Fighting the Semantic Gap on CBIR Systems through New Relevance Feedback Techniques

Abstract

This paper introduces two novel relevance feedback techniques that integrate a new way to implement the query center movement with a suitable weighting on the similarity function. These techniques integrated to a content-based image retrieval (CBIR) system, improves the precision of the results when using texture features up to 42%, and employing at most 5 iterations. Thus, the user satisfaction with the system is increased as our experiments demonstrated. Besides being effective, the new RF techniques are very fast as they take less than one second to reprocess the queries at each iteration. The experiments also show that with three iterations the users are satisfied with the query results, and the major gain in precision happens in the first iteration, achieving improvements of up to 30%, what lessens the user efforts and anxiety.