Enhanced Cross-Modal Hashing via Hybrid Distillation and Structural Refinement.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Cross-modal hashing is vital for managing large multimedia data due to low storage and computation needs.
• Unsupervised hashing is limited by the lack of accurate supervisory data.
• Supervised hashing requires extensive, costly data annotation.

Purpose of the Study:

• To propose a novel semi-supervised cross-modal hashing method (HDSR) addressing data annotation limitations.
• To enhance the accuracy and efficiency of hash code generation for cross-modal retrieval.
• To leverage partially labeled data for improved hashing performance.

Main Methods:

• HDSR learns inter-modal and inter-instance similarities using pointwise semantic alignment and listwise similarity partial order learning.
• It constructs higher-order affinity matrices by fusing inter-modal similarity for stable self-supervised training.
• Hybrid distillation transfers refined structural representations from labeled to unlabeled data branches.

Main Results:

• HDSR effectively extracts refined structural representations from partially labeled data.
• Momentum fusion strategies facilitate stable self-supervised training on unlabeled data.
• Experimental results on benchmark databases demonstrate HDSR's efficiency and superiority over state-of-the-art methods.

Conclusions:

• HDSR significantly enhances cross-modal hash learning performance.
• The method generates compact and accurate hash codes, outperforming existing deep cross-modal hashing techniques.
• HDSR offers an effective solution for semi-supervised cross-modal hashing challenges.