An Approach of Content-Based Image Retrieval Based on Image Salient Region

The traditional CBIR (Content based image retrieval) global descriptor can't extract local features effectively, while the SIFT (Scale Invariant Feature Transform) algorithm only focus on local object features matching, and can not satisfy the need of fuzzy CBIR queries based on visual feeling. In view of this contradiction, an approach is proposed based on image salient region, in which the SIFT algorithm and the traditional CBIR color descriptors are combined, the entropy of salient point is used to distinguish the salient and non-salient regions of image. We use centroid weighted histogram and dominant color descriptor of MPEG-7 to describe the salient and non-salient region respectively, which not only effectively extract the local features of image objects, but also overcome the SIFT algorithm for high dimension, and vulnerable to interference with the global features. Experiments show that for general image retrieval application, the proposed method has better retrieval performance, with better robustness for geometric and perspective transform.