Frequency stepwise distillation to improve image classification with common corruptions

Xiaohong Zhang, Shengwu Xiong, Jianwen Xiang

Abstract: Current deep learning approaches for the image corruption classification often prioritize learning the high-energy low-frequency components, however, they neglect the importance of robust features which come from the low-energy high-frequency bands of model filter. To address this limitation, we propose the Frequency Stepwise Distillation (FSD) method, a novel framework that progressively enhances the learning of high-frequency components through the robustness prompt of distillation architecture. Specifically, first the frequency knowledge is generated under the division of model frequency filter and frequency attention strategy, this refines the knowledge of different properties at each generation network. Second, the role-changed mechanism of stepwise distillation constructs the virtual teacher with different model frequency filters, the frequency knowledge is expressed progressively and robustly under the relative robustness prompt supervision of previous generation network. Extensive experiments on the three image corruption classification datasets demonstrate that, compared with the state-of-the-arts our FSD method gets the better performance in both natural clean and corrupted images, especially for the noise, blur, weather and digital categories.