The existing algorithms of content based image retrieval (CBIR) extract global features in the whole image to query, which have redundant calculation and will undoubtedly reduce the efficiency of the retrieval. In the light of this problem, an algorithm based on the combination of Harris-Laplace corners and support vector machine (SVM) relevance feedback is proposed in this paper. First, image corners are extracted by Harris-Laplace corner detector and the salient region is obtained by the density ratio in each distributed area of image corners. Then, color and shape in the salient region are fused for the initial retrieval. Finally, relevance feedback based on SVM classification is introduced into CBIR. The simulation results show that, the method proposed in this paper performs well in evaluation indexes of average precisions.