Encyclopedia of Data Warehousing and Mining

Transcription

1 Encyclopedia of Data Warehousing and Mining Second Edition John Wang Montclair State University, USA Volume II Data Pro-I Information Science reference Hershey New York

2 Director of Editorial Content: Director of Production: Managing Editor: Assistant Managing Editor: Typesetter: Cover Design: Printed at: Kristin Klinger Jennifer Neidig Jamie Snavely Carole Coulson Amanda Appicello, Jeff Ash, Mike Brehem, Carole Coulson, Elizabeth Duke, Jen Henderson, Chris Hrobak, Jennifer Neidig, Jamie Snavely, Sean Woznicki Lisa Tosheff Yurchak Printing Inc. Published in the United States of America by Information Science Reference (an imprint of IGI Global) 701 E. Chocolate Avenue, Suite 200 Hershey PA Tel: Fax: Web site: and in the United Kingdom by Information Science Reference (an imprint of IGI Global) 3 Henrietta Street Covent Garden London WC2E 8LU Tel: Fax: Web site: Copyright 2009 by IGI Global. All rights reserved. No part of this publication may be reproduced, stored or distributed in any form or by any means, electronic or mechanical, including photocopying, without written permission from the publisher. Product or company names used in this set are for identification purposes only. Inclusion of the names of the products or companies does not indicate a claim of ownership by IGI Global of the trademark or registered trademark. Library of Congress Cataloging-in-Publication Data Encyclopedia of data warehousing and mining / John Wang, editor. -- 2nd ed. p. cm. Includes bibliographical references and index. Summary: "This set offers thorough examination of the issues of importance in the rapidly changing field of data warehousing and mining"--provided by publisher. ISBN (hardcover) -- ISBN (ebook) 1. Data mining. 2. Data warehousing. I. Wang, John, QA76.9.D37E dc British Cataloguing in Publication Data A Cataloguing in Publication record for this book is available from the British Library. All work contributed to this encyclopedia set is new, previously-unpublished material. The views expressed in this encyclopedia set are those of the authors, but not necessarily of the publisher. If a library purchased a print copy of this publication, please go to for information on activating the library's complimentary electronic access to this publication.

3 Section: Facial Recognition 857 Facial Recognition Rory A. Lewis UNC-Charlotte, USA F Zbigniew W. Ras University of North Carolina, Charlotte, USA INTRODUCTION Over the past decade Facial Recognition has become more cohesive and reliable than ever before. We begin with an analysis explaining why certain facial recognition methodologies examined under FERET, FRVT 2000, FRVT 2002, and FRVT 2006 have become stronger and why other approaches to facial recognition are losing traction. Second, we cluster the stronger approaches in terms of what approaches are mutually inclusive or exclusive to surrounding methodologies. Third, we discuss and compare emerging facial recognition technology in light of the aforementioned clusters. In conclusion, we suggest a road map that takes into consideration the final goals of each cluster, that given each clusters weakness, will make it easier to combine methodologies with surrounding clusters. BACKGROUND The National Institute of Standards and Technology (NIST) sponsored 2006 Face Recognition Vendor Test (FRVT) which is the most recent large scale independent synopsis of the state-of-the-art for face recognition systems. The previous tests in the series were the FERET, FRVT 2000, and FRVT The following organizations participated in the FRVT 2006 evaluation: Animetrics, Inc., Carnegie Mellon University, Cognitec Systems GmbH, Diamond Information Systems (DIS), Geometrix, Inc., Guardia, Identix, Inc., Neven Vision, New Jersey Institute of Technology (NJIT), Nivis, LLC, Old Dominion University, Panvista Limited, Peking University, Center for Information Science, PeopleSpot Inc., Rafael Armament Development Authority Ltd., SAGEM SA, Samsung Advanced Institute of Technology (SAIT), Tsinghua University, Tili Technology Limited, Toshiba Corporation, University of Houston, and Viisage. It should be noted that while the FRVT 2006 was conducted by the National Institute of Standards and Technology (NIST), it was jointly sponsored by five other U.S. Government agencies which share NIST s interest in measuring the improvements in face recognition technologies: Federal Bureau of Investigation, National Institute of Justice, National Institute of Standards and Technology, U.S. Department of Homeland Security, and the Transportation Security Administration. The FRVT 2006 measured the progress of facial recognition systems including commercial systems that used Windows or Linux based algorithms. The sequestered data comprised a large standard dataset of full frontal pictures provided to NIST by the U.S. State Department using non-conforming pixel resolutions and lighting angles of 36,000 pictures of persons applying for non-immigrant visas at U.S. consulates in Mexico. The tests evaluated 4 dimensions of facial recognition: high resolution still imagery, 3D facial scans, multi-sample still facial imagery, and pre-processing algorithms that compensate for pose and illumination. The results of the best of the 13 groups that entered have improved remarkably; the best algorithms in the FRVT 2002 computed 20% false rejections compared to only 1% false rejections in the FRVT 2006 tests. However, some of the groups that entered FRVT 2006 had results no better than that of In the tests, the rejection was less palatable: 12% for the best algorithms which still it is better than the 29% rejection rate of the 2002 tests. FRVT tests digress from the traditional facial recognition tests of the 1990 s in two ways: First, speed was not the issue in the tests, some of the algorithms took hundreds of hours to find matches in the database. The correct identification (precision) is the issue. Secondly, rather than the traditional ID searches of comparing a face in the camera with every face in the database for a match, the FRVT tests comprised security verification: is the face of the person standing in front of the Copyright 2009, IGI Global, distributing in print or electronic forms without written permission of IGI Global is prohibited.

4 Facial Recognition Figure 1. The reduction in error rates of facial recognition algorithms camera claiming to be Mr. Whomever indeed the Mr. Whomever whose picture is in the database? THE STATE-OF-THE-ART OF THE 13 GROUPS IN FRVT 2006 The three well known facial recognition corporations, Google owned Neven Vision, Viisage Technology (owned by L-1 Identity Solutions), and Cognitec Systems of Germany performed the best. The two universities that excelled were the University of Houston and Tsinghua University in China. The four methodology clusters are (i) Support Vector Machines, (ii) Manifold/Eigenface, (iii) Principal Component Analysis with Modified Sum Square Error (PCA/SSE), and Pure Eigenface technology: 1. Cognitec Systems - Cognitec Systems FaceVACSincorporates Support Vector Machines (SVM) to capture facial features. In the event of a positive match, the authorized person is granted access to a PC (Thalheim, 2002). 2. Neven Vision - In 2002, Neven s Eyematic team achieved high scores in the Face Recognition Vendor Test (Neven, 2004). Neven Vision incorporates Eigenface Technology. The Neven methodology trains an RBF network which constructs a full manifold representation in a universal Eigenspace from a single view of an arbitrary pose. 3. L-1 Identity Solutions Inc L-1 s performance ranked it near or at the top of every NIST test. This validated the functionality of the algorithms that drive L-1 s facial and iris biometric solutions. L-1 was formed in 2006 when Viisage Technology Inc. bought Indentix Inc. The NIST evaluations of facial and iris technology covered algorithms submitted by Viisage and Indentix, both of which utilize variations of Eigenvalues, one called the principal component analysis (PCA)-based face recognition method and the other, the modified sum square error (SSE)-based distance technique. 4. M A. Turk and A P. Pentland they wrote a paper that changed the facial recognition world - Face 858

5 Facial Recognition Recognition Using Eigenfaces (Turk & Pentland, 1991) which won the Most Influential Paper of the Decade award from the IAPR MVA. In short, the new technology reconstructed a face by superimposing a set of Eigenfaces. The similarity between the two facial images is determined based on the coefficient of the relevant Eigenfaces. The images consist of a set of linear eigenvectors wherein the operators are non-zero vectors which result in a scalar multiple of themselves (Pentland, 1993). It is the scalar multiple which is referred to as the Eigenvalue associated with a particular eigenvector (Pentland & Kirby, 1987). Turk assumes that any human face is merely a combination of parts of the Eigenfaces. One person s face may consist of 25% from face 1, 25% from face 2, 4% of face 3, n% of face X (Kirby, 1990). Figure 2. Original picture set The authors propose that even though the technology is improving, as exemplified by the FRVT 2006 test results, the common dominator for problematic recognition is based on faces with erroneous and varying light source conditions. Under this assumption, 3-D models were compared with the related 2-D images in the data set. All four of the aforementioned primary methodologies of facial recognition across the board, performed poorly in this context: for the controlled illumination experiments, performance of the very-high resolution dataset was better than the high-resolution dataset. Also, from comparing the photos of faces in the database to the angle of view of the person in 3d/real life, an issue FRVT did not even consider, leads one to believe that light source variation on 3-d models compromises the knowledge discovery abilities of all the facial recognition algorithms at FRVT Figure 3. RGB normalization Figure 4. Grayscale and posturization 859 F

6 Facial Recognition INNERTRON TECHNOLOGY Major components of interactive visual data analysis and their functions that make knowledge extraction more effective are the current research theme in this field. Consider the 8 images of the subject photographed in Figure 2. These photos were taken at four different times over the period of 18 months under 4 varying sets of light sources. To the human brain this is clearly the same person without a doubt. However, after normalizing the light in RGB in each of the four sets, we yield the results displayed in Figure 3. Note that this normalization process is normalizing each group of four until the desired, empirical normalization matrix in RGB is achieved as in the manner that a machine would render the normalization process. Note that in Figure 2 the bottom right hand figure is not darkened at all but after normalization in Figure 3 it appears darker. In Figure 4, a simple Grayscale and 6-level posturization is performed. In Figure 5, as in the FRVT 2006 tests, shadow is added to compensate for distance measures. It is evident that varying light conditions change the dynamics of a face in a far more complex manner than facial surgery could. This is evident when one looks at the 2-d images in terms of what is white and what is black, measuring and storing where they are black and white. Innertron facial recognition technology proposes a solution to the core, common denominator problem of the FRVT 2006 tests as illustrated in Figures 2 through 5, where classical lighting and shadow normalization can corrupt lighting on some good images in order to compensate for bad lighting at the level of the median lighting parameters. The Lambertian model states that the image of a 3D object is subject to three factors, namely the surface normal, albedo, and lighting. These can be factorized into an intrinsic part of the surface normal and albedo, and an extrinsic part of lighting (Wang 2004). Figure 5. Distance configuration Figure 6. Difference Eigen 860

7 Facial Recognition where p is the surface texture (albedo) of the face, n ( x, y ) T represents the 3-D rendering of the normal surface of the face, which will be the same for all objects of the class, and s is the position of the light source, which may vary arbitrarily. The quotient image Q y of face y against face a is defined by where u and v range over the image, I y is an image of object y with the illumination direction s y, and x j are combining coefficients estimated by Least Square based on training set (Shashua, 2001), (Raviv, 1999), and I 1, I 2, I 3 are three non-collinearly illuminated images. This is shown in Figure 6 where we superimpose, using negative alpha values, face y onto face a with the higher level image additionally set to a basic Eigenvector exposing the differences of itself a and the image below it y. This cancels out erroneous shadows, and only keeps constant data. The color of the images has changed from that of the original images in Figure 2. This is because of the effect the Eigenvector has on two samples from almost the same place and time. One can also see in Figure 7 that the large amount of shadows that were present in Figures 3 through 4 are almost gone. In Figure 7 and Figure 8 the innertron based strategy has normalized the face. See details in (Lewis, Ras 2005). In our approach, we use training database of human faces described by a set of innertron features to built classifiers for face recognition. FUTURE TRENDS F Figure 7. Calculate tilt Figure 8. Normalize - grid There are many projects focused on creation of intelligent methods for solving many difficult high-level image feature extraction and data analysis problems in which both local and global properties as well as spatial relations are taken into consideration. Some of these new techniques work in pixel and compressed domain (avoiding the inverse transform in decompression), thus speeding up the whole process (Delac & Grgic, 2007), (Deniz et al., 2001), (Du et al., 2005). The development of these new high-level image features is essential for the successful construction of new classifiers for the face recognition. Knowledge discovery techniques are especially promising in providing tools to build classifiers for recognizing emotions associated with human faces (happy, sad, surprised,..). But this area is still at the very beginning stage. CONCLUSION Overcoming the challenges of shadow elimination seems promising with the use of negative alpha values set to basic Eigenvectors particularly when used with the Lambertian methodology. Mining these images in a 2-D database is easier because the many levels of grey will be normalized to a specific set allowing a 861

8 Facial Recognition more precise vector value for distance functions. Pure Eigenvector technology, also fails when huge amounts of negatives are superimposed onto one another, once again pointing to normalization as the key. Data in the innertron database need to be consistent in tilt and size of the face. This means that before a face is submitted to the database, it also needs tilt and facial size normalization. REFERENCES Delac, K., Grgic, M. (2007). Face Recognition, I-Tech Education and Publishing, Vienna, 558 pages. Deniz, O., Castrillon, M., Hernández, M. (2001). Face recognition using independent component analysis and support vector machines, Proceedings of Audio- and Video-Based Biometric Person Authentication: Third International Conference, LNCS 2091, Springer. Du, W., Inoue, K., Urahama, K. (2005). Dimensionality reduction for semi-supervised face recognition, Fuzzy Systems and Knowledge Discovery (FSKD), LNAI , Springer, Kirby, M., Sirovich, L. (1990). Application of the Karhunen-Loeve procedure for the characterization of human faces, IEEE Transactions on Pattern Analysis and Machine Intelligence 12(1), Lewis, R.A., Ras, Z.W. (2005). New methodology of facial recognition, Intelligent Information Processing and Web Mining, Advances in Soft Computing, Proceedings of the IIS 2005 Symposium, Gdansk, Poland, Springer, Neven, H. (2004). Neven Vision Machine Technology, See: neven.html Pentland, L., Kirby, M. (1987). Low-dimensional procedure for the characterization of human faces, Journal of the Optical Society of America 4, Pentland, A., Moghaddam, B., Starner, T., Oliyide, O., Turk, M. (1993). View-Based and modular eigenspaces for face recognition, Technical Report 245, MIT Media Lab. Riklin-Raviv, T., Shashua, A. (1999). The Quotient image: class based recognition and synthesis under varying illumination, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Shashua, A., Riklin-Raviv, T. (2001). The quotient image: class-based re-rendering and recognition with varying illuminations, Transactions on Pattern Analysis and Machine Intelligence 23(2), Turk M., Pentland, A. (1991). Face recognition using eigenfaces, Proc. IEEE Conference on Computer Vision and Pattern Recognition, IEEE Computer Society Conference, Wang, H., Li, S., Wang, Y. (2004). Face recognition under varying lighting conditions using self quotient image, Proceedings of the IEEE International Conference on Automatic Face and Gesture Recognition, KEY TERMS Classifier: Compact description of a set of instances that have common behavioral and structural features (attributes and operations, respectively). Eigenfaces: Set of eigenvectors used in the computer vision problem of human face recognition. Eigenvalue: Scalar associated with a given linear transformation of a vector space and having the property that there is some nonzero vector which when multiplied by the scalar is equal to the vector obtained by letting the transformation operate on the vector. Facial Recognition: Process which automatically identifies a person from a digital image or a video frame from a video source. Knowledge Discovery: Process of automatically searching large volumes of data for patterns that can be considered as knowledge about the data. Support Vector Machines (SVM): Set of related supervised learning methods used for classification and regression. They belong to a family of generalized linear classifiers. A special property of SVMs is that they simultaneously minimize the empirical classification error and maximize the geometric margin; hence they are also known as maximum margin classifiers. 862