A scalable re-ranking method for content-based image retrieval

Content-based Image Retrieval (CBIR) systems consider only a pairwise analysis, i.e., they measure the similarity between pairs of images, ignoring the rich information encoded in the relations among several images. However, the user perception usually considers the query specification and responses in a given context. In this scenario, re-ranking methods have been proposed to exploit the contextual information and, hence, improve the effectiveness of CBIR systems. Besides the effectiveness, the usefulness of those systems in real-world applications also depends on the efficiency and scalability of the retrieval process, imposing a great challenge to the re-ranking approaches, once they usually require the computation of distances among all the images of a given collection. In this paper, we present a novel approach for the re-ranking problem. It relies on the similarity of top-k lists produced by efficient indexing structures, instead of using distance information from the entire collection. Extensive experiments were conducted on a large image collection, using several indexing structures. Results from a rigorous experimental protocol show that the proposed method can obtain significant effectiveness gains (up to 12.19% better) and, at the same time, improve considerably the efficiency (up to 73.11% faster). In addition, our technique scales up very well, which makes it suitable for large collections.