With the increasing concerns on security breaches and transaction fraud, highly reliable and convenient personal verification and identification technologies are more and more requisite in our social activities and national services. Biometrics, which use the distinctive physiological and behavioural characteristics to recognize the identity of an individual, are gaining ever-growing popularity in an extensive array of governmental, military, forensic, and commercial security applications.
The beginning of biometrics can be traced back to centuries ago, from when fingerprint has been used for forensics. Automated biometrics, however, has only 40 years' history. In the early 1960s, the FBI (Federal Bureau of Investigation) began to put more effort in developing automated fingerprint acquisition and identification systems. With the advances in hardware, sensor, pattern recognition, signal and image processing technologies, a number of biometric technologies, such as face, iris, retina, voice, signature, hand geometry, keystroke, ear, and palm print recognition, have been developed, and novel biometrics, such as dental, odor, and skin reflectance, have also been investigated to overcome some of the limitations of current biometric recognition technologies.
Historically, the development of biometric technologies is originated from different disciplines. For example, the beginning of fingerprint recognition research is an interaction of forensics and pattern recognition. Voice recognition technology, on the contrary, came from signal processing, and face recognition started from computer vision. This multi-discipline characteristic, however, makes it very challenging to establish an infrastructural theory framework for developing biometric recognition technologies.
Generally, a biometric system can be regarded as a pattern recognition system, where a feature set is first extracted from the acquired data, and then compared with the stored template set to make a decision on the identity of an individual. A biometric system can be applied to two fields, verification and identification. In verification mode, the decision is whether a person is “who he claims to be?” In identification mode, the decision is “whose biometric data is this?” A biometric system is thus formalized into a two-class or multi-class pattern recognition system.
A biometric system usually includes four major modules: data acquisition, feature extraction, matching, and system database, where feature extraction and matching are two of the most challenging problems in biometric recognition research, and have attracted researchers from different backgrounds: biometrics, computer vision, pattern recognition, signal processing, and neural networks. In this book, we focus on two kinds of biometric recognition technologies, biometric data discrimination and multi-biometrics. Biometric data discrimination technology, which extracts a set of discriminant features by using classical or improved discriminant analysis approaches, is of course one kind of biometric recognition technology. Multi-biometrics, which integrates information from multiple biometric traits to enhance the performance and reliability of the biometric system, is another kind of biometric recognition technology.
The book begins with the topic of biometric data discrimination technologies. Discriminant analysis, which aims at dimensionality reduction while retaining the statistical separation property between distinct classes, is a natural choice for biometric feature extraction. From the late 1980s, many classical discriminant analysis technologies are borrowed and applied to deal with biometric data or features. Among them, principal component analysis (PCA, or K-L transform) and Fisher linear discriminant analysis (LDA) turns out to be effective, in particular for face representation and recognition. Other linear approaches, such as independent component analysis (ICA), canonical correlation analysis (CCA), and partial least squares (PLS), have been investigated and applied to biometric recognition. Recently, non-linear projection analysis technology represented by kernel principal component analysis (KPCA), kernel Fisher discriminant (KFD), and manifold learning, also show great potential in dealing with biometric recognition problems.
Biometric data discriminant analysis is not a simple application of discriminant analysis to biometric data. Biometric data is usually high dimensional and their within-class variations should not be neglected, as the neglect will cause the serious performance degradation of classical discriminant analysis approaches. Various improvements to discriminant analysis techniques have been proposed to make it more suitable for biometric data. Now discriminant analysis has been widely applied to face, ear, fingerprint, gait recognition, and multi-biometrics. Further, the present demand for more reliable biometric recognition technologies is also contributing to the development and improvement on linear/nonlinear discriminant analysis technologies.
The form of biometric data and features is diverse. Biometric data mainly exists in the following three forms: 1D waveform (e.g. voice, signature data), 2D images (e.g. face images, fingerprints, palm prints) or image sequences (i.e., video), and 3D geometric data (such as 3-D facial or hand geometric shapes). Because of the diversity in biometric data and feature forms, it is difficult to develop a universal feature extraction technology which is capable to process all kinds of biometric data.
To deal with the diversity in biometric data forms, a family of tensor discriminant analysis technologies has been investigated. A tensor is a higher order generalization of a vector or a matrix. In fact, a vector is a first-order tensor and a matrix is a tensor of order two. Furthermore speaking, tensors are multilinear mapping over a set of vector spaces. If we have data in three or more dimensions, then we mean to treat them as higher-order tensors. In this way, tensor technologies present a generalized representation and analysis of biometric data discrimination technologies. Nowadays, tensor principal component analysis, tensor discriminant analysis, tensor independent component analysis, and other tensor analysis approaches have been successfully applied to face, palm print, and gait recognition.
In advance of the application of biometric data discrimination technologies, we should determine the appropriate representation of the biometric data. Generally, biometric data discrimination technologies can be performed in either data space or feature space. On the one hand, discrimination technologies can be used to derive the discriminative features directly from the original biometric data. On the other hand, a set of salient features are first extract from biometric data, and discrimination technologies are then performed in the feature space for the second feature extraction. It should be noted that the feature space may be implicit or infinite dimensional for kernel-based methods. In this way, biometric data discrimination technologies provide a general means to integrate different kinds of features for effective biometric feature extraction and matching.
In general, biometric data discrimination technologies share the following characteristics: 1) high dimensionality, which makes direct classification in original data space almost impossible; 2) difference in sample quality, where robust discrimination technologies should be developed to address biometric data with poor quality; 3) small sample size, where the data dimensionality is much higher than the size of the training set, resulting in the singularity and poor estimation of the scatter matrix. Recently, with the efforts of related researchers, progresses have been achieved in solving these problems.
From the lately 1990s, the authors of the book have been devoted to biometric data discrimination research, and developed a series of novel and effective discriminant criteria and discriminant algorithms, which cover both vector-based and tensor-based discrimination technologies. The class of new methods includes: 1) Extensions of Fisher’s discriminant criterion: we present three classification-oriented extensions of Fisher’s discriminant criterion: large margin linear projection (LMLP), minimum norm minimum squared-error, and maximum scatter difference. All these three criteria are classification-oriented and designed to deal with the small sample size problem; 2) Orthogonal discriminant analysis: we investigate two novel orthogonal discriminant analysis methods: Orthogonalized Fisher discriminant (OFD) and Fisher discriminant with Schur decomposition (FDS). Theoretical analysis and experimental results indicate that both OFD and FDS are optimal solutions to multiple Fisher discriminant criterion; 3) Parameterized discriminant analysis: we introduce three parameterized linear discriminant analysis methods. The first is parameterized direct linear discriminant analysis. The second is weighted nullspace linear discriminant analysis, and the third is weighted linear discriminant analysis in the range of within-class scatter matrix; 4) New facial feature extraction methods: we present two novel facial feature extraction methods, multiple maximum scatter difference (MMSD) and discriminant based on coefficients of variances (DCV). MMSD is an extension of binary linear discriminant analysism and DCV is a generalization of the null-space LDA; 5) Tensor PCA: we introduce the algorithms and discuss the properties of a group of tensor PCA methods; 6) Tensor LDA: we investigate the algorithms and discuss the properties of a group of tensor linear discriminant analysis and tensor locality preserving projection methods.
The second topic of the book is multi-biometrics, which combines information from multiple biometric traits to improve the accuracy and reliability of biometric systems. Individual biometric system, the biometric system using a single biometric characteristic, usually suffers from some inherent limitations and can not provide satisfactory recognition performance. For example, manual workers with damaged or dirty hands may not be able to provide high-quality fingerprint images, and thus failure to enrol would happen for single fingerprint recognition system. Multi-biometric systems, which integrate information from multiple biometric traits, provide some effective means to enhance the performance and reliability of the biometric system. In recent years, multi-biometric technologies have received considerable interests in biometric recognition research. Several true multi-modal databases have released for testing multi-biometric recognition algorithms. To facilitate multi-biometric research, the (USA) National Institute of Science and Technology (NIST) presents an open resource of Biometric Scores Set - Release 1 (BSSR1), which includes true multi-model matching scores generated by face and fingerprint recognition algorithms.
To combine information from individual biometric traits, there are three levels of fusion strategies in general, feature level fusion, matching score level fusion, and decision level fusion. In feature level fusion, the data obtained from each sensor is used to compute a feature vector. As the feature extracted from one biometric trait is independent of that extracted from the other, a new feature vector can be constructed using the concatenation rule, the parallelity rule, or the competitive rule for performing multi-biometric based personal authentication. It should be noted that the new feature vector may have a higher dimensionality than the original feature vector generated from each sensor, and feature reduction techniques may be further employed to extract useful features from the set of the new feature vector. In matching score level fusion, each subsystem using one biometric trait of the multi-biometric system provides a matching score indicating the proximity of the feature vector with the template vector. These scores can be combined to assert the veracity of the claimed identity. A number of transformation-based, classifier-based, and density-based score fusion methods have been used to combine scores of multiple scores. In decision level fusion each sensor first acquire one of multiple biometric traits and the resulting feature vectors are individually classified into the two decisions - accept or reject the claimed identity. Then a scheme that exploits the known decisions to make the final decision is used. So far, boolean conjunctions, weighted decision methods, classical inference, Bayesian inference, Dempster–Shafer method, and voting have been proposed to make the final recognition decision. In the field of multi-biometrics, a great number of studies of feature level fusion, matching score level fusion and decision level fusion have been made. Though fusion of multi-biometric are generally recognized as three classes as described above, in real-world applications of multi-biometric system it is possible that the “Fusion Process” may be simultaneously involved in different levels such as in both the matching score level and the decision level.
In recent years, the authors and their collaborators have insisted on research on multi-biometric technologies. Our investigations cover all the three categories of multi-biometric technologies, which include: 1) Feature level fusion: we investigate two novel feature level fusion methods, a pixel level fusion method to combine face and palm print traits, and a feature level fusion method to combine phase and orientation information of palm print images for effective personal recognition; 2) Matching score level fusion: We present a practical application example to study the effectiveness of matching score level fusion in face and palm print recognition; 3) Decision level fusion: We introduce a group decision-making combination approach to combine decisions of multiple face recognition algorithms.
The book is organized into three main parts. As an overview of the book, Chapter 1 describes the basic concepts necessary for a premier understanding of biometric data discrimination and multi-biometric technologies. Part I explores some advanced biometric data discrimination technologies developed for the small sample size (SSS) problem, where we first provide a brief introduction of linear discriminant analysis and the SSS problem, and then describes our solutions to SSS by developing extensions to Fisher’s discriminant criterion and novel improved discriminant analysis approaches. Part II describes several tensor-based biometric data discrimination technologies, including tensor principal component analysis, tensor linear discriminant analysis, and tensor locality preserving projections. Other recently developed tensor approaches, such as tensor independent component analysis, tensor non-negative matrix factorization, tensor canonical correlation analysis, and tensor partial least squares, are also introduced in this part. Part III deals with the second topic of the book, multi-biometrics. We first introduce the fundamental conception and categories of multi-biometrics technologies, and then describe three kinds of multi-biometric technologies, feature level fusion, matching score level fusion, and decision level fusion by providing several implementation examples.
This book includes fifteen chapters. Chapter 1 briefly introduces two advanced biometric recognition technologies, biometric data discrimination and multi-biometrics, to enhance the recognition performance of biometric systems. In Section 1.1, we discuss the necessity, importance, and applications of biometric recognition technology. A brief introduction of main biometric recognition technologies are then presented in Section 1.2. In Section 1.3, we describe biometric data discrimination and multi-biometric technologies. Section 1.4 outlines the history of related work and highlights each chapter of this book.
Chapter 2 is a brief introduction to biometric data discriminant analysis technologies. Section 2.1 describes two kinds of LDA methods: classification-oriented LDA and feature extraction-oriented LDA. Section 2.2 discusses LDA for the small-sample-size problem. Section 2.3 briefly introduces the organization of Part I.
In Chapter 3, we present three novel classification-oriented linear discriminant criteria. The first one is the large margin linear projection criterion, which makes full use of the characteristic of the SSS problem. The second one is the minimum norm minimum squared-error (MNMSE), which is a modification of the classical minimum squared-error (MSE) discriminant criterion. The third one is the maximum scatter difference, which is a modification of the Fisher’s discriminant criterion.
In Chapter 4, we first give a brief introduction to Fisher’s linear discriminant, Foley-Sammon discriminant, orthogonal component discriminant, and application strategies for solving the SSS problem. We then present two novel orthogonal discriminant analysis methods: one is orthogonalized Fisher discriminant; the other is Fisher discriminant with Schur decomposition. At last, we compare the performances of several main orthogonal discriminant analysis methods under various SSS strategies.
Chapter 5 describes three kinds of weighted LDA methods. The first is parameterized direct linear discriminant analysis. The second is weighted nullspace linear discriminant analysis. The third is weighted LDA in the range space of within-class scatter matrix. At last, we give a summery of the chapter.
Chapter 6 introduces two novel facial feature extraction methods. The first is multiple maximum scatter difference (MMSD), which is extension of a binary linear discriminant criterion, i.e. maximum scatter difference. The second is discriminant analysis based on coefficients of variances (DCV) which can be viewed as a generalization of null space LDA. At last, we give a summery of the chapter.
Chapter 7 gives the background for developing tensor-based discrimination technologies. Section 7.2 introduces some basic notations in tensor space. Section 7.3 discusses several tensor decomposition methods. Section 7.4 introduces the tensor rank.
Chapter 8 presents some variants of classical PCA, and discusses the properties of tensor principal component analysis. Section 8.1 gives the background and development of tensor principal component analysis. Section 8.2 introduces tensor principal component analysis. Section 8.3 discusses some potential applications of tensor principal component analysis in biometric systems. We summarize this chapter in Section 8.4.
In Chapter 9, classical LDA and its several variants are introduced. In some sense, the variants of LDA can avoid the singularity problem and achieve computational efficiency. Experimental results on biometric data show the usefulness of LDA and its variants in some cases.
In Chapter 10, we describe two subspace analysis technologies: tensor independent component analysis and tensor non-negative matrix factorization, which can be used in many fields like face recognition and other biometric systems. Section 10.1 gives the background and development of two subspace analysis. Section 10.2 introduces tensor independent component analysis. Section 10.3 gives tensor nonnegative factorization. Section 10.4 discusses some potential applications of these two subspace analysis in biometric systems. We summarize this chapter in Section 10.5.
Chapter 11 deals with two tensor-based classifiers, tensor canonical correlation analysis and tensor partial least squares, which can be used in many fields such as biometric systems. Section 11.1 briefly surveys the history in developing tensor-based classifiers. Section 11.2 introduces tensor-based classifiers. Section 11.3 gives tensor canonical correlation analysis and tensor partial least squares. We summarize this chapter in Section 11.4.
Chapter 12 describes the basic concepts of biometrics and motivation of multi-biometrics. Section 12.2 provides the definitions and notations of biometric and multi-biometric technologies. Section 12.3 presents two feature extraction approaches for biometric images, where one of which is a linear feature extraction approach and the other is a nonlinear feature extraction approach. Section 12.4 briefly indicates some techniques associated with each kind of multi-biometric technologies.
Chapter 13 mainly presents the basic concepts and two examples of feature level fusion methods. As the beginning of this chapter, Section 13.1 provides an introduction to feature level fusion. Section 13.2 briefly surveys current feature level fusion schemes. Section 13.3 shows a feature level fusion example that fuses face and palm print. Section 13.4 shows a feature level fusion example that fuses multiple feature presentations of a single biometric trait. Section 13.5 offers brief comments.
In Chapter 14, we describe several basic aspects of matching score level fusion. Section 14.1 provides a brief introduction of basic characteristics of matching score fusion. Section 14.2 describes a number of matching score fusion rules. Section 14.3 presents the normalization procedures of raw matching scores. Section 14.4 provides several brief comments on matching score fusion.
Chapter 15 deals with the decision fusion rules, the classifier selection approach, and a case study of face recognition based on decision fusion, as well as a summary of multi-biometric technologies. In a multi-biometric system, classifier selection techniques may be associated with the decision fusion as follows: classifier selection is first carried out to select a number of classifiers from all classifier candidates. Then the selected classifiers make their own decisions and the decision fusion rule is used to integrate the multiple decisions to produce the final decision. Section 15.1 provides an introduction to decision level fusion. Section 15.2 presents some simple and popular decision fusion rules such as the AND, OR, RANDOM, and Voting rules, as well as the weighted majority decision rule. Section 15.3 introduces a classifier selection approach based on correlations between classifiers. Section 15.4 presents a case study of group decision-based face recognition. Section 15.5 offers several comments on the three levels of multi-biometric technologies.
Chapter 16 summarizes the book from a holistic viewpoint. Section 16.1 summarizes the contents of the book and indicates the relationship between different chapters in each part. Section 16.2 reveals that how the methods and algorithms described in different parts can be applied to different data forms of biometric traits. Section 16.3 provides comments on the development of multi-biometrics.
In summary, this book is a comprehensive introduction to theoretical analysis, algorithms, and practical applications of two kinds of advanced biometric recognition technologies, biometric data discrimination and multi-biometrics. It would serve as a textbook or as a useful reference for graduate students and related researchers in the fields of computer science, electrical engineering, systems science, and information technology. Researchers and practitioners in industry and R&D laboratories working on security system design, biometrics, computer vision, control, image processing, and pattern recognition would also find much of interest in this book.
In the preparation of this book, David Zhang organizes the contents of the book and is in charge of Chapters 1, 2, 7, 12 and 16. Fengxi Song handles Chapters 3-6. Zhizhen Liang and Yong Xu write Chapters 8-11 and Chapters 13-15, respectively. Finally, David Zhang looks through the whole book and examines all chapters.
Finally, the authors are full of gratitude to our co-workers and students for their persistent support to our research. Our sincere thank goes to Prof. Zhaoqi Bian of Tsinghua University, Beijing, Prof. Jingyu Yang of Najing Polytechnic University, Najing, China, and Prof. Pengfei Shi of Shanghai Jiaotong University, Shanghai for their invaluable advice throughout this research. We would like to thank our team members, Dr. Wangmeng Zuo, Prof. Jian Yang, Prof. Xiaoyuan Jing, Dr. Guangming Lu, Prof. Kuanquan Wang, Dr. Xiangqian Wu, Dr. Jingqi Yan, and Prof. Jie Zhou for their hard work and unstinting support. In fact, this book is the collaborative result of their many contributions. We would also like to express our gratitude to our research fellows, Feng Yue, Laura Liu, Dr. Ajay Kumar, Dr. Lei Zhang, Dr. Hongzhi Zhang, Bo Huang, Denis Guo, and Qijun Zhao for their invaluable help and support. Thanks are also due to Martin Kyle, for his help in the preparation of this book. The financial support from the HKSAR Government, the central fund from the Hong Kong Polytechnic University, the NFSC funds (Nos. 60332010, 60602038, and 60402018), and the 863 fund (No. 2006AA01Z193) in China are of course also greatly appreciated. We owe a debt of thanks to Jan Travers and Kristin Roth of Idea Group Publication, for their valuable suggestions and keeping us on schedule for the preparation and publishing of the book.
- David Zhang
- Fengxi Song
- Yong Xu
- Zhizhen Liang