A Deep Supervised Hashing Algorithm for Image Retrieval

Traditional image retrieval algorithms, which are based on hand-crafted feature, are greatly influenced by the designed features and are not suitable for retrieving images with the same semantics. To solve this problem, an end-to-end deep learning network is designed to learn features automatically and to retrieve images with same semantics. A union loss function is designed to maximize the likelihood function and the discrete constraints, which result in hard optimization, are relaxed to be continuous by regularization. The experiments on CfF AR10 dataset showed that the proposed algorithm achieved high mean average precision and recall compared with several state-of-the-art algorithms.