Nearest-prototype relevance feedback for content based image retrieval

Effective retrieval of images from databases can be attained by adopting relevance feedback mechanisms. The vast majority of such mechanisms that have been proposed so far are based on modifying either the query point, or the feature space, or the similarity measure, so that the average similarity between pairs of relevant images is as minimum as possible. In this paper, a relevance feedback technique based on the nearest-neighbor rule is proposed. For each image of the database, a relevance score is computed as the ratio between the distances from the nearest non-relevant and relevant images respectively. Relevance is thus related to "local" properties of the images rather than to the estimation of global properties. Reported results on the Corel dataset show that the proposed mechanism allows attaining large improvements in retrieval precision compared to other mechanisms.