Unsupervised Contrastive Cross-Modal Hashing.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised cross-modal hashing (CMH) methods often suffer performance degradation due to binary optimization.
• False-negative pairs (FNPs) negatively impact the learning process in CMH by misclassifying similar items.

Purpose of the Study:

• To enhance unsupervised cross-modal hashing (CMH) by integrating contrastive learning (CL).
• To overcome performance limitations in CMH caused by binary optimization and false-negative pairs.

Main Methods:

• A novel momentum optimizer is proposed to make hashing operations learnable within contrastive learning, avoiding binary-continuous relaxation.
• A Cross-modal Ranking Learning (CRL) loss is introduced to mitigate the impact of FNPs by considering all discrimination rather than just hard negatives.

Main Results:

• The proposed method demonstrates improved retrieval performance compared to existing techniques.
• Experiments on five datasets show the effectiveness of the novel approach against 13 state-of-the-art methods.

Conclusions:

• The developed method represents a significant advancement in contrastive hashing.
• The approach offers a more robust and effective solution for unsupervised cross-modal hashing tasks.