Authors: Manjeet Rege and Rajeev Agrawal
Relevance feedback has been long used as a tool for improving the performance of image retrieval. We present a Bayesian framework that captures and synthesizes the user feedback at an object or image region level. By exploiting the statistical structure of images, our system is able to discover the object of user interest. First, all the images in the database are segmented and the image regions are clustered into different region clusters. Next, for every region in the query image, we find the representative cluster that has the highest posterior probability, given the image region. As feedback is received, the cluster priors change, leading to different clusters competing for a image region. We have integrated our region based feedback mechanism into a image retrieval system.Preliminary experiments performed on general purposeimages demonstrates the promise of the proposed framework.