Face recognition is an important problem in pattern recognition and computer vision area. Due to various adverse factors such as expression, lighting, pose, low-resolution and heterogeneous modality variations, the performance of most existing face recognition algorithms is far behind satisfactory and hence limits its application in real world. In this thesis, we study on some issues of face recognition and propose several novel descriptors and recognition methods to improve the face recognition performance. The main contributions of this thesis include following issues: 1. To reduce the effect of illumination variation, we propose two feature extraction methods. One is a local Gabor texton based face representation method, which incorporates the advantages of Gabor and texton and therefore is robust to illumination variation. Another one is named Gabor volume based local binary pattern descriptor. Different from the existing method, it exploits the face information in space, frequency and orientation domains simultaneously to make the face representation more sufficient. Further, it utilizes CMI and LDA method to reduce the feature dimension to improve the face recognition performance efficiently and effectively. 2. In order to improve the performance of low resolution face recognition, we propose a local frequency descriptor (LMD) for face representation. The LMD properly formulates the magnitude and phase information in low frequency band of image. Due to its nearly blur-invariant property, it is suitable for low resolution face recognition. 3. To deal with the heterogeneous image matching problem, we propose coupled spectral regression framework which utilizes two projections to project the heterogeneous data respectively onto the common discriminative subspace to be classified. An efficient solution by using graph embedding and spectral regression is presented. Thanks to the proper regularization technique, it demonstrates good generalization performance. 4. To improve the robustness of system to pose variation, we propose a 3D face shape recovery method from a single 2D face image. By utilizing tensor analysis and canonical correlation analysis (CCA) method, we successfully build the 2D and 3D space and learn the relationship between them. Experimental results show the proposed method is effective and efficient for 3D face shape recovery. With the help of this technique, one could produce virtual samples in registration phase so as to improve the...
修改评论