F 3 KDH: Fine-Grained Feature Fusion Knowledge Distillation Hashing for Cross-Modal Retrieval

Le résumé fourni par la source

In recent years, deep cross-modal hashing retrieval methods have attracted extensive attention due to their strong ability to learn efficient binary encodings. However, they still face two main challenges: 1) insufficient fusion of cross-modal information, leading to low semantic quality of the generated hash codes. 2) limited scalability for deployment in lightweight frameworks. To address these issues, this paper proposes a novel cross-modal hashing method based on knowledge distillation, termed Fine-Grained Feature Fusion Knowledge Distillation Hashing (F3KDH). Specifically, the teacher model incorporates the AFM that thoroughly fuses heterogeneous information, effectively capturing salient semantic cues in cross-modal data to enhance the weighting of important regions. Moreover, we employ a cross-modal knowledge distillation strategy to transfer the rich semantic knowledge captured by the teacher to a deliberately lightweight student model, enabling it to learn compact yet expressive hash representations while greatly reducing computational complexity. Experimental results on multiple benchmark datasets demonstrate that, compared to existing cross-modal hashing methods, our approach achieves significant model compression without sacrificing retrieval accuracy.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.