Noise2Variance

This repository contains the official code of the paper published by IET Image Processing:

Noise2Variance: Dual networks with variance constraint for self‐supervised real‐world image denoising [Paper PDF]

Highlights: This study demonstrates that a straightforward loss design, concentrating on variance, can effectively train a standard CNN denoiser in a self-supervised fashion.

Benchmark results of the paper are available at:

• DND Benchmark Results（sRGB）
• SIDD Benchmark Results（sRGB）

Installation

Clone this repository into any place you want.

git clone https://github.com/TomHeaven/noise2variance_pytorch

Dependencies

• Python
• PyTorch>=2.0
• numpy
• Pillow
• torchvision
• scipy

Expriments

SIDD validation dataset

To train and evaluate the model directly please visit SIDD website and download the original Noisy sRGB data and Ground-truth sRGB data from SIDD Validation Data and Ground Truth and place them in data/SIDD_Small_sRGB_Only folder.

Test and Train

You can now go to src folder and test our model by:

cd src
sh test.sh

The pretrained model configuration and weight of SIDD validation dataset is located at models/Noise2Variance/0320-121732.

or you can train it by yourself as follows:

sh train.sh

Citation

If you use our paper or code, please cite the the paper:

@article{tan2024noise2variance,
  title={Noise2Variance: Dual networks with variance constraint for self-supervised real-world image denoising},
  author={Tan, Hanlin and Liu, Yu and Zhang, Maojun},
  journal={IET Image Processing},
  year={2024},
  publisher={Wiley Online Library}
}

Credits

• Thanks for CVF-SID to share code framework of the repo.