Domain generalization for image classification based on simplified self ensemble learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Domain generalization aims to apply models to new, unseen data domains.
• Existing methods struggle with complex, evolving cross-domain discrepancies.
• High data complexity hinders effective knowledge transfer in current approaches.

Purpose of the Study:

• To develop a method that enhances model adaptability for improved performance in unseen domains.
• To address the limitations of existing domain generalization techniques in complex scenarios.
• To improve the reliability and safety of AI systems in diverse real-world applications.

Main Methods:

• Framed domain generalization as an optimization problem balancing domain discrepancies and sample complexity.
• Proposed a self-ensemble learning framework with a single feature extractor and multiple classifiers.
• Incorporated focal loss and complex sample loss weighting for hard-to-learn instances.
• Utilized a dynamic loss adaptive weighted voting strategy for robust predictions.

Main Results:

• Achieved up to 3.38% improvement in generalization performance over existing methods.
• Demonstrated effectiveness on benchmark datasets (OfficeHome, PACS, VLCS).
• Showcased practical utility in complex domains like autonomous driving and medical imaging.

Conclusions:

• The proposed approach effectively enhances cross-domain generalization beyond simply minimizing discrepancies.
• The self-ensemble learning framework with adaptive weighting improves handling of complex samples and diverse domains.
• This method offers a more reliable and robust solution for real-world AI applications requiring generalization.