This notebook allows creating datasets of images downscaled by random kernels. This is used for Blind Super-Resolution challenges where the downscaling kernel should be predicted.
This code was used to create the dataset for the paper:
KernelGAN: Blind Super-Resolution Kernel Estimation using an Internal-GAN (Bell-Kligler, Shocher, Irani)
Repository of KernelGAN: https://github.com/sefibk/KernelGAN
Project page: http://www.wisdom.weizmann.ac.il/~vision/kernelgan/
The kernels are basically unisotropic gaussians but there is also an option for multiplicative noise that makes the deviate from a pure gaussian. You can control the downscale factors, even changing aspect-ratio, and the probablistic properties of sampling the kernels from the first cell in the notebook.
The code is mostly based on parts from my Resizer repository for resizing images:
https://github.com/assafshocher/Resizer
There is a whole nesty issue of shifting the center of mass of the kernel so that there is no missalignment. Please see the comments in the kernel_shift function in the notebook.
Eventually images and kernels are saved to a wanted path and you get a preview at the bottom of the notebook