Towards General Low-Light Raw Noise Synthesis and Modeling (ICCV 2023)

Technical Report:

LRD is a general framework for low-light raw noise synthesis and modeling. Specifically, we synthesize the signal-dependent and signal-independent noise in a physics- and learning-based manner, respectively. In this way, our method can be considered as a general model, that is, it can simultaneously learn different noise characteristics for different ISO levels and generalize to various sensors.

🐍Synthesis Pipeline

📂LRD Dataset

The LRD dataset is available for download at Baidudisk (vj53) or Dropbox or One Drive

✔️Pretrained Models

The pretrained models for raw denoise is available at Baidudisk (ujzm) or Google Drive.

Due to the business license, pre-trained models for noise synthesis are not available now.

📑Intallation

Download the ELD dataset and SID dataset at the following links:

ELD (official project): download (11.46 GB)
SID (official project): download (25 GB)

• Install the conda environment

conda create -n lrd python=3.8
conda activate lrd

• Install Pytorch

conda install pytorch==1.9 torchvision cudatoolkit=10.2 -c pytorch

• Install Packages for Raw Image

pip install ExifRead
pip install h5py

Note that the rawpy version is using ELD's customized rawpy, which can be downloaded from GoogleDrive or Baidudisk (0lby). To build rawpy from source, please first compile and install the LibRaw library following the official instructions, then type pip install -e . in the customized rawpy directory.

• Install other packages

pip install tqdm
pip install lmdb
pip install glob
pip install imageio
pip install PyYAML
pip install timm
pip install patchify
conda install -c conda-forge scipy
pip install opencv-python
pip install tensorboardx
pip install scikit-image
pip install colour
pip install pylab-sdk
pip install pillow

🚗Run

• Quick start to raw denoising tests for SID dataset

python test_denoise_SID.py

• Quick start to raw denoising tests for ELD dataset

python test_denoise_ELD.py

• Quick start to raw denoising tests for LRD dataset

python test_denoise_LRD.py

Due to the business license, the source code for noise synthesis is not available now.

📖 Citation

If you find our LRD model useful for you, please consider citing 📣

@InProceedings{Zhang_2023_ICCV,
    author    = {Zhang, Feng and Xu, Bin and Li, Zhiqiang and Liu, Xinran and Lu, Qingbo and Gao, Changxin and Sang, Nong},
    title     = {Towards General Low-Light Raw Noise Synthesis and Modeling},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {10820-10830}
}

📧Contact

If you have any question, feel free to email [email protected].