Subdomain adaptation method based on transferable semantic alignment and class correlation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep unsupervised domain adaptation (UDA) faces challenges due to domain shift.
• Existing methods often struggle with precise cross-domain semantic alignment.

Purpose of the Study:

• To propose a subdomain adaptation framework for UDA.
• To enhance cross-domain semantic alignment and classification accuracy.

Main Methods:

• A joint subdomain distribution alignment mechanism to reduce intra-class divergence and enlarge inter-class disparities.
• A domain-adaptive semantic consistency loss for clustering similar samples and separating dissimilar ones.
• Temperature-based label smoothing and a class correlation matrix to improve pseudo-label quality and exploit inter-class relationships.

Main Results:

• The proposed method achieves superior average classification accuracy on multiple public datasets compared to existing approaches.
• Demonstrates the effectiveness of semantic alignment and class correlation modeling in mitigating domain shift.

Conclusions:

• The framework effectively mitigates domain shift by modeling intra-class coherence and inter-class distinction.
• Enhances semantic alignment and recognition performance in the target domain without additional architectural complexity.
• Offers a robust solution for deep unsupervised domain adaptation.