Frequency domain image augmentation for domain generalized image classification

Official URL: https://dx.doi.org/10.1109/SIU66497.2025.11112033

Domain generalization (DG) remains a major challenge in computer vision, in which models trained on source domain(s) perform poorly on unseen target domains. One way to address this issue is by using augmentation techniques to generate synthetic images from the original ones, aiming to make them more similar to the target domain images. To improve this technique, a new augmentation method called Amplitude-Phase Augmentation (APA) was investigated. The aim of this method is to increase the robustness of models by exposing them to a greater level of input variations. APA relies on the product of amplitudes in the frequency domain. By applying a specific form of amplitude multiplication, it generates cross-domain augmented images with a wider range of transformations in the frequency space. Experimental results showed that APA leads to superior performance compared to competitive counterparts.