Dynamic Style Transferring and Content Preserving for Domain Generalization

Although convolutional neural networks (CNNs) have shown remarkable ability in different computer vision tasks, they do not cope well with domain shifts. Recent studies show that the domain shift mainly results from the style or texture variation of images rather than the content. Inspired by this, we propose dynamic style transferring to overcome the style bias of CNNs. Specifically, we design a knowledge-injected attention mechanism for learning adaptive fusion weights and embedding the style knowledge of dynamic chosen images in latent space. So the extent of transferred style is controlled, and we can retain content-related information. Furthermore, we introduce the content preserving module, which builds an adversarial structure with the encoder to make the extracted style information more precise. For balancing the adversarial relationship between encoder and auxiliary predictor, we also introduce a consistency loss to empower the style-biased predictor and indirectly boost the encoder’s ability by extending the back-propagation process. We conduct extensive experiments on PACS and Office-Home datasets to evaluate the effectiveness of our method. Experiment results show remarkable performance over the state-of-the-art methods in the domain generalization.