Exploring Diverse Color Spaces in Frequency Domain-Based Image Augmentation for Corruption Robustness

This study investigates the application of the Random Erasing in the Frequency Domain (REF) technique to enhance corruption robustness in image classification. The REF method involves manipulating images using the Discrete Fourier Transform (DFT) and the Inverse Discrete Fourier Transform (IDFT) to erase specific frequency components. While REF has proven effective, this research explores the potential benefits of applying REF in different color spaces, such as YUV and HSV, beyond the conventional RGB space. We hypothesize that using diverse color spaces with REF can effectively address previously unexplored dimensions and improve corruption robustness. Through a series of experiments and analysis, we identify the optimal color space for maximizing improvements in corruption robustness, contributing to a deeper understanding of this technique's potential in image classification.