Abstract: Recently,sparse representation based classification(SRC) has attracted much attention in face recognition tasks.SRC forms the dictionary by directly using all the training samples.When giving lots of training samples,the speed of the subsequent sparse solver can be very slow.To alleviate this problem,a new local SRC,which is based on the similarities of sparse coefficients of both training samples and test samples,was presented.According to this similarity,a certain number of training samples are selected to form the over-complete dictionary,and then the test sample is decomposed using this dictionary.In contrast to original LSRC,which is based on kNN to choose neighbors of test samples,the proposed approach can steadily achieve better performance.Experimental results obtained on the ORL database,Yale database and AR database indicate that the proposed method is superior to both SRC and LSRC.

Key words: Sparse representation based classification(SRC),Local SRC(LSRC),Sparse coefficients,Similarity,Face recognition