Attentive deep network for blind motion deblurring in dynamic scenes

TensorFlow Implementation of CVIU paper "Attentive deep network for blind motion deblurring in dynamic scenes"
Refer to https://csyhquan.github.io/category/c_publication.html for our more publications.

Dataset

We trained our model using the dataset from DeepDeblur_release. Please put the training dataset into training_set/, and testing set into testing_set/.

Test on our pretrain model

Our code is easy to go with:

python run_model.py --phase test --height 720 --width 1280 --gpu gpu_id

The quantitative results of PSNR and SSIM is calculted using MATLAB based on the deblurring results. Here we can get a PSNR result of about 31.22dB with python codes.

Training

Training our model is easy to go with:

python run_model.py --phase train --batch batch_size --lr 0.0001 --epoch 4000

Defocus Deblurring

Our model also works well on defocus deblurring, we train our model with DPDD datatset and obtain a SOTA performance (about 25.22dB on PSNR). The pretrained model is placed in checkpoints/defocus/. We only use the single color image as input, i.e. train_c and test_c of DPDD dataset.

To test our defocus deblurring performance on DPDD testing set, you can easy to go with:

python run_model.py --phase test --height 1120 --width 1680 --gpu gpu_id --model defcous --steps 105000 --input_path input_dir --output_path out_dir

Citation

If you think this work is useful for your research, please cite the following paper.

@article{XU2021103169,
title = {Attentive deep network for blind motion deblurring on dynamic scenes},
journal = {Computer Vision and Image Understanding},
volume = {205},
pages = {103169},
year = {2021},
issn = {1077-3142},
doi = {https://doi.org/10.1016/j.cviu.2021.103169},
url = {https://www.sciencedirect.com/science/article/pii/S1077314221000138},
author = {Yong Xu and Ye Zhu and Yuhui Quan and Hui Ji}
}

Acknowledgement

Many parts of this code is adapted from SRN

Thanks the authors for sharing codes for their great works