HDFLStyler: Hierarchical domain-invariant feature learning for source-free domain generalization

Source-Free Domain Generalization (SFDG) aims to generalize a model to unknown domains without using any specific source domain data. Currently, SFDG methods mainly use the vision-language large models to extract different style features from text prompts and use different style features to train linear classifiers, thereby eliminating the model's dependence on source domain images. However, a challenging problem in the source-free domain generalization classification is how to generate as diverse styles as practicable solely from text prompts and learn domain-invariant features from different styles. In this paper, we propose a hierarchical domain-invariant feature learning method (HDFLStyler) for SFDG to improve the classification accuracy. HDFLStyler is mainly composed of diverse style generation and domain-invariant feature learning. Diverse style generation dynamically generates as many styles as practicable through random distribution adjustment and adaptive mixing strategies. Domain-invariant feature learning comprehensively learns domain-invariant features of diverse styles by combining global and local approaches. In addition, to better learn domain-invariant features, we also design a domain-invariant consistency loss to improve the classification performance of HDFLStyler. Extensive experiments demonstrate that HDFLStyler achieves excellent classification performance.