Information preservation-based hashing for image retrieval

Abstract

With the great advance of information techniques, we are witnessing the tremendous growth of data generation. How to retrieve desirable objects quickly and effectively from such massive data is still a challenging issue. Hash learning provides a feasible solution for fast retrieval, due to its potential advantages. In this paper, we develop a novel hashing model based on an information-preserving metric for image retrieval. First, we theoretically reveal the relationships of the Hamming distances between the binary representations to the angles between their corresponding original data in the Euclidean space. We further demonstrate that the Hamming distances are unbiased estimations of the angles, especially when the length of binary representations is long enough. Based on this property, we introduce an information-preserving metric and take it into account within our hashing model, so that the information embedded within the original data can be preserved as much as possible during the encoding stage of hash learning. To verify the superiority of our hashing model, we conducted comprehensive experiments on three public image collections. The experimental results show that the proposed hashing algorithm can achieve competitive performance compared to the representative hashing algorithms.