Skip to content

onwn/C2N

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

8 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

C2N: Practical Generative Noise Modeling for Real-World Denoising - Official PyTorch release

This is an official PyTorch release of the paper "C2N: Practical Generative Noise Modeling for Real-World Denoising" from ICCV 2021.

architecture

If you find C2N useful in your research, please cite our work as follows:

@InProceedings{Jang_2021_ICCV,
    author    = {Jang, Geonwoon and Lee, Wooseok and Son, Sanghyun and Lee, Kyoung Mu},
    title     = {C2N: Practical Generative Noise Modeling for Real-World Denoising},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {2350-2359}
}

[PDF] [Supp] [arXiv]


Setup

Dependencies

  • Python 3.9.6
  • numpy >= 1.16.4
  • cudatoolkit >= 10. (if using GPU)
  • PyTorch 1.2.0
  • opencv-python
  • scikit-image >= 0.15.0
  • tqdm
  • pillow
  • pyyamml
  • imutils

Data

You can place any custom images in ./data and image datasets in subdirectory ./data/[name_of_dataset]

For the SIDD and DND benchmark images, you can find them at SIDD Benchmark and DND Benchmark. Convert them into .png images and place them in each subdirectory.

Pre-trained Models

Download following pre-trained models:

Generator Clean Noisy config Pre-trained
C2N SIDD SIDD C2N_DnCNN model
C2N SIDD DND C2N_DnCNN model
Denoiser Generator Clean Noisy Clean (denoiser train) config Pre-trained
DnCNN C2N SIDD SIDD SIDD C2N_DnCNN model
DIDN C2N SIDD SIDD SIDD C2N_DIDN model
DIDN C2N SIDD DND SIDD C2N_DIDN model

Demo (Quick start)

test_generate.py:

  • config: Name of the configuration.
  • ckpt: Name of the checkpoint to load. Choose between 'C2N-SIDD_to_SIDD' and 'C2N-DND_to_SIDD' depending on the noisy images it is trained on.
  • mode: 'single' or 'dataset'.
  • data: Filename of clean image if mode is 'single', dataset of clean images if mode is 'dataset'.
  • gpu: GPU id. Currently this demo only supports single-GPU or CPU device.

Examples:

# Generate on single clean image
python test_generate.py --ckpt C2N-SIDD_to_SIDD.ckpt --mode single --data clean_ex1.png --gpu 0
python test_generate.py --ckpt C2N-DND_to_SIDD.ckpt --mode single --data clean_ex2.png --gpu 0

# Generate on clean images in a dataset
python test_generate.py --ckpt C2N-SIDD_to_SIDD.ckpt --mode dataset --data SIDD_clean_examples --gpu 0
python test_generate.py --ckpt C2N-DND_to_SIDD.ckpt --mode dataset --data SIDD_clean_examples --gpu 0

test_denoise.py:

  • config: Name of the configuration. Choose between 'C2N_DnCNN' and 'C2N_DIDN' depending on the denoiser to be used.
  • ckpt: Name of the checkpoint to load.
    • Name format: '[denoiser model]-[C2N train noisy set]to[C2N train clean set]-on_[denoiser train set]'
      • denoiser model: 'DnCNN' or 'DIDN'.
      • C2N train noisy set: Dataset that the noisy images for C2N training are sampled from.
      • C2N train clean set: Dataset that the clean images for C2N training are sampled from.
      • denoiser train set: Dataset that the clean->noisy images are generated from, to train the denoiser.
  • mode: 'single' or 'dataset'.
  • data: Filename of noisy/generated image if mode is 'single', dataset of noisy/generated images if mode is 'dataset'.
  • gpu: GPU id. Currently this demo only supports single-GPU or CPU device.

Examples:

# Denoise single noisy image
python test_denoise.py --config C2N_DnCNN --ckpt DnCNN-SIDD_to_SIDD-on_SIDD --mode single --data noisy_ex1_SIDD.png --gpu 0
python test_denoise.py --config C2N_DIDN --ckpt DIDN-SIDD_to_SIDD-on_SIDD --mode single --data noisy_ex1_SIDD.png --gpu 0
python test_denoise.py --config C2N_DIDN --ckpt DIDN-SIDD_to_DND-on_SIDD --mode single --data noisy_ex2_DND.png --gpu 0

# Denoise noisy images in a dataset
python test_denoise.py --config C2N_DnCNN --ckpt DnCNN-SIDD_to_SIDD-on_SIDD --mode dataset --data SIDD_benchmark --gpu 0
python test_denoise.py --config C2N_DIDN --ckpt DIDN-SIDD_to_SIDD-on_SIDD --mode dataset --data SIDD_benchmark --gpu 0
python test_denoise.py --config C2N_DIDN --ckpt DIDN-SIDD_to_DND-on_SIDD --mode dataset --data DND_benchmark --gpu 0

# Denoise the generated images from C2N
# You may copy the generated images in `results/[input_clean_data_path*]` to `data/[input_clean_data_path*]_generated.png`, for example.
python test_denoise.py --config C2N_DIDN --ckpt DIDN-SIDD_to_SIDD-on_SIDD --mode single --data clean_ex1_generated.png --gpu 0
python test_denoise.py --config C2N_DIDN --ckpt DIDN-SIDD_to_DND-on_SIDD --mode single --data clean_ex2_generated.png --gpu 0

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages