Long-term learning of semantic grouping from relevance-feedback

Abstract

Relevance-Feedback is a powerful paradigm for incorporating semantic information in content-based image retrieval. Various relevance-feedback methods have been proposed. They were evaluated according to their success-rate in individual image retrieval sessions, where each session is considered independently of other sessions. In this paper we propose a method for accumulating semantic grouping information from multiple relevance-feedback sessions. We show that such information enables gradual improvements in image retrieval, enabling the current session to benefit from knowledge acquire in previous sessions