Hashing with Cauchy Graph

Abstract

Approximate nearest neighbor search within large scale image datasets strongly demands efficient and effective algorithms. One promising strategy is to compute compact bits string via the hashing scheme as representation of data examples, which can dramatically reduce query time and storage requirements. In this paper, we propose a novel Cauchy graph-based hashing algorithm for the first time, which can capture more local topology semantics than Laplacian embedding. In particular, greater similarities are achieved through Cauchy embedding mapped from the pairs of smaller distance over the original data space. Then regularized kernel least-squares, with its closed form solution, is applied to efficiently learn hash functions. The experimental evaluations over several noted image retrieval benchmarks, MNIST, CIFAR-10 and USPS, demonstrate that performance of the proposed hashing algorithm is quite comparable with the state-of-the-art hashing techniques in searching semantic similar neighbors, especially in quite short length hash codes, such as those of only 4, 6, and 8 bits.