Enhanced dynamic feature representation learning framework by Fourier transform for domain generalization

Domain generalization is an active research topic for mitigating distribution shifts and improving model generalization by transferring domain-invariant knowledge learned from source domains to a target domain. Most existing methods that use static parameters attempt to obtain domain-invariant knowledge by aligning all the sample distributions. However, such static models lack adaptability, and it is difficult to resolve cross-domain conflicts. Moreover, these approaches align all representations to a latent space, where the aligned non-transferable features are prone to cause negative transfer and reduce the discriminability of the algorithms. To solve these problems, we develop an enhanced dynamic feature representation learning framework (DFRL) using Fourier transform for domain generalization. The framework consists of two parts: Fourier-based dynamic residual feature representation learning (DRFR) and dynamic factor. Specifically, DRFR is implemented by developing a dynamic residual module based on Fourier-based sample amplitude mixing to capture the coarse-grained and fine-grained dynamic feature knowledge. Then, we construct a dynamic factor to quantitatively trade-off the alignment and discriminability, preventing model performance degradation due to excessive of pursuit of any aspects. Extensive experiments have been carried on several standard public benchmarks, including Digits-DG, VLCS, PACS and Office-Home, which demonstrate that our framework achieves significant results.