Face Recognition Based On Discriminant Sparsity Preserving Projection

Face recognition is a kind of mature biometric identification technology,which is based on the identification and discrimination of human faces.Face features are the unique and the most natural and friendly.In the field of face recognition,data is often encountered in the high-dimensional,generally leads to dimension disaster problem.In recent years,sparse representation(sparse representation,SR)in dealing with the face recognition shows certain effectiveness.It uses the linear combination of the least nonzero elements to achieve the correct representation of the data.However,this method does not contain discriminant information.Then sparse preserving projections(sparse preserving projections,SPP)algorithm which is keeping the data structure of sparse representation successfully applied to face recognition.It contains some discriminant information and can automatically select the nearest neighbor information,but it can not use the sample label information and the computational complexity is high.Therefore,in order to deal with above problems,the research of the dissertation is concentrated by the following aspects:Firstly,to overcome the SPP algorithm failed to effectively use the class label information problem,this paper reconstructs objective function of the SPP algorithm by using the maximum scatter difference criterion(Maximum scatter difference criterion,MSDC).Secondly,the paper proposed a new algorithm based on the improved the maximum scatter difference criterion and sparse preserving projections(DSPP).It can effectively reduce the complexity of the SPP algorithm and improve the efficiency of face recognition.Thirdly,an algorithm MSPCA(a method based on ModulePCA and discriminant sparsity preserving projection)for the single sample face images is proposed.Primarily,the method of ModulePCA is used to expand the number of samples.Then,the transformed images and the original images will be used as a new training set for sparse representation.In this dissertation,we have done sufficient experiments on ORL,Yale-B,CAS-PEAL and IMM Database,the results show that our algorithms are comparable to others.