Semantic image clustering using relevance feedback

This paper describes an image clustering approach to grouping semantically similar images. In this approach, the similarity between images is estimated using users' relevance feedback information recorded in the user log of an image retrieval system. An algorithm similar to CAST (Cluster Affinity Search Technique) is used to identify clusters of semantically related images. It is a two-stage clustering method: the pre-classification partitions the images into closely related groups; within each group, the fine clustering mines semantically related clusters of images. Experiments on more than 10,000 images demonstrate the effectiveness of this approach.