Non-Local Recurrent Network for Image Restoration (NeurIPS 2018)

Paper | Bibtex

This is the original version of the NLRN code. A newer version can be found here.

Usage

Denoising

Preparing 400 images from BSDS500 for training

Under the root directory of this repository

mkdir -p data

Download the compressed 400 training images in grayscale here. They are converted from the color images in BSDS500. Move the compressed file BSDS500.tar.gz to ./data Uncompress them and generate the training file name list:

cd data
tar -zxf BSDS500.tar.gz
cd BSDS500
find train_gray_rgb2gray/*.png test_gray_rgb2gray/*.png > ../train.list

Preparing Set12 and BSD68 for evaluation

These two datasets can be downloaded from here. Move them to ./data

Training on 400 images (train and test) of BSD500

Under the root directory of this repository

bash train.sh

Released models

Noise level (sigma): 15 25 50

Prediction on Set12 and BSD68

Unzip the downloaded files and move them under ./checkpoints

bash test.sh

Image Super-Resolution

Released models

The model can be downloaded here.

Preparing Set5 and Set14 for evaluation

These two datasets can be downloaded from here. Move them to ./data

Prediction on Set5 and Set14

Unzip the downloaded files and move them under ./checkpoints

bash test_sr.sh

Dependencies

• Python 2.7
• TensorFlow 1.10

Bibtex

@inproceedings{liu2018non,
  title={Non-Local Recurrent Network for Image Restoration},
  author={Liu, Ding and Wen, Bihan and Fan, Yuchen and Loy, Chen Change and Huang, Thomas S},
  booktitle={Advances in Neural Information Processing Systems},
  pages={1680--1689},
  year={2018}
}