Proximal denoiser for convergent plug-and-play optimization with nonconvex regularization

Code for the paper "Proximal denoiser for convergent plug-and-play optimization with nonconvex regularization" published at ICML 2022.

[Paper]

Samuel Hurault, Arthur Leclaire, Nicolas Papadakis.
Institut de Mathématiques de Bordeaux, France.

Prerequisites

The code was computed with Python 3.8.10, PyTorch Lightning 1.2.6, PyTorch 1.7.1

pip install -r requirements.txt

Prox-Denoiser (Prox-DRUNet)

The code relative to the Proximal (Gradient Step) Denoiser can be found in the GS_denoising directory.

Training

• Download training dataset from https://plmbox.math.cnrs.fr/f/4f56db2f0f7d49a88663/?dl=1 and unzip DRUNET in the datasets folder

• Realize a first baseline training of the Gradient Step denoiser without constraining the spectral norm (1200 epochs) :

cd GS_denoising
python main_train.py --name GS_DRUNet

Checkpoints, tensorboard events and hyperparameters will be saved in the GS_denoising/logs/GS_DRUNet subfolder.

• Save the trained model in the ckpts directory :

cp logs/GS_DRUNet/version_0/checkpoints/* ckpts/GS_DRUNet.ckpt

We also provide pretrained checkpoint at https://plmbox.math.cnrs.fr/f/414fbb3e754840978ef8/?dl=1

• Finetune previous training constraining the spectral norm (15 epochs) :

python main_train.py --name Prox_DRUNet  --resume_from_checkpoint --pretrained_checkpoint ckpts/GS_DRUNet.ckpt --jacobian_loss_weight 1e-3 

• Save the trained model in the ckpts directory :

cp logs/Prox_DRUNet/version_0/checkpoints/* ckpts/Prox_DRUNet.ckpt

Testing

• Download pretrained checkpoint from https://plmbox.math.cnrs.fr/f/faf7d62213e449fa9c8a/?dl=1 and save it as GS_denoising/ckpts/Prox_DRUNet.ckpt
• For denoising the whole CBSD68 dataset at input Gaussian noise level 25 :

cd PnP_restoration
python denoise.py --dataset_name CBSD68 --noise_level_img 25

Add the argument --extract_images the save the output images.

Plug-and-Play Image Restoration (Prox-PnP)

Deblurring

• If not already done, download pretrained checkpoint from https://plmbox.math.cnrs.fr/f/faf7d62213e449fa9c8a/?dl=1 and save it as GS_denoising/ckpts/Prox_DRUNet.ckpt
• Chose the PnP algorithm in {PGD,DRS,DRSdiff} and the input Gaussian noise level in {2.55,7.65,12.75}
• For instance, for deblurring the CBSD68 images with the algorithm DRS, at input Gaussian noise level 7.65, sequentially blurred with the 10 different kernels exposed in the paper:

cd PnP_restoration
python deblur.py --dataset_name CBSD68 --PnP_algo DRS --noise_level_img 7.65 

Add the argument --extract_images the save the output images and --extract_curves the save convergence curves.

Super-resolution

For performing super-resolution of CBSD10 images, downscaled with scale sf, with the algorithm DRS, at Gaussian noise level 7.65, and sequentially blurred with the 4 different kernels exposed in the paper:

cd PnP_restoration
python SR.py --dataset_name CBSD68 --PnP_algo DRS --noise_level_img 7.65 --sf 2

Acknowledgments

This repo contains parts of code taken from :

• Deep Plug-and-Play Image Restoration (DPIR) : https://github.com/cszn/DPIR
• Gradient Step Denoiser for convergent Plug-and-Play (GS-PnP) : https://github.com/samuro95/GSPnP

Citation

@article{hurault2022proximal,
  title={Proximal denoiser for convergent plug-and-play optimization with nonconvex regularization},
  author={Hurault, Samuel and Leclaire, Arthur and Papadakis, Nicolas},
  journal={arXiv preprint arXiv:2201.13256},
  year={2022}
}