Content-preserved augmentation and style-transferred normalization for single domain generalization

Although large progress has been achieved, current domain generalization methods still suffer from the domain gap between the source-domain and target-domain samples, e.g. the style cues of the source and target domains may differ much, which is particularly evident in the scenario with a single source domain. Data augmentation is one of prevailing approaches for producing diverse source-domain samples and narrowing this domain gap. However, the content cues of the augmentation samples by current data augmentation methods may be distorted too much, which makes network learning biased toward the classification-insensitive style cues. To address the content degradation and domain style gap, we propose the modules of Adversarial Content Consistent Normalization (ACCN) and style reconstruction-based Amplitude Normalization (SRAN) to reduce this domain gap while preserving content cues from a perspective of frequency domain. Our algorithm has two merits, (1) our ACCN can restore the content cues damaged in data augmentation, by regularizing the content representations of augmented samples to approach those of the original; (2) our SRAN can reduce the style gap between the source and target domains by modeling instance-aware style cues of source domain and transferring them onto the target-domain features. Our ACCN and SRAN are simple yet effective, and extensive results on five public benchmarks further demonstrate the effectiveness of our method in enhancing the generalization performance on unseen target domains.