Supervised discrete cross-modal hashing with exploiting semantic correlations

Hashing has attracted considerable attention for large-scale cross-modal retrieval due to its excellent performance in terms of retrieval speed and storage cost. Although numerous supervised hashing approaches have been developed to deal with cross-modal tasks, it is still difficult to fully excavate potential correlation information among multimedia data. Moreover, most existing methods directly project the multimodal data into Hamming space using a single transformation matrix, which often leads to the loss of valuable information during the transition from continuous to discrete representations. To this end, this article presents a novel hierarchical transformation and semantic correlation hashing (HTSCH) model. Specifically, we introduce a label correlation strategy to generate more discriminative binary codes, and design a hierarchical transformation scheme to alleviate the effect of information loss. Furthermore, we leverage the supervised semantic information and multimodal data to construct the fusion similarity matrix, thereby preserving richer cross-modal relationships. Finally, the proposed HTSCH can efficiently solve the discrete constraints of hash codes by the given discrete optimization iterative algorithm. Experiments on three large-scale datasets show the efficacy of our HTSCH.