Fast Deep Asymmetric Hashing for Image Retrieval

Abstract

Recently, by exploiting asymmetric learning mechanism, asymmetric hashing methods achieve superior performance in image retrieval. However, due to the discrete binary constraint, these methods typically rely on a special optimization strategy of discrete cyclic coordinate descent (DCC), which is time-consuming since it must learn the binary codes bit by bit. To address this problem, we propose a novel deep supervised hashing method called Fast Deep Asymmetric Hashing (FDAH), which learns the binary codes of training and query sets in an asymmetric way. FDAH designs a novel asymmetric hash learning framework using the inner product of the output of deep network and semantic label regression to approximate the similarity and minimize the discriminant reconstruction error between the deep representation and the binary codes. Instead of using the DCC optimization strategy, FDAH avoids using the quadratic term of binary variables and the binary code of all bits can be optimized simultaneously in one step. Moreover, by incorporating the semantic information in binary code learning and the quantization process, FDAH can obtain more discriminative and efficient binary codes. Extensive experiments on three well-known datasets show that the proposed FDAH can achieve state-of-the-art performance with less training time.