Research On Face Recognition Algorithm Based On Kernel Subspace Low-rank Representation

With the great digital development of human society in the last century,Face Recognition(FR)as the representative of Pattern Recognition(PR)and Machine Learning(ML),has widely received researchers' attention,showing its huge practical value.Sparse and Low-rank Representation as the representative of reconstructive representation algorithms,by virtue of its excellent performance and robustness to external noise,show their superiority and attract researchers' attention among many newly-proposed face recognition methods.The purpose of this thesis is to focus on improving the ability of fusing local and global features of LRRC and providing a more powerful nonlinear feature learning algorithm for linear LRRC.First of all,in chapter one,the introduction is presented including research background,significance and the domestic and foreign research status of face recognition technology.Then,it is introduced that the problems and challenges of face recognition in the current stage.At the end of this chapter,the content arrangement of this thesis is shown.In the second chapter,a face recognition algorithm based on Kernel Locality Preserving Low-rank Representation(KLP-LRR)is proposed.This method can be regarded as a nonlinear extended version of LRRC.In order to improve the LRRC's fusion ability of global and local feature information,with the introduction of Tikhonov regularization constraint,KLP-LRR is able to explore and learn the local flow structure information hidden in the face samples.In the third chapter,a face recognition algorithm named Low-rank representation based on Twin Tensor Kernel Dictionary Learning(LRR-TTK)is created.LRR-TTK is based on the theory of tensor and firstly introduces tensor theory into reconstructive representation methods.Through the use of twin tensor kernel,LRR-TTK can extract and reconstruct the high discriminative facial features image at the same time.In order to improve the uniformity of the proposed algorithm,we further propose a Twin Tensor Kernel Locality Preserving Projection(TT-LPP)algorithm to make a seamless connection with our proposed algorithm.Then,an algorithm named Kernel Low-rank Embedded Dictionary Learning for face recognition(KLEDL)is presented in the fourth chapter.In view of the classification ability of dictionary learning depending on global and local feature and structure in the original data space,the proposed KLEDL can construct a high discriminative dictionary to improve the ability of LRRC,which purpose is to seek the lowest rank solution in the original sample space.Finally,KLEDL can get a better recognition accuracy rate.At the end chapter of this thesis,all the proposed algorithms are summarized.Besides,a prospection of face recognition technology is made with the limited knowledge of the author.