This repository provides a simple Pytorch implementation of plug and play ADMM with examples.
The notebook pnp_admm_example.ipynb contains a example in which a pretrained convnet gaussian denoiser is downloaded (120MB) and plugged into ADMM for solving a motion deblur, inpainting, and super-resolution problem. You can also view in Google Colab.
The notebook denoiser_training.ipynb contains code for training a denoiser from scratch using a subset of ImageNet as trainingset. The trainingset will be download automatically (250MB). You can also view in Google Colab.
| Degraded | PnP output | Ground truth | |
|---|---|---|---|
| Motion deblur |  |
 |
 |
| Inpainting |  |
 |
 |
| Super-resolution |  |
 |
 |
References:
S. Venkatakrishnan, C. Bouman, and B. Wohlberg, “Plug-and-play priors for model based reconstruction,” in Proc. IEEE Global Conference on Signal and Information Processing, 2013, pp. 945–948.