Joint Denoising and Demosaicking with Green Channel Prior for Real-world Burst Images

Implement of our GCP-Net.

Arxiv: https://arxiv.org/abs/2101.09870

IEEE Final Version: https://ieeexplore.ieee.org/document/9503334

Testing

pretrain model:

• store in gcpnet_model/600000_G.pth

  Testing Vid4 and REDS4:

• set data_mode in test.py to 'REDS4' and 'Vid4', the default noise level is set as the 'high noise level' mentioned in the paper.

python /codes/test.py

• To Note that: we only put a subset of REDS4 and Vid4 to save space, please download the full testset in official website, RED and Vid. More detail can refer to data preparation

Testing on real captured images:

• SC_burst (Smartphone burst) Dataset: we captured 16 burst images using smartphones, and put one burst of Scene 00 in sub_SC_burst. We unified raw format and saved SC_burst in ".MAT", where the raw data and metadata are stored.
• Whole dataset: BaiduYun with password d8u8.
• Bayer pattern: Our model is trained only use RGGB. Thus when testing raw images with other patterns (e.g., GRBG), don't forget to unified bayer pattern to RGGB by padding or flipping.

  python /code/test_real.py

Training

• training data preparation: Please refer to the "Video Super-Resolution" part of data preparation. To create LMDB dataset, please run create_lmdb.py.

• change training options in train_GCP_Net.yml

  python -m torch.distributed.launch --nproc_per_node=2 --master_port=4540 train.py -opt options/train/train_GCP_Net.yml --launcher pytorch

Environment

• Refer to the requirement.txt
• We utilize pytorch 1.2 and the deformable version does not support pytorch > 1.3. Thus when you use newest pytorch, please replace deformable version to newest (refer to BasicSR).

Citation

@article{guo2021joint,
  title={Joint Denoising and Demosaicking with Green Channel Prior for Real-world Burst Images},
  author={Guo, Shi and Liang, Zhetong and Zhang, Lei},
  journal={arXiv preprint arXiv:2101.09870},
  year={2021}
}

Acknowledgement

This repo is built upon the framework of EDVR, and we borrow some code from Unprocessing denoising, thanks for their excellent work!