The official implementation of ACMMM 2022 Award Candidates "Learnability Enhancement for Low-light Raw Denoising: Where Paired Real Data Meets Noise Modeling" 【Arxiv / ACM DL】
Presentation video for the Best Paper Session: YouTube / BiliBili
Interested readers are also referred to our official Note about this work in Zhihu (Chinese).

Please switch to the TPAMI branch to get the latest version of PMN
TPAMI branch is compatible with this branch on evaluation, but has little difference on training

(2023.08.30): 📰 We have written an official blog post (in Chinese) for the TPAMI version of the paper. I believe this blog will help you gain a deeper understanding of our work.
(2023.08.03): 🎉 Our paper was accepted by IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 2023~~

• We present a learnability enhancement strategy to reform paired real data according to noise modeling. Our strategy consists of two efficient techniques: shot noise augmentation (SNA) and dark shading correction (DSC).
• SNA improves the precision of data mapping by increasing the data volume. Benefiting from the increased data volume, the mapping can promote the denoised images with clear texture.
• DSC reduces the complexity of data mapping by reducing the noise complexity. Benefiting from the reduced noise complexity, the mapping can promote the denoised images with exact colors.

• Python >=3.6, PyTorch >= 1.6
• Requirements: opencv-python, rawpy, exifread, h5py, scipy
• Platforms: Ubuntu 16.04, cuda-10.1
• Our method can run on the CPU, but we recommend you run it on the GPU

Please download the ELD dataset and SID dataset first, which are necessary for validation (or training).
ELD (official project): download (11.46 GB)
SID (official project): download (25 GB)

1. use get_dataset_infos.py to generate dataset infos

# Evaluate
python3 get_dataset_infos.py --dstname ELD --root_dir /data/ELD --mode SonyA7S2
python3 get_dataset_infos.py --dstname SID --root_dir /data/SID/Sony --mode evaltest
# Train
python3 get_dataset_infos.py --dstname SID --root_dir /data/SID/Sony --mode train

2. evaluate

# If you don't want to save pictures, please add '--save_plot False'. This option will save your time and space.
# ELD & SID
python3 trainer_SID.py -f runfiles/Ours.yml --mode evaltest
# ELD only
python3 trainer_SID.py -f runfiles/Ours.yml --mode eval
# SID only
python3 trainer_SID.py -f runfiles/Ours.yml --mode test

3. train

python3 trainer_SID.py -f runfiles/Ours.yml --mode train

The parameter sampling of SNA is implemented as the raw_wb_aug_torch function in the file data_process/process.py. The complete process of SNA has the CPU version in the Mix_Dataset class in data_process/real_datasets.py and the GPU version in the preprocess function in trainer_SID.py.

Both dark shading calibration and noise calibration require massive dark frames. We provide the calibration results directly. The calibration results for dark shading are stored in the resources folder.
The raw noise parameters at each ISO are stored in the get_camera_noisy_params_max function in process.py, which can be used to calibrate the noise parameters based on a noise model (P-G or ELD).

HINT: The calibration is based on a SonyA7S2 camera, which has the same sensor as the public datasets but not the same camera.

Note: The quantitative results on the SID dataset is different from the provided results in ELD(TPAMI) because only the central area is compared in ELD(TPAMI) on the SID dataset.

If you find our code helpful in your research or work please cite our paper.

@inproceedings{feng2022learnability,
    author = {Feng, Hansen and Wang, Lizhi and Wang, Yuzhi and Huang, Hua},
    title = {Learnability Enhancement for Low-Light Raw Denoising: Where Paired Real Data Meets Noise Modeling},
    booktitle = {Proceedings of the 30th ACM International Conference on Multimedia},
    year = {2022},
    pages = {1436–1444},
    numpages = {9},
    location = {Lisboa, Portugal},
    series = {MM '22}
}

@ARTICLE{feng2023learnability,
  author={Feng, Hansen and Wang, Lizhi and Wang, Yuzhi and Fan, Haoqiang and Huang, Hua},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={Learnability Enhancement for Low-Light Raw Image Denoising: A Data Perspective}, 
  year={2024},
  volume={46},
  number={1},
  pages={370-387},
  doi={10.1109/TPAMI.2023.3301502}
}

• SID (CVPR 2018)
• ELD (CVPR 2020 / TPAMI 2021)
• SFRN (ICCV 2021)

If you would like to get in-depth help from me, please feel free to contact me ([email protected] / [email protected]) with a brief self-introduction (including your name, affiliation, and position).