Theoretical design principles of self-supervised denoising networks

Hayeong Yu* · Seungjae Han* · Young-Gyu Yoon

(* equal contribution)

WACV 2025 Oral

Paper

We report the theoretical design principles of self-supervised denoising networks. We show that a U-Net-shaped blind spot network (U-BSN), whose design is derived by following these principles, achieves superior denoising performance at a low computational cost.

Installation

Clone the repository and create an anaconda environment using

pip install -r requirements.txt

Dataset

We follow the dataset setup in AP-BSN. Please click this link for a detailed preparation description.

Evaluation

To evaluate our model, run:

python test.py -c UBSN_SIDD_val -g 0 --pretrained 'UBSN_SIDD_pretrained.pth' -s UBSN_SIDD_val
python test.py -c UBSN_SIDD_bench -g 0 --pretrained 'UBSN_SIDD_pretrained.pth' -s UBSN_SIDD_bench

-c refers the configuration name (*.yaml), -g refers the GPU ID, --pretrained refers to the name of pretrained model file (in the ckpt/ directory), -s refers to the session name for saving the results.

Pretrained Models

Pretrained model is in the ckpt/UBSN_SIDD_pretrained.pth.

Citation

If you find our code or paper useful, please cite

@inproceedings{yu2025design,
  title={Design Principles of Multi-Scale J-invariant Networks for Self-Supervised Image Denoising},
  author={Yu, Hayeong and Han, Seungjae and Yoon, Young-Gyu},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={1309--1318},
  year={2025}
}

Acknowledgements

This project is built upon AP-BSN. We thank all the authors for their great work and repos.