Code for ICML 2017 paper "From Patches to Images: A Nonparametric Generative Model"
Python
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.

README.md

hdp-grid-image-restoration

This repository contains pre-trained models and demo code for ICML 2017 paper "From Patches to Images: A Nonparametric Generative Model"

Prerequisites

bnpy

Please first intall the latest version of our bnpy package. Instructions could be found at https://github.com/bnpy/bnpy.

Pillow

Pillow is a maintained fork of Python Imaging Library (PIL). It could be installed with pip by running

pip install Pillow

Installation

Clone the repository

git clone https://github.com/bnpy/hdp-grid-image-restoration.git

Running the demo code

python demo.py

It will run the three demos in the demo.py file.

The first demo is written in function demo_eDP(), in which the Barbara image is first polluted by additive white Gaussian noisy with standard deviation 25, and then gets denoised by our external DP Grid method. The denoised image would be saved in png format, and should match the middle right plot shown in Figure 3 of our paper.

The second demo is written in function demo_HDP(), in which the airplane image is polluted by the same amount of noise as above. It would get denoised by our HDP Grid method. The saved output should match the bottom right plot in Figure 8.

The last demo is written in function demo_inpainting(), where the HDP Grid method would inpaint the New Orleans image, and the saved output should match the bottom right plot shown in Figure 7.

Depending on the speed of your computer, the two denoising demos may each take up to 15~30 minutes to run, and the inpainting one could take about two hours.

Reference

@inproceedings{ji2017patches,
    title={From Patches to Images: A Nonparametric Generative Model},
    author={Ji, Geng and Hughes, Michael C and Sudderth, Erik B},
    title={International Conference on Machine Learning},
    year={2017}
}

For questions, please contact: Geng Ji (gji@cs.brown.edu).