DS-N2N

Dual-Sampling Noise2Noise: Efficient Single Image Denoising

How to run

1. Dependences

• python == 3.8
• pytorch == 2.0.1
• skimage == 0.19.0
• tqdm == 4.50.2

2. Train DS-N2N

python main.py \
  --image_folder (your nosiy images folder, type = str) \
  --device (cuda or cpu, type = str, default = 'cuda') \
  --max_epoch (max number of epochs, type = int, default = 10000) \
  --lr (learning rate, type = float, default = 0.001) \
  --step_size (step size of changing the learning rate, type = int, default = 9000) \
  --gamma (factor by which learning rate decays, type = float, default = 0.5) \
  --embedding (image convolution embedding channels, type = int, default = 48) \
  --noise_type (guass or poiss, type = str, default = 'guass') \
  --noise_level (guass: σ or poiss: λ, type = int, default = 25) \
  --psnr_data_save_path (psnr data save path, type = str, default = './data/25g_psnr.txt') \
  --ssim_data_save_path (ssim data save path, type = str, default = './data/25g_ssim.txt') \
  --loss_data_save_path (loss data save path, type = str, default = './data/25g_loss.txt')

# If you want to save psnr, ssim, and loss data, please release the save file comments under train.py：
* denosied_psnr = Test_Denosie_PSNR(self.model, noise_image, clean_image)
* denosied_ssim = Test_Denosie_SSIM(self.model, noise_image, clean_image)
* progress_bar.set_postfix(loss=loss.item(), denoising_PSNR=denosied_psnr, denosied_SSIM=denosied_ssim)
* save_to_txt(denosied_psnr, self.psnr_txt_path)
* save_to_txt(denosied_ssim, self.ssim_txt_path)
* save_to_txt(loss.item(), self.loss_txt_path)
Please note that saving the data will significantly increase the denoising processing time！

If you have referenced the content of this repository, please cite the following papers:

@ARTICLE{10927623,
  author={Bai, Jibo and Zhu, Daqi and Chen, Mingzhi},
  journal={IEEE Transactions on Instrumentation and Measurement}, 
  title={Dual-Sampling Noise2Noise: Efficient Single-Image Denoising}, 
  year={2025},
  volume={74},
  number={},
  pages={1-12},
  keywords={Noise reduction;Noise;Noise measurement;Training;Image denoising;Image resolution;Convolutional neural networks;Computational modeling;Residual neural networks;Gaussian noise;DsNet;dual sampler;dual-sampling Noise2Noise (DS-N2N);single-image denoising method},
  doi={10.1109/TIM.2025.3551427}
}