Non-official Python Implementation of Polyblur

Slightly out-of-focus image	Deblurred result with Polyblur

This repository contains a Python implementation of Polyblur: Removing mild blur by polynomial reblurring in IEEE Transaction on Computational Imaging 2021 by Mauricio Delbracio, Ignacio Garcia-Dorado, Soungjoon Choi, Damien Kelly and Peyman Milanfar. We provide Pytorch and Numpy non-official implementations reproducing the quantitative and qualitative results of the original paper.

A description an analysis of the algorithm can be found in a companion IPOL paper: Breaking down Polyblur: Fast blind Correction of Small Anisotropic Blurs in Image Processing OnLine 2022 by Thomas Eboli, Jean-Michel Morel and Gabriele Facciolo.[paper] [demo].

Installation

---

First install the requirements with

pip install -r requirements.txt

Second, install the main module with

python setup.py install

The Pytorch implementation of this code runs ONLY with torch 1.10+ versions (because of torch.fft for the computation of the gradients with the pytorch implementation).

Alternatively, you can install the package from pypi.org:

pip install polyblur

Test

---

After installation from sources or via pip, you can test the blind deblurring technique with

python main.py --impath ./pictures/peacock_defocus.png --N 3 --alpha 6 --beta 1

You can modify several parameters, e.g. the number of Polyblur iterations and the deconvolution filter's parameters $alpha$ and $beta$. You may further overlead the default calibration parameters $c$ and $b$ if you need even more aggressive deblurring.

Import into your projects

---

After installation, the package is imported to any project with

import polyblur

It contains a main function polyblur_deblurring that is agnostic to the input argument img being a (H,W,C) or (H,W) Numpy array or (B,C,H,W) Pytorch tensor. An example reads:

from polyblur import polyblur_deblurring

im_restored = polyblur_deblurring(img, n_iter=3, alpha=2, beta=4, c=0.352, b=0.768)

It runs three iterations of Polyblur with default parameters and returns and Numpy array or a Pytorch tensor depending on the input.

A Pytorch module PolyblurDeblurring that inherits from torch.nn.Module is also provided. The module and the functional version above are fully differentiable. They thus can be be put within neurals networks are training losses. An example reads:

from polyblur import PolyblurDeblurring

deblurrer = PolyblurDeblurring()
deblurrer.to(device)

im_restored = deblurrer(img, n_iter=3, alpha=2, beta=4, c=0.352, b=0.768)

It runs three iterations of Polyblur with default parameters and returns a Pytorch tensor.

You'll most likely have to play with $n_{iter}$, $\alpha$ and $\beta$ to get the best deblurring results on one example, or finding a default set of parameter that works for most images. You can also tweack $c$ and $b$ that are the affine model parameters. The default values are the one calibrated with the code below that are similar to that reported in the original paper. You may change them too, but it is not recommanded. Please refer to our [paper] for details on the parameter tuning.

Calibration

---

We also provide a code for reproducing the calibration curves of the original paper with our FFT-based implementation of the gradients, or any other one one can image. To compute the slope and intercept of the model, please run

python calibration_blur_parameters.py

Below is an example of the predicted curves for 1% additive Gaussian noise. If you use the exact same implementation of the gradients as in this repo, you should find something like (0.362, -0.468) for the affine model parameters. You should first provide the path to the DVI2K validation set.

Principal blur direction	Orthogonal blur direction

Roadmap

---

This code reproduces the results of the TCI'22 article but is not as fast. Even though it is not the goal of this repository, here are some elements I will enhance in a near future:

• polybur.domain_transform should be implemented with a C++ or CUDA-based normalized convolution to be faster.
• the convolution are too slow with FFT. I should implement it with separable filters as in the TCI paper.

Contact

---

If you encounter any problem with the code, please contact me at [email protected].

Citation

---

If this code or the demo are useful for your research, please cite our paper.

@article{eboli22polyblur,
  title={Breaking down Polyblur: Fast Blind Correction of Small Anisotropic Blurs},
  author={Eboli, Thomas and Morel, Jean-Michel and Facciolo, Gabriele},
  journal={Image Processing On Line},
  volume={12},
  pages={435--456},
  year={2022},
}