Linear Dimensionality Reduction Technology For Face And Palmprint Feature Extraction

Automatic identification and authentication systems that make use of biometric data, that is, distinctive anatomical and behavioral characteristics, are becoming ever more widely used for access control, surveillance, and system security. Face and palmprint are two representative biometric characteristics. Because of its wide applications and theoretical values, face recognition has been a focus of research in the areas of computer vision and pattern recognition. Palmprint recognition too has attracted considerable interest, partially because palmprint images can be captured at a low cost, and palmprints contain many unique features for personal identification.Feature extraction or dimensionality reduction is a basic part of any biometric system. Biometric data is high dimensional, and always contains information that is less discriminative or that is not useful for recognition. Dimensionality reduction allows this information to be efficiently suppressed without losing discriminative information. Since of its simplicity, effectiveness, and generalization, linear dimensionality reduction technology have been extensively applied to biometric data feature extraction.When applied to face and palmprint recognition, linear dimensionality reduction technologies, such as principal component analysis (PCA) and linear discriminant analysis (LDA), usually encounters the Small Sample Size problem, where the data dimensionality is much higher than the size of the training set, leading to the singularity and poor estimation of the scatter matrix (the generalization problem). Besides, biometric systems capture, detect and recognizing biometric data automatically, making it inevitable that face or palmprint images will sometimes be noisy, or partially occluded. For PCA and LDA, with the increase of the degree of noise and partial occlusion, the recognition performance would deteriorate severely (the robustness problem).In this dissertation, we investigate the generalization and the robustness of linear dimensionality reduction technologies, and propose several novel PCA-based and LDA-based dimensionality reduction methods, including: (1) Bi-Directional Principal Component Analysis: To improve the generalization performance, we propose a BDPCA feature extraction method by offering a number of significant advantages. First, BDPCA is directly performed on the image matrix, while PCA requires mapping an image matrix to a high dimensional vector in advance. Second, BDPCA can circumvent the over-fitting problem associated with PCA. Third, the feature dimension of BD-PCA is much less than that of 2DPCA.(2) Assembled Matrix Distance (AMD): To reflect the fact that the BD-PCA feature is a matrix, we define an AMD distance to measure the distance between two feature matrices. We have theoretically proved that the AMD distance is a matrix metric. Using the nearest neighbor classifier or the nearest feature line classifier, the AMD distance measure would achieve a lower error rate than the Frobenius or the Yang distance measures.(3) BDPCA Plus LDA (BDPCA+LDA): BDPCA+LDA adopts a two-stage framework, where LDA is performed in the BDPCA subspace. With this framework, the BDPCA+LDA method has a number of advantages over the popular PCA+LDA approach, such as less computational requirement, less memory requirement, and higher recognition accuracy.(4) Post-Processed Discriminant Analysis: In this method, 2D-Gaussian filtering is used for post-processing the discriminant vectors. Post-processing technique can be used to combine with other discriminant analysis methods, such as the enhanced Fisher's discriminant model and the complete Fisher discriminant framework, to further improve the recognition performance. Theoretically, post-processing technique is equivalent to the image Euclidean distance, and needs less computational requirement in the recognition stage.(5) Combination of BDPCA+LDA and Post-Processed Fisherfaces: We propose a Combined LDA method for fusing BDPCA+LDA and Post-Processed Fisherfaces. Because post-Processed Fisherfaces and BDPCA+LDA use two distinctly different projectors and achieve similar recognition performance, Combined LDA can be used to further improve the recognition accuracy. Experimental results show that Combined LDA outperforms post-Processed Fisherfaces, BDPCA+LDA, and other popular LDA-based methods. (6) Iteratively Reweighted Fitting of Eigenfaces: To address the performance degradation of Eigenfaces against noise and partially occlusion, we propose an iteratively reweighted fitting of the Eigenfaces method (IRF-Eigenfaces). IRF-Eigenfaces first defines a generalized objective function and then uses the iteratively reweighted least-squares fitting algorithm to extract the feature vector by minimizing the generalized objective function. Experimental results show that IRF-Eigenfaces is far superior to both Eigenfaces and to the local probabilistic method in recognizing noisy and partially occluded faces.