Research On Methods Of The Discriminant Feature Extraction In Face Recognition

Feature extraction is one of the elementary problems in the area of pattern recognition. It is the key to the classifier problems such as face identification. Linear discriminant analysis including principal component analysis (PCA) (or K-L transform) and Fisher linear discriminant analysis (FDA), and nonlinear discriminant analysis based kernel trick are classical and widely used techniques for feature extraction. Some of their theories and algorithms are researched further in this dissertation, in which some proposed algorithms work well for human face recognition.How to get the optimal Fisher discriminant vectors efficiently in the case of small number samples is a very difficult and critical problem. Based on the generalized Fisher linear discriminant criterion, this problem is solved by combining projection transform, isomorphic mapping and compressed transform. A new algorithm called PPCA+FDA is developed, which is thought to be effective for feature extraction. By the proposed algorithm, in fact, the optimal discriminant vectors can be derived from a low dimensional transformed subspace. Experiments are performed on ORL face image database with respect to three kinds of resolution, and the experimental results indicate that PPCA+FDA is robust, efficient, and the optimal discriminant vectors extracted by the proposed algorithm have powerful ability for feature extraction by the general minimum distance classifier.By the virtue of PCA, Modular PCA, a human face recognition technique, is presented in this paper. First, in the proposed approach, the original images are divided into modular images, which are also called sub-images. Then, the well-known PCA method can be directly used to the sub-images obtained from the previous step. There are two advantages for this way: 1. lbcal feature of the images can be extracted efficiently; 2. singular value decomposition of matrix may be avoided in the process of feature extraction, which is simpler than that of previous technologies such as PCA. Moreover, PCA is a special case of modular PCA. Following the modular PCA, another method of feature extraction named M2PCA+FDA is presented in this paper. There are two steps for M2PCA+FDA. Modular PCA is firstly used on the original images to get lower dimensional patterns corresponding to the original images. Then the well-known fisher-faces method is followed on the lower dimensional patterns to finish the pattern classification. That is to say that M2PCA+FDA is equivalent to Modular PCA plus fisher-faces method. Both Modular PCA and M2PCA+FDA are tested with recognition performance on two human face image databases respectively. The experimental results indicate that the performances of two proposed methods are obviously superior to that of PCA.In this paper, modular two-dimensional principal component analysis (M2DPCA,or Modular 2DPCA), which is the generalization of two-dimensional principal component analysis (2DPCA), is...