This repository includes the implentation of the Restore-to-Classify GANs (R2C-GANs) proposed in R2C-GAN: Restore-to-Classify Generative Adversarial Networks for Blind X-Ray Restoration and COVID-19 Classification.
R2C-GANs jointly learn inverse and forward mappings between poor and high quality image domains
and predict the class labels.
Software environment:
# Python with the following version and libraries.
conda create -n r2c-gan python=3.7.11
conda activate r2c-gan
conda install tensorflow-gpu=2.4.1
conda install scikit-image tqdm scikit-learn pydot
conda install -c conda-forge oyaml
pip install tensorflow-addons==0.13.0
pip install numpy==1.19.2
If you use the provided method in this repository, please cite the following paper:
@article{r2cgan,
title = {R2C-GAN: Restore-to-Classify Generative Adversarial Networks for blind X-ray restoration and COVID-19 classification},
journal = {Pattern Recognition},
volume = {156},
pages = {110765},
year = {2024},
issn = {0031-3203},
doi = {https://doi.org/10.1016/j.patcog.2024.110765},
url = {https://www.sciencedirect.com/science/article/pii/S0031320324005168},
author = {Mete Ahishali and Aysen Degerli and Serkan Kiranyaz and Tahir Hamid and Rashid Mazhar and Moncef Gabbouj}
}
We collect the dataset consisting of unpaired poor and good quality X-ray samples. The dataset is available here, please download the file and unzip it:
unzip dataset.zip
The proposed R2C-GAN method can be trained as follows,
python train.py --method operational --q 3
There are three R2C-GAN models available for training: i. using operational layers --method operational --q 3 where --q is the polynomial order (see Operational Layers), ii. deep R2C-GAN model with convolutional layers --method convolutional, and iii. compact model with convolutional layers --method convolutional-light.
The model checkpoints for R2C-GAN (Q = 3), R2C-GAN (CNN), and R2C-GAN (CNN-compact) are provided in output.zip file, please unzip the file and place the checkpoints under outputs/checkpoints/ folder to reproduce the reported results in the paper. The evaluation of the methods can be performed using the provided test.py script as follows,
python test.py --method operational --q 3
python test.py --method convolutional
python test.py --method convolutional-light
Optionally you can use --saveImages flag, e.g., python test.py --method operational --q 3 --saveImages True to store the restored images by the selected model under output/samples_testing folder.