A Two-Step Similarity Ranking Scheme for Image Retrieval
Di Wu, Jun Wu, Mingyu Lu, Chunli Wang · 2014
similarity ranking is one of the keys of a content-based image retrieval (CBIR) system. Among various methods, manifold ranking (MR) is popular for its application to relevance feedback in CBIR. Most existing MR methods only take the visual features into account in the similarity ranking, however, which is not accurate enough to reflect the intrinsic semantic structure of a given image database. In this paper, we propose a two-step similarity ranking scheme that aims to preserve both visual and semantic resemblance in the similarity ranking. Concretely, in the first step it derives an initial visual-based similarity rank through a self-tuning MR solution. In particular, the Gaussian kernel used in our scheme is refined by using a point-wise bandwidth. In the second step, the rank of each database image is further adjusted to achieve semantic consistency by mining the query log. An empirical study shows that using two-step similarity ranking in CBIR is beneficial, and the proposed scheme is more effective than some existing MR approaches.