Face Recognition And Its Applications

Face recognition is a hot topic in the field of artificial intelligence and pattern recognition. Face detection, image preprocessing, feature extraction and recognition are included in a wide sense of the face recognition. But in the strict sense, face recognition includes feature extraction and recognition only. Feature extraction is a process of modeling facial features, the purpose of which is extracting low dimensional features from the high dimensional facial patterns. And then, the following recognition becomes easy. The core step of face recognition is feature extraction, which influences the performance of recognition directly.In this dissertation, a more detailed study on feature extraction is carried out, as follows:(1) Training samples are mapped into the kernel space by KECA, in which projection matrix is selected by the entropy contribution. Although features extracted by KECA reflect the human face better, the amount of calculation is increased caused from the kernel method. We proposed ECA method inspired by KECA and PCA. Covariance matrix is formed in the input space directly, and the projection matrix is selected by the entropy contribution. The accurate recognition rate of ECA is higher than that of PCA, while the efficiency of ECA is higher than that of KECA. In order to extract features from the two dimensional image matrix, we proposed an algorithm of 2DECA. The efficiency of 2DECA is higher than that of ECA according to the experimental results.(2) The (?)(?) norm is sensitive to outliers. Some face recognition methods based on none (?)(?) norm were proposed, such as R1-PCA. R1-PCA calculates projection matrix by an iterative method. More iteration numbers are needed for the convergence caused by the high dimensional image vector. We proposed 2DR1-PCA to extract features from the image matrix directly. The iteration numbers are reduced. As a result, the efficiency and recognition rates are improved, respectively.(3) Human face is approximately symmetrical. Main facial organs, such as eyes, nose and mouse, can be separated into different parts based on the symmetric property. We proposed a new modular 2DPCA method for obtaining local and holistic features of a human face. At last, every patch is weighted to the minimum distance classifier. The proposed method obtained a better performance even on training set which has few training samples per person.(4) LBP is an effective texture extraction operator. Images obtained by LBP are insensitive to the illumination condition. We proposed a new method for the image contrast enhancement inspired by LBP. Skins are weakened and organs are emphasized in the enhanced images. The difference of facial organs between images could be expressed by the city-block distance between corresponding enhanced images. Then we weight it to the minimum distance classifier. A better performance is obtained in the experiments.(5) Single mode images would not express face exactly and comprehensively. We proposed a new feature level image fusion method to fuse infrared images and visible images. Advantages of infrared and visible images are retained in the fused images. As details and character are increased, the fused images lead to better performance in the face recognition. We extract features from these three mode images by ECA and 2DECA, and then recognition performance is processed. We got a higher recognition rates in the experiments.