TY - JOUR

T1 - Optimized Projection for Hashing

AU - Chu, Chaoqun

AU - Gong, Dahan

AU - Chen, Kai

AU - Guo, Yuchen

AU - Han, Jungong

AU - Ding, Guiguang

N1 - This is the author’s version of a work that was accepted for publication in Pattern Recognition Letters. Changes resulting from the publishing process, such as peer review, editing, corrections, structural formatting, and other quality control mechanisms may not be reflected in this document. Changes may have been made to this work since it was submitted for publication. A definitive version was subsequently published in Pattern Recognition Letters, 117, 2019 https://www.sciencedirect.com/journal/pattern-recognition-letters

PY - 2019/1

Y1 - 2019/1

N2 - 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.

AB - 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.

KW - Hashing

KW - Quantization

KW - Algorithm

U2 - 10.1016/j.patrec.2018.04.027

DO - 10.1016/j.patrec.2018.04.027

M3 - Journal article

VL - 117

SP - 169

EP - 178

JO - Pattern Recognition Letters

JF - Pattern Recognition Letters

SN - 0167-8655

ER -