Re-ranking in the context of CBIR: A comparative study

Abstract

There are many efforts put, in the last years, on the re-ranking mechanism, in the context of content based-image retrieval (CBIR), aiming to improve the results answered after the first search based on image features. In this paper, we address this scheme of re-ranking categorized here into two directions: reranking based on pseudo relevance feedback and re-ranking with relevance feedback information. Each of the cited categories contains four kinds of re-ranking: (i) re-ranking through refinement of the initial query, (ii) re-ranking through updating the weights of utilized signatures/similarities, (iii) re-ranking through classification (non-supervised or supervised) and (iv) re-ranking through rerating algorithms. The comparative study revealed that MVRA is the best in terms of pseudo relevance feedback and that Incremental KNN outperforms the considered methods for relevance feedback.