Experience-based relevance feedback for image retrieval

This paper presents a novel search technique, ´Experience-based Relevance Feedback´ (XRF), formed from the fusion of keyword-based image retrieval (KBIR) and content-based image retrieval (CBIR). The technique significantly reduces the semantic-visual content gap apparent in most existing image retrieval systems. Information about previous searches is recorded and analysed in order to improve the current performance. Semantic knowledge is incorporated in the retrieval process, however, unlike KBIR the actual annotation is totally transparent to the user. As with many CBIR systems, XRF utilises relevance feedback (RF) for supervised learning and is able to retrieve visually similar images. It learns about the user´s preferences during the search process and utilises past experience to enhance current and future retrieval operations.