Uncertainty-Aware Unsupervised Image Deblurring with Deep Residual Prior

This is the official Python implementation of the CVPR 2023 paper Uncertainty-Aware Unsupervised Image Deblurring with Deep Residual Prior

The repository contains reproducible PyTorch source code for computing the deblurred image and residual given a single kernel and blurry image.

Here are the Paper and Supplementary material.

Citation

@InProceedings{Tang_2023_CVPR,
    author    = {Tang, Xiaole and Zhao, Xile and Liu, Jun and Wang, Jianli and Miao, Yuchun and Zeng, Tieyong},
    title     = {Uncertainty-Aware Unsupervised Image Deblurring With Deep Residual Prior},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {9883-9892}
}

Visualization of the kernel induced error (residual)

Test on real blurry images from Lai dataset

Robustness to the kernel error

If the inputs contain only blurry images, you will need to run some outsourcing kernel estimation algorithm to obtain the kernel. The dataset folder includes some estimated kernel and corresponding blurry images for test.

Before runing UaUDeblur_demo.py, the first thing you have to do is to pre-process the (estimated) kernel with Fourier transformation and save it as a .mat file. We provide an example for this procdure, see KernelGen_Inaccu.m; where the FFT of blur kernel is done via the function extendHforConv.m.

If you have any problem, contact me at [email protected]