Master Assignment
Fourier-basis noise image data augmentation for CNN robustness
Type: Master CS
Period: TBD
Student: (Unassigned)
If you are interested please contact :
Background:
Traditional image augmentations include but are not limited to rotation, crop and flip, which mainly operate in the spatial domain. By using image data augmentation, the variability in the data increases, and thus can improve model robustness. Instead of operating in the spatial domain, this research aims at exploring whether augmenting the frequency information of images helps improve model robustness. To augment frequency information, Fourier basis noise is used [1], which is additive noise in a specific frequency. The frequency-augmented images are further used for training computer vision models, for which the robustness towards image corruption [2] is evaluated and compared with state-of-the-art approaches. Optimization of the augmentation policy will further be investigated, regarding computation efficiency, training time, complexity of frequency combinations, and so on.
References:
[1] Yin, D., Lopes, R. G., Shlens, J., Cubuk, E. D., & Gilmer, J. (2019). A Fourier Perspective on Model Robustness in Computer Vision. doi:10.48550/ARXIV.1906.08988
[2 Hendrycks, D., & Dietterich, T. (2019). Benchmarking Neural Network Robustness to Common Corruptions and Perturbations. doi:10.48550/ARXIV.1903.12261