Domain adaptation with category-level contrastive learning for semi-supervised cross-modal hashing
Cross-modal hashing(CMH) is a key technique in information retrieval, valued for its efficiency, low dimensionality, and minimal storage requirements. Despite notable progress in this field, challenges persist, particularly the reliance on large labeled datasets. This paper presents a novel domain a...
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2025
|
| Online Access: | http://psasir.upm.edu.my/id/eprint/120694/ http://psasir.upm.edu.my/id/eprint/120694/1/120694.pdf |
| Summary: | Cross-modal hashing(CMH) is a key technique in information retrieval, valued for its efficiency, low dimensionality, and minimal storage requirements. Despite notable progress in this field, challenges persist, particularly the reliance on large labeled datasets. This paper presents a novel domain adaptation framework that leverages a limited set of labeled data from the source domain to guide the training of a large quantity of unlabeled data in the target domain. Our approach incorporates pseudo-label generation to iteratively refine semantic representations in the target domain, progressively narrowing the semantic gap between domains. Additionally, we propose a category-level contrastive learning(CLCL) method to address class conflict issues common in traditional instance-based contrastive learning. By generating category prototype representations, we enhance the model’s ability to discriminate between categories effectively. Moreover, our framework includes a comprehensive optimization objective that integrates pseudo-label generation loss, contrastive learning loss, and hash code learning loss, ensuring that the generated hash codes are both discrete and discriminative. Experimental results on benchmark datasets demonstrate the superiority of our approach over existing CMH methods. |
|---|