Improving domain generalization by hybrid domain attention and localized maximum sensitivity

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Domain generalization (DG) aims to train models on source domains for effective performance on unseen target domains.
• Current DG methods often process all visual features equally, leading to suboptimal generalization.
• Human generalization relies on focusing on salient features and ignoring irrelevant ones.

Purpose of the Study:

• To develop a domain generalization method that mimics human-like feature selection.
• To improve model generalization by focusing on label-relevant features.
• To enhance model robustness against input perturbations for better generalization.

Main Methods:

• A channel-wise and spatial-wise hybrid domain attention mechanism was proposed.
• This mechanism forces models to prioritize important, label-associated features.
• Localized maximum sensitivity to input perturbations was reduced to improve robustness.

Main Results:

• The proposed method demonstrated improved generalization performance.
• Experiments on PACS, VLCS, and Office-Home datasets validated the approach.
• The hybrid attention mechanism effectively focused on discriminative features.

Conclusions:

• The proposed hybrid domain attention mechanism significantly enhances domain generalization.
• Reducing sensitivity to input perturbations improves network robustness and generalization capability.
• The method offers a promising direction for developing more effective domain generalization techniques.