Learning latent hash codes with discriminative structure preserving for cross-modal retrieval

Due to the low storage cost and computational efficiency, hashing approaches have drawn considerable interest and gained great success in multimodal retrieval. However, most existing works study the local geometric structure in the original space, which suffers from intra- and inter-modality ambiguity, resulting in low discriminative hash codes. To address this issue, we propose a novel cross-modal hashing approach by taking inter- and intra-modality structure preserving into consideration, dubbed discriminative structure preserving hashing (DSPH). Specifically, DSPH explores the intra- and inter-modality in the latent structure of the constructed common space. In addition, the local geometric consistency is improved by a supervised shrinking scheme. DSPH learns the hash codes and latent features based on factorization coding scheme. The objective function includes common latent subspace learning and inter- & intra-modality structure embedding. We devise an alternative optimization scheme, where the hash codes are solved by a bitwise scheme, and the large quantization error can be avoided. Owing to the merit of DSPH, more discriminative hash codes can be generated. The extensive experimental results on several widely used databases demonstrate that the proposed algorithm outperforms several state-of-art cross-media retrieval methods.

Automated summary

A novel cross-modal hashing approach, DSPH, is proposed in this study, which generates more discriminative hash codes by considering intra- and inter-modality structure preserving, as well as improving local geometric consistency. Extensive experimental results demonstrate that the proposed algorithm outperforms several state-of-art cross-media retrieval methods.