Supervised Hashing Based on the Dimensions’ Value Cardinalities of Image Descriptors
Dimitrios Rafailidis · IEEE Signal Processing Letters · 2016
This letter presents a hashing method, in which the key idea is to exploit the discriminative power of image descriptors' dimensions' value cardinalities (DVC), that is, the number of distinct values that occur at the dimensions of the descriptors. DVC are an inherit characteristic of image descriptors, capable of boosting the search accuracy for approximate nearest neighbor search. However, previous DVC-based search strategies function in an unsupervised manner. To account for the fact that the semantic information of images can significantly leverage the search accuracy, this letter proposes an efficient supervised hashing strategy based on DVC. Given a set of training data, the proposed approach first calculates a consensus sparse matrix to consider both the DVC-based similarities and the sparse semantic information of images. Then, it formulates an objective function as a joint minimization problem, to jointly compute 1) the binary codes of the training data and 2) the projection matrix to map external queries to the Hamming space. The joint problem is solved via an efficient alternating optimization algorithm. Experiments on a benchmark dataset demonstrate the superiority of the proposed approach over other state-of-the-art supervised hashing and DVC-based search strategies.