DCT2net: An Interpretable Shallow CNN for Image Denoising (IEEE TIP'22)

Sébastien Herbreteau and Charles Kervrann

Requirements

The repo supports python 3.8 + pytorch 1.8.1 + numpy 1.21.2 + skimage 0.19.2 + PIL 8.2.0.

Install

To install in an environment using pip:

python -m venv .dct2net_env
source .dct2net_env/bin/activate
pip install /path/to/DCT2net

Datasets

All the models are trained on BSD400 and tested on BSD68 and Set12. Simply modify the argument --in_folder for training on other datatets. Only one model is trained for all noise levels between 1 and 55.

Results

Gray denoising

(Training on BSD400)

Model	Params	sigma=15	sigma=25	sigma=50
DnCNN	556k	31.72	29.23	26.23
BM3D	-	31.07	28.57	25.62
DCT2net	29k	31.09	28.64	25.68

Run the Code

To train a new model for gray denoising:

python ./trainer.py --in_folder /path/to/dataset

Pretrained model

To denoise an image with DCT2net (remove --add_noise if it is already noisy):

python ./dct2net_denoiser.py --sigma 25 --add_noise --in ./test_images/102061.png --out ./denoised.png --model_name ./saved_models/dct2net.p

To denoise an image with DCT/DCT2net (remove --add_noise if it is already noisy):

python ./dct-dct2net_denoiser.py --sigma 25 --add_noise --in ./test_images/102061.png --out ./denoised.png --model_name ./saved_models/dct2net.p

Acknowledgements

This work was supported by Bpifrance agency (funding) through the LiChIE contract. Computations were performed on the Inria Rennes computing grid facilities partly funded by France-BioImaging infrastructure (French National Research Agency - ANR-10-INBS-04-07, “Investments for the future”).

We would like to thank R. Fraisse (Airbus) for fruitful discussions.

Citation

@ARTICLE{9799727,
  author={Herbreteau, Sébastien and Kervrann, Charles},
  journal={IEEE Transactions on Image Processing}, 
  title={DCT2net: An Interpretable Shallow CNN for Image Denoising}, 
  year={2022},
  volume={31},
  number={},
  pages={4292-4305},
  doi={10.1109/TIP.2022.3181488}
}