DeGAN_noise_distribution (practice)

Reference

Qiongshuai Lyu, Min Guo, Zhao Pei, DeGAN: Mixed noise removal via generative adversarial networks, Applied Soft Computing, Volume 95, 2020, 106478, ISSN 1568-4946,

Data

• three types of noise

  • White Gaussian noise(AWGN)
  • Salt-and-Paper noise(SPIN)
  • random-valued noise(RVIN)

• Training Data

  • AWGW+SPIN

  • 

  • AWGW+RVIN

  • 

  • AWGW+SPIN+RVIN

  • 

• Dataset in this paper

  • Joseph Chet Redmon, Pascal VOC Dataset Mirror (VOC2007).

  https://pjreddie.com/projects/pascal-voc-dataset-mirror/

  • K. Zhang, Datasets, 2016.

  https://drive.google.com/drive/u/0/folders/0B-_yeZDtQSnobXIzeHV5SjY5NzA

• Our noise dataset

  • https://drive.google.com/file/d/1xfFTD_gd_u-LqqYI3qkblhZGrSBnKCVV/view
  • https://drive.google.com/file/d/1OUqzojlkCKBVhheZjd3KslY-d9ZEY4fg/view

Architecture

I change the unet, because the paper's information can not match the paper's picture, so i use the most common unet to my generator. And I also change the loss function, because I thought that the ssim loss is more important than mse, so I give some different weight to different loss.

Requirement

• numpy==1.19.2
• Pillow==8.1.0
• torch==1.7.1+cu110
• torchaudio==0.7.2
• torchvision==0.8.2+cu110
• typing-extensions==3.7.4.3

pip install -r requirements.txt 

Our result

PSNR = 69.36

SSIM = 0.746

noise image

denoise image

ground truth

Proeblem

noise distribution only for case-by-case