Feature Extraction Based On Neighborhood Structure And Its Applications To Face Recognition

Face recognition has become a hot research topic in real-world biometric application. It has recently received significant attention, especially during the past several years. Meanwhile, as one kind of widely used biometric identification technology, face recognition has been applied in wide range of both public security and daily life. Feature extraction is a key step in the procedure of face recognition, and is always an important research topic. To obtain more effective discriminant features, many feature extraction algorithms have been proposed. Among them, manifold learning algorithms from new angle of view, enrich the existing feature extraction techniques, and own important theoretical and practical value.Based on the analysis and summary of existing manifold learning algorithms, this dissertation focuses on the feature extraction method using neighborhood structure. The major research works and contributions of this dissertation are summarized as follows:(1) To solve the problems existed in locality preserving projection (LPP) such as ignoring the class information of samples and similarity weight, we propose a novel manifold learning method called classification-membership preserving projection (CPP). The proposed CPP can not only maintain the local structure of all samples, but also take the class information into account and redefine the similarity weights of intra-class samples. After projection, neighborhood relationship of the intra-class samples which possess more classification probability can be preserved.(2) To overcome the limitation of marginal fisher analysis (MFA) that it can not find the optimized projection axis for enlarging the margin between different classes, we propose a novel algorithm named maximum margin learning projection (MMLP) algorithm based on redefined within-class scatter matrix and between-class scatter matrix, which are constructed by employing the relationship between each data point and its neighboring points. The proposed MMLP can not only maintain the samples of the original distribution information in low dimensional feature space, but also separate the different samples in a small neighborhood from each other.(3) To solve the common problem existed in locality preserving projection (LPP) and other related algorithms that the similarity weight can only measure the distance between samples and cannot provide other valuable information for classification, we propose a novel algorithm called classification probability preserving discriminant analysis (CPPDA) algorithm based on a new similarity measure mechanism. The proposed CPPDA can not only contain the neighborhood information of sample, but also reflect the probability that the samples are correctly classified by k-nearest neighbor classifier. CPPDA seeks to find an optimal projection matrix such that the distance between each sample and the within-class centers are minimized, the distances between each sample and the between-class centers are maximized simultaneously.(4) To reduce the influence on recognition performace caused by the variations of illumination, facial expression and poses, we propose a novel algorithm called membership-degree preserving discriminant analysis (MPDA) algorithm by combining LPP and LDA. MPDA employs fuzzy k-nearest neighbor method to calculating the membership-degree, and takes the membership-degree to characterize the similarity weight between the sample and the center of a class. The proposed MPDA can not only well reflect the neighborhood information of the original data samples, but also make full use of the class information of each sample.