Gated Texture CNN

This repository contains the code for the following paper:

Gated Texture CNN for Efficient and Configurable Image Denoising [arXiv]

Kaito Imai, Takamichi Miyata

If you find this work useful in your research and use this code, please cite:

@article{imai2020gated,
title={Gated Texture CNN for Efficient and Configurable Image Denoising},
author={Imai, Kaito and Miyata, Takamichi},
journal={arXiv preprint arXiv:2003.07042},
year={2020}
}

Quick Modification Demo

You can run a demo local machine or colab.

To run the test/train code, please clone this repo and see below instructions.

Dependencies

We provide requirements.txt.

• Python 3 (Recommend to use Anaconda)

• numpy

• PyTorch >= 1.0

• torchvision

• NVIDIA GPU + CUDA(>10)

• opencv

• tqdm

• PyYAML

• OmegaConf

• Comet.ml,(for logging)

• Apex (option for train)

Evaluation

2. Download all test set in ./Dataset/test/.

   • set12
   • bsd68
   • cbsd68
   • urban100

3. Run the test script.

   sh run_test.sh
   

You can change test set like

python main.py confs/GTCNN experiment.test_only=True experiment.best_model=Color50_GTCNN_D1 dataset.test_set=[YourDataset]

More detailed see the How to Edit Config.

Note that our model was trained in the BGR.

Demo of Texture Modification

2. Open the jupyter notebook Demo_Texture_Modification

3. Run Demo!

Training

1. Download Div2K train and val set in ./Dataset/train/ and ./Dataset/val/.

2. Run the training script.

   sh run_train.sh
   

If you do not have enough RAM (required at least 64GB), you could use random crop options (We don't confirm the performance, expect not bad). Also, We recommend use apex if the VRAM of GPU doesn't enough to training.

• random crop: sh run_random_crop_train.sh
• Train with apex: sh run_apex_train.sh

Note you should set the flag model.GTL_stage_option=outconv_slim, if the model depth is more than 5.

How to Edit Config

If you want to change the configuration of experiments (e.g,. batch size, depth, apex), you can edit the default config (GTCNN.yaml) by OmegaConf.

OmegaConf docs

OmegaConf live tutorial

GTCNN.yaml contains config of experiment, dataset and model .

You can edit the config like below

python main.py confs/GTCNN experiment.color=1 experiment.device_ids=[0,1,2,4] experiment.opt_level=O1 experiment.saveImage=True dataset.train_set=[YourDataset] dataset.test_set=[YourDataset1,YourDataset2] model.depth=5

• Examples
  • Change color to gray or color: experiment.color=1 or =3
  • Change the noise level: experiment.sigma=40
  • Use multi or change GPU: experiment.device_ids=[0] or =[3] or =[0,1,2,3]
  • Run test: experiment.test_only=True
  • Load your pretrain: experiment.bestmodel =FileName
  • save the denoised images in the local folder: experiment.saveImage=True
  • Use Apex: :experiment.opt_level=O1 or =O2
  • Change BatchSize: experiment.batchsize=256
  • Change the depth of model: model.depth=5
  • Use your testset: dataset.test_set=[YourDataset1,YourDataset2]
  • Use comet log: experiment.comet_disabled=False
  • For Depth>5: model.GTL_stage_option=outconv_slim

And more details see the GTCNN.yaml