Supervised deep hashing with a joint deep network

Abstract

Hashing has gained great attention in large-scale image retrieval due to efficient storage and fast search. Recently, many deep hashing approaches have achieved good results since deep neural network owns powerful learning capability. However, these deep hashing approaches can perform deep features learning and binary-like codes learning synchronously, the information loss between binary-like codes and binary codes will increase due to the binarization operation. A further deficiency is that binary-like codes learning based on deep feature representations is a shallow learning procedure, which cannot fully exploit deep feature representations to generate hash codes. To solve the above problems, we propose a Deep Learning Supervised Hashing (DLSH) method which adopts deep structure to learn binary codes based on deep feature representations for large-scale image retrieval. Specifically, we integrate deep features learning module, deep mapping module and binary codes learning module in one unified architecture. The network is trained in an end-to-end way. In addition, a new objective function is designed to preserve the balancing property and semantic similarity of binary codes by incorporating the semantic similarity term and the balanceable property term. Experimental results on four benchmarks demonstrate that the proposed approach outperforms several state-of-the-art hashing methods.