Abstract: In the problem of image classification,the traditional sparse representation algorithms only consider the whole information of the sample,neglect the local structure information.Besides,those sparse representation algorithms require uniform alignment of the testing face samples and the training face samples,when the posture and angle change,the classification effect decreases obviously.Aimed at these problems,an improved L2 regularization sparse representation algorithm is proposed in this paper.Firstly,the sparse coefficient matrix is obtained by solving the least square method.Then,the residuals of the reconstructed samples are calculated and the sparsity between samples is quantified by the sparse metric formula.Finally,while keeping the sparsity of samples,the local discriminant information between sample classes and classes is increased,to improve the classification performance of samples.To verify the effectiveness of the proposed method,many experiments are conducted on human face datasets ORL,FERET,FEI and visual datasets Stanford 40 Actions and Caltech-UCSD Birds (CUB200-2011).The experimental results show that the proposed method is superior to the traditional sparse representation algorithm and other common classification algorithms on different face databases and extended visual data sets.