Research On Appearance-based Statistical Face Recognition

Face recognition is a research task full of theoretical and application value, and study on it now is quite immature. Firstly, the research history and development status of it is reviewed and several mainstream face recognition algorithms are discussed. Secondly, the basic theory of feature extraction and selection is presented. Thirdly, it is emphatically discussed on both linear and non-linear feature extraction methods.As a linear feature extraction method, Principal Component Analysis (PCA) is based on data description, and it is the optimal transform in the sense of minimum mean squared error. Linear Discriminant Analysis (LDA) is a method based on data classification, which looks for the mapping direction with maximum distances between classes and minimum distances within class, and also keeps differentiation information of the original samples as much as possible. The face recognition performance of two linear methods is compared by experiment and it is found that PCA can have a better face recognition performance than LDA with a small number of samples, the explanation is given. Non-lineal feature extraction method of face recognition is also studied, and an improved method based on KPCA-ICA feature extraction is proposed. Kernel Principal Component Analysis (KPCA) is a non-lineal generalization of PCA, which extracts non-linear information from human facial image. Independent Component Analysis (ICA), a generalized of PCA, is based on high-order statistics, which is able to extract high-order statistic information of the human face image pixels, and can codes parts of images. Combination of the two methods maybe a feasible method to extract non-linear features of human faces images. Result turns out that the improved method we proposed has a good recognition capability.