Face Recognition Based On Statistical Learning Methods

As one of the most popular research fields of biometric person recognition at present, face recognition has become the most successful application area of image processing, pattern recognition, computer vision and so on, and has a wide range of promising applications, such as security surveillance, personal identity identification. The present automatic face recognition system (AFRS) is based on 2D image, but disturbed by pose, illumination, expression and other factors, effectiveness and robust has been greatly weakened. 3D model based face recognition is one of the effective methods to solve the pose problem, but 3D model based face recognition is still based on 2D image, and the reality of the model, the problem of model optimization are still far from application. Therefore, the techniques required by AFRS are far from application, and a successful AFRS has a long way to go. The success of FR will be heavily dependent on the interdisciplinary research of computer vision, pattern recognition, cognition, psychology, neural science and so on.The present face recognition methods mainly are statistical learning models, including based on global features and local features. The canonical methods based on global features contains Eigenfaces, Fisherfaces, Laplacianfaces and etc, while HMM based face recognition belongs to the methods based on local features. The dissertation mainly addresses several problems of statistical learning methods applied in face recognition based on 2D image. The main contribution of the dissertation is summarized as follows:1) A comprehensive study on the survey of face recognition is summarized. Detail kinds of face recognition approaches, including feature-based method, elastic graph match, neuro network, hiddern markov model, subspace learning, manifold learning, 3D model. With comparing pros and cons of different approaches, the thesis shed some new light on the improvement and developing orientations of these approaches.2) Propose PPCA in DCT domain to extract features fed into EHMM model as...