Deep Supervised Hashing by Classification for Image Retrieval

Hashing has been widely used to approximate the nearest neighbor search for image retrieval due to its high computation efficiency and low storage requirement. With the development of deep learning, a series of deep supervised methods were proposed for end-to-end binary code learning. However, the similarity between each pair of images is simply defined by whether they belong to the same class or contain common objects, which ignores the heterogeneity within the class. Therefore, those existing methods have not fully addressed the problem and their results are far from satisfactory. Besides, it is difficult and impractical to apply those methods to large-scale datasets. In this paper, we propose a brand new perspective to look into the nature of deep supervised hashing and show that classification models can be directly utilized to generate hashing codes. We also provide a new deep hashing architecture called Deep Supervised Hashing by Classification (DSHC) which takes advantage of both inter-class and intra-class heterogeneity. Experiments on benchmark datasets show that our method outperforms the state-of-the-art supervised hashing methods on accuracy and efficiency.