The Research And Application Of Face Recognition

Face recognition is one of the most important research topics in computer vision, machine learning and biometrics. In recent years, with the efforts of the researchers around the world, face recognition has made tremendous progress and obtained remarkable research achievements, but it is still a big challenge to make the face recognition methods more robust to misalignment, occlusions and variations(e.g., pose, illumination and expression), which also needs further studies. In order to improve the robustness of face recognition, we focus our attention on the related theories and algorithms in our dissertation. We did some researches about the methods of global and local features, shearlet transform features, sparse representation, dictionary learning, low-rank representation and so on, and we also proposed some algorithms for face recognition. The main work and contributions of our dissertation are summarized as follows:(1) The PCA method extracts global features of the original images, and it does not consider the local features. In contrast, Modular PCA method extracts the important local features. However, vectorization in PCA or modular PCA often causes "curse of dimensionality". In order to extract features from matrix or higher-order tensor objects directly, multilinear principal component analysis(Multilinear PCA) is developed. Multilinear PCA can avoid "curse of dimensionality", meanwhile it would not destroy the original data structure. Inspired by Modular PCA and Multilinear PCA, we propose a new method called Modular Multilinear Principal Component Analysis(M2PCA) for face recognition. Experiments were conducted on the face databases respectively, and experimental results indicate that, under the same condition of sub-blocks, the proposed method is obviously superior to the general Modular PCA.(2) In order to extract richer texture features of face images to improve face recognition accuracy, a new face recognition algorithm based on the Shearlet_ULBP feature which extracts the histogram of Uniform Local Binary Pattern(ULBP) from the Shearlet coefficients and collaborative representation(Shearlet_ULBP CRC) is proposed. First, Shearlet transform is used to extract the multi-orientation facial information, and the average fusion method is exploited to fuse the original Shearlet features of the same scale. Second, the fused image is divided into several nonoverlapping blocks, then face image is described by the histogram sequence extracted from all the blocks with the ULBP operator. Finally, the extracted features are fed into the collaborative representation based classifier. The proposed method can extract richer information about edges and texture features. Several experiments have been conducted on face databases, and the results show that the new method is robust to the illumination, pose and expression variations, as well as occlusions.(3) In order to handle the problem that both training and testing images are corrupted, we proposed a novel semi-supervised method based on the theory of the low-rank matrix recovery and the incoherent dictionary for face recognition, the incoherent dictionary is learned by incorporating a correlation penalty into the dictionary learning model. We can learn discriminative low-rank and sparse representations for both training and testing images simultaneously with the method. Experimental results obtained on several face image databases show the effectiveness of our method, i.e., the proposed method is robust to the illumination, expression and pose variations, as well as images with noises such as block corruptions or uniform noises.