llnl / gpp Goto Github PK

Exploiting the patch manifold for inverse imaging

Generative Patch Priors for Practical Compressive Image Recovery -- (pdf)

Rushil Anirudh, Suhas Lohit, Pavan Turaga. In WACV, 2021. Image of size 1536×1024 recovered with GPP at a measurement rate of 10% using a GAN trained on CIFAR-C images of size 32x32.

Overview of the approach for patch-based compressive sensing

Dependencies

There are two versions of GPP, with python 3.6:

• Pytorch 1.6.0 (also works with 1.4.0+)
• Tensorflow 1.8.0 We have included the corresponding patch-generators trained on CIFAR-32 for each framework. There are some performance differences; we report results from Tensorflow in the paper, but the PyTorch numbers are better on most examples (!!).

The code also has the option of using BM3D as part of the inverse patch transform in order to mitigate some of the patching artifacts. Any implementation should work, we used two of them -- pybm3d and bm3d. GPP does not need it to work, but will work better with BM3D.

This is ongoing work, if you find errors or bugs please let us know!

Description

This section will be updated in the coming days. Please see the paper for details about GPP and its workings.

Citation

If you find this code useful in your work, please consider citing our paper:

@inproceedings{Anirudh2021GPP,
  title={Generative Patch Priors for Practical Compressive Image Recovery},
  author={Anirudh, Rushil and Lohit, Suhas and Turaga, Pavan},
  booktitle={WACV},
  year={2021}
}

License

This code is distributed under the terms of the MIT license. All new contributions must be made under this license. LLNL-CODE- 812404 SPDX-License-Identifier: MIT