RELEVANCE FEEDBACK BASED ON QUERY REFINING AND FEATURE DATABASE UPDATING IN CBIR SYSTEM

摘要

Relevance feedback (RF), which introduces human visual perception into the retrieval process gradually, is an efficient improvement for narrowing down the gap between low-level visual feature representation of an image and its semantic meaning in content-based image retrieval (CBIR). In this paper, a new relevance feedback approach based on query refining and feature database updating in CBIR system is proposed. To make the new query more representative for the query concept in each round of feedback iteration, a new query-refining scheme is put forth, which is based on the different contributions among positive samples to the formation of query concept. In similarity measure, a nonlinear exponential mapping for the coefficient corresponding to different feature component is adopted to reduce the bias, which is caused due to small number of user labeled samples. In addition, an updateable feature-database strategy is also proposed to gradually accumulate the useful semantic information from past rounds of feedback iteration for next round. We test the proposed algorithm on the Corel natural image database and the final experimental results show that the proposed approach greatly improves the retrieval performance.