Home

Awesome

Photon Limited Non-Blind Deblurring Using Algorithm Unrolling

Pytorch code for Photon Limited Non-Blind Deblurring Using Algorithm Unrolling - published at Transactions on Computational Imaging

<img src="https://user-images.githubusercontent.com/20774419/177593608-9b5ccba2-ca3d-485a-9542-5f08df8e081a.png" width="800">

Pretrained model here

Instructions

  1. Create a local copy of repository using the following commands

    foor@bar:~$ git clone https://github.com/sanghviyashiitb/poisson-deblurring.git
foor@bar:~$ cd poisson-deblurring
foor@bar:~/poisson-deblurring$

  2. Download the pretrained model into model_zoo from the link here

  3. To test the network using synthetic data, run the file

    foo@bar:~/poisson-deblurring$ python3 demo_synthetic.py

    Output:

    <img src="results/demo_synthetic.png" alt="demo_synthetic" width="400"/>

  4. Download the zip file containing real dataset into the main directory and unzip using the following command:

    foo@bar:~/poisson-deblurring$ unzip real_data.zip -d data/

  5. To test the network using real data, run the file

    foo@bar:~/poisson-deblurring$ python3 demo_synthetic.py  --idx=11

    (Variable idx represents the file index and can be any integer from [0,29] )

    demo_real

    Output: PSNR: 29.08, SSIM: 0.696

Training

Before running train.py, add clean images (for example Flickr2K) in the data/training and data/val folders.

Citation

@ARTICLE{9903556,
 author={Sanghvi, Yash and Gnanasambandam, Abhiram and Chan, Stanley H.},
 journal={IEEE Transactions on Computational Imaging}, 
 title={Photon Limited Non-Blind Deblurring Using Algorithm Unrolling}, 
 year={2022},
 volume={8},
 number={},
 pages={851-864},
 doi={10.1109/TCI.2022.3209939}}

Feel free to ask your questions/share your feedback at sanghviyash95@gmail.com