Dissimilarity Representation of Images for Relevance Feedback in Content-Based Image Retrieval

Relevance feedback mechanisms are adopted to refine image-based queries by asking users to mark the set of retrieved images as being relevant or not. In this paper, a relevance feedback technique based on the "dissimilarity representation" of images is proposed. Each image is represented by a vector whose components are the similarity values between the image itself and a "representation set" made up of the images retrieved so far. A relevance score is then assigned to each image according to its distances from the sets of relevant and non-relevant images. Three techniques to compute such relevance scores are described. Reported results on three image databases show that the proposed relevance feedback mechanism allows attaining large improvements in retrieval precision after each retrieval iteration. It also outperforms other techniques proposed in the literature.