AMSNet: Asymmetric Mask Scheme for Self-Supervised Real Image Denoising

April 1, 2025 · View on GitHub

This repository is the official implementation of ECCV 2024 "Asymmetric Mask Scheme for Self-Supervised Real Image Denoising"

Asymmetric Mask Scheme for Self-Supervised Real Image Denoising
Xiangyu Liao*, Tianheng Zheng, Jiayu Zhong, Pingping Zhang, Chao Ren

Environment prepare

Dependencies

The project depends on the following packages:

PyTorch
TorchVision
OpenCV
Loguru
TIMM (PyTorch Image Models)
PyTorch Lightning
Weights & Biases (wandb)
Rich

Prerequisites

Before you start, make sure you have the following installed:

Python (3.10, 3.11)
pip (Python package installer)

Setup Steps

Create a virtual environment (optional but recommended, otherwise you are likely to encounter unknown and unnecessary errors)
```
conda create -n myenv python=3.10
conda activate myenv
```
Upgrade pip
```
pip install --upgrade pip
```

Install the dependencies

pip install torch torchvision opencv-python loguru timm pytorch-lightning wandb rich

Additional Notes

PyTorch Installation: Depending on your system's configuration, you might want to install PyTorch with specific CUDA versions. Check the PyTorch official website for more details.
OpenCV Installation: If you need specific OpenCV modules or face issues, refer to the OpenCV installation guide.
Weights & Biases Configuration: For using Weights & Biases (wandb), you might need to login to your account. Run wandb login and follow the instructions.

By following the above steps, you should have your environment set up and ready for development. If you encounter any issues, refer to the documentation of each package or seek help from the community.

Downloading Pretrained Weights

Download Pretrained Weights:
- Navigate to this link to access the pretrained weights.
Save to weights Directory:
- Once downloaded, place the pretrained weights into the weights directory of your project like this
.

Dataset prepare

Downloading the SIDD Medium Dataset for train

Download the SIDD Medium Dataset
- Visit the SIDD Medium Dataset website.
- Follow the instructions on the website to download the dataset.
Extract the Dataset

Once the download is complete, extract the contents of the dataset to a directory of your choice. For example:
```
zip -FF SIDD_Medium_Srgb_Parts.zip --out combined.zip
unzip combined.zip  
```

Downloading the SIDD Validation Dataset for validation

To evaluate your models, you need to download the SIDD Validation dataset.

Download the Validation Files
- Visit the SIDD Dataset page.
- Download the following files:
  - ValidationGtBlocksSrgb.mat
  - ValidationNoisyBlocksSrgb.mat

Organize the Validation Files

Place the downloaded files in a folder. The folder structure should be as follows:

datasets/
└── SIDD_Validation/
    ├── ValidationGtBlocksSrgb.mat
    ├── ValidationNoisyBlocksSrgb.mat

Using Your Own Noisy Images

If you do not want to use the SIDD Medium dataset, you can prepare your own noisy images.

Organize Your Noisy Images

Place your noisy images in a folder. The folder structure should be as follows:

datasets/
└── custom_noisy_images/
    ├── image1.png
    ├── image2.png
    ├── ...

Verify the Images

Ensure that your images are in a readable format (e.g., PNG, JPEG) and that they are accessible from your script or notebook.

Converting train Data to LMDB

To facilitate fast reading, you can convert your dataset into an LMDB format using the divide_sub_image.py script.

Example Command

python divide_sub_image.py --hw 512 512 --lmdb --suffix .PNG .bmp --path <Noisy PATH> --re "NOISY" --name <your_dataset_name> --output <your_output_path> --size 40 --step 256 256

Verify the lmdb file

The lmdb folder structure should be as follows:
```
datasets/
└── lmdb_folder/
    ├── data.mdb
    ├── lock.mdb
    ├── medium_keys.txt
    ├── meta_info.pkl
    └── meta_info.txt
```

Train

modify the config file

To train the denoise model, you need to modify the paths in the configuration file. Open the configs/restormer.yaml file and update the paths as follows:

data:
  train_dataset:
    class_path: codes.data.UnpairDataset
    init_args:
      path: your_path
      datatype: lmdb or images, make yourself happy
      max_len: any number you want
      crop_size: 320
      augment: True

more detail in codes/data.py

train the model

just run the command(take restormer as an example):

python main.py fit --config configs/restormer.yaml

Validation

modify the config file

To validate the denoise model, you need to modify the paths in the configuration file. Open the configs/restormer.yaml file and update the paths as follows:

  val_dataset:
    class_path: codes.data.SIDD_validation or codes.data.SIDD_benchmark
    init_args:
      sidd_val_dir: your_path
      len: 1280
  test_dataset:
    class_path: codes.data.SIDD_validation or codes.data.SIDD_benchmark
    init_args:
      sidd_val_dir: your_path
      len: 128

  val_dataset:
    class_path: codes.data.PairDataset
    init_args:
      please read the codes/data.py and input correct cfg
  test_dataset:
    class_path: codes.data.PairDataset
    init_args:
      please read the codes/data.py and input correct cfg

more detail in codes/data.py, and just run the command(take restormer as an example):

python main.py test/validate --config configs/restormer.yaml

Please note that the test or validate does not require the training dataset, just comment out the corresponding content in the configuration file

denoise for single image or imges

just run the command:

python main.py predict --config ./configs/restormer.yaml --data.predict_dataset.init_args.path=./images/test/test_sample.png

or 

python main.py predict --config ./configs/restormer.yaml --data.predict_dataset.init_args.path=./images/test/

Citation

If you find this code useful for your research, please consider citing:

@inproceedings{liao2024asymmetric,
  title={Asymmetric Mask Scheme for Self-supervised Real Image Denoising},
  author={Liao, Xiangyu and Zheng, Tianheng and Zhong, Jiayu and Zhang, Pingping and Ren, Chao},
  booktitle={European Conference on Computer Vision},
  pages={199--215},
  year={2024},
  organization={Springer}
}

Acknowledgements

This project is built on source codes shared by AP-BSN, SpatiallyAdaptiveSSID, CVF-SID, DeamNet, Restormer, NAFNet, SCPGabNet, timm and pytorch.