In recent years,face recognition technology has been rapidly development,and has been applied to various fields.Compared with the other biometric technologies,face recognition has the following advantages,face information is easily collected,and noncontact form is easily accepted by people.Thus face recognition is popular in various application fields.However,there are enormous challenges in practical application.Both kinds of environmental changes and the degree of subject’s cooperation lead to the large gap between the gallery samples and probe samples in face recognition system,and affect the recognition accuracy.The purpose of this thesis is to establish the representation model of the nonideal probe samples and to extract the discriminant features.The research method for face recognition is the variation feature representation and local variation patterns.The main contribution of this dissertation is summarized as follows:(1)Propose the face recognition method based on Variational Feature Representation Classification(VFRC).VFRC uses the information of the general training samples,and refers to the information of the gallery samples at the same time.The VFRC method obtains the variational part and the normal part of the probe sample by using the linear regression model.The combination of two information will be used in the classification model for deciding the identify of the probe sample.The experimental results show that the VFRC method not only spends a few time,but also achieves the high recognition accuracy under the complex conditions,such as the exaggerated expressions,the head posture change and the obvious occlusions.(2)Propose the Customized Sparse Representation model with the Mixed norm(CSR-MN).In the CSR-MN model,the representation coefficient includes two parts: the part of corresponding to the gallery data and another part of corresponding to the variation dictionary.Because of the complex features of the variation dictionary,this thesis proposes the mixed norm corresponding to the new distribution that matches the practice better.Through the reasonable transformation,the complex CSR-MN model becomes the l1-minimization model to be easily solved by homotopy algorithm.Because the CSR-MN model very well simulates the actual situation of face recognition,the model shows the obvious advantage in a lot of numerical experiments.(3)Propose Customized Dictionary-based Face Recognition with Extended Joint Sparse Representation.The extended joint sparse representation is used in the classification stage of the method.Instead of using the general training set,the method directly learns the variation dictionary from the face data of the current probe subject and fits well the intra variation of the probe subject.In the classification stage,the extended joint sparse representation model not only takes full advantage of the customized dictionary and the group structure to enhance the recognition accurate.(4)Propose Gabor-scale Binary Pattern(GSBP)based on Gabor wavelet and local binary pattern(LBP).GSBP takes full account of the relationship between the neighborhoods from the spatial domain and the frequency domain.Compared with other related methods,GSBP extracts the strong discriminant features by refining the Gabor coefficients and reusing the LBP operator.In addition,the Gabor filter with small scale(such as 2 or 3)achievers high recognition rates while reduces the computational complexity.(5)Propose to the improved Color local binary pattern(ICLBP).A new sampling rule and a strategy for determining threshold are use to extract the color image feature.For this,the k-uniform pattern is proposed,which not only extends the classic uniform LBP pattern theoretically,but also improves the effects of color face recognition. |