Research on Discrete Hash Algorithm Based on Deep Semantics

Abstract

With the rapid growth of image and video data on the network, hash technology has been widely studied in the field of image and video search in recent years. Benefiting from the latest progress in deep learning, deep hash method has achieved good results in image retrieval. However, the previous deep hash method has the limitation that the semantic information is not fully utilized. In this paper, we develop a discrete hash algorithm based on deep supervision, assuming that learning binary code should be an ideal choice of classification. The pair tag information and classified information are used to learn hash codes within a framework. The output of the last layer is restricted to binary code directly, which is rarely studied in deep hash algorithm. Due to the discrete properties of hash codes, the alternate minimization method is used to optimize the target function. The proposed algorithm is proved to be better than the other supervised hash methods in two public image retrieval databases CIFAR-10 and NUS-WIDE.