Research On Subspace Analysis-based Face Recognition Methods

With the development of network and communications technologies and the constantly expanding of physical and virtual space of the human being, the demand on the safety for information is growing. Biometrics utilizes computer to automatic identify a person based on his/her distinct physiological or behavioral characteristics. Biological characteristics possess many desirable features. For example, they are innate for everyone, different individuals have different characteristics and they will remain unchanged for a long time. Moreover, biological characteristics can not be forgotten or lost and possess innate convenience and efficiency in technology.As an important part of the biometrics, face recognition that exploits the effective information from the facial image and video for automated personal authentication. Compared with other biometrics, face recognition possess many virtues: convenient and contactless data acquisition, without any damage to the users, and interactable. So face recognition is being a hotspot of biometrics and has a wide application prospect.Among the present various facial feature extraction and face recognition algorithms, the subspace analysis-based algorithms are aroused wide concerns due to theirs favorable properties, such as convenient computation and effectiveness for identification. The subspace analysis-based algorithm has become the main method of facial feature extraction and face recognition. We have mainly studied the subspace-based feature extraction and recognition algorithms and the information fusion methods and proposed a series of efficient facial feature extraction and face recognition algorithms. The main creative work in the dissertation is:1. Propose a Complete Fisher discriminant analysis (CFDA) algorithm to solve the small sample size problem (SSSP) in the linear discriminant analysis (LDA) based face recognition algorithms. At First, a fact is proved that all the discriminative information based on Fisher criterion holds inΨ_t, the range space of the total scatter matrix. So the optimal discriminant features can be found in the lower dimensional spaceΨ_t without loss of effective discriminatory information. At second, two kinds of discriminant features, DF1 and DF2, are extracted from the range space and null space of the within-class scatter matrix, respectively. These features provide the face sample a complete representation that based on Fisher criterion. 2. Propose a Global and Local Complete Fisher Analysis and Fuzzy Integral (GLCFDA-FI) based method for face recognition. Eyes image holds lots of discriminatory information and is robust with the variation of facial expression. Moreover, eyes image is unaffected by the variation of mouth opening or closing, having or not beards and mustaches, and with or without respirator. So eyes image is the important complementarity to the entire face image for face recognition. GLCFDA-FI algorithm firstly crops the entire face image and eyes image from the original image based on the eye locations. The global and local features are then projected using CFDA, respectively and four kinds of discriminant features, i.e. global DF1, global DF2, local DF1, and local DF2, are obtained. Finally, the four kinds of features are fused by fuzzy measure and fuzzy integral for classification. GLCFDA-FI face recognition algorithm can convert the small sample size problem into advantage, and to get a good recognition result by adjusting the values of fuzzy densities when facial expression or lighting is varying.3. Propose a Wavelet decomposition based Complete Fisher Analysis and Fuzzy Integral (WCFDA-FI) algorithm for face recognition. In WCFDA-FI algorithm, the face image is decomposition by suitable level 2-D wavelet transform at first, and the low frequency component and appropriate high frequency component are extracted to form the low frequency and high frequency feature respectively. The low frequency and high frequency features are then projected using CFDA, and four kinds of discriminant features, i.e. low frequency DF1, low frequency DF2, high frequency DF1, and high frequency DF2, are obtained. Finally, the four kinds of features are fused by fuzzy measure and fuzzy integral for classification. Since the variation of expression mainly affects the high frequency components and the variation of lighting mainly affects the low frequency component, to adjust the values of fuzzy densities can obtain a good result for face recognition.4. Propose a Global and Local Complete Kernel Fisher Analysis and Fuzzy Integral (GLCKFDA-FI) based method for face recognition. Being a main trend face recognition method, the kernel based nonlinear subspace analysis algorithm can solve the nonlinear distribution problem of the face samples in the original input space. Moreover, kernel techniques can be introduced to avoid the difficulty of intensive computation. Complete Kernel Fisher Discriminant Analysis (CKFDA) carries out CFDA in the high dimensional feature space F, and GLCKFDA-FI algorithm is the nonlinear version of GLCFDA-FI algorithm. At first, the global and local features are projected using CKFDA, respectively and four kinds of discriminant features, i.e. global KDF1, global KDF2, local KDF1, and local KDF2, are obtained. At second, the four kinds of features are fused by fuzzy measure and fuzzy integral for classification. It is indicated experimentally that GLCFDA-FI face recognition algorithm can obtain a good recognition result under the complex environment in which facial expression or lighting is varying.5. To solve the problem that the face samples are nonlinear embedded in the original sample space, a novel face recognition algorithm called Local Fisher Discriminant Analysis (LFDA) is proposed in this paper. Different from Fisher Discriminant Analysis (FDA) which defines the global-based within-class and between-class scatter, LFDA defines the local within-class and local between-class scatter which describe the properties of the local stricture of face sample data. Based on the Fisher criterion, a projection is to be found to improve the reparability of the samples, which maximized the local between-class scatter as well as minimized the local within-class scatter. To solve the small sample size problem (SSSP), we propose two face recognition algorithms based on LFDA: (1) Local Fisherface (L-Fisherface) algorithm, which uses PCA at first, to reduce the dimension of original samples, then Local Fisher Discriminant Analysis is carried out in the PCA subspace. (2) Local Complete Discriminant Analysis (LCFDA) algorithm, which extracts the discriminant features from the range space and null space of the local within-class scatter matrix at same time with the purpose of never loss of the discriminative information. In LCFDA algorithm, the two kinds of discriminant features are fused in feature level for face recognition. Experimental results suggest that the proposed L-Fisherface and LCFDA algorithms provide better representation and achieve higher accuracy in face recognition.