A novel image representation and learning method using SVM for region-based image retrieval

Support vector machines (SVM) is gaining a considerable attention as an approach to improvement performance of the content-based image retrieval (CBIR). Most SVM for CBIR rely on global feature, which length of the feature representation is fixed. However, region-based image retrieval (RBIR) use variable length representation, and common kernel utilize the inner product or l_p norm in input space, they are infeasible for RBIR. In this paper, we present a SVM-Based relevance feedback techniques for region-based image retrieval including: (1) introducing an image segmentation algorithm and devising a compact and computational effective representation for the color content of a region of an image; (2) using earth mover's distance and hybrid feature including color, texture and shape as feature vector to match image; (3) developing a generalized SVM as a learning machines kernel for region-based image retrieval. Experimental results on a database of 1000 real images demonstrate the efficacy and robustness of the proposed method.