The influence of the similarity measure to relevance feedback

In this paper we discuss the influence of the similarity measure in the context of content-based image retrieval. To bridge inherent descriptor gap, we propose a new relevance feedback (RF) approach which uses a hierarchical clustering strategy. It has the advantage of performing on the initial set of retrieved images, instead of performing additional queries, as most approaches do. The images are clustered with respect to the positive/negative examples provided by the user, in a continuous manner, as user successively browses through the retrieved images. Comparative experimental tests conducted using state-of-the-art content descriptors and distance measures show that the proposed approach provides significant improvement in retrieval performance while the choice of the distance metric plays a decisive role on system performance.