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

WIP: fast evaluation with custom ops

An older version of the NLRN code can be found here.

Usage

Denoising

Preparing BSD500 for training

mkdir -p data/bsd500
wget -O data/bsd500/BSR_bsds500.tgz http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/BSR/BSR_bsds500.tgz
`cd data/bsd500 && tar -xvf BSR_bsds500.tgz`
mkdir -p data/bsd500/flist1
find data/bsd500/BSR/BSDS500/data/images/train/*.jpg data/bsd500/BSR/BSDS500/data/images/test/*.jpg > data/bsd500/flist1/train.flist
find data/bsd500/BSR/BSDS500/data/images/val/*.jpg > data/bsd500/flist1/eval.flist

Preparing Set12 and BSD68 for evaluation

git clone https://github.com/cszn/DnCNN.git data/denoise
find data/denoise/testsets/Set12/*.png > data/set12.flist
find data/denoise/testsets/BSD68/*.png > data/bsd68.flist

Training on flist1 (train and test) of BSD500

python trainer.py --dataset denoise --train-flist data/bsd500/flist1/train.flist --eval-flist data/bsd500/flist1/eval.flist --model nlrn --job-dir debug
# or incremental trainer by number of recurrent states
python incremental_trainer.py --dataset denoise --train-flist data/bsd500/flist1/train.flist --eval-flist data/bsd500/flist1/eval.flist --model nlrn --job-dir debug

Pre-trained models

12 recurrent states/with correlation propagation: sigma 15, sigma 25, sigma 50.

15 recurrent states/without correlation propagation: sigma 15, sigma 25, sigma 50.

Prediction on Set12 and BSD68

python -m datasets.denoise --noise-sigma SIGMA --model-dir MODEL_DIR --input-dir data/denoise/testsets/Set12 --output-dir ./output/Set12
python -m datasets.denoise --noise-sigma SIGMA --model-dir MODEL_DIR --input-dir data/denoise/testsets/BSD68 --output-dir ./output/BSD68

MODEL_DIR is the directory of tf.saved_model and located in export/Servo/ of job_dir.

Super-resolution

Preparing Set5 Set14 BSD100 Urban100 for evaluation

wget -O data/SR_testing_datasets.zip http://vllab.ucmerced.edu/wlai24/LapSRN/results/SR_testing_datasets.zip
`cd data/ && unzip SR_testing_datasets.zip`

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}
}

Ding-Liu / NLRN