No abstract

Image retrieval systems adopt a combination of multiple features and then total distance measures of particular features for ranking the results. Therefore, the top-ranked images with smallest total distance measures are returned to the users. However, images with smallest partial distance measures which are suitable for users’ purpose may not be included in these results. Therefore, partial distance measure should be considered. In this paper, we propose to adopt the Pareto set in the distance measure space. This set assures that the returned results contain not only points with smallest total distance obtained by linear combinations, but also other points have smallest partial distance measures which cannot be found by the linear combination in the distance measure space. Especially, the searching space based on the distance measures is compacted by our algorithm, namely PDFA. This algorithm collects all the Pareto set with different depths, and is efficient for the classification and regression tree (CART). The experimental results on three image collections show the effectiveness of our proposed method.

Relevance feedback as a bridge between high level semantic concepts and low features. It is  important to improve the performance of content based image retrieval (CBIR) is preprocessing image features and refining distance  measures for query based on user information needs. We propose a novel method to normilize features and distance for CBIR using combination features. In addition, we also use relevant feedback from users and learning from low features to update weights distance measures and refine query. Experimental results over the benchmark Corel dataset demonstrate the effectiveness of this propose method.