Research on the improvement of domain generalization by the fusion of invariant features and sharpness-aware minimization

Domain generalization (DG) aims to enhance the model’s ability to generalize from the source domains to unseen domains by addressing distribution shifts. A common approach in DG is to learn invariant features across domains. However, limited data availability restricts the model’s generalization capability when relying solely on domain-invariant features. To address this limitation, this paper proposes a straightforward yet effective method called combined feature and sharpness, which converges the minimum loss to a flat region while learning invariant features. Specifically, instead of considering domains as the basis for feature invariance, this paper focuses on learning invariant features at a more granular level between samples to reduce their dependence on specific domains. Additionally, we improve the sharpness-aware minimization method by minimizing both empirical risk and the surrogate gap, preventing minimum loss convergence into sharp valleys. Extensive experiments and ablation studies are conducted on four DG datasets (VLCS, PACS, OfficeHome, and TerraIncognita) using the DomainBed benchmark. The results demonstrate that our proposed method outperforms other DG methods in terms of generalization performance across unseen domains.