Efficient discrete latent semantic hashing for scalable cross-modal retrieval

Hashing has been widely exploited in the cross-modal retrieval applications in recent years for its low storage cost and high retrieval efficiency. However, most existing cross-modal hashing methods either fail to capture the discriminative semantics of multi-modal data or suffer from the relatively high training cost. To address these limitations, we propose an efficient Discrete Latent Semantic Hashing (DLSH) method. DLSH first learns the latent semantic representations of different modalities, and then projects them into a shared Hamming space to support the scalable cross-modal retrieval. Because DLSH directly correlates the explicit semantic labels with binary codes, it can enhance the discriminative capability of the learned hashing codes. Furthermore, to obtain binary codes, traditional methods often relax the discrete constraint, resulting in relatively high computation cost as well as quantization loss. In contrast, DLSH directly learns the binary codes with an efficient discrete hash optimization, and thus increases efficiency and reduces the quantization loss in hash optimization. Extensive experiments on several public datasets show that, DLSH outperforms several state-of-the-art cross-modal hashing methods.