A Spectrum Filtering Framework for Domain Generalization

Domain generalization aims to address the distribution shift problem inherent in neural networks, wherein a misalignment between test data distribution and training data distribution leads to significant performance degradation. This paper introduces Fourier Style Restitution (FSR), a Fourier-transform-driven method for domain generalization. FSR integrates the principles of Fourier augmentation and style disentanglement with feature reconstruction, enhancing model generalizability to unseen domains. The framework implements a cross-domain filtering enhancement strategy based on Fourier transform, leveraging frequency domain filtering to bolster model robustness against distributional variations. Through this paradigm, each sample transcends source domain constraints to derive optimized domain-invariant feature representations tailored to its intrinsic characteristics. The framework further incorporates style regularization to distill consistency signals from stylized images and employs prototype compensation to recover lost domain-invariant features. Extensive experiments demonstrate state-of-the-art performance on benchmark datasets. The method’s efficacy stems from feature enhancement and style reconstruction through Fourier-based operations for robust domain generalization.