Face Recognition Via Sparse Representation And Linear Regression

Face recognition technology is widely used in the fields of public safety and finance and information security since the 1960 s because of its confidentiality,security,friendliness and convenience.At present,in the strict control of light intensity and angle and other variable conditions,as well as the degree of cooperation with the user,the system achieved excellent recognition result.However,in the practical application,the face image presents a rich kind of intra-class diversity and high inter-class similarity.Face recognition system is facing severe challenge.Based on sparse representation method and linear regression method,we carried out a series of research work,for the problems of illumination,facial expression variances and occlusion in face recognition.The main contents are as follows:(1)There is a lot of nonlinear structure information caused by the illumination and expression variances in human face,which the linear method cannot form effective mapping and identification.In addition,it is necessary to reduce the dimension of the sample data in the high-dimensional kernel space,to improve the efficiency of classification.Consequently,we proposed a Kernel Linear Collaborative Discriminant Regression Classification(K-LCDRC)algorithm,to solve the problems mentioned above.Firstly,the samples are mapped into a high-dimensional kernel space by a nonlinear mapping,and the reconstruction errors and intra-class reconstruction errors based on cooperative representation are constructed.By maximizing the difference between between-class reconstruction error and within-class reconstruction error we solved the optimal projection matrix.Then,the obtained projection matrix is used to reduce the dimension of data in the kernel feature space.Finally,the regression coefficients are solved by the Least-Squares Estimation(LES),and the test samples are classified into the categories with the least reconstruction errors.K-LCDRC uses the Kernel Maximum Margin Criterion to construct the objective function,which can effectively avoid the performance degradation of the algorithm due to the small sample problem.(2)A practical solution is proposed for the large area of face occlusion in the real scene,named Structed Sparse Face Recognition based on Markov Random Fields(MRF-SSRC).Firstly,we use the Markov random field to accurately and effectively estimate the occlusion area in the occlusion face image.By removing the information of the non-occlusion area,we obtain the occlusion samples and use them to build a occlusion dictionary.Secondly,a structured sparse occlusion face recognition model is established,which combines the occlusion dictionary with the face dictionary to form a compound dictionary,introduces the structured sparse constraint,and accurately models the occlusion problem.The physical meaning of structured sparse constraints on such methods is clear.By solving the structured sparse representation coefficients,the face identity and occlusion categories can be identified separately.(3)According to the research content of this paper,we realized a prototype system of face recognition.The methods proposed in this paper are verified.