Abstract
Face recognition is one of the most intensively studied topics in computer vision and pattern recognition. There are three essential issues to be dealt with in the research of face recognition; namely, pose, illumination, and expression variations. The recognition rate will drop considerably when the head pose or illumination variation is too large, or when there is expression on the face.Although many researches were focused on overcoming these challenges, few were focused on how to robustly recognize expressional faces with one single training sample per class. In this thesis, we modify the regularization-based optical flow algorithm by imposing constraints on some given point correspondences to compute precise pixel displacements and intensity variations. The constrained optical flow computation can be efficiently computed by applying the modified ICPCG algorithm.By using the optical flow computed from the input expression-variant face image with respect to a reference neutral face image, we can remove the expression from the face image by elastic image warping to recognize the subject with facial expression. On the other hand, the optical flow can be computed in the opposite direction, which is from the neutral face image to the input expression-variant face image. By combining information from the computed intra-person optical flow and the synthesized face image in a probabilistic framework, an integrated face recognition system is proposed, which can be robust against facial expressions with a limited size of training database. Experimental validation on the Binghamton University 3D Face Expression (BU-3DFE) Database is given to show that the proposed expression normalization algorithm significantly improves the accuracy of face recognition on expression variant faces.A possible solution for overcoming the pose variation problem in face recognition is also presented in this thesis. The ideal solution is to reconstruct a 3D model from the input images and synthesize the virtual image with the corresponding pose, which might be too complex to be implemented in a real-time application. By formulating this kind of solution as a nonlinear pose normalization problem, we propose an algorithm that integrates the nonlinear kernel function and the linear regression method, which makes the solution resemble to the ideal one. Some discussions and experiments on CMU PIE database are carried out, and the experimental results show that the proposed method is robust against pose variations.