A Web image retrieval system using relevance feedback with image clustering (follow-up)

Web image retrieval is performed based on the terms appearing near images in HTML files. However, the relationship between the terms and the images are not so acculate, the precision tends to be low. In order to improve the performance, relevance feedback is one of the promising approach. The user’s burden, however, is the inherent defect of the approach. Moreover, the user must browse several pages in order to modify the query to improve performance, and so the burden to do it is another inherent problem. So, this research tries to lessen the burden by using relevance feedback based on the clusters of the retrieved images and to show the query terms extracted from the relevant Web pages referring to the images in the cluster. The initial prototype has been reported in [1]. This paper is a follow-up to the paper and describes improvements in the clustering method and the candicate selection mechanism from the selected positive cases, and explains the preliminary effectiveness study of the tool with different search engines.