Reduct-based result set fusion for relevance feedback in CBIR

Relevance feedback (RF) is a widely used technique to deal with the issues of user subjectivity and the semantic gap in Content-Based Image Retrieval (CBIR). We build on existing work that outlined a rough set based general framework called CAFe for RF and proposed a re-weighting strategy based on a rough set theoretic analysis of the user feedback. This paper presents a method that uses the approximation of the information need distilled from the user classification as the busis for multiple distinct retrievals. The final result set that is presented as the subsequent iteration to the user is obtained by fusing the result sets from the different retrievals. The method is demonstrated in the context of a simple test image collection for clarity. An analysis of the sample iterations of feedback is presented. The method presented remains independent of the retriever, relies on a conceptually appealing model of the user feedback and serves to establish the utility of the gene ral framework.