Frequency Domain Image Augmentation For Domain Generalized Image Classifıcation

Domain Shift remains a major challenge in Domain Generalization (DG), wheremodels trained on source domain(s) tend to perform poorly on unseen target domains.One effective approach to address this problem is the use of data augmentationtechniques that synthetically enhance domain diversity. In this thesis, Iintroduce a frequency-domain augmentation method called Amplitude-Phase Augmentation(APA). APA operates by multiplying the amplitude components of sourceimages with those from other domains in the frequency domain, while preservingthe original phase information. This controlled mixing leads to the creation ofcross-domain images that retain semantic structure but carry varied textural cues,increasing the robustness of models to distributional changes. I evaluate APA ontwo standard DG benchmarks: PACS and VLCS, using three diverse backbonearchitectures—ResNet-50, T2T-ViT-14, and DeiT-Small. APA is implemented ontop of a standard Empirical Risk Minimization (ERM) framework and is also testedin conjunction with existing DG strategies. Extensive experiments show that APAimproves generalization performance across both datasets and three backbones. Notably,APA achieves competitive results compared to strong baselines and recentaugmentation-based methods on PACS dataset and superior results on VLCS acrossall three backbones. In addition to performance evaluations, I conduct detailed ablationstudies on the amplitude mixing strategy and its effect on model robustness.These results demonstrate the practical effectiveness and adaptability of APA as alightweight and domain-agnostic augmentation method for DG tasks. Code availableat https://github.com/sina-nuel/APA.