MSc. Thesis Defense: Sina Salehnia, Frequency Domain Image Augmentation for Domain Generalized Image Classification

Frequency Domain Image Augmentation for Domain Generalized Image Classification

Sina Salehnia
Computer Science, MSc. Thesis Dissertation, 2025

Thesis Jury

Assoc. Prof. Öznur Tatan (Thesis Advisor)

Assoc. Prof. Hüseyin Özkan

Prof. Dr. Pınar Duygulu Şahın

Date & Time: July 17th, 2025 –  10:00 AM

Place: FENS G025

Keywords: Domain generalization, Frequency domain augmentation, domain shift,

fast Fourier transform

Abstract

Domain Shift remains a major challenge in Domain Generalization (DG), where models 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 augmentation techniques that synthetically enhance domain diversity. In this thesis, I introduce a frequency-domain augmentation method called Amplitude-Phase Augmentation (APA). APA operates by multiplying the amplitude components of source images with those from other domains in the frequency domain, while preserving the original phase information. This controlled mixing leads to the creation of cross-domain images that retain semantic structure but carry varied textural cues, increasing the robustness of models to distributional changes.

I evaluate APA on two standard DG benchmarks: PACS and VLCS, using three diverse backbone architectures—ResNet-50, T2T-ViT-14, and DeiT-Small. APA is implemented on top of a standard Empirical Risk Minimization (ERM) framework and is also tested in conjunction with existing DG strategies. Extensive experiments show that APA improves generalization performance across both datasets and three backbones. Notably, APA achieves competitive results compared to strong baselines and recent augmentation-based methods on PACS dataset and superior results on VLCS across all three backbones.

In addition to performance evaluations, I conduct detailed ablation studies on the amplitude mixing strategy and its effect on model robustness. These results demonstrate the practical effectiveness and adaptability of APA as a lightweight and domain-agnostic augmentation method for DG tasks.