Research On Key Problems Of Face Detection And Recognition

Face recognition is the most natural, direct, and nonintrusive method among biometrics recognition methods. Face detection and recognition has been widely applied in image recognition tasks, such as identity authentication, electronic commerce, video surveillance, and human machine interaction, and it has become a challenging and hot research point in pattern recognition and artificial intelligence domain. The research of face detection and recognition has very large theoretical and practical values. In this dissertation, several key problems of face detection and recognition, including feature extraction, non-uniform illumination compensation, and the small sample size (SSS) problem are discussed. The main contents of this dissertation can be summarized as follows:1) Face feature extraction based on adaptive signal decomposition is investigated. The function dictionary of the matching pursuit algorithm is composed of the intrinsic mode functions (IMFs) of mean faces from training samples. Using matching pursuit algorithm for face feature extraction, a face detection algorithm and a face gender classification algorithm based on empirical mode decomposition (EMD) are proposed, respectively.2) Face detection based on time-frequency transform is discussed. The time-frequency transform outputs of mean faces from training samples are used as the projection vectors. For image feature extraction, the inner products between the input image and the projection vectors are calculated. Then, face detection algorithms based on wavelet decomposition and Gabor transform are presented, respectively.3) Face non-uniform illumination compensation based on wavelet transform is discussed. In the logarithm domain, for face illumination compensation, the low frequency coefficients of face image are discarded by calculating wavelet transform, wavelet packet transform and multiwavelet transform, respectively. At the same time, three face non-uniform illumination compensation algorithms are given.4) Alleviating small sample size problem of face recognition based on Gabor transform is examined. All output images of face Gabor transform are regarded as independent new samples. Therefore, it is equivalent to multiple the sample size of each subject while keep the sample dimension invariable. After that, several novel face recognition algorithms which can alleviate small sample size problem are developed.All the algorithms above are experimentally validated and analyzed, and the validity of the algorithms is then proved.