Feature Stylization Adversarial Domain Generalization

Although deep learning (DL) systems have achieved remarkable performance in many fields, the problem of severe performance slippage caused by domain shift still needs to be solved. To overcome this problem, domain generalization (DG) aims to learn a generalized model for arbitrary unseen domains leveraging data from multiple source domains. In this paper, we propose a novel DG approach, FSADG. FSADG consists of two components for DG: a domain discriminator and a set of feature style randomization modules. The feature style randomization modules aim to learn the latent distribution of image styles and generate diverse stylized features. Furthermore, adversarial training is conducted between the feature extractor of the target model and the domain discriminator for domain invariant representation learning. This paper also introduces applying FSADG for a more challenging task, DG with the data sparsity problem. We evaluate our method on PACS and OfficeHome datasets on image classification tasks. The experimental results demonstrate the effectiveness of FSADG.