Asymmetric Supervised Deep Discrete Hashing Based Image Retrieval

Hashing has been widely used for image retrieval tasks. In view of the limitations of existing deep supervised hashing methods, a new asymmetric supervised deep discrete hashing method is proposed to maintain the semantic structure between different categories and generate binary codes. Firstly, a deep network is used to extract image features and reveal the similarity between each pair of images according to their semantic labels. To enhance the similarity between binary codes and ensure the retention of multi-label semantics, this paper design an asymmetric hashing method that utilizes a multi-label binary code mapping to make the hash codes have multi-label semantic information. In addition, the bit balance of the binary code is introduced to balance each bit, which encourages the number of -1 and +1 to be approximately similar among all training samples. Experimental results on two benchmark datasets show that the proposed method is superior to other methods in image retrieval.