Multi-Block based Random Erasing in the Frequency Domain

Random Erasing in the Frequency domain (REF) is a data augmentation method that performs random erasing in the frequency domain. In order to evaluate the performance, it utilizes two metrics such as the corruption error, which measures the error rate for a corruption dataset, and the clean error, which measures the error rate for a clean dataset. While REF is effective in reducing the corruption error, it shows the increment in clean error. To address this increment, this paper proposes Random Erasing in the Frequency domain using multi-block. Additionally, it considers both clean error and corruption error to identify the best multi-block configuration. The training image is divided into n×n blocks, and REF is applied to each block to generate various transformations. Experimental results show that utilizing a 2×2 block configuration reduces clean error by up to 0.466%. Moreover, the best multi-blocks configuration results in a decrease of clean error by up to 0.046% and a reduction of corruption error by up to 5.246%.