Comprehensive disentanglement with fine-grained feature mitigation for domain generalization.

Abstract

Domain generalization is proposed as an approach capable of solving the domain shift challenge, which aims at generalizing knowledge learned from multiple source domains with different distributions to the target domain that is invisible during the training process. A range of domain generalization methods include unstable domain-specific features when performing domain-invariant representation learning. Our method is dedicated to comprehensive and explicit feature disentanglement, which realizes the independence of domain-invariant and domain-specific features and reduces the spurious reliance on domain-specific features with pursuing sufficient stable semantics. In this regard, the novel learning paradigm of Source Split-flow Disentanglement with Smoothness-Fine-grained Feature Mitigation (SSDS-FFM) is presented. Firstly, we propose the source split-flow structure where the domain-invariant feature extractor and the domain-specific feature extractor share the same shallow layer and are split into two independent flows. Mutual information minimization is utilized to separate the two features. At the same time, we avoid the highly confident domain classifier and introduce domain label smoothing to predict the corresponding soft probabilities, which is combined with structural design to ensure the learning of domain-invariant representations. Secondly, to further enhance class discriminability, we propose fine-grained feature mitigation to perform selective reverse contrastive learning, which can address local domain misalignment and alleviate over-compressed feature space with obtaining sufficient stable semantics. Our paradigm is logical and can achieve comprehensive feature disentanglement to preform stable domain-invariant representation learning, promoting the improvement of generalization ability. Extensive experimental results on PACS, VLCS, Office-Home and DomainNet datasets verify the effectiveness and superiority of the proposed SSDS-FFM.