Deep Uncoupled Discrete Hashing via Similarity Matrix Decomposition

Hashing has been drawing increasing attention in the task of large-scale image retrieval owing to its storage and computation efficiency, especially recent asymmetric deep hashing methods. These approaches treat query database an way can take full advantage whole training data. Though it achieved state-of-the-art performance, methods still suffer from large quantization error efficiency problem on datasets due tight coupling between database. In this article, we propose a novel method, called D eep U ncoupled iscrete H ashing (DUDH), for approximate nearest neighbor search. Instead directly preserving similarity database, DUDH first exploits small similarity-transfer set transfer underlying semantic structures implicitly keep desired similarity. As result, matrix is decomposed into two relatively ones decoupled Then both codes are learned during optimization. The only exists process set. By uncoupling cost optimizing CNN model no longer related size Besides, further accelerate process, optimize with constant-approximation solution. doing so, be almost ignored. Extensive experiments four widely used benchmarks demonstrate that achieve performance remarkable reduction (30× - 50× relative).