SRECA: Style Remodeling and Essence Content Augmentation for Domain Generalization

Deep learning models have demonstrated a significant success across various visual tasks. However, their performance often significantly deteriorates when applied to unseen data distributions due to domain shift. In this paper, we investigate the combined aspects of style and content in the holistic visual representation of images, introducing a new method called Style Remodeling and Essence Content Augmentation (SRECA) for domain generalization. SRECA consists primarily of two components: Novel Style Learning (NSL) and Geometric Skeleton Generation (GSG). NSL identifies the most salient directions of style transition, crafting a diverse and plausible style distribution to maximize the generalization efficacy of style enhancement. Simultaneously, GSG emphasizes robust skeletal representations within the content to mitigate the adverse effects of spurious correlations, focusing on stable attributes conducive to accurate classification. This framework enhances model generalization by synergistically generating diverse styles while preserving essential content integrity. Extensive experimentation across two benchmark datasets validates the efficacy of our method.