Performance Enhancement Method of Frequency Component Modification with The Middle Region for Corruption Robustness

Frequency Component Modification (FCM) is a data augmentation that modifies the frequency domain and has shown effectiveness in enhancing model robustness. However, it has a limitation in that it introduces unnatural transitions between frequency components. To address this issue, we propose Frequency Component Modification with Mid-Region Tuning (FCM-MRT), which encourages smoother transitions by incorporating a middle frequency region. The frequency domain is randomly divided into low, mid, and high frequency regions, and the middle frequency component is modified using various strategies: random values sampled from a uniform distribution, linearly varying values, and nonlinearly varying values. To evaluate performance, we use two metrics: corruption error, measured on corrupted dataset, and clean error, measured on clean dataset. Experimental results demonstrate that on CIFAR10, the linear variation method yields the best performance, reducing the corruption error by 1.376%, and on CIFAR100, the uniform distribution method achieves the best improvement, reducing the corruption error by 0.877%.