Collaborative Learning for Extremely Low Bit Asymmetric Hashing

Hashing techniques are in great demand for a wide range of real-world applications such as image retrieval and network compression. Nevertheless, existing approaches could hardly guarantee satisfactory performance with the extremely low-bit (e.g., 4-bit) hash codes due to severe information loss shrink discrete solution space. In this article, we propose novel Collaborative Learning strategy that is tailored generating high-quality codes. The core idea jointly distill bit-specific informative representations group pre-defined code lengths. learning short among can benefit from manifold shared other long codes, where multiple views different provide supplementary guidance regularization, making convergence faster more stable. To achieve that, an asymmetric hashing framework two variants multi-head embedding structures derived, termed Multi-head Asymmetric (MAH), leading efficiency training querying. Extensive experiments on three benchmark datasets have been conducted verify superiority proposed MAH, shown 8-bit generated by MAH 94.3 percent MAP 1 1. Mean Average Precision (MAP) score CIFAR-10 dataset, which significantly surpasses 48-bit state-of-the-arts tasks.