-members…” is perhaps the most ubiquitous statement<br />
in the field <strong>of</strong> linear spectral unmixing. This is simply because<br />
most <strong>of</strong> the conventional unmixing techniques are<br />
based on least squares, 1 convex geometry, 2 or combination<br />
<strong>of</strong> both and the number <strong>of</strong> end-members (unknowns) is<br />
dependent on the dimensionality (equations) <strong>of</strong> the data.<br />
Least squares can unmix as many end-members as up to the<br />
dimensionality <strong>of</strong> the data, and at the very best exceed by<br />
one when the unity constraint is en<strong>for</strong>ced. In convex geometry,<br />
the number <strong>of</strong> determinable end-members (at the<br />
unmixing stage) is equal to the number <strong>of</strong> vertices <strong>of</strong> the<br />
data simplex, and this number exceeds the dimensionality <strong>of</strong><br />
the data by one. After extracting the end-member spectra,<br />
most <strong>of</strong> the convex geometry-based techniques apply the<br />
least squares approach (combined case) in computing the<br />
fractions <strong>of</strong> the end-members.<br />
Some linear spectral unmixing techniques include Sequential<br />
Maximum Angle Convex Cone (SMACC) Spectral<br />
Tool, 3 (Generalized) Orthogonal Subspace Projection, 4,5<br />
Convex Cone Analysis, 6 N-FINDR, 7 Orasis, 7 and Iterative<br />
Error Analysis. 7 Keshava 8 gives a detailed account <strong>of</strong> spectral<br />
unmixing techniques. A number <strong>of</strong> commercially available<br />
s<strong>of</strong>tware, including ENVI, IDRISI Kilimanjaro, PCI, and<br />
ERDAS Imagine, have linear spectral unmixing modules.<br />
The greatest fundamental commonality <strong>of</strong> all conventional<br />
linear spectral unmixing techniques is that none <strong>of</strong> them can<br />
derive multiple end-members 2 + from a single band. The<br />
object <strong>of</strong> the mix-unmix concept is to overcome this problem<br />
and unmix as many end-members as can be deciphered<br />
from the reference data and without introducing any sort <strong>of</strong><br />
restrictions, impracticalities, or trade-<strong>of</strong>fs that may be associated<br />
with conventional unmixing techniques.<br />
DESCRIPTION OF THE MIX-UNMIX CONCEPT<br />
As the term implies, the model consists <strong>of</strong> two branches,<br />
namely, mixing and unmixing. The mixing branch entails<br />
development <strong>of</strong> hypothetical mixed pixels on the basis <strong>of</strong><br />
desired end-members’ actual digital numbers (DNs).<br />
Unmixing involves determination <strong>of</strong> each real image pixel’s<br />
DN’s contributory end-members and their fractions by<br />
back-propagating through the mixing branch using a pixel<br />
<strong>of</strong> the same DN in the hypothetical image as a proxy. This<br />
preliminary study demonstrates the concept on a single<br />
simulated band.<br />
Mixing Branch<br />
Nominally, the end-members are paired up hierarchically<br />
into a single hypothetical mixed class (Figure 1;<br />
EM1=end-member 1, EM1.2=combined end-members 1<br />
and 2). Essentially, in pairing up, each and every DN from a<br />
member <strong>of</strong> a pair is combined with each and every DN from<br />
the other member, at complementary percentages ranging<br />
from 0% to 100%, giving rise to various “mixture tables”<br />
(MTs) whose number depends on the ranges <strong>of</strong> training<br />
DNs <strong>of</strong> the two members.<br />
Theory <strong>of</strong> the mixing branch and <strong>for</strong>mation <strong>of</strong> mixture<br />
tables (MTs)<br />
The number <strong>of</strong> possible DN combinations, MTs, <strong>of</strong> two<br />
members, A and B, <strong>of</strong> a pair is equal to the product <strong>of</strong> their<br />
training DN ranges, i.e.,<br />
MTs = A max DN − A min DN + 1 B max DN<br />
Received Dec. 5, 2006; accepted <strong>for</strong> publication Mar. 22, 2007.<br />
1062-3701/2007/514/360/8/$20.00.<br />
where<br />
− B min DN + 1,<br />
360
Ngigi and Tateishi: Solving under-determined models in linear spectral unmixing <strong>of</strong> satellite images: mix-unmix concept<br />
Figure 1. Bottom-up pairing up <strong>of</strong> end-members as well as the resultant super-end-members—pairing up<br />
involves mixing the two members-<strong>of</strong>-a-pair’s DN ranges at all complementary fractions. In case the number <strong>of</strong><br />
super-end-members is even but not a multiple <strong>of</strong> four, one level <strong>of</strong> mixing is skipped <strong>for</strong> one pair as indicated<br />
by EM5 and EM6 <strong>for</strong> six end-members. For an odd number, one end-member is simply carried <strong>for</strong>ward to the<br />
next level individually as indicated by EM7 <strong>for</strong> seven end-members. In this case, EM5.6 and EM7 have the<br />
same hierarchical status as EM1.2.3.4. At the base <strong>of</strong> the branch are training DN ranges and assumed<br />
fractions <strong>of</strong> the end-members—the DNs are known by in situ observation, from spectral libraries, or identification<br />
<strong>of</strong> pure pixels in the image to be unmixed, etc. At the top <strong>of</strong> the branch are hypothetical pixels’ DN values<br />
resulting from mixing all the end-members’ spectra at all possible complementary fractions.<br />
A max DN = maximum DN <strong>of</strong> A,<br />
A min DN = minimum DN <strong>of</strong> A,<br />
Subsequently, the total number <strong>of</strong> possible DNs and percentages<br />
combinations <strong>of</strong> the two is<br />
MTs N%s.<br />
The expression also gives the total number <strong>of</strong> possible mixture<br />
pixels <strong>of</strong> the two.<br />
Thus all pixels, in a hypothetical band, composed <strong>of</strong><br />
only two end-members, EM1 and EM2, would be defined by<br />
the following expression—discussed assuming: that the endmembers’<br />
training DNs range from, respectively, 10–89 and<br />
90–150 in the band; a mixture interval <strong>of</strong> 10%, and assuming<br />
linear mixing.<br />
B max DN = maximum DN <strong>of</strong> B,<br />
B min DN = minimum DN <strong>of</strong> B.<br />
The number <strong>of</strong> possible percentages combinations, N%s, <strong>of</strong><br />
the two members is given by<br />
1<br />
N%s=100 % ÷ MI +1,<br />
where<br />
MI = adopted mixture interval.<br />
where<br />
• f 1,i =percentage <strong>of</strong> EM1 in pixel i Table Ia 1st<br />
column,<br />
• f 2,i =percentage <strong>of</strong> EM2 in pixel i Table Ia 2nd<br />
column,<br />
• f 1 +f 2 =100%,<br />
• DN 1,i =DN <strong>of</strong> EM1 in pixel i Table Ia 2nd row,<br />
• DN 2,i =DN <strong>of</strong> EM2 in pixel i Table Ia 3rd row,<br />
• DN 1,2,i =mixture DN <strong>of</strong> DN 1,i and DN 2,i in pixel i<br />
Table Ia all cells excluding the first two columns<br />
and first three rows,<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 361
Ngigi and Tateishi: Solving under-determined models in linear spectral unmixing <strong>of</strong> satellite images: mix-unmix concept<br />
Table I. a MTs <strong>of</strong> EM1 and EM2. The EM1.2 DNs are computed as: EM1.2 DN=EM1 % EM1 DN +EM2 % EM2 DN. Note that EM1% and<br />
EM2% are complementary.<br />
MT= 1 2 3−<br />
4 880 4 881<br />
4 879<br />
EM1 DN= 10 11 88 89<br />
EM2 DN= 90 90 150 150<br />
EM1% EM2% EM1.2 DN<br />
0 100 90 90 150 150<br />
10 90 82 82 143 143<br />
20 80 74 74 137 137<br />
30 70 66 66 131 131<br />
40 60 58 58 125 125<br />
50 50 50 51 119 119<br />
60 40 42 43 113 113<br />
70 30 34 35 107 107<br />
80 20 26 27 100 101<br />
90 10 18 19 94 95<br />
100 0 10 11 88 89<br />
b Min-max MTs <strong>of</strong> EM1 and EM2.<br />
c Min-max LUTs <strong>of</strong> EM1 and EM2.<br />
Min-MT<br />
EM1 DN= 10<br />
EM2 DN= 90<br />
Min-LUT<br />
Max-MT<br />
EM1 DN= 89<br />
EM2 DN= 150<br />
Max-LUT<br />
EM1 % EM2 %<br />
EM1.2 DN<br />
Amount <strong>of</strong><br />
overlap with<br />
EM1.2 vector<br />
90–150<br />
0 100 90 150 60<br />
10 90 82 143 53<br />
20 80 74 137 47<br />
30 70 66 131 41<br />
40 60 58 125 35<br />
50 50 50 119 29<br />
60 40 42 113 23<br />
70 30 34 107 17<br />
80 20 26 101 11<br />
90 10 18 95 5<br />
100 0 10 89 0<br />
• min EM1DN=minimum DN <strong>of</strong> EM1,<br />
• max EM1DN=maximum DN <strong>of</strong> EM1.<br />
Bounding mixture tables, and various numbers <strong>of</strong> endmembers<br />
From Table I(a), it is very clear that, given constant fractions<br />
<strong>of</strong> EM1 and EM2, the mixture class DNs (EM1.2 DNs) always<br />
fall between the values in the first and last MTs, thus<br />
the two MTs fully give the ranges <strong>of</strong> all possible DNs <strong>of</strong><br />
EM1.2. Hereinafter, the two are referred to as min-MT and<br />
max-MT, respectively, and min-max MTs collectively [Table<br />
I(b)].<br />
Similarly, min-max MTs <strong>of</strong> the other paired endmembers<br />
are generated: <strong>for</strong> EM3 and EM4; in Eq. (1) EM1,<br />
EM2, and EM1.2 are replaced with EM3, EM4, and EM3.4,<br />
respectively. Table II(a) shows min-max MTs <strong>of</strong> EM3 and<br />
EM4 shows the DN ranges are 151–180 and 181–210, respectively.<br />
Next, second level min-max MTs are developed from<br />
the above first level MTs: <strong>for</strong> EM1.2 and EM3.4; in Eq. (1)<br />
EM1, EM2, and EM1.2 are replaced with EM1.2, EM3.4, and<br />
EM1.2.3.4, respectively. Table III(a) shows min-max MTs <strong>of</strong><br />
EM1.2, and EM3.4. Since EM1.2 represents EM1 and EM2,<br />
and EM3.4 represents EM3 and EM4, subsequently, the second<br />
level min-max MTs inherently represent all the possible<br />
DN outcomes <strong>of</strong> mixing all the end-members EM1, EM2,<br />
EM3, and EM4 at all possible complementary fractions.<br />
362 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Ngigi and Tateishi: Solving under-determined models in linear spectral unmixing <strong>of</strong> satellite images: mix-unmix concept<br />
Table II. a Min-max MTs <strong>of</strong> EM3 and EM4. The EM3.4 DNs are computed as: EM3.4<br />
DN=EM3 % EM3 DN +EM4 % EM4 DN. Note that EM3% and EM4% are<br />
complementary. b Min-max LUTs <strong>of</strong> EM3 and EM4.<br />
Min-MT<br />
EM3 DN= 151<br />
EM4 DN= 181<br />
Min-LUT<br />
Max-MT<br />
EM3 DN= 180<br />
EM4 DN= 210<br />
Max-LUT<br />
Amount <strong>of</strong><br />
overlap with<br />
EM3.4 vector<br />
172–201<br />
EM3 % EM4 % EM3.4 DN<br />
0 100 181 210 20<br />
10 90 178 207 29<br />
20 80 175 204 29<br />
30 70 172 201 29<br />
40 60 169 198 26<br />
50 50 166 195 23<br />
60 40 163 192 20<br />
70 30 160 189 17<br />
80 20 157 186 14<br />
90 10 154 183 11<br />
100 0 151 180 8<br />
For more end-members, the process is successively repeated<br />
as shown in Fig. 1. For three end-members in Eq. (1)<br />
EM1, EM2, and EM1.2 are replaced with EM1.2, EM3, and<br />
EM1.2.3, respectively.<br />
Unmixing Branch<br />
This is similar to the mixing branch (Fig. 1) but with the<br />
arrows (processing) reversed and the MTs renamed look-uptables<br />
(LUTs)—Tables I(c), II(b), and IIIb. As discussed below,<br />
a real image pixel DN is fractionalized into two highest<br />
level super-end-members, each <strong>of</strong> which is then split into its<br />
two constituent end-members. The process continues until<br />
the finest level (end-members <strong>of</strong> interest) from which the<br />
mixing branch was constructed (see Figure 2).<br />
Fractionalization<br />
This discussion demonstrates the unmixing process on a<br />
single-band image composed <strong>of</strong> the four end-members outlined<br />
in the Mixing Branch section. For each DN in the<br />
band, all the vectors in which it lies are identified, e.g., Table<br />
III(b) EM.1.2.3.4 italicized DNs give all the possible vectors<br />
<strong>for</strong> DN 180, with the lower nodes located in Table III(b-1)<br />
and the upper nodes in Table III(b-2)—the first vector is<br />
172–204 (bold). Each one <strong>of</strong> these vectors is a combination<br />
<strong>of</strong> two minor vectors, one apiece from EM1.2 and EM3.4<br />
(italicized); e.g., <strong>for</strong> the vector 172–204, the constituent vectors<br />
are 90–150 (bold) from EM1.2 and 181–210 (bold)<br />
from EM3.4.<br />
The most probable vector (MPV) in which the DN 180<br />
lies is computed as<br />
Figure 2. Top-bottom fractionalization <strong>of</strong> a pixel; first into two highest level super-end-members, then effectively<br />
into second highest level four super-end-members by fractionalizing each <strong>of</strong> the highest level super-endmembers<br />
into two. The process is repeated successively until the lowest level end-members that were used to<br />
build up the mixing branch. At the top <strong>of</strong> the branch is a universe <strong>of</strong> values encompassing all the DNs in the<br />
image to be unmixed—all: assuming that the image is composed <strong>of</strong> only the end-members used in the mixing<br />
branch. At the base <strong>of</strong> the branch are estimated contributory percentages fractions <strong>of</strong> the end-members cf.<br />
Fig. 1.<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 363
Ngigi and Tateishi: Solving under-determined models in linear spectral unmixing <strong>of</strong> satellite images: mix-unmix concept<br />
Table III. a-1 Min-MT <strong>of</strong> EM1.2 and EM3.4. The EM1.2.3.4 DNs are computed as: EM1.2.3.4 DN=EM1.2% EM1.2 DN +EM3.4% EM3.4 DN. Note that EM1.2% and EM3.4% are complementary.<br />
b-1 Min-LUT <strong>of</strong> EM1.2 and EM3.4.<br />
EM3.4 DN b<br />
EM1.2<br />
DN a<br />
EM1.2<br />
%<br />
EM3.4<br />
%<br />
181 178 175 172 169 166 163 160 157 154 151<br />
EM1.2.3.4 DN<br />
90 0 100 181 178 175 172 169 166 163 160 157 154 151<br />
90 10 90 172 169 167 164 161 158 156 153 150 148 145<br />
90 20 80 163 160 158 156 153 151 148 146 144 141 139<br />
90 30 70 154 152 150 147 145 143 141 139 137 135 133<br />
90 40 60 145 143 141 139 137 136 134 132 130 128 127<br />
90 50 50 136 134 133 131 130 128 127 125 124 122 121<br />
90 60 40 126 125 124 123 122 120 119 118 117 116 114<br />
Rows 8 to 117<br />
10 70 30 61 60 60 59 58 57 56 55 54 53 52<br />
10 80 20 44 44 43 42 42 41 41 40 39 39 38<br />
10 90 10 27 27 27 26 26 26 25 25 25 24 24<br />
10 100 0 10 10 10 10 10 10 10 10 10 10 10<br />
a-2 Max-MT <strong>of</strong> EM1.2 and EM3.4.<br />
b-2 Max-LUT <strong>of</strong> EM1.2 and EM3.4.<br />
EM3.4 DN d<br />
EM1.2<br />
DN c<br />
EM1.2<br />
%<br />
EM3.4<br />
%<br />
210 207 204 201 198 195 192 189 186 183 180<br />
EM1.2.3.4 DN<br />
150 0 100 210 207 204 201 198 195 192 189 186 183 180<br />
150 10 90 204 201 199 196 193 191 188 185 182 180 177<br />
150 20 80 198 196 193 191 188 186 184 181 179 176 174<br />
150 30 70 192 190 188 186 184 182 179 177 175 173 171<br />
150 40 60 186 184 182 181 179 177 175 173 172 170 168<br />
150 50 50 180 179 177 176 174 173 171 170 168 167 165<br />
150 60 40 174 173 172 170 169 168 167 166 164 163 162<br />
Rows 8 to 117<br />
89 70 30 125 124 124 123 122 121 120 119 118 117 116<br />
89 80 20 113 113 112 111 111 110 110 109 108 108 107<br />
89 90 10 101 101 100 100 100 99 99 99 99 98 98<br />
89 100 0 89 89 89 89 89 89 89 89 89 89 89<br />
a Column 1 elements are from EM1 and EM2 min-MT Table Ib, column 3<br />
b Row 1 elements are from EM3 and EM4 min-MT Table IIa, column 3<br />
c Column 1 elements are from EM1 and EM2 max-MT Table Ib, column 4<br />
d Row 1 elements are from EM3 and EM4 max-MT Table IIa, column 4<br />
where<br />
n<br />
lower nodes<br />
i=1<br />
cMNxDN =<br />
n<br />
, 2<br />
cMNyDN =<br />
n<br />
upper nodes<br />
i=1<br />
n<br />
, 3<br />
• cMNxDN=lower node <strong>of</strong> DN 180 MPV from combined<br />
classes M and N (EM1.2 or EM3.4),<br />
• cMNyDN=upper node ditto,<br />
• lower nodes=all the EM1.2.3.4 italicized DN in Table<br />
III(b-1),<br />
• upper nodes=ditto Table III(b-2),<br />
• n=number <strong>of</strong> EM1.2.3.4 italicized DN vectors=count<br />
<strong>of</strong> EM1.2.3.4 italicized DN nodes in Table III(b-1) or<br />
Table III(b-2).<br />
From Eqs. (2) and (3), cMNxDN=156 and cMNyDN<br />
=190. From Table III(b), the pair <strong>of</strong> nodes most close to the<br />
pair 156/190 is 156/191 and it is adopted as the MPV <strong>for</strong><br />
the DN 180. This vector 156–191 [Table III(b) bold and<br />
underlined] lies at the intersection <strong>of</strong> EM1.2 vector 90–150<br />
364 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Ngigi and Tateishi: Solving under-determined models in linear spectral unmixing <strong>of</strong> satellite images: mix-unmix concept<br />
given rise to the EM1.2 vector 90–150 is taken to be proportional<br />
to the amount <strong>of</strong> overlap with it. Each probability is<br />
also taken to be the probability <strong>of</strong> the corresponding paired<br />
percentages (PPs) having given rise to the EM1.2 vector 90–<br />
150 since the PV was developed from them (PPs). Similarly<br />
<strong>for</strong> Table II(b) PVs and PPs in the case <strong>of</strong> EM3.4 vector<br />
172–201. There are seven possible overlap scenarios as depicted<br />
by Figure 3. Table I(c) last column gives the weights<br />
<strong>of</strong> the EM1 and EM2 vectors to EM1.2 vector 90–150, and<br />
Table II(b) ditto EM3 and EM4 vectors to EM3.4 vector<br />
172–201.<br />
From Fig. 3 and Tables I(c) and II(b), the most probable<br />
percentage contribution (MPPC) <strong>of</strong> each daughter class to<br />
its mother class is computed as<br />
Figure 3. Possible universal overlap scenarios. A is EMx.y’s most probable<br />
vector, e.g., EM1.2 MPV 90-150 or EM3.4 MPV 172-201; all/<br />
some <strong>of</strong> the other non-arrowed lines B-H are vectors contained in<br />
EMx.y’s min-max LUTs e.g., Table Ic <strong>for</strong> EM1.2, or Table IIb <strong>for</strong><br />
EM3.4; arrowed lines are the respective overlaps.<br />
and EM3.4 vector 172–201. There<strong>for</strong>e, by extension, the DN<br />
180 most probably resulted from these EM1.2 and EM3.4<br />
vectors as the combination most probably gave rise to the<br />
DN 180 MPV 156–191.<br />
Further, percentages-wise, the DN 180 could have resulted<br />
from any <strong>of</strong> the paired percentages associated with the<br />
EM1.2.3.4 italicized DNs vectors. The most probable contributory<br />
paired percentages (MPPC) are computed as<br />
where<br />
MPPC x.y =<br />
n<br />
i% x.y p i <br />
i=1<br />
±<br />
n<br />
n p i<br />
i=1<br />
n<br />
i=1<br />
p i v i<br />
2<br />
n<br />
n 2 <br />
i=1<br />
p i<br />
,<br />
• i% x.y =ith paired percentages <strong>of</strong> x EM1.2<br />
and y EM3.4,<br />
• p i =weight <strong>of</strong> i% x.y =count <strong>of</strong> i% x.y ’s EM1.2.3.4<br />
italicized DNs,<br />
• n=count <strong>of</strong> probable contributory paired<br />
percentages,<br />
• v=i% x.y −MPPC x,y . The second term in Eq. 4<br />
is computed after the first one.<br />
From Eq. (4), EM1.2% =16.67% ±2.34% and EM3.4%<br />
=83.33% ±2.34%. Hence, the DN 180 most probably resulted<br />
from these EM1.2 and EM3.4 percentages combinations<br />
as the pair most probably gave rise to the DN 180<br />
MPV 156-191.<br />
Next, the EM1.2 90-150 and EM3.4 172-201 vectors are<br />
checked against the lower level min-max LUTs, Tables I(c)<br />
and II(b), respectively, and all the vectors with which they<br />
(EM1.2 and EM3.4 vectors) overlap <strong>for</strong>m the universe <strong>of</strong><br />
possible vectors (PVs) from which they (EM1.2 and EM3.4<br />
vectors) or, in other words, a fraction <strong>of</strong> the value 180, arose.<br />
The probability (weight) <strong>of</strong> each <strong>of</strong> the Table I(c) PVs having<br />
4<br />
where<br />
cM %=<br />
q<br />
cM% i p i<br />
i=1<br />
±<br />
q<br />
q p i<br />
i=1<br />
q<br />
i=1<br />
p i v i<br />
2<br />
q<br />
q 2 <br />
i=1<br />
p i<br />
,<br />
5<br />
• cM% =MPPC <strong>of</strong> daughter class cM (EM1 or EM2) to<br />
its mother class (EM1.2). EM3 or EM4 <strong>for</strong> EM3.4;<br />
• cM% i =percent <strong>of</strong> cM’s ith probable vector—Table I(c)<br />
columns 1 and 2 <strong>for</strong> EM1 and EM2, respectively; Table<br />
II(b) columns 1 and 2 <strong>for</strong> EM3 and EM4, respectively;<br />
• p i =overlap range <strong>of</strong> cM’s ith probable vector with its<br />
cM mother’s MPV. Table I(c) last column <strong>for</strong> EM1<br />
and EM2. Table II(b) last column <strong>for</strong> EM3 and EM4;<br />
• q=count <strong>of</strong> probable paired-percentages;<br />
• v=cM%-cM% i . The second term in Eq. (5) is computed<br />
after the first one.<br />
From Eq. (5) and Tables I(c) and II(b), the MPPCs <strong>of</strong><br />
EM1 and EM2 to EM1.2, and EM3 and EM4 to EM3.4 are;<br />
EM1=71% ±2.84%, EM2=29% ±2.84%, EM3<br />
=59% ±2.42%, and EM4=41% ±2.42%.<br />
The MPPC <strong>of</strong> an end-member to the original pixel DN<br />
is simply the product <strong>of</strong> all MPPCs along the path from the<br />
end-member itself to the pixel DN. Hence, <strong>for</strong> end-member:<br />
• 1=EM1% EM1.2% =71% 16.67%<br />
=11.84±1.73%,<br />
• 2=EM2% EM1.2% =29% 16.67%<br />
=04.83±0.83%,<br />
• 3=EM3% EM3.4% =59% 83.33%<br />
=49.16±2.44%,<br />
• 4=EM4% EM3.4% =41% 83.33%<br />
=34.17±2.23%.<br />
The standard deviation <strong>of</strong> product AB is computed as<br />
+ 2<br />
B<br />
, 6<br />
B<br />
AB = AB 2<br />
A<br />
A<br />
where k =standard deviation <strong>of</strong> k.<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 365
Ngigi and Tateishi: Solving under-determined models in linear spectral unmixing <strong>of</strong> satellite images: mix-unmix concept<br />
Figure 4. Left location <strong>of</strong> the study area; right 30 m resolution Landsat<br />
ETM+ data RGB=342.<br />
Figure 5. Left reference data: 30 m resolution spectral classes. The<br />
spectral classes correspond to broad in<strong>for</strong>mation classes dense vegetation<br />
C1, dense vegetation/bare land C2, bare land/dense vegetation<br />
C3, and bare land C4; right raw image: 240 m resolution first<br />
principal component. The rectangles show locations <strong>of</strong> reference black<br />
and mix-unmix red training sites.<br />
STUDY TEST<br />
Four spectral end-members are mapped from a single-band<br />
simulated raw image.<br />
Study Area and Data<br />
Landsat ETM+ data, <strong>of</strong> 21 February 2000, covering<br />
southern-central Kenya is utilized to generate both the reference<br />
and raw data. The land covers in the area transition<br />
from dense vegetation (<strong>for</strong>est) to bare land (Figure 4).<br />
Simulation <strong>of</strong> Reference and Raw Data<br />
K-mean classification is run on the ETM+ data bands 1, 2,<br />
3, 4, 5, and 7 to produce four spectral classes. The spectral<br />
classes are adopted as reference data. The six bands data is<br />
resampled to 240 m resolution (mimics moderate resolution<br />
data, e.g., MODIS bands 1 and 2–250 m resolution) and<br />
then principal components trans<strong>for</strong>mation (PCT) executed<br />
on the new data set. Each <strong>of</strong> the resampled bands and PCs is<br />
taken as a candidate raw image.<br />
Selection <strong>of</strong> Band to Unmix, and its Unmixing<br />
A section from the 30 m resolution reference spectral classes<br />
image, black rectangle in Figure 5, hereinafter referred to as<br />
Figure 6. a Comparison <strong>of</strong> DNs’ distribution curves <strong>of</strong> C1, C2, C3,<br />
and C4 in mix-unmix training site across original bands and principal<br />
components PCs. The least overlap between the curves occurs in PC1<br />
and, thus, it is adopted as the raw band to unmix. Y axes=frequencies,<br />
and X axes=DNs—but values not shown. b Training DNs <strong>of</strong> EM1,<br />
EM2, EM3, and EM4.<br />
reference training site, is geographically overlaid on each<br />
candidate raw image (240 m resolution) and pure pixels in<br />
the overlay section, red rectangle in Fig. 5, hereinafter referred<br />
to as mix-unmix training site, <strong>of</strong> the candidate raw<br />
image <strong>for</strong> each spectral class identified. A pixel in the mixunmix<br />
training site is pure if the geographically corresponding<br />
8-pixel8-pixel block in the reference training site is<br />
composed <strong>of</strong> a single class, 8 is the ratio <strong>of</strong> the two resolutions.<br />
Figure 6(a) compares the four spectral classes’ purepixels’<br />
DNs’ distribution curves in the mix-unmix training<br />
site. Since the four spectral classes exhibit the highest spectral<br />
dissimilarity between themselves in the first PC, it is<br />
adopted as the raw image to be unmixed. The training DN<br />
ranges <strong>of</strong> the four spectral classes (now denoted as endmembers,<br />
EMs) are as shown in Fig. 6(b). The raw image<br />
(first PC) is unmixed under the mix-unmix concept on the<br />
basis <strong>of</strong> the training DNs into the four end-members.<br />
Mix-Unmix Fraction Images versus Reference Fraction<br />
Images<br />
Reference fraction images <strong>of</strong> the four spectral classes (Fig. 5)<br />
are generated by computing the percentage coverage <strong>of</strong> each<br />
class in every 8-pixel8-pixel block (each block is 240 m<br />
366 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Ngigi and Tateishi: Solving under-determined models in linear spectral unmixing <strong>of</strong> satellite images: mix-unmix concept<br />
Figure 7. Data processing flowchart.<br />
240 m). The reference fraction images are compared with<br />
the mix-unmix fraction images. Figure 7 outlines the entire<br />
image processing flow, and Figure 8 compares the fraction<br />
images. The correlation coefficients <strong>of</strong> the mix-unmix image<br />
fractions versus the reference image fractions <strong>of</strong> the four<br />
end-members are 0.88 (EM1), 0.80 (EM2), 0.67 (EM3), and<br />
0.78 (EM4).<br />
DISCUSSION<br />
The mix-unmix fraction images show similar transition patterns<br />
(highest to lowest concentration levels) as the reference<br />
fraction images <strong>for</strong> all the end-members; though the correlation<br />
coefficients are not “very high.” Although a mixture<br />
interval <strong>of</strong> 10% is used throughout this study, any value that<br />
as a divisor <strong>of</strong> 100 gives a whole number can be used. We<br />
cannot use 100 itself as it would mean that each pixel contains<br />
just a single end-member.<br />
FUTURE<br />
As discussed in the Mixing Branch section, only the extreme<br />
DN values (i.e., bounding mixture tables) are used in this<br />
study. Also, the training DNs <strong>of</strong> end-members are assumed<br />
to be “frequency-less.” As the mix-unmix s<strong>of</strong>tware develops,<br />
all mixture tables and training DNs’ distribution curves will<br />
be incorporated.<br />
The effect <strong>of</strong> adopted mixture interval and overlap <strong>of</strong><br />
training DNs on accuracy <strong>of</strong> the concept will be addressed<br />
on implementation <strong>of</strong> the above. Also, per<strong>for</strong>mance <strong>of</strong> the<br />
concept across different numbers <strong>of</strong> end-members, different<br />
resolutions, and different geographical scales will be tested.<br />
Figure 8. Reference fraction images upper row and Mix-unmix fraction<br />
images lower row <strong>of</strong> four end-members first column=EM1, second<br />
=EM2, third=EM3, fourth=EM4. White and black are background,<br />
i.e., 0%.<br />
CONCLUSIONS<br />
This preliminary investigation shows that the mix-unmix<br />
concept is capable <strong>of</strong> addressing the problem <strong>of</strong> underdetermination<br />
in linear spectral unmixing—a very revolutionary<br />
dimension in data processing as the number <strong>of</strong> endmembers<br />
is not pegged on that <strong>of</strong> available bands. It is the<br />
only method that truly solves the problem <strong>of</strong> underdetermination.<br />
Sequential Maximum Angle Convex Cone<br />
(SMACC) Spectral Tool does not work on a single band, and<br />
Generalized Orthogonal Subspace Projection cannot generate<br />
additional bands from a single band. Further, the mixunmix<br />
concept creates a processing environment that allows<br />
any pixel to be unmixed without any sort <strong>of</strong> restrictions<br />
(e.g., minimum determinable fraction), impracticalities (e.g.,<br />
negative fractions), or trade-<strong>of</strong>fs (e.g., either positivity or<br />
unity sum) that may be associated with conventional unmixing<br />
techniques.<br />
REFERENCES<br />
1 Y. E. Shimabukuro and J. A. Smith, “The least squares mixing methods<br />
to generate fraction images derived from remote sensing multispectral<br />
data”, IEEE Trans. Geosci. Remote Sens. 29, 16 (1991).<br />
2 J. W. Boardman, “Geometric mixture analysis <strong>of</strong> imaging spectrometry<br />
data”, Proc. Int. Geosci Remote Sens Symposium 4, 2369 (1994).<br />
3 J. Gruninger, A. J. Ratkowski, and M. L. Hoke, “The sequential<br />
maximum angle convex cone (SMACC) endmember model”, Proc. SPIE<br />
5425 1 (2004).<br />
4 R. Hsuan and C. Yang-Lang, “Error Analysis <strong>for</strong> Band Generation in<br />
Generalized Process Orthogonal Subspace Projection”, IEEE Geoscience<br />
and Remote Sensing Symposium Proceedings, IGARSS (IEEE Press,<br />
Piscataway, NJ, 2005).<br />
5 I. Emmett, “Hyperspectral Image Classification Using Orthogonal<br />
Subspace Projections: Image Simulation and Noise Analysis”, http://<br />
www.cis.rit.edu/~ejipci/Reports/osp_paper.pdf (2001).<br />
6 A. Ifarraguerri and C. Chang, “Multispectral and Hyperspectral Image<br />
Analysis with Convex Cones”, IEEE Trans. Geosci. Remote Sens. 37, 756<br />
(1999).<br />
7 M. E. Winter and E. M. Winter, “Comparison <strong>of</strong> approaches <strong>for</strong><br />
determining end-members in hyperspectral data”, Proc. IEEE Aerospace<br />
Conference (IEEE Press, Piscataway, NJ, 2000).<br />
8 N. Keshava, “A Survey <strong>of</strong> Spectral Unmixing Techniques”, Lincoln Lab.<br />
J. 14, 55 (2003).<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 367
<strong>Journal</strong> <strong>of</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology® 51(4): 368–379, 2007.<br />
© <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology 2007<br />
Color Shift Model-Based Segmentation and Fusion<br />
<strong>for</strong> Digital Aut<strong>of</strong>ocusing<br />
Vivek Maik<br />
Image Processing and Intelligent Systems Lab, Graduate School <strong>of</strong> Advanced <strong>Imaging</strong> <strong>Science</strong>, Multimedia<br />
and Film, Chung Ang University, Seoul 156-756, South Korea<br />
E-mail: vivek5681@wm.cau.ac.kr<br />
Dohee Cho<br />
Digital/Scientific <strong>Imaging</strong> Lab, Graduate School <strong>of</strong> Advanced <strong>Imaging</strong> <strong>Science</strong>, Multimedia and Film, Chung<br />
Ang University, Seoul 156-756, South Korea<br />
Jeongho Shin<br />
Department <strong>of</strong> Web In<strong>for</strong>mation Engineering, Hankyong National University, Anseong 456-749,<br />
South Korea<br />
Donghwan Har<br />
Digital/Scientific <strong>Imaging</strong> Lab, Graduate School <strong>of</strong> Advanced <strong>Imaging</strong> <strong>Science</strong>, Multimedia and Film, Chung<br />
Ang University, Seoul 156-756, South Korea<br />
Joonki Paik<br />
Image Processing and Intelligent Systems Lab, Graduate School <strong>of</strong> Advanced <strong>Imaging</strong> <strong>Science</strong>, Multimedia<br />
and Film, Chung Ang University, Seoul 156-756, South Korea<br />
Abstract. This paper proposes a novel color shift model-based<br />
segmentation and fusion algorithm <strong>for</strong> digital aut<strong>of</strong>ocusing <strong>of</strong> color<br />
images. The source images are obtained using new multiple filteraperture<br />
configurations. We shift color channels to change the focal<br />
point <strong>of</strong> the given image at different locations. For each respective<br />
location we then select the optimal focus in<strong>for</strong>mation and, finally,<br />
use s<strong>of</strong>t decision fusion and blending (SDFB) to obtain fully-focused<br />
images. The proposed aut<strong>of</strong>ocusing algorithm consists <strong>of</strong>: (i) color<br />
channel shifting and alignment <strong>for</strong> varying focal positions; (ii) optimal<br />
focus region selection and segmentation using sum modified Laplacian<br />
(SML); and (iii) SDFB, which enables smooth transition<br />
across region boundaries. By utilizing segmented images <strong>for</strong> different<br />
focal point locations, the SDFB algorithm can combine images<br />
with multiple, out-<strong>of</strong>-focus objects. Experimental results show per<strong>for</strong>mance<br />
and feasibility <strong>of</strong> the proposed algorithm <strong>for</strong> aut<strong>of</strong>ocusing<br />
images with one or more differently out-<strong>of</strong>-focus objects. © 2007<br />
<strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology.<br />
DOI: 10.2352/J.<strong>Imaging</strong>Sci.Technol.200751:4368<br />
Received Sep. 25, 2006; accepted <strong>for</strong> publication Mar. 22, 2007.<br />
1062-3701/2007/514/368/12/$20.00.<br />
INTRODUCTION<br />
Demand <strong>for</strong> digital aut<strong>of</strong>ocusing techniques is rapidly increasing<br />
in many visual applications, such as camcorders,<br />
digital cameras, and video surveillance systems. Until now,<br />
most focusing ef<strong>for</strong>ts have been put on gray scale images.<br />
Even with specialized color processing techniques, each color<br />
channel is processed independently <strong>for</strong> aut<strong>of</strong>ocusing applications.<br />
In this paper, a novel aut<strong>of</strong>ocusing algorithm utilizing<br />
color shift property is proposed, which can restore an<br />
image with multiple, differently focused objects. We propose<br />
a new filter-aperture (FA) model <strong>for</strong> aut<strong>of</strong>ocusing color images.<br />
The proposed method overcomes the fusion with multiple<br />
source images as it uses a single input image. The FA<br />
model separates and distributes the out-<strong>of</strong>-focusing blur<br />
into different color channels. The multiple FA models also<br />
make it possible to generate as many source images as necessary<br />
<strong>for</strong> fusion-based aut<strong>of</strong>ocusing. Multiple focal points<br />
are spotted on the image and color channel shifting aligns<br />
each channel with the respective focal point. For each alignment<br />
the sum modified Laplacian (SML) operator is used to<br />
obtain a numerical measure indicating the degree <strong>of</strong> focus <strong>of</strong><br />
that image. The in-focus pixels are selected and combined at<br />
each process using s<strong>of</strong>t decision fusion and blending (SDFB)<br />
to produce the in-focus image with maximum focus metric.<br />
The SML operator can also be used to estimate a number <strong>of</strong><br />
focal points starting from the minimum degree <strong>of</strong> focus in<br />
the input image. The proposed algorithm does not use any<br />
restoration filter, which usually results in undesired artifacts,<br />
such as ringing, reblurring, and noise clustering.<br />
The rest <strong>of</strong> the paper is organized as follows. The following<br />
section summarizes existing techniques, and presents<br />
the major contribution <strong>of</strong> the proposed work. The section<br />
titled “Multiple FA model” gives a detailed description <strong>of</strong> the<br />
multiple FA method and “Digital Aut<strong>of</strong>ocusing Algorithm”<br />
describes the proposed aut<strong>of</strong>ocusing algorithm. “Experi-<br />
368
Maik et al.: Color shift model-based segmentation and fusion <strong>for</strong> digital aut<strong>of</strong>ocusing<br />
EXISTING STATE-OF-THE-ART AUTOFOCUSING<br />
METHODS<br />
FA Model<br />
The conventional photo sensor array uses micro lenses in<br />
front <strong>of</strong> every pixel to concentrate light onto the photosensitive<br />
region. 1 In this paper, we can interpret the optical<br />
design in a gradual step that we are able to make the multiple<br />
detectors beneath the each micro lens, instead <strong>of</strong> multiple<br />
arrays <strong>of</strong> detectors. The artificial compound eye sensor<br />
(insect eyes) is composed <strong>of</strong> a micro lens array and a photo<br />
sensor. 2 However, the imaging quality <strong>of</strong> these optical designs<br />
is fundamentally inferior to a camera system with a<br />
large single lens; the resolution <strong>of</strong> these small lens arrays is<br />
severely limited by diffraction. The “wave front coding”<br />
system 3 is similar to the proposed system (see Figure 1) in<br />
that it provides a way to decouple the trade-<strong>of</strong>f between<br />
aperture size and depth <strong>of</strong> field, but their design is very<br />
different. Rather than collecting and resorting rays <strong>of</strong> light,<br />
they use aspheric lenses that produce images with a depthindependent<br />
blur. Deconvolution <strong>of</strong> these images retrieves<br />
image details at all depths as shown in Figure 2.<br />
Figure 1. Block diagram <strong>of</strong> the proposed algorithm.<br />
mental Results” shows the simulation results and comparisons<br />
with existing methods. Finally, we have the concluding<br />
remarks.<br />
Aut<strong>of</strong>ocusing Methods<br />
The traditional aut<strong>of</strong>ocusing system in a camera usually<br />
consists <strong>of</strong> two different modules: analysis and control. The<br />
analysis module estimates a degree-<strong>of</strong>-focus <strong>of</strong> an image<br />
projected onto the image plane. The control module per<strong>for</strong>ms<br />
focusing functions by moving the lens assembly to the<br />
optimal focusing position according to the degree-<strong>of</strong>-focus<br />
in<strong>for</strong>mation estimated in the analysis module. There are five<br />
different focusing techniques, such as manual focusing<br />
(MF), infrared aut<strong>of</strong>ocusing (IRAF), through-the-lens<br />
aut<strong>of</strong>ocusing (TTLAF), semi-digital aut<strong>of</strong>ocusing (SDAF),<br />
and fully digital aut<strong>of</strong>ocusing (FDAF). 4–7 Table I briefly<br />
summarizes and compares those techniques.<br />
The FDAF systems usually involve restoration and fusion<br />
methods in the control module which operates using<br />
prior in<strong>for</strong>mation like point spread function (PSF), gradients,<br />
multiple source inputs, etc. to obtain the details about<br />
out-<strong>of</strong>-focus blur in images. Image fusion-based aut<strong>of</strong>ocus-<br />
Figure 2. Representation <strong>of</strong> the schematic <strong>of</strong> the a wave front coding system, b proposed FA system.<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 369
Maik et al.: Color shift model-based segmentation and fusion <strong>for</strong> digital aut<strong>of</strong>ocusing<br />
Table I. Comparison <strong>of</strong> conventional AF systems with the proposed system.<br />
Aut<strong>of</strong>ocusing<br />
technique Analysis module Control module<br />
Focusing<br />
accuracy<br />
Hardware<br />
specifications<br />
Manual Human decision Manual Subject to human operation Low shutter speed f/2-f/8.0<br />
IRAF Calculating the time <strong>of</strong> IR travel Moving the focusing lens High High shutter speed f/3.5-f/5.6<br />
TTLAF Minimizing a phase difference Moving the focusing lens Very high under good conditions High shutter speed f/2-f/11<br />
SDAF Calculating high frequency <strong>of</strong> image Moving the focusing lens Acceptable Medium shutter speed f/2-f/28<br />
FDAF Estimating PSF, blur models Restoration filters and fusion methods Acceptable NIL<br />
Proposed method Color channel shifting Multiple filter aperture FA Acceptable 30 to 1 / 4,000 sec. f/5.6-f/22<br />
ing methods have focused on operation <strong>of</strong> multiple source<br />
images using wavelet or discrete cosine trans<strong>for</strong>mations<br />
(DCT) 8,9 with a priori obtained camera PSF. Other methods<br />
use pyramid-based representation to decompose the source<br />
images into different spatial scales and orientations. 10,11<br />
Similar results, although with more artifacts and less visual<br />
stability, can be achieved by using a set <strong>of</strong> basis functions. 12<br />
Another technique similar to pyramid representation approach<br />
has been based on wavelet trans<strong>for</strong>m to decompose<br />
the image into various subbands. 13,14 The output is generated<br />
by selecting one <strong>of</strong> the decomposed subbands such that<br />
the selected subband has maximum energy. Restorationbased<br />
techniques have been carried out to overcome the out<strong>of</strong>-focus<br />
problem. However, restoration <strong>of</strong> images with different<br />
depth <strong>of</strong> fields tend to cause reblurring and ringing<br />
artifacts in the region with low depth <strong>of</strong> field or in-focus<br />
regions. 15,16 Even with equal depth <strong>of</strong> field the nature <strong>of</strong><br />
restoration poses a serious limitation to the visual quality <strong>of</strong><br />
the restored images. Another drawback is the slow convergence<br />
process <strong>of</strong> the iterative framework.<br />
The main contribution <strong>of</strong> the proposed method is listed<br />
below:<br />
(a) Multiple apertures and corresponding sensors can<br />
enhance depth in<strong>for</strong>mation.<br />
(b) Focusing process is inherently designed in accordance<br />
with color in<strong>for</strong>mation.<br />
(c) Neither image restoration nor blur identification is<br />
necessary.<br />
Figure 3. General single aperture model.<br />
(d) Set <strong>of</strong> images with multiple apertures and focus<br />
settings can be generated using a single image with<br />
channel shifting,<br />
(e) Fusion algorithm involves separate feature-based<br />
fusion and color blending consistency to preserve<br />
the channel dependencies.<br />
(f) Proposed algorithm does not need trans<strong>for</strong>mation<br />
or convolution operations.<br />
Recently, images obtained at different shutter speeds<br />
were combined into an image in which full dynamic range is<br />
preserved. 17 The proposed approach extends and generalizes<br />
the standard fusion approach to color images. The proposed<br />
approach does not need multiple source images captured at<br />
different aperture settings. Instead we derive different source<br />
images from a single out-<strong>of</strong>-focus image to obtain various<br />
positions <strong>of</strong> focal points. 18–20 For each focal point three color<br />
channels are aligned and the corresponding images are used<br />
<strong>for</strong> fusion.<br />
MULTIPLE FILTER-APERTURE (FA) MODEL<br />
An aperture <strong>of</strong> a lens can adjust the amount <strong>of</strong> incoming<br />
light accepted through the lens. It can also control the focal<br />
length, camera-to-object distance, and depth <strong>of</strong> field. Generally,<br />
the center <strong>of</strong> an aperture is aligned on the optical axis<br />
<strong>of</strong> the lens. Any controlled aperture accepts light from various<br />
angles depending on the object position. Correspondingly,<br />
the convergence pattern on the imaging plane <strong>for</strong>ms<br />
either a point or a circular region as shown in Figure 3. For<br />
370 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Maik et al.: Color shift model-based segmentation and fusion <strong>for</strong> digital aut<strong>of</strong>ocusing<br />
Figure 4. Aperture shifted from the center.<br />
Figure 5. Multiple aperture set-up <strong>for</strong> red and blue channel filters.<br />
objects placed at near-, mid-, and far-focal distance the image<br />
convergence takes place either in front, on, or behind the<br />
CCD/CMOS sensor. However, this image convergence as<br />
well as the blur in<strong>for</strong>mation can only be represented in a<br />
bi-axial plane as shown in Fig. 3.<br />
An interesting alternative <strong>for</strong> tri-axial representation <strong>of</strong><br />
the image and out-<strong>of</strong>-focus blur was found to be achieved<br />
using non-centric aperture as shown in Figure 4. For a noncentric<br />
aperture located either on the upper or lower part <strong>of</strong><br />
the optical axis, the convergence pattern was found to be<br />
split between these axes. The split difference between the<br />
patterns will give another dimension to the conventional biaxial<br />
plane making it a tri-axial representation. For the objects<br />
at the same positions (near, in, far focal distances), the<br />
convergence pattern <strong>of</strong> the channel aperture <strong>for</strong>m an overlapping<br />
convergence on the CCD/CMOS sensor. For instance,<br />
the near focal distance object converges on the upper<br />
part <strong>of</strong> the optical axis where, at the same position, the far<br />
focal distance object converges on the lower part. If these<br />
overlapping channels are exactly aligned, then we will have a<br />
focused pattern in the image.<br />
An extension <strong>of</strong> the above approach will be to use a lens<br />
with two apertures on either side <strong>of</strong> the optical axis. An<br />
interesting phenomenon that can be observed is that, <strong>for</strong> the<br />
near and far focused objects, the convergence pattern lies on<br />
opposite sides <strong>for</strong> each aperture in reverse order <strong>for</strong> each<br />
channel. For example, the red aperture can have nearfocused<br />
convergence on the top and far-focused convergence<br />
on the bottom whereas the blue aperture has far-focused<br />
convergence on the top and near-focused convergence at the<br />
bottom, as shown in Figure 5. This phenomenon is called<br />
the filter-aperture (FA) extraction. The out-<strong>of</strong>-focus blur is<br />
now distributed among the color channels <strong>for</strong>ming the<br />
image.<br />
Now we extend the above multiple aperture convergence<br />
to a typical RGB image scenario. To obtain an RGB<br />
image using the multiple aperture configurations we need to<br />
obtain R, G, and B channel convergence patterns separately.<br />
This can be done using three apertures in a Bayer pattern<br />
where the images are individually obtained on the sensor <strong>for</strong><br />
the three apertures and later combined to <strong>for</strong>m the RGB<br />
image. Evidently, multiple apertures provide additional<br />
depth in<strong>for</strong>mation <strong>of</strong> objects at different distances. Since any<br />
color image is composed <strong>of</strong> three channels, we have used<br />
three apertures and, correspondingly, three filters (see<br />
Figure 6).<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 371
Maik et al.: Color shift model-based segmentation and fusion <strong>for</strong> digital aut<strong>of</strong>ocusing<br />
Figure 6. Multiple FA model showing the convergence pattern <strong>for</strong> the R, G, and B color channels.<br />
The main advantage <strong>of</strong> the FA model is that it can provide<br />
an alternative method <strong>for</strong> blur estimation in aut<strong>of</strong>ocusing<br />
applications. Images acquired by using a normal lens<br />
have uni<strong>for</strong>m or spatially variant blur confined on all channels.<br />
However, in the proposed algorithm, by using three<br />
filtered sensors the aut<strong>of</strong>ocusing problem turns into the<br />
alignment <strong>of</strong> R, G, and B channels with various depths <strong>of</strong><br />
field. The out-<strong>of</strong>-focusing phenomenon with single and<br />
multiple aperture lenses are compared in Figure 7. As shown<br />
in Fig. 7(b) the out <strong>of</strong> focus blur is modeled as a misalignment<br />
<strong>of</strong> three color channels <strong>of</strong> R, G, and B.<br />
DIGITAL AUTOFOCUSING ALGORITHM<br />
The proposed algorithm uses the image obtained from the<br />
multiple FA configurations <strong>for</strong> the aut<strong>of</strong>ocusing application.<br />
The proposed aut<strong>of</strong>ocusing algorithm consists <strong>of</strong> the following<br />
procedures to obtain a well-restored image: (i) salient<br />
feature computation, (ii) color channel shifting and alignment<br />
<strong>for</strong> selected pixels, and (iii) s<strong>of</strong>t decision fusion and<br />
blending.<br />
Salient Focus Measure<br />
The feature saliency computation process contains a family<br />
<strong>of</strong> functions that estimate saliency in<strong>for</strong>mation. In practice,<br />
these functions can operate on individual pixels or on a local<br />
region <strong>of</strong> pixels. When combining images having different<br />
focus measures, <strong>for</strong> instance, a desirable saliency measure<br />
would provide a quantitative measure that increases when<br />
features are better focused. Various saliency measures, including<br />
variance and gradients, have been employed and<br />
Figure 7. Comparison <strong>of</strong> out-<strong>of</strong>-focus blurs <strong>for</strong> a single aperture model and the proposed multiple aperture<br />
models: a and b out-<strong>of</strong>-focus image captured using ordinary camera and proposed FA system under same<br />
focal settings, c restored result using the regularized restoration method, d restored result using the proposed<br />
channel shifting and fusion algorithm.<br />
372 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Maik et al.: Color shift model-based segmentation and fusion <strong>for</strong> digital aut<strong>of</strong>ocusing<br />
validated <strong>for</strong> related applications. The saliency function only<br />
selects the frequencies in the focused image that will be attenuated<br />
due to defocusing. One way to detect a high frequency<br />
component is to apply the following absolute<br />
Laplacian operator as<br />
2 L k =<br />
2 2 L k<br />
+ 2 2 L k<br />
.<br />
x y<br />
The second derivatives in the x and y directions <strong>of</strong>ten have<br />
opposite signs and tend to cancel each other. In the case <strong>of</strong><br />
textured images, this phenomenon may frequently occur and<br />
the Laplacian behaves in an unstable manner. However, this<br />
problem can be overcome by using the absolute Laplacian as<br />
in Eq. (1). In order to accommodate <strong>for</strong> possible variations<br />
in the size <strong>of</strong> texture elements, we compute the partial derivative<br />
using a variable spacing between pixels <strong>for</strong> computing<br />
the derivatives. Hence a discrete approximation to the<br />
modified Laplacian, ML k i,j, <strong>for</strong> pixel intensity, Ii,j, is<br />
given as<br />
ML k i,j = 2Ii,j − Ii −1,j − Ii +1,j<br />
1<br />
+ 2Ii,j − Ii,j −1 − Ii,j +1. 2<br />
Finally, the focus measure at a point i,j is computed as the<br />
sum <strong>of</strong> modified Laplacian values, in a small window around<br />
i,j, that are greater than a prespecified threshold value,<br />
i+N<br />
fi,j = <br />
j+N<br />
<br />
p=i−N q=j−N<br />
ML k p,q <strong>for</strong> ML k p,q T 1 .<br />
The heuristically determined threshold value T 1 in the range<br />
40–60 provides acceptable results in most cases. The parameter<br />
N represents the window size <strong>for</strong> computing the focus<br />
measure. In contrast to region-based aut<strong>of</strong>ocusing methods,<br />
we typically use a smaller window <strong>of</strong> size, e.g., N=1. Equation<br />
(3) can be referred to as sum modified Laplacian (SML)<br />
which is used as an intermediate image estimate <strong>for</strong> determining<br />
focus in<strong>for</strong>mation.<br />
3<br />
Figure 8. Schematic <strong>of</strong> channel alignment procedure <strong>for</strong> R, G, and B<br />
channels.<br />
Color Channel Shift and Alignment<br />
For shifting and aligning color channels we need to find the<br />
optimal pixel-<strong>of</strong>-interest at different positions in the image<br />
according to their focal measures. These pixels-<strong>of</strong>-interest<br />
can be referred to as a focal point pixels. The term “focal<br />
point pixel” refers to a pixel-<strong>of</strong>-interest around which channel<br />
shifting and alignment is carried out. For a given image,<br />
the SML measure can be used to determine the focal point<br />
region whose focal measure is significantly lower than other<br />
regions <strong>of</strong> the image. Then <strong>for</strong> a given region, we select the<br />
focal point pixel either from the center <strong>of</strong> region or the pixel<br />
with the lowest focus measure. Similar operations can be<br />
per<strong>for</strong>med <strong>for</strong> different selected focal point regions in different<br />
neighborhood. Hence<strong>for</strong>th, <strong>for</strong> a corresponding focal<br />
point pixel, we per<strong>for</strong>m channel alignment and remove the<br />
out-<strong>of</strong>-focus blur in that given neighborhood (see Figure 8).<br />
For a given particular image captured by using FA configuration,<br />
the out-<strong>of</strong>-focus blur was just confined to channels<br />
on either side <strong>of</strong> the green channel as shown in Figure 9.<br />
As can be seen from the figure, the green channel suffers<br />
minimal blur distortion as the sensor was placed at the center<br />
whereas the red and the blue channels have maximal blur<br />
distortion. The proposed aut<strong>of</strong>ocusing technique uses the<br />
green channel as the reference and aligns the red and the<br />
blue channels to the green channel <strong>for</strong> any particular location,<br />
such as<br />
I RGB = S r,c I R + I B + I G ,<br />
where S r,c represents the shift operator and the shift vector<br />
r,c represents the amount <strong>of</strong> shift in row and column directions<br />
<strong>for</strong> the respective red and blue channels with respect<br />
to the reference focal point on the green channel. If the shift<br />
vectors are not identical, we can generalize the above equation<br />
as<br />
I RGB = SI R r 1 ,c 1 + I B r 2 ,c 2 + I G .<br />
The shift vectors on the same sensor filter are linearly dependent.<br />
For a particular reference channel it is possible to<br />
estimate the exact number <strong>of</strong> shift vectors using the sensor<br />
filter configurations. For example, in our experiments the<br />
green channel has been used as reference, hence the red and<br />
blue pixels are misaligned by a pattern corresponding to the<br />
sensor filter as shown in Figure 10.<br />
S<strong>of</strong>t Decision Fusion and Blending<br />
In order to merge images with multiple focal point planes,<br />
image fusion is required on multiple channel images. Un<strong>for</strong>tunately,<br />
when the channel-shifted images are directly fused,<br />
misalignment or misregistration is unavoidable. The pixels<br />
<strong>of</strong> different channel aligned images, when fused together,<br />
may sometimes tend to overlap or get missed because <strong>of</strong> the<br />
channel shifting. This problem can be overcome by applying<br />
an inverse shift operation to the images with respect to a<br />
reference image. The reference image has to be chosen from<br />
one <strong>of</strong> the several channel shifted images extracted using<br />
channel shifting and alignment. In the proposed approach<br />
we choose the reference image as the one that will have a<br />
focal point located approximately in the center <strong>of</strong> the image,<br />
−1<br />
I k = S r,cIr I k ,I r ,<br />
where I r represents the reference image <strong>for</strong> registration and<br />
S −1 represents the inverse shift operation. After selection <strong>of</strong><br />
4<br />
5<br />
6<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 373
Maik et al.: Color shift model-based segmentation and fusion <strong>for</strong> digital aut<strong>of</strong>ocusing<br />
Figure 10. Row and column shift vectors <strong>for</strong> color channel shifting and<br />
alignment with reference G channel.<br />
Figure 9. Multiple FA model image: a R, b G, and c B channels,<br />
respectively.<br />
the focal point using the SML operator and channel alignment<br />
by channel shifting, the appropriate regions need to be<br />
selected from the given image. Given the location <strong>of</strong> the<br />
pixel-<strong>of</strong>-interest <strong>for</strong> channel alignment, we simply select an<br />
approximate region area that is defined on its neighborhood.<br />
But <strong>for</strong> a more efficient fusion process we could isolate the<br />
region around that pixel using neighborhood connectivity.<br />
For a given pixel-<strong>of</strong>-interest and the eight-neighborhood<br />
connectivity, we can extract the region more accurately <strong>for</strong><br />
the purpose <strong>of</strong> image fusion as<br />
Figure 11. Experimental set up: a digital camera with the FA model;<br />
b interior configuration <strong>of</strong> the FA; and c the R, G, and B sensor filters.<br />
374 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Maik et al.: Color shift model-based segmentation and fusion <strong>for</strong> digital aut<strong>of</strong>ocusing<br />
Table II. A sample set <strong>of</strong> shift vectors estimated <strong>for</strong> different locations.<br />
Table III. Hardware configuration <strong>for</strong> the multiple FA system.<br />
Region Red channel r ,c Blue channel r ,c<br />
Hardware title<br />
Specifications<br />
R 1 upper left 10,4 8,5<br />
R 2 upper middle 7,6 8,5<br />
R 3 upper right 10,5 7,6<br />
R 4 center left 9,4 10,5<br />
R 5 center middle 9,3 9,3<br />
R 6 center right 8,6 10,2<br />
R 7 lower left 10,4 9,4<br />
R 8 lower middle 11,6 9,4<br />
R 9 lower right 11,5 8,3<br />
i+N<br />
F k = <br />
j+N<br />
<br />
x=i−N y=j−N<br />
f p xs i , ...,s i+k ,yt i , ...,t j+k ,<br />
where F k represents the region around the pth focal point<br />
pixel, f p and s,t represent the neighborhood connectivity.<br />
Even though the SML operator can provide an accurate<br />
measure, we need to extract the specific region from the<br />
7<br />
Digital camera Nikon D-100<br />
R, G, B filters Green-Kodak-Wratten Filter No. 58<br />
Blue-Kodak-Wratten Filter No. 47<br />
Red-Kodak-Wratten Filter No. 25<br />
Focusing<br />
APO-Symmar-L-150-5.6,11,22<br />
f-5.6, f-11, f-22<br />
Sensor<br />
23.7 15.6 mm RGB CCD; 6.31 million total<br />
pixels<br />
Lens mounting<br />
Schneider Apo-Tele-Xenar<br />
Relative aperture focal length −5.6/ 250<br />
Shutter speed<br />
30 to 1 / 4,000 sec. and bulb<br />
Color mode<br />
Triple mode <strong>for</strong> R, G, and B channels<br />
image <strong>for</strong> fusion. One <strong>of</strong> disadvantages <strong>of</strong> the FA model is<br />
that, <strong>for</strong> the channel-aligned images with closely located focal<br />
points, the SML operator does not always per<strong>for</strong>m well.<br />
Hence, we used a color-based region segmentation algorithm<br />
<strong>for</strong> extracting selective regions from the channel-<br />
Figure 12. Experimental results: a the source image; b the focal point location <strong>for</strong> channel shifting and<br />
alignment; c the image after channel shifting to new focal point location; and d the final image fused from<br />
a and c.<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 375
Maik et al.: Color shift model-based segmentation and fusion <strong>for</strong> digital aut<strong>of</strong>ocusing<br />
Figure 13. Experimental results: a the source image; b objects closer to the left side <strong>of</strong> the image are<br />
focused and object to the right side are out-<strong>of</strong>-focus; c similar set up with the focus on the right side; and d<br />
the final image fused from b and c.<br />
aligned images if the SML results are not good enough.<br />
When fusing color images, features such as edge and textures<br />
should be preserved, and also the color blending consistency<br />
should be maintained. The fusion process is per<strong>for</strong>med on<br />
each level <strong>of</strong> the channel-aligned images in conjunction with<br />
SML to generate the composite image C. The reconstruction<br />
process integrates in<strong>for</strong>mation from different levels as<br />
and<br />
I ck = F k · I ak + 1−F k I bk<br />
−1<br />
I a = S r,cIr I a ,I r ,<br />
8<br />
9<br />
where I ck represents the reconstructed image from two input<br />
images I ak and I bk . The variable k represents the regions<br />
extracted based on their respective focal measure. The inverse<br />
shifting operation is described in Eq. (9) where I r represents<br />
the reference image and r,cI r the corresponding<br />
shift vectors with respect to I r . A typical problem <strong>of</strong> image<br />
fusion is the appearance <strong>of</strong> unnatural borders between the<br />
different decisions regions due to overlapping blur at focus<br />
boundaries. To combat this, s<strong>of</strong>t decision blending can be<br />
employed using smoothing or low pass filtering <strong>of</strong> the saliency<br />
parameter F k . In this paper Gaussian smoothing has<br />
been used to obtain the desired effect <strong>of</strong> blending. This creates<br />
weighted decision regions where a linear combination <strong>of</strong><br />
pixels in the two images A and B are used to generate corresponding<br />
pixels in the fused image C. We then have<br />
376 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Maik et al.: Color shift model-based segmentation and fusion <strong>for</strong> digital aut<strong>of</strong>ocusing<br />
Figure 14. Experimental results: a the source image; b–c channel shifted images <strong>for</strong> focal points at near<br />
and center positions; and d the final image obtained by fusing a–c.<br />
I ck = F˜k · I ak + 1−F˜kI bk ,<br />
10<br />
where F˜k is now a smoothed version <strong>of</strong> its <strong>for</strong>mer self. At<br />
times there can be missing pixels in the fused image which<br />
are not selected using the SML. The number <strong>of</strong> missing<br />
pixels varies from image to image, but is always confined to<br />
a very small portion <strong>of</strong> the entire image. The missing pixels<br />
have to be replaced with pixels from any one <strong>of</strong> the available<br />
channel aligned images. One simple way to find an appropriate<br />
replacement is to get the location <strong>of</strong> the missing pixel<br />
in the image and then match it with the image whose focal<br />
point pixel is nearest to the respective missing pixel:<br />
Ix,y = min disIx,y,f p x,y.<br />
11<br />
EXPERIMENTAL RESULTS<br />
Dataset Simulation and Experiments<br />
In order to demonstrate the per<strong>for</strong>mance <strong>of</strong> the proposed<br />
algorithm, we used test images captured using the proposed<br />
multiple FA model with multiple out-<strong>of</strong>-focus objects in the<br />
background. The experimental setup is shown in Figure 11<br />
which represents the camera used <strong>for</strong> the experiments along<br />
with the multiple FA configurations <strong>of</strong> the camera and the<br />
sensor filter. The hardware specifications used <strong>for</strong> the system<br />
are listed in Table III. Experiments were per<strong>for</strong>med on an<br />
RGB image <strong>of</strong> size 640480. Here, each image contains<br />
multiple objects at different distances from the camera.<br />
Figure 12(a) represents a test image with low depth-<strong>of</strong>-field,<br />
where focus is on the objects close to the camera lens. The<br />
channels aligned <strong>for</strong> the focal point are shown in Fig. 12(b).<br />
The image after channel shifting is shown in Fig. 12(c). The<br />
blue object in the back <strong>of</strong> the astronaut was out <strong>of</strong> focus in<br />
Fig. 12(a), which is now in-focus in Fig. 12(c), whereas the<br />
other regions <strong>of</strong> the image tend to get defocused. The fused<br />
image <strong>of</strong> Figs. 12(a) and 12(c) is shown in Fig. 12(d). Similar<br />
results with multiple objects are shown in Figures 13 and<br />
14. The selected focal point <strong>for</strong> the channel alignment<br />
and shifting are represented in Figures 15(f) and 15(g).<br />
Table IV. Image quality comparisons <strong>for</strong> the various aut<strong>of</strong>ocusing methods.<br />
Aut<strong>of</strong>ocusing<br />
method<br />
Prior<br />
in<strong>for</strong>mation<br />
Mode<br />
Input<br />
frames Operation RMSE PSNR<br />
Wiener filter PSF Gray 1 Pixel based 12.35 23.36<br />
Iterative filter NIL Gray 1 Pixel based 8.56 26.32<br />
Constrained least<br />
Edge Gray 1 Pixel based 9.56 25.10<br />
square filter<br />
Pyramid fusion NIL Gray, Color At least 2 Window based 5.68 28.42<br />
and Pixel based<br />
Wavelet fusion NIL Gray, Color At least 2 Window based 5.02 29.95<br />
and Pixel based<br />
Proposed NIL Color 1 Window based<br />
and Pixel based<br />
8.06 26.41<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 377
Maik et al.: Color shift model-based segmentation and fusion <strong>for</strong> digital aut<strong>of</strong>ocusing<br />
Figure 15. Experimental results: a the source image; b–d channel shifted images <strong>for</strong> focal points in right,<br />
center, and left positions; e–g the SML results <strong>for</strong> selected regions; and h the final image obtained fusing<br />
b–d.<br />
Figures 15(b)–15(d) illustrate the results <strong>of</strong> SML operator<br />
<strong>for</strong> a selected region. These figures represent images which<br />
have different out-<strong>of</strong>-focus regions obtained from single<br />
source images using channel shifting. Figure 15(e) represents<br />
the fused image from Figs. 15(b)–15(d). The resulting fused<br />
image contains in-focus regions from respective images. The<br />
above set <strong>of</strong> results illustrates the feasibility <strong>of</strong> the proposed<br />
fusion based algorithm <strong>for</strong> aut<strong>of</strong>ocusing.<br />
378 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Maik et al.: Color shift model-based segmentation and fusion <strong>for</strong> digital aut<strong>of</strong>ocusing<br />
Per<strong>for</strong>mance Comparison<br />
For measuring the per<strong>for</strong>mance <strong>of</strong> the multiple FA configurations,<br />
various test images were captured using the proposed<br />
system as well as the ordinary Nikon D-100 camera.<br />
The test images were then processed <strong>for</strong> out-<strong>of</strong>-focus removal<br />
using the proposed channel shifting and fusion algorithm<br />
and the ordinary camera images were restored using<br />
some state <strong>of</strong> the art restoration methods, including Wiener<br />
filter, regularized iterative restoration, constrained least<br />
squares filter, as well as some existing fusion-based methods<br />
including pyramid decomposition and wavelet methods. The<br />
per<strong>for</strong>mance metric in the <strong>for</strong>m <strong>of</strong> PSNR and RMSE were<br />
obtained <strong>for</strong> the test images using the above algorithms as<br />
given in Table IV. As can be seen in the table, the images<br />
captured using the multiple FA configurations tend to have<br />
some degradation when compared to conventional camera<br />
images when there is no out-<strong>of</strong>-focus blurs. But with the<br />
out-<strong>of</strong>-focus blur the image quality <strong>of</strong> the conventional camera<br />
images tend to drastically reduce due to processing by<br />
restoration and is more or less comparable with the restored<br />
images using the color channel shifting and fusion. However,<br />
the fusion methods tend to give slightly higher image quality,<br />
but they require multiple source input images <strong>for</strong> achieving<br />
good per<strong>for</strong>mance, whereas the proposed method can<br />
achieve it with just a single source input image making our<br />
method more suitable and efficient <strong>for</strong> increasing potential<br />
applications.<br />
For aligning the blue channel with the green channel<br />
the pixels have to be shifted in an upward direction and<br />
towards the left or diagonally to the left and vice versa <strong>for</strong><br />
the red channel. In our experiments we tried precomputing<br />
the shift vectors at nine different locations on a test image<br />
manually using the above convention. We found that the<br />
shift vectors differ slightly <strong>for</strong> different regions on the image,<br />
as shown in Table II. These shift vectors were then used<br />
accordingly <strong>for</strong> various test images based on the location <strong>of</strong><br />
the focal point pixel in one <strong>of</strong> the nine regions. The corresponding<br />
shift vectors were then used to align the channels.<br />
CONCLUSIONS<br />
In this paper, we proposed an aut<strong>of</strong>ocusing algorithm which<br />
restores an out-<strong>of</strong>-focus image with multiple, differently out<strong>of</strong>-focus<br />
objects. A novel FA configuration is proposed <strong>for</strong><br />
modeling out-<strong>of</strong>-focus blur in images. The proposed algorithm<br />
starts with a single input image and multiple source<br />
images with different apertures are generated using channel<br />
shifting. The fusion is carried out <strong>for</strong> segmented regions<br />
from each source image using the SML operator. The s<strong>of</strong>t<br />
decision fusion algorithm overcomes undesired artifacts in<br />
the region <strong>of</strong> merging in the fused images. Experimental<br />
results show that the proposed algorithm works well <strong>for</strong> the<br />
images with multiple out-<strong>of</strong>-focus objects.<br />
ACKNOWLEDGMENTS<br />
This research was supported by Seoul Future Contents Convergence<br />
(SFCC) Cluster established by Seoul R&BD Program<br />
and by the Korea <strong>Science</strong> and Engineering Foundation<br />
(KOSEF) through the National Research Laboratory Program<br />
funded by the Ministry <strong>of</strong> <strong>Science</strong> and Technology<br />
(M103 0000 0311-06J0000-31110).<br />
REFERENCES<br />
1 Y. Ishihara and K. Tanigaki, “A high photosensitivity IL-CCD image<br />
sensor with monolithic resin lens array”, in Proc. IEEE Integrated<br />
Electronic and Digitial Microscopy (IEEE Press, Piscataway, NJ, 1983) pp.<br />
497–500.<br />
2 J. Tanida, R. Shogenji, Y. Kitumara, K. Yamada, M. Miyamoto, and S.<br />
Miyatake, “Color image with an integrated compound imaging system”,<br />
Opt. Express 18, 2109–2117 (2003).<br />
3 E. R. Dowski and G. E. Johnson, “Wavefront coding system: amodern<br />
method <strong>of</strong> achieving high per<strong>for</strong>mance and/or low cost imaging<br />
systems”, Proc. SPIE 3779, 137–145 (1999).<br />
4 S. Kim and J. K. Paik, “Out-<strong>of</strong>-focus blur estimation and restoration <strong>for</strong><br />
digital aut<strong>of</strong>ocusing system”, Electron. Lett. 34, 1217–1219 (1998).<br />
5 J. Shin, V. Maik, J. Lee, and J. Paik, “Multi-object digital aut<strong>of</strong>ocusing<br />
using image fusion”, Lect. Notes Comput. Sci. 3708, 806–813 (2005).<br />
6 V. Maik, J. Shin, and J. Paik, “Regularized image restoration by means <strong>of</strong><br />
fusion <strong>for</strong> digital aut<strong>of</strong>ocusing”, Lecture Notes in Artificial Intelligence<br />
3802, 929–934 (2005).<br />
7 G. Ligthart and F. Groen, “A comparison <strong>of</strong> different aut<strong>of</strong>ocus<br />
algorithms”, IEEE Int. Conf. Pattern Recognition (IEEE Press,<br />
Piscasataway, NJ, 1992) pp. 597–600.<br />
8 M. Subbarao, T. C. Wei, and G. Surya, “Focused image recovery from<br />
two defocused images recorded with different camera settings”, IEEE<br />
Trans. Image Process. 4, 1613–1628 (1995).<br />
9 M. Matsuyama, Y. Tanji, and M. Tanka, “Enhancing the ability <strong>of</strong><br />
NAS-RIF algorithm <strong>for</strong> blind image deconvolution”, Proc. IEEE Int.<br />
Conf. Circuits and Systems, vol. 4 (IEEE Press, Piscataway, NJ, 2000) pp.<br />
553–556.<br />
10 A. Tekalp and H. Kaufman, “On statistical identification <strong>of</strong> a class <strong>of</strong><br />
linear space-invariant image blurs using non minimum-phase ARMA<br />
models”, IEEE Trans. Acoust., Speech, Signal Process. ASSP36,<br />
1360–1363 (1988).<br />
11 K. Kodama, H. Mo, and A. Kubota, “All-in-focus image generation by<br />
merging multiple differently focused images in three-dimensional<br />
frequency domain”, Lect. Notes Comput. Sci. 3768, 303–314 (2005).<br />
12 K. Aizawa, K. Kodama, and A. Kubota, “Producing object-based special<br />
effects by fusing multiple differently focused images”, IEEE Trans.<br />
Circuits Syst. Video Technol. 10, 323–330 (2000).<br />
13 L. Bogoni and M. Hansen, “Pattern selective color image fusion”, Int. J.<br />
Pattern Recognit. Artif. Intell. 34, 1515–1526 (2001).<br />
14 S. Li, J. T. Kwok, and Y. Wang, “Combination <strong>of</strong> images with diverse<br />
focuses using the spatial frequency”, Int. J. Inf. Fusion 2, 169–176<br />
(2001).<br />
15 Z. Zhang and R. S. Blum, “A categorization <strong>of</strong> multiscaledecomposition-based<br />
image fusion schemes with a per<strong>for</strong>mance study<br />
<strong>for</strong> a digital camera application”, Proc. IEEE 87, 1315–1326 (1999).<br />
16 V. Maik, J. Shin, J. Lee, and J. Paik, “Pattern selective image fusion <strong>for</strong><br />
multi-focus image reconstruction”, Lect. Notes Comput. Sci. 3691,<br />
677–684 (2005).<br />
17 A. Kubota, K. Kodama, and K. Aizawa, “Registration and blur estimation<br />
method <strong>for</strong> multiple differently focused images”, Proc. IEEE Int. Conf.<br />
Image Processing, vol. 2 (IEEE Press, Piscataway, NJ, 1999) pp. 515–519.<br />
18 S. K. Lee, S. H. Lee, and J. S. Choi, “Depth measurement using frequency<br />
analysis with an active projection”, Proc. IEEE Conf. Image Processing,<br />
vol. 3 (IEEE Press, Piscataway, NJ, 1999) 906–909.<br />
19 G. Piella, “A general framework <strong>for</strong> multi resolution image fusion: From<br />
pixels to regions”, J. Inf. Fusion 4, 259–280 (2003).<br />
20 T. Adelson and J. Y. Wang, “Single lens stereo with a plenoptic camera”,<br />
IEEE Trans. Pattern Anal. Mach. Intell. 2, 99–106 (1992).<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 379
<strong>Journal</strong> <strong>of</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology® 51(4): 380–385, 2007.<br />
© <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology 2007<br />
Error Spreading Control in Image Steganographic<br />
Embedding Schemes Using Unequal Error Protection<br />
Ching-Nung Yang, Guo-Jau Chen and Tse-Shih Chen<br />
CSIE Dept., National Dong Hwa University, #1, Da Husueh Rd., Sec. 2, Hualien, Taiwan<br />
E-mail: cnyang@mail.ndhu.edu.tw<br />
Rastislav Lukac<br />
The Edward S. Rogers Sr. Department <strong>of</strong> ECE, University <strong>of</strong> Toronto, 10 King’s College Road, Toronto,<br />
Ontario, M5S 3G4 Canada<br />
Abstract. A steganographic scheme proposed by van Dijk and<br />
Willems can alter a relatively small amount <strong>of</strong> bits to hide the secret<br />
compared to other schemes while reducing the distortion and improving<br />
the resistance against the steganalysis. However, one bit<br />
error in the embedding scheme by van Dijk and Willems may result<br />
in multibit error when extracting the hidden data. This problem is<br />
called as error spreading. It is observed that only some single-bit<br />
errors suffer from error spreading. In this paper, we propose a new<br />
steganographic solution which takes advantage <strong>of</strong> unequal error<br />
protection codes and allows <strong>for</strong> the different protection <strong>of</strong> different<br />
secret bits. Thus the proposed solution can effectively protect bits<br />
which could suffer from error spreading. In addition, it saves parity<br />
bits, thus greatly reducing the amount <strong>of</strong> bit alterations compared to<br />
the relevant previous schemes. Experimentation using various test<br />
images indicates that the proposed solution achieves the trade<strong>of</strong>f<br />
between the per<strong>for</strong>mance and the protection <strong>of</strong> the embedded<br />
secret. © 2007 <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology.<br />
DOI: 10.2352/J.<strong>Imaging</strong>Sci.Technol.200751:4380<br />
INTRODUCTION<br />
Steganography is a method <strong>of</strong> hiding or embedding the secret<br />
message into a cover media to ensure that an unintended<br />
party will not be aware <strong>of</strong> the existence <strong>of</strong> the embedded<br />
secret. Popular steganographic techniques <strong>for</strong> visual<br />
data protection embed the secret message such as a binary<br />
image by manipulating the least significant bit (LSB) plane<br />
<strong>of</strong> a cover image, thus producing the so-called stegoimage.<br />
In the simplest <strong>for</strong>m, the data embedding can be realized by<br />
c 0 s with c 0 denoting the least significant bit <strong>of</strong> a pixel from<br />
the cover image and s denoting the secret bit.<br />
A more efficient steganographic method was proposed<br />
by van Dijk and Willems. 1 Their coded LSB method attempts<br />
to reduce the distortion when the noise (e.g., due to faulty<br />
communication channels) or active attacks by the third party<br />
(intentional modifications <strong>of</strong> some insignificant bits in a<br />
cover image to prevent the extraction <strong>of</strong> hidden secret by the<br />
authorized user) are introduced into the stegoimage. However,<br />
in some situations, one error bit produced during the<br />
stegoimage transmission phase or by active attacks <strong>of</strong>ten results<br />
in two or more errors when decoding (extracting) the<br />
Received Oct. 28, 2006; accepted <strong>for</strong> publication Mar. 22, 2007.<br />
1062-3701/2007/514/380/6/$20.00.<br />
embedded message. This phenomenon, known as an error<br />
spreading problem, affects the clearness <strong>of</strong> the extracted secret<br />
image. A straight<strong>for</strong>ward application <strong>of</strong> error correction,<br />
i.e., using stegoencoding to hide the secret first and then<br />
adding the parity to provide the error correcting (EC) capability,<br />
will inevitably increase the required amount <strong>of</strong> bit<br />
alterations. Thus, the risk <strong>of</strong> being detected will increase.<br />
Zhang and Wang 2 proposed a new stegoencoding approach<br />
combining the coded LSB and EC capability simultaneously<br />
to address the error spreading problem. Their solution has<br />
the same error correcting capability <strong>for</strong> all protected bits, but<br />
according to our observation the error spreading does not<br />
affect all spatial locations <strong>of</strong> the secret image.<br />
In this paper, we propose a more reasonable solution,<br />
unequal error protection (UEP) codes, to obtain the different<br />
protection ability <strong>for</strong> nonaffected and affected bits and to<br />
save parity bits. It will be shown that the proposed solution<br />
outper<strong>for</strong>ms the previous relevant solutions in terms <strong>of</strong> the<br />
trade<strong>of</strong>f between the per<strong>for</strong>mance and the protection <strong>of</strong> the<br />
embedded secret.<br />
The rest <strong>of</strong> this paper is organized as follows. In the<br />
Coded LSB Scheme Section, the coded LSB scheme is described.<br />
In the EC Codes Based Error Correction: Zhang-<br />
Wang Scheme Section, Zhang-Wang stegoencoding based on<br />
EC codes is presented to show a solution <strong>for</strong> the error<br />
spreading problem. Our scheme based on UEP codes is proposed<br />
in the UEP Codes Based Error Correction the Proposed<br />
Scheme Section. Motivation and design characteristics<br />
are discussed in detail. In the Comparison and Experimental<br />
Results Section, the proposed method is tested using a variety<br />
<strong>of</strong> test images. The effect <strong>of</strong> UEP codes-based data embedding<br />
is evaluated and compared with the previous approaches.<br />
Finally, conclusions are drawn in the Conclusions<br />
Section.<br />
CODED LSB SCHEME<br />
In the plain LSB embedding scheme, the secret bits are hidden<br />
by simply replacing LSBs <strong>of</strong> the cover pixels. Due to the<br />
noiselike appearance <strong>of</strong> the LSB plane <strong>of</strong> natural images,<br />
embedding n bits implies, in average, the alteration <strong>of</strong> n/2<br />
original LSBs. To reduce the number <strong>of</strong> altered LSBs and<br />
380
Yang et al.: Error spreading control in image steganographic embedding schemes using unequal error protection<br />
Table I. Eight cosets <strong>for</strong> the coded LSB scheme with l=3, n=7 using 7,4 Hamming code with Gx=x 3 +x 2 +1.<br />
L 0<br />
L 1<br />
L 2<br />
L 3<br />
L 4<br />
L 5<br />
L 6<br />
L 7<br />
0000000, 0001101, 0011010, 0010111, 0110100, 0111001, 0101110, 0100011,<br />
1101000, 1100101, 1110010, 1111111, 1011100, 1010001, 1000110, 1001011<br />
0000001, 0001100, 0011011, 0010110, 0110101, 0111000, 0101111, 0100010,<br />
1101001, 1100100, 1110011, 1111110, 1011101, 1010000, 1000111, 1001010<br />
0000010, 0001111, 0011000, 0010101, 0110110, 0111011, 0101100, 0100001,<br />
1101010, 1100111, 1110000, 1111101, 1011110, 1010011, 1000100, 1001001<br />
0000011, 0001110, 0011001, 0010100, 0110111, 0111010, 0101101, 0100000,<br />
1101011, 1100110, 1110001, 1111100, 1011111, 1010010, 1000101, 1001000<br />
0000100, 0001001, 0011110, 0010011, 0110000, 0111101, 0101010, 0100111,<br />
1101100, 1100001, 1110110, 1111011, 1011000, 1010101, 1000010, 1001111<br />
0000101, 0001000, 0011111, 0010010, 0110001, 0111100, 0101011, 0100110,<br />
1101101, 1100000, 1110111, 1111010, 1011001, 1010100, 1000011, 1001110<br />
0000110, 0001011, 0011100, 0010001, 0110010, 0111111, 0101000, 0100101,<br />
1101110, 1100011, 1110100, 1111001, 1011010, 1010111, 1000000, 1001101<br />
0000111, 0001010, 0011101, 0010000, 0110011, 0111110, 0101001, 0100100,<br />
1101111, 1100010, 1110101, 1111000, 1011011, 1010110, 1000001, 1001100<br />
preserve the original features such as edges and fine details<br />
<strong>of</strong> the cover image, the coded LSB scheme <strong>of</strong> Ref. 1 divides<br />
the secret image into chips <strong>of</strong> l bits. Each l-bit chip is then<br />
embedded into LSBs <strong>of</strong> n pixels using n,k cyclic codes<br />
where n is the in<strong>for</strong>mation code length and l=n−k.<br />
k<br />
Let G 1 x= i=0 g 1,i x i l<br />
and G 2 x= i=0 g 2,i x i be two binary<br />
polynomials with degree k and l, respectively, so that<br />
G 1 x·G 2 x=x n +1. Using n,k cyclic codes with the generating<br />
function Gx=G 2 x, it is possible to construct 2 l<br />
code sets which consist <strong>of</strong> unique 2 k codewords. Thus, each<br />
code set can be used to describe one l-bit secret chip by<br />
choosing the nearest codeword to represent the embedded<br />
secret, as depicted in algorithm 1.<br />
Algorithm 1<br />
Inputs: secret message <strong>of</strong> l bits s l−1 s l−2 ...s 0 , and cover image<br />
O with n LSBs c n−1 c n−2 ...c 0 .<br />
Output: stegoimage O with n LSBs c n−1 c n−2 ...,c 0 .<br />
Step 1: Choose one cyclic n,k code with the generating<br />
function Gx=g l x l + ¯ +g 1 x+g 0 and then select any<br />
k-tuples as the input to construct a code set (coset) <strong>of</strong> 2 k<br />
codewords. Choose one n-tuple codeword that does not<br />
appear in this coset, and then add an unused n-tuple to<br />
all the codewords in the coset to construct another coset.<br />
Step 2: Repeat step 1, until all 2 n codewords are used. The<br />
process generates 2 l different cosets L 0 ,L 1 ,...,L 2 l −1 that<br />
include 2 k codewordsineachcoset.<br />
Step 3: Encrypt the secret bits s l−1 ,s l−2 ,...,s 0 by choosing<br />
the coset L i with i= l−1 i=0 s i 2 i . Then, find the codeword<br />
c n−1 c n−2 ...c 0 in the coset L i such that the Hamming distance<br />
between c n−1 c n−2 ...c 0 and c n−1 c n−2 ...c 0 is minimum.<br />
Step 4: Deliver n LSBs c n−1 c n−2 ...c 0 to the corresponding<br />
pixels in the embedded stegoimage O.<br />
As shown in Ref. 2, the efficiency <strong>of</strong> the steganographic<br />
schemes can be demonstrated using the so-called embedding<br />
rate (ER) which is defined as the number <strong>of</strong> embedded bits<br />
per pixel, i.e., ER=l/n. Another suitable criterion is the socalled<br />
embedding efficiency (EE) which is calculated as the<br />
number <strong>of</strong> altered bits per pixel, i.e., EE=l/l alt where l alt<br />
denotes the average LSB alteration when l secret bits are<br />
embedded into n LSBs. The value l alt can be calculated from<br />
all codewords in the cosets by a computer program. The ER<br />
parameter is suitable <strong>for</strong> discussing the embedded capacity<br />
whereas the EE parameter is used to evaluate the distortion<br />
in the cover image. It is obvious that <strong>for</strong> the plain LSB embedding<br />
scheme ER=l/n=n/n=100% and EE=1/l alt<br />
=n/n/2=200%. For the comparison purposes, the values<br />
corresponding to algorithm 1 (coded LSB scheme with<br />
l=3 and n=7) are provided below.<br />
Let us assume algorithm 1 with n=7 and l=3, i.e., the<br />
objective is to embed three secret bits into seven LSBs. The<br />
above setting implies that k=4, resulting in x 7 +1=x 4 +x 3<br />
+x 2 +1x 3 +x 2 +1 and Gx=x 3 +x 2 +1. After the first two<br />
steps in algorithm 1, we construct eight cosets with sixteen<br />
codewords in each coset (see Table I). Suppose that 101<br />
denotes the secret and 0001110 denotes the original set <strong>of</strong><br />
LSBs. We use the coset L 5 to find the codeword 1001110 that<br />
has the minimum Hamming distance equal to one from<br />
0001110 while altering only one LSB to embed three secret<br />
bits. The embedding rate and embedding efficiency are<br />
ER=3/7=42.9% and EE=3/0.875=343%, respectively.<br />
Note that l alt =0.875 <strong>for</strong> Table I. It is easy to see that the plain<br />
LSB embedding scheme ER=100% ,EE=200% has larger<br />
embedded capacity than the coded LSB scheme. On the<br />
other hand, the coded LSB scheme modifies fewer bits, thus<br />
reducing the distortion in the cover image.<br />
However, the coded LSB scheme suffers from the error<br />
spreading problem. The occurrence <strong>of</strong> one error bit in the<br />
encoded LSBs may cause more than one bit error during the<br />
secret message extraction process. Considering the above example,<br />
we can use 0000000 to carry the secret 000. Suppose<br />
that there is one error bit, <strong>for</strong> example, 0010000. Then, according<br />
to Table I, the extracted secret is 111, i.e., there are<br />
three error bits. However, employing the error pattern<br />
0000001 results in the extracted secret 001 which corresponds<br />
to only one error bit (no error spreading). This sug-<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 381
Yang et al.: Error spreading control in image steganographic embedding schemes using unequal error protection<br />
Table II. Eight cosets <strong>for</strong> the EC-based coded LSB scheme Zhang-Wang Scheme with l=3,N E =11 and 1-error correcting capability using 7,4<br />
Hamming code and 11,7 shorten Hamming code.<br />
L 0<br />
L 1<br />
L 2<br />
L 3<br />
L 4<br />
L 5<br />
L 6<br />
L 7<br />
00000000000, 10010001101, 10110011010, 00100010111, 11110110100, 01100111001<br />
01000101110, 11010100011, 01111101000, 11101100101, 11001110010, 01011111111<br />
10001011100, 00011010001, 00111000110, 10101001011<br />
11100000001, 01110001100, 01010011011, 11000010110, 00010110101, 10000111000<br />
10100101111, 00110100010, 10011101001, 00001100100, 00101110011, 10111111110<br />
01101011101, 11111010000, 11011000111, 01001001010<br />
01010000010, 11000001111, 11100011000, 01110010101, 10100110110, 00110111011<br />
00010101100, 10000100001, 00101101010, 10111100111, 10011110000, 00001111101<br />
11011011110, 01001010011, 01101000100, 11111001001<br />
10110000011, 00100001110, 00000011001, 10010010100, 01000110111, 11010111010<br />
11110101101, 01100100000, 11001101011, 01011100110, 01111110001, 11101111100<br />
00111011111, 10101010010, 10001000101, 00011001000<br />
10100000100, 00110001001, 00010011110, 10000010011, 01010110000, 11000111101<br />
11100101010, 01110100111, 11011101100, 01001100001, 01101110110, 11111111011<br />
00101011000, 10111010101, 10011000010, 00001001111<br />
01000000101, 11010001000, 11110011111, 01100010010, 10110110001, 00100111100<br />
00000101011, 10010100110, 00111101101, 10101100000, 10001110111, 00011111010<br />
11001011001, 01011010100, 01111000011, 11101001110<br />
11110000110, 01100001011, 01000011100, 11010010001, 00000110010, 10010111111<br />
10110101000, 00100100101, 10001101110, 00011100011, 00111110100, 10101111001<br />
01111011010, 11101010111, 11001000000, 01011001101<br />
00010000111, 10000001010, 10100011101, 00110010000, 11100110011, 01110111110<br />
01010101001, 11000100100, 01101101111, 11111100010, 11011110101, 01001111000<br />
10011011011, 00001010110, 00101000001, 10111001100<br />
gests that if the error falls in the right three positions, i.e.,<br />
0000100, 0000010, and 0000001, then there is still only one<br />
error in the extracted secret and the damage is not expanded.<br />
However, <strong>for</strong> the other four error patterns 1000000,<br />
0100000, 0010000, and 0001000, the decoded secret is 110,<br />
011, 111, and 101, respectively, indicating that the extracted<br />
secret suffers from more than one error bit. Since such errors<br />
affect the quality <strong>of</strong> the extracted secret message, the scheme<br />
should be improved by using the error correcting mechanism,<br />
such as one described below based on EC codes.<br />
EC CODES BASED ERROR CORRECTION:<br />
ZHANG-WANG SCHEME<br />
Following the previous approach, a coded LSB scheme with<br />
ER=l/n is constructed using the generating function<br />
Gx=G 2 x. Then, 2 k n-tuple vectors in each coset are encoded<br />
into an N E ,n cyclic EC code with N E denoting the<br />
code length <strong>of</strong> EC codes. Although in this improved coded<br />
LSB scheme the embedding rate is reduced to ER=l/N E , the<br />
scheme now has the error correcting capability <strong>of</strong> N E ,n<br />
cyclic EC codes.<br />
Let us consider the previous example with one-error<br />
correcting capability and parameters n=7, l=3 and N E =11.<br />
A 7,4 Hamming code is used to embed three secret bits<br />
and a 11,7 shorten Hamming code is used to achieve one<br />
error correcting capability. For example, embedding the secret<br />
000 into the 7-tuple 0001101 first and then appending<br />
the parity 1001 <strong>for</strong>ms the codeword 10010001101 which is<br />
listed as the second codeword in the coset L 0 (see Table II<br />
listing all eight generated cosets). If one error occurs in the<br />
sixth position, resulting in the codeword 10010101101, then<br />
the minimum Hamming distance is associated with the<br />
codeword 10010001101 from L 0 and the extracted secret is<br />
000. Because <strong>of</strong> the error correcting capability <strong>of</strong> the 11,7<br />
shorten Hamming code, the error is always corrected no<br />
matter where the error bit occurs, thus overcoming the error<br />
spreading problem. On the other hand, the approach is less<br />
efficient than the conventional method, as it reduces the<br />
embedding rate from 42.9% to 3/11=27.3% and also decreases<br />
the embedding efficiency from 343% to<br />
3/2.625=114% (note that l alt =2.625 <strong>for</strong> Table II). There<strong>for</strong>e,<br />
the different correction mechanism is needed. Since<br />
only the error in the first k bits <strong>of</strong> an n-tuple will produce<br />
additional errors in the secret extraction phase, it should be<br />
sufficient to ensure the validity <strong>of</strong> the first k bits instead <strong>of</strong><br />
all n bits.<br />
UEP CODES BASED ERROR CORRECTION: THE<br />
PROPOSED SCHEME<br />
UEP codes, a category <strong>of</strong> EC codes, allow different protection<br />
<strong>for</strong> different bit locations. 3,4 In practice, some in<strong>for</strong>mation<br />
bits are protected against a greater number <strong>of</strong> errors<br />
than other, less significant, in<strong>for</strong>mation bits. Basically, a UEP<br />
code can be denoted as n,k,d 1 ,d 2 ,...,d k . By employing<br />
UEP codes to protect the message, the occurrence <strong>of</strong> no<br />
more than ⌊d i −1/2⌋ errors in the transmitted codeword<br />
does not affect the correctness <strong>of</strong> the ith bit in the decoded<br />
message.<br />
It was noted that the first k bits <strong>of</strong> vectors need an<br />
enhanced protection to prevent the error spreading. There<strong>for</strong>e,<br />
we propose to apply UEP codes to assure the correctness<br />
<strong>of</strong> these k bits and reduce the number <strong>of</strong> redundant<br />
parity bits. The main difference between the UEP-based<br />
scheme and the EC-based scheme relates to the use <strong>of</strong><br />
382 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Yang et al.: Error spreading control in image steganographic embedding schemes using unequal error protection<br />
Table III. Eight cosets <strong>for</strong> the UEP-based coded LSB scheme the proposed scheme with l=3, N U =11 using 10,7,3333222 UEP code.<br />
L 0<br />
L 1<br />
L 2<br />
L 3<br />
L 4<br />
L 5<br />
L 6<br />
L 7<br />
0000000000*11000011011000011010010001011110001101000100111001<br />
000010111011001000110001101000110110010110011100100101111111<br />
1001011100010101000100010001101101001011<br />
0010000001*11100011001010011011011001011010101101010110111000<br />
001010111111101000100011101001111110010010111100110111111110<br />
1011011101011101000000110001111111001010<br />
010000001010000011111100011000000001010111001101100000111011<br />
010010110010001000010101101010100110011111011100000001111101<br />
1101011110000101001101010001001001001001<br />
011000001110100011101110011001001001010011101101110010111010<br />
011010110110101000000111101011101110011011111100010011111100<br />
1111011111001101001001110001011011001000<br />
1000000100*01000010010000011110110001001100001100001100111101<br />
100010101001001001111001101100010110000100011101101101111011<br />
0001011000110101010110010000100101001111<br />
101000010101100010000010011111111001001000101100011110111100<br />
101010101101101001101011101101011110000000111101111111111010<br />
0011011001111101010010110000110111001110<br />
110000011000000010110100011100100001000101001100101000111111<br />
110010100000001001011101101110000110001101011101001001111001<br />
0101011010100101011111010000000001001101<br />
111000011100100010100110011101101001000001101100111010111110<br />
111010100100101001001111101111001110001001111101011011111000<br />
0111011011101101011011110000010011001100<br />
N U ,n,d 1 ,d 2 ,...,d k UEP codes with N U denoting the<br />
code length <strong>of</strong> UEP codes instead <strong>of</strong> N E ,n EC codes. Since<br />
the value <strong>of</strong> N U is smaller than N E , the proposed UEP-based<br />
coded LSB scheme will have the higher embedding rate<br />
while still providing the same protection <strong>of</strong> the secret to the<br />
error spreading.<br />
As be<strong>for</strong>e, let us consider the scenario with one error<br />
correcting capability and parameters n=7, l=3, and<br />
N U =10. Suppose that the 10,7,3333222 UEP code is<br />
used to ensure the protection against errors. The corresponding<br />
eight cosets are listed in Table III. Assuming that<br />
the embedded secret and the encoded result are, <strong>for</strong> instance,<br />
000 and 0000000000, respectively, one error can occur in the<br />
following cases:<br />
• The presence <strong>of</strong> the error in the 7th bit (from right)<br />
implies the codeword 0001000000. As shown in Table<br />
III, there is only one codeword 0000000000 in the coset<br />
L 0 with the unit Hamming distance to 0001000000. In<br />
this case, the recovered secret is 000, i.e., no processing<br />
error.<br />
• If the error affects the third bit (from right) beyond the<br />
error correcting capability <strong>of</strong> the considered UEP code,<br />
then 0000000000 in the coset L 0 and 1000000100 in the<br />
coset L 4 are the two codewords with the unit Hamming<br />
distance to the codeword 0000000100 under consideration.<br />
The decoding process can result in the extracted<br />
secret 000 or 100, respectively. Thus, even in the latter<br />
situation (i.e., 100), there is still only one error, suggesting<br />
no error spreading.<br />
• Finally, the alteration <strong>of</strong> the 8th bit (from right) due to<br />
the error implies 0010000000 which will be decoded as<br />
0000000000 in the coset L 0 or 0010000001 in the coset<br />
L 1 . This suggests that the secret can be extracted as 000<br />
or 001, respectively. Similar to the previous case, even<br />
when 001 is used as the extracted secret, the proposed<br />
method still overcomes the error spreading problem.<br />
It is evident that when no more than one error falls in<br />
the first four bits <strong>of</strong> the original 7-bit vector, the use <strong>of</strong> UEP<br />
codes ensures that the first four bits will be correctly decoded<br />
and the error will be corrected, as ⌊3−1/2⌋=1, i.e.,<br />
the Hamming distance between the first four bits <strong>of</strong> two<br />
codewords is three providing one error correcting capability<br />
<strong>for</strong> the first four bits. However, no error spreading is also<br />
observed in the situations when one error occurs in other<br />
places. This is due to the fact that a single error in other<br />
places will result, in the worst case, in a single error in the<br />
decoded secret bits. The achieved embedding rate and embedding<br />
efficiency are ER=l/N U =30% and EE=l/l alt<br />
=133%. Note that l alt =2.25 <strong>for</strong> Table III.<br />
By employing the familiar representation used in UEP<br />
codes, the 11,7 shorten Hamming code can be represented<br />
as 11,7,3333333. There<strong>for</strong>e, if we protect the first four<br />
bits only, then we can save one redundant checking bit by<br />
using the 11,7,3333222 UEP code. Note that the errorcorrecting<br />
capability <strong>of</strong> N U ,n,d 1 ,d 2 ,...,d k UEP codes is<br />
not better compared to N E ,n,d EC codes. However, UEP<br />
codes have better embedding rate and embedding efficiency<br />
and also overcome the error spreading problem.<br />
COMPARISON AND EXPERIMENTAL RESULTS<br />
Different analytical tools, such as the sample pair analysis 5<br />
and image quality metrics, 6 are used to analyze the<br />
steganographic solutions. To resist the various attacks on the<br />
stegoimage while still providing the required per<strong>for</strong>mance,<br />
an ideal steganographic scheme should be constructed by<br />
considering the relation between the cover image and the<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 383
Yang et al.: Error spreading control in image steganographic embedding schemes using unequal error protection<br />
Table IV. Comparison <strong>of</strong> the conventional, EC-based, and UEP-based coded LSB schemes <strong>for</strong> n=2 to 12.<br />
Conventional scheme Zhang-Wang EC-based scheme Proposed UEP-based scheme<br />
n,k,l ER=l/n N E ,n,d ER=l/N E N U ,n,d 1 ,d 2 ,...,d k ER=l/N U<br />
2,1,1 50.0% 5,2,3 20.0% 4,2,32 25.0%<br />
4,1,3 75.0% 7,4,3 42.9% 6,4,3222 50.0%<br />
5,4,1 20.0% 9,5,4 11.1% 8,5,33332 12.5%<br />
6,4,2 33.3% 10,6,4 20.0% 9,6,333322 22.2%<br />
7,4,3 42.9% 11,7,4 27.3% 10,7,3333222 30.0%<br />
8,4,4 50% 12,8,4 33.3% 11,8,33332222 36.4%<br />
9,4,5 55.6% 13,9,4 38.5% 12,9,333322222 41.7%<br />
10,4,6 60% 14,10,4 42.9% 13,10,3333222222 46.2%<br />
11,4,7 63.6% 15,11,4 46.7% 14,11,33332222222 50.0%<br />
12,4,8 66.7% 17,12,5 47.1% 15,12,333322222222 53.3%<br />
stegoimage (or the relations between the stegopixel and the<br />
original pixel) and simultaneously the scheme should allow<br />
to achieve the high embedding rates. There<strong>for</strong>e, the schemes<br />
under consideration are evaluated here in terms <strong>of</strong> the embedding<br />
rate, embedding efficiency, and the peak-signal-tonoise<br />
(PSNR) ratio calculated using the original cover image<br />
and its stegoversion.<br />
Table IV shows the embedding rates and codes used in<br />
the conventional scheme, EC-based scheme and UEP-based<br />
scheme <strong>for</strong> n=2 to 12. As it can be seen from the listed<br />
results, all UEP-based schemes have the shorter code length<br />
than the EC-based schemes. Both these schemes have the<br />
ability to correct the first k bits when no larger than one<br />
error occurs, and avoid the error spreading problem. Table V<br />
shows the detail comparison <strong>for</strong> these three schemes with<br />
n,k,l=7,4,3. Code-based schemes address the error<br />
spreading problem at the cost <strong>of</strong> their smaller ER and EE.<br />
The UEP-based scheme, with ER=30.0% and EE=133%,<br />
needs 3334 pixels in a cover image to embed 1000<br />
=333430% secret bits while altering 750 LSBs<br />
=33342.25/12 within these embedded pixels; and, as<br />
can be seen, it outper<strong>for</strong>ms the EC-based scheme.<br />
In order to compare the distortion caused by the<br />
schemes under consideration, the well-known 259259 test<br />
gray-scale images “Baboon”, “Barb”, “Boat”, “Elaine”,<br />
“Mena”, and “Peppers” have been used as the cover images.<br />
The secret NDHU (National Dong Hwa University) “logo”<br />
and “text” gray-scale images to be embedded are shown in<br />
Figure 1. Secret images with size 5959 left, 4747 middle, and<br />
4949 pixels right.<br />
Table V. ER and EE <strong>for</strong> the conventional, EC-based, and UEP-based coded LSB schemes<br />
with n,k,l=7,4,3.<br />
Coded LSB scheme ER EE Embedding <strong>of</strong> 1000 bits<br />
Number <strong>of</strong> pixels needed Altered LSBs<br />
Conventional 42.9% 342% 2334 292<br />
EC-based 27.3% 114% 3667 877<br />
UEP-based 30.0% 133% 3334 750<br />
Table VI. PSNRdB between the cover image and its stegoimage <strong>for</strong> the conventional,<br />
EC-based, and UEP-based schemes.<br />
Coded LSB scheme Conventional EC-based UEP-based<br />
Cover image<br />
Baboon 57.173 54.671 54.691<br />
Barb 57.148 54.687 54.675<br />
Boat 57.181 54.696 54.677<br />
Elaine 57.143 54.675 54.681<br />
Mena 57.151 54.683 54.710<br />
Peppers 57.179 54.693 54.671<br />
Figure 1. Note that due to the different embedding rates <strong>for</strong><br />
different schemes, the secret images <strong>of</strong> 5959, 4747, and<br />
4949 pixels <strong>for</strong> the conventional, EC-based, and UEPbased<br />
schemes, respectively, have been used to ensure fair<br />
comparisons. The achieved PSNR values are listed in Table<br />
VI. The results indicate that the considered schemes produce<br />
high-quality stegoimages and the highest PSNR was<br />
achieved by the conventional scheme due to the higher EE<br />
(343% versus 114% in the EC-based scheme and 133% in<br />
the UEP-based scheme). This suggests that adding the error<br />
correcting capability does not distort the stegoimage seriously<br />
and that employing the error correcting codes (EC or<br />
UEP) in the coded LSB embedding scheme constitutes a<br />
reasonable and practical solution to overcome the error<br />
spreading problem.<br />
384 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Yang et al.: Error spreading control in image steganographic embedding schemes using unequal error protection<br />
Figure 2. Recovered secret images with BER=2%, 4%, and 8% <strong>for</strong> errors<br />
placed in random positions.<br />
To further study the error spreading problem, the two<br />
types <strong>of</strong> error patterns, namely random errors and worst<br />
errors have been added into the LSBs <strong>of</strong> stegoimages. The<br />
first type (random errors) means that the errors are randomly<br />
distributed in n-bit vector. However, the second type<br />
(worst errors) means that the errors occur in the worst positions<br />
(the first k bits <strong>of</strong> the original n-bit vector) where will<br />
cause error spreading. Figures 2 and 3 show the corresponding<br />
results obtained by extracting the embedded secret images<br />
from the noise corrupted stegoimages. Visual inspection<br />
<strong>of</strong> the results reveals that in a noisy environment the<br />
schemes based on UEP and EC codes have comparable per<strong>for</strong>mance<br />
and clearly outper<strong>for</strong>m the conventional coded<br />
LSB scheme. Moreover, since the proposed UEP-based<br />
scheme has higher ER than the EC-based scheme, it can be<br />
concluded that our solution provides a trade<strong>of</strong>f between the<br />
data embedding per<strong>for</strong>mance and the protection <strong>of</strong> the embedded<br />
secret.<br />
CONCLUSIONS<br />
A refined steganographic solution was introduced. Using<br />
UEP codes, we overcame the error spreading problem in the<br />
coded LSB steganographic scheme originally proposed by<br />
van Dijk and Willems. Our solution has the same correction<br />
effect as the Zhang-Wang EC-based scheme while allowing<br />
<strong>for</strong> lower embedding rates. This suggests that the solution<br />
Figure 3. Recovered secret images with BER=2%, 4%, and 8% <strong>for</strong> errors<br />
placed in the worst positions.<br />
proposed in this paper embeds the same secret message with<br />
the higher efficiency and produces less distortion in the generated<br />
stegoimage. The proposed solution is suitable <strong>for</strong> applications,<br />
such as transmission <strong>of</strong> the private digital materials<br />
(e.g., documents or signature images) through public<br />
and wireless networks, where data hiding and protection<br />
against communication errors are required or recommended.<br />
ACKNOWLEDGMENT<br />
This work was supported in part by TWISC@NCKU, National<br />
<strong>Science</strong> Council under the Grants NSC 94-3114-P-<br />
006-001-Y.<br />
REFERENCES<br />
1 M. van Dijk and F. Willems, “Embedding in<strong>for</strong>mation in grayscale<br />
images”, Proc. 22nd Symp. In<strong>for</strong>m. Theory in the Benelux (Elsevier,<br />
Netherlands, 2001), pp. 147–154.<br />
2 X. Zhang and S. Wang, “Stego-encoding with error correction<br />
capability”, IEICE Trans. Fundamentals E88-A, 3663–3667 (2005).<br />
3 W. J. van Gils, “Two topics on linear unequal error protection codes:<br />
bounds on their length and cyclic code classes”, IEEE Trans. Inf. Theory<br />
29, 866–876 (1983).<br />
4 M. C. Lin, C. C. Lin, and S. Lin, “Computer search <strong>for</strong> binary cyclic UEP<br />
codes <strong>of</strong> odd length up to 65”, IEEE Trans. Inf. Theory 36, 924–935<br />
(1990).<br />
5 S. Dumitrescu, X. Wu, and Z. Wang, “Detection <strong>of</strong> LSB steganography<br />
via sample pair analysis”, IEEE Trans. Signal Process. 51, 1995–2007<br />
(2003).<br />
6 I. Avcibas, N. Memon, and B. Sankur, “Steganalysis using image quality<br />
metrics”, IEEE Trans. Image Process. 12, 221–229 (2003).<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 385
<strong>Journal</strong> <strong>of</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology® 51(4): 386–390, 2007.<br />
© <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology 2007<br />
In Situ X-ray Investigation <strong>of</strong> the Formation <strong>of</strong> Metallic<br />
Silver Phases During the Thermal Decomposition <strong>of</strong> Silver<br />
Behenate and Thermal Development <strong>of</strong><br />
Photothermographic Films<br />
B. B. Bokhonov, M. R. Sharafutdinov and B. P. Tolochko<br />
Institute <strong>of</strong> Solid State Chemistry and Mechanochemistry, Russian Academy <strong>of</strong> <strong>Science</strong>, Kutateladze 18,<br />
Novosibirsk 630128, Russia<br />
E-mail: bokhonov@solid.nsk.su<br />
L. P. Burleva and D. R. Whitcomb <br />
Health Group, Eastman Kodak Company, Oakdale, Minnesota 55128<br />
Abstract. Metallic silver <strong>for</strong>mation, resulting from the thermal decomposition<br />
<strong>of</strong> silver behenate, AgBe, and from the thermally induced<br />
reduction <strong>of</strong> AgBe incorporated into a photothermographic<br />
imaging construction, has been compared by in situ x-ray investigation.<br />
In the case <strong>of</strong> the thermal decomposition <strong>of</strong> individual AgBe<br />
crystals, the main factor that determines the growth <strong>of</strong> the silver<br />
particles is the change in the AgBe crystal structure, leading to the<br />
<strong>for</strong>mation <strong>of</strong> intermediate mesomorphic phases that still retain characteristic<br />
layer structure. By contrast, development <strong>of</strong> AgBecontaining<br />
photothermographic films generates silver particles by<br />
the reduction <strong>of</strong> intermediate silver complexes, which are in a liquid<br />
state during the development process. The silver nanoparticles resulting<br />
from these processes exhibit different sizes and morphologies<br />
that are important <strong>for</strong> optimizing the optical properties <strong>of</strong><br />
photothermographic films. © 2007 <strong>Society</strong> <strong>for</strong> <strong>Imaging</strong> <strong>Science</strong> and<br />
Technology.<br />
DOI: 10.2352/J.<strong>Imaging</strong>Sci.Technol.200751:4386<br />
INTRODUCTION<br />
Silver behenate, AgO 2 C 22 H 43 2 , is one <strong>of</strong> the fundamental<br />
components <strong>of</strong> photothermographic materials because it<br />
provides the silver ions <strong>for</strong> reduction in the thermal development<br />
process that leads to the <strong>for</strong>mation <strong>of</strong> a visible<br />
image. 1–4 In the literature, there are a large number <strong>of</strong> reports<br />
devoted to the investigation <strong>of</strong> the phase changes <strong>of</strong><br />
long, saturated-chain silver carboxylates, including silver<br />
behenate, in thermal systems as well as the effect <strong>of</strong> individual<br />
components added to “dry silver” photothermographic<br />
<strong>for</strong>mulations. 5–11 The x-ray investigation <strong>of</strong><br />
silver carboxylates with carbon atoms from 2 to 22 12 showed<br />
that all <strong>of</strong> these crystal structures fall into the triclinic class<br />
and contain two molecules in the unit cell. Among the<br />
dominant characteristics <strong>of</strong> the silver carboxylate crystal<br />
structures, which are defined by their significant anisotropic<br />
physical and chemical properties, 1,13 is the presence <strong>of</strong> a lay-<br />
<br />
IS&T Member<br />
Received Jan. 17, 2007; accepted <strong>for</strong> publication Mar. 22, 2007.<br />
1062-3701/2007/514/386/5/$20.00.<br />
ered structure in which a double layer <strong>of</strong> silver ions separates<br />
a double layer <strong>of</strong> long methylene chains. For example, the<br />
solid-state crystal structure <strong>of</strong> silver stearate (AgSt,<br />
AgO 2 C 22 H 43 2 ) shows that the molecules are actually<br />
dimers connected together <strong>for</strong>ming a polymer. 3<br />
Thermally induced phase changes in the silver carboxylate<br />
crystals have been investigated by various analytical<br />
methods, such as NMR, IR, conductivity, DSC, and XRD.<br />
The temperatures <strong>of</strong> the multiple-phase transitions <strong>for</strong> silver<br />
carboxylates having various chain lengths have been<br />
characterized. 5,14–17 Upon transition from the crystalline<br />
state to the isotropic liquid, the silver carboxylates undergo<br />
up to six to seven phase changes <strong>of</strong> the following sequence:<br />
crystal state→curd→super curd SUC→sub-waxy SW<br />
→waxy W→super waxy SUW→sub-neat SN→neat<br />
N→isotropic liquid. 5,18 It may be relevant that the phase<br />
changes in the silver carboxylate from the crystalline state<br />
into the super curd (SUC) or sub-waxy (SW) phase occur in<br />
the 120–125°C range, the temperature at which the thermal<br />
development in photothermography is normally carried out.<br />
X-ray diffraction, calorimetric, and IR methods were<br />
used in the investigation <strong>of</strong> the structural changes in the<br />
silver stearate crystal lattice. 5 It was shown that, upon heating,<br />
the silver stearate structure proceeded through a series<br />
<strong>of</strong> mesomorphic states. That is, the first phase transition<br />
occurred at 122°C, which was associated with a packing<br />
disorder <strong>of</strong> the aliphatic chains, manifested by a significant<br />
decrease in the separation between silver ion layers. It was<br />
proposed that increasing the temperature above 130°C leads<br />
to further disorder and the breakup <strong>of</strong> the silver ion layers,<br />
and it is responsible <strong>for</strong> the onset <strong>of</strong> the thermal decomposition<br />
reaction <strong>of</strong> the silver stearate, resulting in the <strong>for</strong>mation<br />
<strong>of</strong> metallic silver and paraffin byproducts.<br />
The structure trans<strong>for</strong>mation <strong>of</strong> polycrystalline silver<br />
behenate was also studied by x-ray diffraction during in situ<br />
heating. 15 In contrast to the results reported in Ref. 5, these<br />
authors 15 observed an increase in the interlayer spacing dur-<br />
386
Bokhonov et al.: In situ x-ray investigation <strong>of</strong> the <strong>for</strong>mation <strong>of</strong> metallic silver phases during the thermal decomposition <strong>of</strong> silver behenate...<br />
ing the heating <strong>of</strong> silver behenate crystals. The authors also<br />
indicated that heating silver behenate over 120°C irreversibly<br />
trans<strong>for</strong>ms it from a crystalline to an amorphous state.<br />
Further, at 138–142°C, the first phase changes are observed,<br />
established by the appearance <strong>of</strong> diffraction peaks at the<br />
smaller 2 Bragg angles, which correspond to an increase in<br />
interlayer distance in the silver behenate structure. In agreement<br />
with these results, upon heating above 145°C, the silver<br />
behenate crystals trans<strong>for</strong>m into a liquid-crystalline state<br />
and generate metallic silver phases at 180°C. The authors <strong>of</strong><br />
this report consider that the initial stage <strong>of</strong> heating is the<br />
disordering <strong>of</strong> the silver behenate aliphatic chains. However,<br />
despite the agreement in the explanation <strong>of</strong> the structure<br />
trans<strong>for</strong>mations occurring in the silver behenate 15 and silver<br />
stearate, 5 there is a significant difference in the explanation<br />
<strong>of</strong> the subsequent structure changes in the phase trans<strong>for</strong>mations<br />
<strong>of</strong> these silver carboxylates. While heating silver<br />
stearate decreases its interlayer spacing, 5 heating silver<br />
behenate crystals initially proceeds through an amorphous<br />
phase followed by an increased distance between the layers. 15<br />
Such a contradictory heating behavior between the silver<br />
stearate and behenate seems to be quite surprising because<br />
<strong>of</strong> the close similarities between the silver stearate and<br />
behenate structures (C 18 and C 22 chain length, respectively)<br />
and their phase-trans<strong>for</strong>mation temperatures. In addition,<br />
we have recently reported the thermal decomposition <strong>of</strong> silver<br />
myristate, AgMy, AgO 2 C 14 H 27 2 under conditions<br />
similar to the AgSt and noted similar behavior. 5 Considering<br />
the importance <strong>of</strong> AgBe as a material <strong>for</strong> photothermographic<br />
imaging products and the contradiction between<br />
the trend observed <strong>for</strong> thermal decomposition <strong>of</strong><br />
AgMy and AgSt (in solids and in photothermographic films)<br />
relative to the interlayer spacing differences reported <strong>for</strong> the<br />
AgBe, we have continued the systematic investigation <strong>of</strong> the<br />
effect <strong>of</strong> increasing the chain length on the thermal properties<br />
<strong>of</strong> the AgBe component in this series. Once the reasons<br />
<strong>for</strong> the <strong>for</strong>mation <strong>of</strong> the solid products from these chemical<br />
reactions are better understood, novel routes to achieve control<br />
<strong>of</strong> these processes should be possible, and<br />
photothermographic properties can be further improved. In<br />
this work, we show the results <strong>of</strong> our in situ x-ray diffraction<br />
investigation related to the <strong>for</strong>mation <strong>of</strong> metallic silver from<br />
the thermal decomposition <strong>of</strong> pure silver behenate, as well as<br />
from the thermal development <strong>of</strong> photothermographic materials<br />
based on silver behenate.<br />
EXPERIMENTAL<br />
The synthesis <strong>of</strong> silver behenate was carried out by the exchange<br />
reaction between sodium behenate and silver nitrate,<br />
as typically practiced. 5 Photothermographic films were prepared<br />
from pure AgBe (not a mixture <strong>of</strong> chain lengths as is<br />
common in photothermography) and pre<strong>for</strong>med AgBr,<br />
along with the normal additional imaging components, as<br />
described elsewhere. 19<br />
X-ray experiments were carried out at the time-resolved<br />
diffractometry station—channel 5b <strong>of</strong> VEPP-3, BINP<br />
=1.506 Å. Transmission mode was used <strong>for</strong> small-angle<br />
Figure 1. Phthalazine and 4-methylphthalic acid.<br />
scattering (SAXS). X-ray patterns were obtained on the onecoordinate<br />
detector OD-3 with a 0.01° angular resolution<br />
and a 30 s recording time per frame. Samples were heated at<br />
1°C/min in a special tube furnace, and sample temperatures<br />
were controlled by a thermocouple.<br />
RESULTS AND DISCUSSION<br />
Despite the fact that the thermal decomposition <strong>of</strong> silver<br />
carboxylates and the development <strong>of</strong> photothermographic<br />
films both produce solid products <strong>of</strong> metallic silver, the<br />
chemical trans<strong>for</strong>mations occurring within these processes<br />
are completely different. If the thermal decomposition <strong>of</strong><br />
silver carboxylate proceeds according to the following<br />
scheme: 7<br />
AgO 2 C n H 2n−1 2 → 2Ag + 2CO 2 +C 2n−1 H 22n−1<br />
with the <strong>for</strong>mation <strong>of</strong> metallic silver and paraffin, then the<br />
development stages <strong>of</strong> photothermographic films are more<br />
complicated. Thermally induced reduction <strong>of</strong> the silver ions<br />
during film development can be summarized in a more simplified<br />
<strong>for</strong>m as follows:<br />
AgO 2 C n H 2n−1 2 → silver ion intermediates<br />
→ 2Ag + 2HO 2 C n H 2n−1 .<br />
This reduction reaction is the result <strong>of</strong> preliminary exposure<br />
<strong>of</strong> the silver halide in the film, which <strong>for</strong>ms active latent<br />
image centers that catalyze the thermal development step at<br />
110–130°C. The silver ion intermediates are reduced at the<br />
latent image centers, resulting in crystalline silver particles,<br />
which comprise the visible image. It is generally accepted<br />
that the silver ion source <strong>for</strong> silver particle <strong>for</strong>mation at the<br />
latent image center is the silver carboxylate, 1–4 which is not<br />
light sensitive in the visible region <strong>of</strong> the spectrum, and that<br />
the transport <strong>of</strong> silver ions from the silver carboxylate to the<br />
latent image center is from the thermally initiated <strong>for</strong>mation<br />
<strong>of</strong> the various silver complexes obtained with the components<br />
added into the <strong>for</strong>mulation (developers, toners, and<br />
antifoggants). 1–6<br />
A recent investigation into the composition <strong>of</strong> phase<br />
changes occurring during the development process raised<br />
doubts about the source <strong>of</strong> the silver ions <strong>for</strong>ming the visible<br />
image coming only from the silver carboxylate phase. 20 Contrary<br />
to all other literature, 1–4 the reduction <strong>of</strong> a model system<br />
composition (AgBe/AgBr with phthalazine (PHZ,<br />
Figure 1) 4-methyl-phthalic acid (4MPA, Fig. 1), and developer)<br />
resulted in a significant decrease 45% in the x-ray<br />
peak intensity <strong>for</strong> the AgBr phase. This change was proposed<br />
1<br />
2<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 387
Bokhonov et al.: In situ x-ray investigation <strong>of</strong> the <strong>for</strong>mation <strong>of</strong> metallic silver phases during the thermal decomposition <strong>of</strong> silver behenate...<br />
to be related to the contribution <strong>of</strong> silver ions from the AgBr<br />
in the <strong>for</strong>mation <strong>of</strong> the metallic silver particles <strong>of</strong> the image.<br />
As discussed below, in the full photothermographic imaging<br />
<strong>for</strong>mulation, we see no change in the AgBr signal.<br />
The in situ investigation <strong>of</strong> the structural and phase<br />
changes in the thermal decomposition <strong>of</strong> silver behenate and<br />
the development <strong>of</strong> the photothermographic films prepared<br />
from it showed that the processes accompanying the thermal<br />
<strong>for</strong>mation <strong>of</strong> the silver particles are significantly<br />
different.<br />
In Situ Investigation <strong>of</strong> the Phase Formed in the Process<br />
<strong>of</strong> Thermal Decomposition <strong>of</strong> AgBe<br />
The change in the diffraction characteristics <strong>of</strong> AgBe in the<br />
small angle region 2=0.4–10° during in situ heating<br />
from 20–220°C is shown in Figure 2(a).<br />
Increasing the temperature through this range is accompanied<br />
by a change in the AgBe x-ray diffraction pattern that<br />
is due to phase trans<strong>for</strong>mations occurring within the heated<br />
powder. As the temperature is increased, the reflections <strong>of</strong><br />
the high temperature phases shift to the high diffraction<br />
angle regions, which confirm the decrease in the interlayer<br />
spacing in the silver carboxylate structure. Analysis <strong>of</strong> the<br />
diffraction data <strong>of</strong> these phases in this temperature interval<br />
allows us to separate out at least six phases that have different<br />
structural characteristics, Fig. 2(b).<br />
It should be noted that the similar structural changes in<br />
AgBe and AgMy, with the <strong>for</strong>mation <strong>of</strong> intermediate phases,<br />
were established previously. 5,6,21 According to the authors, 5<br />
the first three phases correspond to the transitions within<br />
the AgBe crystalline state, but then at 155°C, silver behenate<br />
trans<strong>for</strong>ms into a liquid-crystalline material.<br />
The x-ray diffraction patterns <strong>of</strong> the intermediate<br />
phases <strong>for</strong>med during heating, Fig. 2(b), include at least two<br />
series <strong>of</strong> layer reflections, which are evidence <strong>for</strong> the <strong>for</strong>mation<br />
<strong>of</strong> a two-dimensional structure. Increasing the temperature<br />
above 230°C leads to the disappearance <strong>of</strong> the diffraction<br />
image <strong>of</strong> a layered structure. At the same time in the<br />
small 2 angle region, an increase in SAXS intensity is observed,<br />
the maximum <strong>of</strong> which is at 1.17°, Fig. 2(c).<br />
Subsequent heating to 250°C corresponds to a change<br />
in the shape <strong>of</strong> the small angle scattering peak, which appears<br />
as a decrease in the intensity <strong>of</strong> the SAXS maximum.<br />
Simultaneously, peaks appear in the small angle scattering<br />
angles at 20.8°.<br />
The in situ x-ray diffraction <strong>of</strong> the thermal decomposition<br />
<strong>of</strong> AgBe in the wide-angle region is (WAXS,<br />
2=25–55°). Figure 3 showed that heating the powder to<br />
140°C does not appear to significantly change the diffraction<br />
pattern. Upon heating to higher than 145°C, the<br />
crystal-phase reflections <strong>of</strong> AgBe disappear, and beginning at<br />
230°C broad reflections are observed because <strong>of</strong> the (111)<br />
and (200) planes <strong>of</strong> metallic silver, Fig. 3, the intensity <strong>of</strong><br />
which increases as the temperature is increased. Given that<br />
the first phase transition in silver behenate occurs at 128°C,<br />
the presence in the x-ray diffraction <strong>of</strong> the crystalline phase<br />
Figure 2. a Change in the x-ray diffraction pattern <strong>of</strong> silver behenate<br />
during in situ heating. b X-ray diffraction pattern <strong>for</strong> the initial 20°C<br />
and intermediate phases <strong>for</strong>med in the heating <strong>of</strong> AgBe. c SAXS <strong>of</strong><br />
AgBe.<br />
<strong>of</strong> AgBe is evidence that the first phase transition occurs<br />
from one crystalline state to another.<br />
388 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
Bokhonov et al.: In situ x-ray investigation <strong>of</strong> the <strong>for</strong>mation <strong>of</strong> metallic silver phases during the thermal decomposition <strong>of</strong> silver behenate...<br />
Figure 3. WAXS <strong>of</strong> AgBe during in situ thermal decomposition.<br />
Figure 5. Change in the x-ray diffraction pattern <strong>of</strong> photothermographic<br />
films during thermal development.<br />
Figure 4. Change in the x-ray diffraction pattern <strong>of</strong> AgBe during the<br />
development <strong>of</strong> photothermographic films: Initial decrease in the AgBe<br />
layer peak intensities with a simultaneous increase in the signal intensity <strong>of</strong><br />
the small-angle scattering peaks.<br />
In situ X-ray Diffraction Investigation <strong>of</strong> Phase<br />
Formation During Development <strong>of</strong> Photothermographic<br />
Films Prepared with AgBe<br />
The in situ x-ray investigation <strong>of</strong> the <strong>for</strong>mation <strong>of</strong> phases<br />
during the development <strong>of</strong> photothermographic films prepared<br />
with AgBe showed that the <strong>for</strong>mation <strong>of</strong> the silver<br />
phases occurs at temperatures significantly lower than the<br />
temperature <strong>of</strong> the first phase trans<strong>for</strong>mation 126°C.<br />
There are no shifts in the AgBe reflections observed during<br />
heating the photothermographic films from 20–80°C. After<br />
80°C, however, the intensity <strong>of</strong> the x-ray diffraction peaks<br />
related to the layered structure <strong>of</strong> AgBe (001) decrease,<br />
which corresponds to the simultaneous increase in the intensity<br />
in the small-angle scattering region <strong>of</strong> SAXS<br />
2=0.4–1.2°, Figure 4.<br />
The in situ x-ray diffraction pattern over 24–54° 2<br />
showed that the reflections that were due to the metallic<br />
silver appear as low as 80°C, Figure 5. Upon increasing the<br />
temperature (or the development time), the peak intensity <strong>of</strong><br />
the silver reflections increases.<br />
It is important to note that the in situ investigation <strong>of</strong><br />
the thermal development <strong>of</strong> films did not reveal any kind <strong>of</strong><br />
additional reflections from intermediate solid phases. More<br />
significantly, the reflection intensity <strong>for</strong> silver bromide (200)<br />
up to and after processing remained completely unchanged,<br />
Fig. 5.<br />
Comparing the half-widths <strong>of</strong> the silver reflections<br />
(111) and (200) recorded during the decomposition <strong>of</strong><br />
AgBe, Fig. 3, and the developed photothermographic films,<br />
Fig. 5, provide clear evidence that the size <strong>of</strong> the silver crystals<br />
in the developed films are significantly larger than that<br />
<strong>for</strong>med in the process <strong>of</strong> the thermal decomposition <strong>of</strong> pure<br />
AgBe, similar to that observed in AgMy. 6<br />
All <strong>of</strong> these results on the thermal decomposition <strong>of</strong><br />
AgBe and thermal development <strong>of</strong> photothermographic<br />
films can be summarized as follows.<br />
The <strong>for</strong>mation <strong>of</strong> metallic silver phases from the thermal<br />
decomposition <strong>of</strong> pure AgBe occurs through the <strong>for</strong>mation<br />
<strong>of</strong> a series <strong>of</strong> intermediate mesomorphic phases. The<br />
<strong>for</strong>mation <strong>of</strong> silver particles, established by the appearance in<br />
the x-ray diffraction pattern <strong>of</strong> in situ heated signal in the<br />
small-angle scattering, proceeds after the destruction <strong>of</strong> the<br />
AgBe layer structure.<br />
The development <strong>of</strong> the photothermographic films<br />
shows the <strong>for</strong>mation <strong>of</strong> silver phases at 80°C that correspond<br />
to the decreasing intensity <strong>of</strong> the silver behenate layer<br />
structure reflections with a simultaneous increase in the intensity<br />
<strong>of</strong> the SAXS. In particular, it must be emphasized<br />
that the <strong>for</strong>mation <strong>of</strong> the silver phase does not proceed<br />
through any change in the intensity <strong>of</strong> the silver halide<br />
peaks, and it is clear that the silver particles <strong>for</strong>m from the<br />
reduction <strong>of</strong> silver ions originating only from the AgBe crystals.<br />
J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007 389
Bokhonov et al.: In situ x-ray investigation <strong>of</strong> the <strong>for</strong>mation <strong>of</strong> metallic silver phases during the thermal decomposition <strong>of</strong> silver behenate...<br />
Overall, all <strong>of</strong> these results are in good agreement with<br />
previous reports. 22,23 That is, the initial stages <strong>of</strong> thermal<br />
decomposition <strong>of</strong> individual silver carboxylate <strong>for</strong>m nanosized<br />
2–5 nm particles <strong>of</strong> silver, which subsequently agglomerate<br />
up to 10–15 nm, crystallizing on the lateral<br />
planes <strong>of</strong> the silver carboxylate crystals. In our opinion, this<br />
stage <strong>of</strong> silver particle growth is the cause <strong>of</strong> the curve shape<br />
changes in the small-angle scattering in which a decrease in<br />
intensity and a shift to the small-angle region <strong>of</strong> the SAXS<br />
maxima was observed. Finally, it should be noted that the<br />
difference between the thermal behavior <strong>of</strong> pure AgBe described<br />
in here and Ref. 16 could be attributed to the differences<br />
in the preparation procedures. The same effect may<br />
influence the x-ray results during the study <strong>of</strong> the role <strong>of</strong><br />
AgBr in the photothermographic process. 15<br />
CONCLUSIONS<br />
The differences in the diffraction data during the development<br />
<strong>of</strong> photographic films and the thermal decomposition<br />
<strong>of</strong> pure AgBe are related to the differences in the chemical<br />
trans<strong>for</strong>mations in these processes: in contrast to the thermal<br />
decomposition <strong>of</strong> pure AgBe, development <strong>of</strong> the<br />
photothermographic films generates silver particles by the<br />
reduction <strong>of</strong> intermediate silver complexes, which are in the<br />
liquid state (not observable by x-ray diffraction). In the case<br />
<strong>of</strong> the thermal decomposition <strong>of</strong> individual AgBe crystals,<br />
the main factor that determines the growth <strong>of</strong> the silver<br />
particles is the change in the structure, leading to the <strong>for</strong>mation<br />
<strong>of</strong> intermediate mesomorphic phases, which still retain<br />
the characteristic layer structure.<br />
ACKNOWLEDGMENT<br />
The authors gratefully thank T. Blanton (Eastman Kodak<br />
Company) <strong>for</strong> helpful discussions.<br />
REFERENCES<br />
1 P. J. Cowdery-Corvan and D. R. Whitcomb, in Handbook <strong>of</strong> <strong>Imaging</strong><br />
<strong>Material</strong>s, edited by A. S. Diamond and D. S. Weiss (Marcel Dekker,<br />
New York, 2002), p. 473.<br />
2 D. H. Klosterboer, in Neblette’s Eighth Edition: <strong>Imaging</strong> Processes and<br />
<strong>Material</strong>s, edited by J. M. Sturge, V. Walworth, and A. Shepp (Van<br />
Nostrand-Reinhold, New York, 1989), Chap. 9, p. 279.<br />
3 V. M. Andreev, E. P. Fokin, Yu. I. Mikhailov, and V. V. Boldyrev, Zhur.<br />
Nauch. Priklad. Fotogr. Kinemat. 24, 311 (1979).<br />
4 T. Maekawa, M. Yoshikane, H. Fujimura, and I. Toya, J. <strong>Imaging</strong> Sci.<br />
Technol. 45, 365 (2001).<br />
5 B. B. Bokhonov, M. R. Sharafutdinov, B. P. Tolochko, L. P. Burleva, and<br />
D. R. Whitcomb, J. <strong>Imaging</strong> Sci. Technol. 49, 389 (2005).<br />
6 B. B. Bokhonov, L. P. Burleva, A. A. Sidelnikov, M. R. Sharafutdinov, B.<br />
P. Tolochko, and D. R. Whitcomb, J. <strong>Imaging</strong> Sci. Technol. 47, 89 (2003).<br />
7 V. M. Andreev, L. P. Burleva, and V. V. Boldyrev, J. Sib. Branch Acad. Sci.<br />
USSR 5(5), 3 (1984).<br />
8 D. R. Whitcomb and R. D. Rogers, J. <strong>Imaging</strong> Sci. Technol. 43, 517–520<br />
(1999).<br />
9 D. R. Whitcomb and M. Rajeswaran, J. <strong>Imaging</strong> Sci. Technol. 47, 107<br />
(2003).<br />
10 D. R. Whitcomb and R. D. Rogers, Inorg. Chim. Acta 256, 263 (1997).<br />
11 P. Z. Velinzon, S. I. Gaft, O. A. Karekina, N. K. Ryasinskaya, and I. G.<br />
Chezlov, Zhur. Nauch. i Priklad. Fotogr. 48(3), 35–45 (2003).<br />
12 V. Vand, A. Aitken, and R. K. Campbell, Acta Crystallogr. 2, 398–403<br />
(1949).<br />
13 A. E. Gvozdev, Ukr. Fiz. Zh. (Russ. Ed.) 24, 1856 (1979).<br />
14 M. Ikeda, Photograph. Sci. Eng. 24(6), 277 (1980).<br />
15 I. Geuens, I. Vanwelkenhuysen, and R. Gijbels, Proc. 2000 International<br />
Symposium on Silver Halide <strong>Imaging</strong> (IS&T, Springfield, VA, 2000) pp.<br />
203–233.<br />
16 K. Binnemans, R. V. Deun, B. Thijs, I. Vanwelkenhuysen, and I. Geuens,<br />
Chem. Mater. 16, 2021 (2004).<br />
17 X. Liu, S. Liu, J. Zhang, and W. Cao, Thermochim. Acta 440, 1 (2006).<br />
18 V. M. Andreev, L. P. Burleva, B. B. Bokhonov, and Y. I. Mikhailov, Izv.<br />
Sib. Otd. AN SSSR, Ser. Khim. Nauk. 2(4), 58 (1983).<br />
19 C. Zou, J. B. Philip, S. M. Shor, M. C. Skinner, and C. P. Zhou, US<br />
Patent No. 5,434,043 (1995).<br />
20 H. Strijckers, J. <strong>Imaging</strong> Sci. Technol. 47, 100 (2003).<br />
21 T. N. Blanton, S. Zdzieszynski, M. Nikolas, and S. Misture, Powder Diffr.<br />
48, 27 (2005).<br />
22 B. B. Bokhonov, L. P. Burleva, and D. R. Whitcomb, J. <strong>Imaging</strong> Sci.<br />
Technol. 43, 505 (1999).<br />
23 B. B. Bokhonov, L. P. Burleva, D. R. Whitcomb, and M. R. V. Sahyun,<br />
Microsc. Res. Tech. 42, 152 (1998).<br />
390 J. <strong>Imaging</strong> Sci. Technol. 514/Jul.-Aug. 2007
IS&T Corporate Members<br />
IS&T Corporate Members provide significant financial support, thereby assisting the <strong>Society</strong> in achieving its goals <strong>of</strong><br />
disseminating in<strong>for</strong>mation and providing pr<strong>of</strong>essional services to imaging scientists and engineers. In turn, the <strong>Society</strong><br />
provides a number <strong>of</strong> material benefits to its Corporate Members. For complete in<strong>for</strong>mation on the Corporate<br />
Membership program, contact IS&T at info@imaging.org.<br />
Sustaining Corporate Members<br />
Adobe Systems Inc.<br />
345 Park Avenue<br />
San Jose, CA 95110-2704<br />
Canon USA Inc.<br />
One Canon Plaza, Lake Success<br />
New York, NY 11042-1198<br />
Eastman Kodak Company<br />
343 State Street<br />
Rochester, NY 14650<br />
Hewlett-Packard Company<br />
1501 Page Mill Road<br />
Palo Alto, CA 94304<br />
Lexmark International, Inc.<br />
740 New Circle Road NW<br />
Lexington, KY 40511<br />
Xerox Corporation<br />
Wilson Center <strong>for</strong> Research and<br />
Technology<br />
800 Phillips Road<br />
Webster, NY 14580<br />
Supporting Corporate Members<br />
Fuji Photo Film Company, Ltd.<br />
210 Nakanuma, Minami-ashigara<br />
Kanagawa 250-0193 Japan<br />
Konica Minolta Holdings Inc.<br />
No. 1 Sakura-machi<br />
Hino-shi, Tokyo 191-8511 Japan<br />
TREK, Inc./TREK Japan KK<br />
11601 Maple Ridge Road<br />
Medina, NY 14103-0728<br />
Pitney Bowes<br />
35 Waterview Drive<br />
Shelton, CT 06484<br />
Donor Corporate Members<br />
ABBY USA S<strong>of</strong>tware House, Inc.<br />
47221 Fremont Blvd.<br />
Fremont, CA 94538<br />
Axis Communications AB<br />
Embdalavägen 14<br />
SE-223 69 Lund, Sweden<br />
Ball Packaging Europe GmbH<br />
Technical Center Bonn<br />
Friedrich-Woehler-Strasse 51<br />
D-53117 Bonn, Germany<br />
Cheran Digital <strong>Imaging</strong> & Consulting, Inc.<br />
798 Burnt Gin Road<br />
Gaffney, SC 29340<br />
Clariant Produkte GmbH<br />
Division Pigments & Additives<br />
65926 Frankfurt am Main Germany<br />
Felix Schoeller Jr. GmbH & Co. KG<br />
Postfach 3667<br />
D-49026 Osnabruck, Germany<br />
Ferrania SpA<br />
Viale Martiri Della Liberta’ 57<br />
Ferrania (Savona) I-17014<br />
GretagMacbeth<br />
Logo GmbH & Co. KG<br />
Westfälischer H<strong>of</strong> Garbrock 4<br />
48565 Steinfurt, Germany<br />
GR8moments Limited<br />
Units 15-16 Town Yard Industrial Estate<br />
Station Street<br />
Leek, Staf<strong>for</strong>dshire<br />
England, ST13 8BF<br />
Hallmark Cards, Inc.<br />
Chemistry R & D<br />
2501 McGee, #359<br />
Kansas City, MO 64141-6580<br />
ILFORD <strong>Imaging</strong> Switzerland GmbH<br />
Route de l’Ancienne Papeterie 1<br />
CH-1723 Marly, Switzerland<br />
MediaTek Inc.<br />
No. 1 Dusing Rd., 1<br />
Hsinchu 300 R.O.C, Taiwan<br />
Pantone, Inc.<br />
590 Commerce Blvd.<br />
Carlstadt, NJ 07072-3098<br />
Quality Engineering Associates (QEA), Inc.<br />
99 South Bed<strong>for</strong>d Street, #4<br />
Burlington, MA 01803<br />
The Ricoh Company, Ltd.<br />
16-1 Shinei-cho, Tsuzuki-ku<br />
Yokohama 224-0035 Japan<br />
Sharp Corporation<br />
492 Minosho-cho, Yamatokoriyama<br />
Nara 639-1186 Japan<br />
Sony Corporation/<br />
Sony Research Center<br />
6-7-35 Kita-shinagawa<br />
Shinagawa, Tokyo 141 Japan<br />
*as 7/1/07
<strong>Journal</strong> <strong>of</strong> the <strong>Imaging</strong> <strong>Society</strong> <strong>of</strong> Japan VOL.46 NO.3<br />
2007<br />
CONTENTS<br />
Original Papers<br />
Analysis <strong>of</strong> the Magnetic Force Acting on the Toner in the Black Image Area and White Image Area in the<br />
Magnetic Printer (2)N. KOKAJI ...1722<br />
A Model <strong>for</strong> Electrostatic Discharge in the Toner Layer during the Transfer Process<br />
M. MAEDA, K. NISHIWAKI, K. MAEKAWA and M. TAKEUCHI ...1788<br />
<strong>Imaging</strong> Today<br />
The Law <strong>of</strong> Environmental Standard, Safety Standard and Energy Saving<br />
Introduction H. YAMAZAKI, T. TAKEUCHI, K. NAGATO, K. MARUYAMA and K. SUZUKI ...18414<br />
The Notification Systems <strong>of</strong> New Chemical Substances in the World T. YAMAMOTO ...18515<br />
The Environmental Regulations in Japan H. SATO ...19222<br />
The Trend <strong>of</strong> European Product Environmental Legislation and Eco-labels<br />
R. IWANAGA, K. FUJISAWA and A. MATSUMOTO ...19929<br />
Practical Side <strong>of</strong> the Environmentally Conscious Technology <strong>for</strong> the Product<br />
T. BISAIJI, K. YASUDA, T. ARAI, K. SUZUKI, K. AKATANI and M. HASEGAWA ...20737<br />
Lectures in <strong>Science</strong><br />
Introduction <strong>of</strong> Optics (I)<br />
The Behavior <strong>of</strong> a Beam <strong>of</strong> Light upon Reflection or Refraction at a Plane Surface <br />
H. MUROTANI ....21646<br />
Meeting Reports 22353<br />
Announcements 22454<br />
Guide <strong>for</strong> Authors 22959<br />
Contents <strong>of</strong> J. Photographic <strong>Society</strong> <strong>of</strong> Japan23060<br />
Contents <strong>of</strong> J. Printing <strong>Science</strong> and Technology <strong>of</strong> Japan23161<br />
Contents <strong>of</strong> J. Inst. Image Electronics Engineers <strong>of</strong> Japan 23262<br />
Contents <strong>of</strong> <strong>Journal</strong> <strong>of</strong> <strong>Imaging</strong> <strong>Science</strong> and Technology 23363<br />
Essays on <strong>Imaging</strong><br />
The <strong>Imaging</strong> <strong>Society</strong> <strong>of</strong> Japan<br />
c/o Tokyo Polytechnic University, 2-9-5, Honcho, Nakano-ku, Tokyo, 1648768 Japan<br />
Phone :033373-9576 Fax :033372-4414 E-mail :
Advanced Measurement Systems <strong>for</strong> All R&D and<br />
Quality Control Needs in Electrophotography,<br />
Inkjet and Other Printing Technologies<br />
PDT ® -2000 series<br />
Electrophotographic characterization,<br />
uni<strong>for</strong>mity mapping, and defect<br />
detection <strong>for</strong> large and small <strong>for</strong>mat<br />
OPC drums<br />
PDT ® -1000L<br />
PDT ® -1000<br />
ECT-100 TM<br />
OPC drum coating thickness<br />
gauge<br />
Electrophotographic<br />
Component Testing<br />
MFA-2000 TM<br />
Magnetic field distribution<br />
analysis in mag roller magnets<br />
DRA-2000 TM<br />
Semi-insulating components testing<br />
including charge rollers, mag rollers,<br />
transfer rollers, transfer belts, and<br />
print media<br />
TFS-1000 TM<br />
Toner fusing latitude testing<br />
Objective Print Quality<br />
Analysis <strong>for</strong> All Digital<br />
Printing Technologies<br />
IAS ® -1000<br />
Fully-automated high volume print<br />
quality testing<br />
Scanner-based high speed print<br />
quality analysis<br />
Scanner IAS ®<br />
Personal IAS ®<br />
Handheld series <strong>for</strong> print quality,<br />
distinctness <strong>of</strong> image (DOI), and<br />
color measurements. Truly portable;<br />
no PC connection required<br />
PocketSpec TM<br />
DIAS TM<br />
Quality Engineering Associates, Inc.<br />
99 South Bed<strong>for</strong>d Street #4, Burlington, MA 01803 USA<br />
Tel: +1 (781) 221-0080 • Fax: +1 (781) 221-7107 • info@qea.com • www.qea.com
imaging.org<br />
your source <strong>for</strong> imaging technology conferences<br />
UPCOMING CONFERENCES<br />
Join us in Anchorage, Alaska!<br />
September 16-21, 2007<br />
NIP23 Sessions<br />
NIP23<br />
23rd International Conference on<br />
Digital Printing Technologies<br />
www.imaging.org/conferences/nip23<br />
• Advanced <strong>Material</strong>s and Nanoparticles in <strong>Imaging</strong><br />
• Color <strong>Science</strong> and Image Processing<br />
• Digital Art<br />
• Electronic Paper and Paper-like Displays<br />
• Environmental Issues<br />
• Fusing Curing and Drying<br />
• Image Permanence<br />
• Ink Jet Printing <strong>Material</strong>s<br />
• Ink Jet Printing Processes<br />
• Media <strong>for</strong> Digital Printing and Displays<br />
Come and learn about...<br />
• Photo-electronic <strong>Material</strong>s and Devices<br />
• Print and Image Quality<br />
• Printing Systems Engineering and Optimization<br />
• Production Digital Printing<br />
• Security and Forensic Printing<br />
• Textile & Industrial Printing<br />
• Thermal Printing<br />
• Toner Based Printing <strong>Material</strong>s<br />
• Toner Based Printing Processes<br />
DF 2007 Sessions<br />
• Industrial and Commercial Applications<br />
• <strong>Material</strong>s and Substrates<br />
• New and Novel Direct Write Methods<br />
• Printed Architectural Components<br />
• Printed Electronics and Devices<br />
• Printing <strong>of</strong> Biomaterials<br />
• Plus: Joint Session Intellectual Property Panel on “Future<br />
and Limitations <strong>of</strong> Ink Jet and Electrophotography”<br />
Digital<br />
Fabrication 2007<br />
www.imaging.org/conferences/df2007<br />
For more in<strong>for</strong>mation visit the conference website or visit: www.imaging.org/conferences; or contact us at info@imaging.org