Introduction

This is the repo for the deep learning image restoration model proposed in S. López-Tapia, J. Mateos, R. Molina and A. K. Katsaggelos, "Deep robust image restoration using the Moore-Penrose blur inverse", IEEE Internaational Conference on Image Processing (ICIP), 2023.

Requirements

• Python >= 3.8
• Numpy
• Scikit-image
• Pytorch == 1.7.1
• pytorch_msssim
• Scipy

Installation

conda create -n ddnet python=3.8 scipy scikit-image
conda activate ddnet
conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch
pip install pytorch_msssim

Usage

First, activate conda environment:

conda activate ddnet

Data preparation training

Download dataset from this link and extract it. Modify TRAIN_FILE_PATH and EVAL_FILE_PATH in config.py to the paths of the train data folder and the validation data folder.

Train

Our systems uses two models:

1. Luminance model or "y model". Script: train_y.py
2. Color model or "cbcr model". Script: train_cbcr.py

Before training, modify both W_PATH_SAVE and W_COLOR_PATH_SAVE to the paths of the folders where you want to save the weights of both models.

python train_y.py
python train_cbcr.py

Val/Test

To process a folder and generate the restores images of its contents, use predict.py. You will need to download our weights from here or provide your own using the training scripts.

python predict.py <image_path> <psf_path> <output_path> <model_y_weights_path> <model_cbcr_weights_path>

• <image_path>: Path to folder containing the blur images. They have to be in png or jpg formats.
• <psf_path>: Path to folder containing the estimated PSFs. They must be saved as a matrix in a npy file. Each one must have the same name as its corresponding blur image with the subfix _psf.npy added.
• <output_path>: Path where the restored images would be saved.
• <model_y_weights_path>: Path to the weights of model y.
• <model_cbcr_weights_path>: Path to the weights of model cbcr.

Citation

@inproceedings{Lopez2023BID,
  title={Deep robust image restoration using the Moore-Penrose blur inverse},
  author={Santiago L\'opez-Tapia and Javier Mateos and Rafael Molina and Aggelos K. Katsaggelos},
  booktitle={IEEE Internaational Conference on Image Processing (ICIP)},
  year={2023}
}