Robustness of ConvNet to High-Frequency Image Corruptions

Image processing techniques such as blurring, JPEG compression are applied to natural images to meet different objectives. Additionally, corruptions such as Gaussian and shot noise appear on images due to digital fluctuations. Unfortunately, standard vision models tend to perform quite poorly under such unavoidable corruptions, i.e., these models are not robust to the distribution shifts induced by these corruptions at test time. The standard approach for overcoming this issue for a known corruption is by augmenting the training data with images perturbed using the corruption of interest. Motivated by settings where the corruption might not be known during training, Gaussian noise is used as an augmentation strategy to gain robustness to high-frequency corruptions. In this paper, we try to understand its properties from a Fourier lens. However, we show that Gaussian augmentation fails to maintain robustness to few high-frequency corruptions at high severity levels. Analyzing the Fourier signature of those corruptions reveal a change in behavior - at high severity they corrupt low frequencies as well. A Gaussian-trained model loses its performance due to this change. Current augmentation strategies for low-frequency corruptions are discussed at the end.