Causal Inference via Style Bias Deconfounding for Domain Generalization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) face challenges with out-of-distribution data, limiting real-world visual application reliability.
• Existing domain generalization methods often neglect the impact of style frequency, leading to spurious correlations and reduced inference reliability.

Purpose of the Study:

• To introduce a novel causal inference-based framework, Style Deconfounding Causal Learning (SDCL), to enhance domain generalization in image modalities.
• To address style as a confounding factor, improving the learning of causal representations over spurious correlations.

Main Methods:

• Constructing a structural causal model (SCM) for domain generalization and applying backdoor adjustment for style influence.
• Designing a style-guided expert module (SGEM) for adaptive clustering of style distributions.
• Implementing a backdoor causal learning module (BDCL) for causal interventions during feature extraction.

Main Results:

• The SDCL framework effectively reduces style bias by ensuring fair integration of confounding styles.
• Experiments demonstrate superior performance in both multi-domain and single-domain generalization scenarios.
• The approach shows efficacy across diverse natural and medical image recognition tasks.

Conclusions:

• SDCL offers a versatile and effective solution for enhancing domain generalization in DNNs by explicitly handling style confounding factors.
• The causal inference-based approach improves model robustness and reliability when faced with domain shifts.