Similarity and Ranking Preserving Deep Hashing for image Retrieval

Hash codes based on deep learning can effectively learn image features. For supervised deep learning methods, the label information of the image can be used to further learn the semantic information of the image. However, the current supervised deep learning methods often use 1 and 0 (or -1) to represent the similarity of two images. In fact, these two extreme values do not fully reflect the similarity between images. Thus, we proposed a novel similarity and ranking preserving deep hashing method (SRPDH). In order to enrich and more comprehensively reflect the semantic information between images, we refine the single-label information into multi-label information, and use Jaccard coefficient model to calculate the similarity between label information. In the loss function model, we use the cross entropy model and consider the loss caused by the binary quantization of the network output. The experimental results show that our method can further improve the mean average precision (MAP) of image retrieval compared with the existing methods.