Chu, Chaoqun and Gong, Dahan and Chen, Kai and Guo, Yuchen and Han, Jungong and Ding, Guiguang (2019) Optimized Projection for Hashing. Pattern Recognition Letters, 117. pp. 169-178. ISSN 0167-8655
1_s2.0_S0167865518301478_main.pdf - Accepted Version
Available under License Creative Commons Attribution-NonCommercial-NoDerivs.
Download (1MB)
Abstract
Hashing, which seeks for binary codes to represent data, has drawn increasing research interest in recent years. Most existing Hashing methods follow a projection-quantization framework which first projects high-dimensional data into compact low-dimensional space and then quantifies the compact data into binary codes. The projection step plays a key role in Hashing and academia has paid considerable attention to it. Previous works have proven that a good projection should simultaneously 1) preserve important information in original data, and 2) lead to compact representation with low quantization error. However, they adopted a greedy two-step strategy to consider the above two properties separately. In this paper, we empirically show that such a two-step strategy will result in a sub-optimal solution because the optimal solution to 1) limits the feasible set for the solution to 2). We put forward a novel projection learning method for Hashing, dubbed Optimized Projection (OPH). Specifically, we propose to learn the projection in a unified formulation which can find a good trade-off such that the overall performance can be optimized. A general framework is given such that OPH can be incorporated with different Hashing methods for different situations. We also introduce an effective gradient-based optimization algorithm for OPH. We carried out extensive experiments for Hashing-based Approximate Nearest Neighbor search and Content-based Data Retrieval on six benchmark datasets. The results show that OPH significantly outperforms several state-of-the-art related Hashing methods.