Multi-feature Fusion-Based Central Similarity Deep Supervised Hashing

The deep image hashing aims to map the input image into simply binary hash codes via deep neural networks. Nevertheless, previous deep supervised hashing methods merely focus on the high-level features of the image and neglect the low-level features of the image. Low-level features usually contain more detailed information. Therefore, we propose a multi-feature fusion-based central similarity deep supervised hashing method. Specifically, a cross-layer fusion module is designed to effectively fuse image features of high and low levels. On top of that, a channel attention module is introduced to filter out the useless information in the fused features. We perform comprehensive experiments on three widely-studied datasets: NUS-WIDE, MS-COCO and ImageNet. Experimental results indicate that our proposed method has superior performance compared to state-of-the-art deep supervised hashing methods.