Paper:
https://arxiv.org/abs/2105.14766 (arXiv)
https://www.ieee-jas.net/en/article/doi/10.1109/JAS.2022.105563 (IEEE/CAA JAS)

The code is for the work:

@article{liang2021BaMBNet,
  title={BaMBNet: A Blur-Aware Multi-Branch Network for Dual-Pixel Defocus Deblurring},
  author={Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma},
  journal={IEEE/CAA Journal of Automatica Sinica},
  volume={},
  number={},
  pages={},
  year={2022},
}

pytorch == 1.7.1
kornia == 0.4.1
opencv == 4.4.0

Please refer to the official repo at Defocus deblurring using dual-pixel data.

Note that the image list of small training dataset used in meta-learning can be found in Google Drive.

The deblurring images of DPDBlur dataset are available at Google Drive

• Step 1: train COC network to estimate the blur amounts of DP data.

  python blur_train.py -opt option/train/COC_Dataset_Train.yaml

• Step 2: prepare the COC maps for deblurring training.

  python blur_test.py -opt option/test/COC_Dataset_Test.yaml

• Step 3: train the deblurred network.

  python train.py -opt option/train/Deblur_Dataset_Trained.yaml

python test.py -opt option/test/Deblur_Dataset_Test.yaml

• Results of DPDD
• Results of Pixel5
• Results of dual_pixel_defocus_estimation_deblurring

This project is under the CC-BY-NC 4.0 license. See LICENSE for details.