icbcbicc/pytorch_epll

PyTorch implementation of Expected Patch Log Likelihood (EPLL) image prior in paper "D. Zoran and Y. Weiss, "From learning models of natural image patches to whole image restoration," ICCV 2011.

PyTorch implementation of Expected Patch Log Likelihood (EPLL) image prior

1. Paper:

The EPLL image prior model is presented in "D. Zoran and Y. Weiss, "From learning models of natural image patches to whole image restoration," 2011 International Conference on Computer Vision, Barcelona, 2011, pp. 479-486, doi: 10.1109/ICCV.2011.6126278."

2. Acknowledgement:

This code is based on the original Matlab implementation, authored by Yair Weiss. The pre-trained GMM model is also borrowed from the original code.

3. Key features:

• Pure PyTorch implementation
• Support color image
• Support batch mode (input.shape=[B, C, H, W])
• Adjustable stride: --stride n (default 4, the stride in the original matlab code is fixed to 1, higher stride will speed up the code and reduce RAM usage)
• Support parallel processing: --parallel True (default True, requires more RAM)

4. Requirements:

• torch-scatter, tested with version 2.0.5
• Tested with pytorch 1.6.0 and cuda 10.2 on Ubuntu 18.04

5. Usage:

Denoising a clean grayscale image 160068.jpg contaminated by gaussian noise of standard deviation 0.1:

python epll_demo.py --im_file 160068.jpg --noise_std 0.1

• Specifying --stride n will speed up the code and reduce RAM usage.
• Specifying --parallel True will speed up the code but this requires more RAM.
• Specifying --use_cuda True will let you use GPU. However, this may consume a lot of GPU RAM.

6. Results:

160068.jpg

new.jpg

7. TODO:

• Support color image
• Support batch processing
• Optimize speed
• Support backpropagation

pytorch_epll's Issues

Training guide

Hi,
Thanks for your great work! I was wondering if you could provide a training guide. Thank you!