A Residual Dense U-Net for Image Denoising

This repository is for the RDUNet model proposed in the following paper:

Javier Gurrola-Ramos, Oscar Dalmau and Teresa E. Alarcón, "A Residual Dense U-Net Neural Network for Image Denoising", IEEE Access, vol. 9, pp. 31742-31754, 2021, doi: 10.1109/ACCESS.2021.3061062.

Citation

If you use this paper work in your research or work, please cite our paper:

@article{gurrola2021residual,
  title={A Residual Dense U-Net Neural Network for Image Denoising},
  author={Gurrola-Ramos, Javier and Dalmau, Oscar and Alarcón, Teresa E},
  journal={IEEE Access},
  volume={9},
  pages={31742--31754},
  year={2021},
  publisher={IEEE},
  doi={10.1109/ACCESS.2021.3061062}
}

Pre-trained models

Link to download the pretrained models.

Dependencies

• Python 3.6
• PyTorch 1.5.1
• pytorch-msssim 0.2.0
• ptflops 0.6.3
• tqdm 4.48.2
• scikit-image 0.17.2
• yaml 0.2.5
• MATLAB (to create testing datasets)

Dataset

For training, we used DIV2K dataset. You need to download the dataset for training the model and put the high-resolution image folders in the './Dataset' folder. You can modify the train_files.txt and val_files.txt to load only part of the dataset.

Training

Default parameters used in the paper are set in the config.yaml file:

patch size: 64
batch size: 16
learning rate: 1.e-4
weight decay: 1.e-5
scheduler gamma: 0.5
scheduler step: 3
epochs: 21

Additionally, you can choose the device, the number of workers of the data loader, and enable multiple GPU use.

To train the model use the following command:

python main_train.py

Test

Place the pretrained models in the './Pretrained' folder. Modify the config.yaml file according to the model you want to use: model channels: 3 for the color model and model channels: 1 for the grayscale model.

Test datasets need to be prepared using the MATLAB codes in './Datasets' folder according to the desired noise level. We test the RDUNet model we use the Set12, CBSD68, Kodak24, and Urban100 datasets.

To test the model use the following command:

python main_test.py

Results

Results reported in the paper.

Color:

Grayscale:

Contact

If you have any question about the code or paper, please contact francisco.gurrola@cimat.mx .