A Unified Deep-Domain Adaptation Framework: Advancing Feature Separability and Local Alignment.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Domain shift, a key challenge in domain adaptation, arises from significant data distribution discrepancies between source and target domains.
• Existing domain alignment methods risk altering intrinsic data properties.
• Effective domain adaptation is crucial for improving model performance on unseen data distributions.

Purpose of the Study:

• To introduce a novel unified deep-domain adaptation framework (DDASLA) to address domain shift.
• To enhance feature extraction and alignment capabilities in deep learning models for domain adaptation.
• To improve model generalization and robustness across different data domains.

Main Methods:

• Incorporation of an attention mechanism, specifically self-attention, into the ResNet18 architecture for improved feature extraction.
• Utilizing a combined loss function comprising angular loss for feature discrimination, Local Maximum Mean Discrepancy (LMMD) for local distribution alignment, and entropy minimization for decision boundary refinement.
• Development of a unified deep-domain adaptation framework (DDASLA).

Main Results:

• DDASLA demonstrated superior performance compared to several state-of-the-art methods on the Office and remote sensing datasets.
• The proposed method effectively improved feature separability and local alignment between domains.
• Experimental results validate the enhanced feature extraction capability of the attention-augmented ResNet18 model.

Conclusions:

• DDASLA offers an effective solution for mitigating domain shift in transfer learning.
• The framework significantly improves model generalization and robustness across diverse domains.
• The findings provide a foundation for future research in deep domain adaptation.