In ROI-based image retrieval, salient point is an important point feature. The classic extraction algorithm of salient points-SPARSE(Salient Points Auto-Reduction using Segmentation) is more complex. This paper presents an improved algorithm, using dynamic threshold segmentation algorithm in the between-class variance and within-class variance of the image segmentation. And then it uses three color features and three texture features of the segmentation to point out the significant features of labeling. Finally it makes similarity measurement with Euclidean distance of feature vectors of salient points. Experimental results show that the improved algorithm’s extraction of salient points for image retrieval has better search results.