Plug-and-play Deep Image Prior

This repository contains the code for plug-and-play deep image prior.

Requirement

• pytorch==1.2.0
• numpy
• matplotlib
• skimage
• glob
• bm3d
• prox_tv

Usage

The directory data contains 25 images from CelebA dataset. We use these data to get the PSNR curves and table in the paper. The directory data_visual has pepper and monarch images, which are used for visual results in the paper.

You can directly run the python files in

• experiments/inpainting
• experiments/inpainting_vis
• experiments/uniform_denoising
• experiments/uniform_denoising_vis
• experiments/superrsolution_vis

The names of python files are the priors they use. For DIP + others, it use our plug-and-play deep image prior method. Without DIP in the name, it uses plug-and-play prior method.

When the directory has the suffix vis, it will use the data at data_visual. Otherwise, it will use the data at data. The result will be stored in the corresponding data directory.