AmpNorm: An Effective Style Normalization Method for Single Domain Generalization

Deep neural networks have demonstrated remarkable efficacy in numerous computer vision tasks. However, due to the training and testing sets of data coming from different domains, the domain gap limits the performances of deep neural networks. To enhance the generalization performance of deep neural networks, we provide a unique amplitude normalization (AmpNorm) approach to decrease the domain gap from the frequency domain perspective. Specifically, ours AmpNorm collects the amplitude spectrum from the source domain during training and then converts the images from the unseen target domain into those with a style similar to the source domain during testing. Our AmpNorm is simple yet effective, as well as plug-and-play, which is readily implemented into the majority of single-domain generalization (SDG) methods. Extensive results on three public benchmarks demonstrate that our AmpNorm can greatly improve these models’ performance on the invisible target domain.