Research On Face Recognition Method Based On Face Alignment And Multi - Feature Fusion

With the rapid development of computer science and network communication technology, persons began to use biological characters of human for identification. Biometric recognition includes face recognition, voice recognition, fingerprint and palmprint recognition, gait recognition and so on. Compared with other biometrics, face recognition has wide applications in social public security and daily life. face recognition has the advantage of feature obtained easy and little time cost, so it has a wide applicate prospect in social public security and daily life. A complete face recognition system includes the following process steps:face detection, face alignment, feature extraction and classification. Among them, face alignment and feature extraction are the two key steps. So, we will research these two problems.The main contributions of this dissertation include:(1) ASM is the most representative face alignment method based on statistical model. However, the feature point search range of ASM is limited. To overcome this problem, we propose an improved face alignment method named the improved ASM based on LBP. This method extends the search region to search the feature points and uses coding of LBP instead of gray value to build the local texture model. The experimental results demonstrate that the improved ASM based on LBP perform better than ASM.(2) We mainly investigate regression based methods for face alignment, including linear regression based method and shape regression based method. We firstly study the dense linear regression method, sparse linear regression method and fully sparse linear regression method. Thus, we evaluated different experiments to verify the differences of above three methods. Compared with regression based methods, explicit shape regression based method has no need of fixed shape model parameters and employs cascaded regressor to update the initial face shape for locating facial landmarks. Shape regression based method employs the linear combination of samples to constrain the shape rather than using PCA to constrain that. Experimental results demonstrate that explicit shape regression based method achieves better performance than linear regression based methods.(3) We mainly investigate the local feature extraction methods (e.g. LBP、SIFT、Gabor、 LGBP) for face representation. Motivated by the idea of LGBP and image decomposition, a novel face representation method is proposed. This method firstly decomposes one image into a series of sub-images according to local structure. We use LBP feature and SIFT feature of all sub-images to construct two feature vectors. For LBP and SIFT, we employ LDA to achieve a low-dimensional and discriminative feature and fuse them using sum score. Experimental results demonstrate advantages of our method over LBP、SIFT and Gabor.