Central similarity joint-learning for cross-domain retrieval

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Cross-domain retrieval is crucial for image retrieval, but current methods struggle with domain and class alignment.
• Existing approaches often neglect semantic information in hash code learning, compromising cross-domain semantic structure preservation.

Purpose of the Study:

• To propose an effective cross-domain retrieval method, Central Similarity Joint-Learning (CSJL), addressing limitations in domain and class alignment.
• To enhance the incorporation of semantic information for robust cross-domain image retrieval.

Main Methods:

• CSJL employs a similarity joint-learning strategy, treating intra-domain and inter-domain relationships distinctly.
• It utilizes class prototypes for semantic-based sample gathering across domains, ensuring accurate class alignment via orthogonal category prototypes.
• Hash code generation integrates multiple semantic sources: feature representations, label supervision, and pairwise similarities.

Main Results:

• CSJL demonstrates effective domain alignment by preserving intra-domain similarity structures and enhancing inter-domain semantic consistency.
• Accurate class alignment is achieved irrespective of domain differences through orthogonal category prototypes.
• The method generates discriminative hash codes that effectively preserve cross-domain semantic structures.

Conclusions:

• CSJL achieves state-of-the-art performance on multiple cross-domain retrieval tasks.
• The proposed method offers a significant advancement in cross-domain image retrieval by effectively integrating domain, class, and semantic information.