Central similarity joint-learning for cross-domain retrieval.

Abstract

Cross-domain retrieval holds significant research value in the field of image retrieval. However, existing cross-domain retrieval methods have the following limitations: 1) excessive pursuit of domain alignment and neglect of class alignment; 2) inadequate incorporation of semantic information in hash code learning, which hinders the preservation of cross-domain semantic structures. To solve the above problems, we propose an effective cross-domain retrieval method named Central Similarity Joint-Learning (CSJL). CSJL achieves effective domain alignment by using the proposed three key components. Firstly, it employs a similarity joint-learning strategy that handles intra-domain and inter-domain relationships separately. This allows CSJL to preserve reliable similarity structures within each domain while enhancing semantic consistency across domains. Secondly, it leverages class prototypes to gather samples from different domains based on their semantic similarities. By constructing orthogonal category prototypes, CSJL enables accurate class alignment regardless of domains. Finally, it incorporates multiple sources of semantic information into hash code generation, including feature representations, label supervision, and pairwise similarity relationships. This comprehensive semantic guidance helps generate discriminative hash codes that effectively preserve cross-domain semantic structure. Experimental results verify that CSJL achieves the state-of-the-art performance on multiple cross-domain retrieval tasks.