Atrous Convolution and Residual Network (ARN)

Introduce

1. This is net can be used in image restoration. It can make images with rain or fog be more clear.

2. The net named ARN, it means 'Atrous Convolution and Residual Network'. It had a group of Atrous Convolution and two blocks of Residual model.

3. In the val of dataset it can get this result in quantitative analysis.

   	PSNR	SSIM	iter_time
   ARN	24.29dB	85.88%	5.7ms

4. These are qualitative analysis

Environmental requirements

1. cuda 10.1
2. cudnn 7.6.5
3. python=3.7
4. pytorch=1.4.0
5. torchvision=0.5.0
6. scikit-image
7. numpy, scipy, tqdm, pillow, opencv (maybe loss something, this is my first time writing doc.)

Model

checkpoints/ARN_model.pth

Training environment

	Detail
System	Ubuntu 16.04
CPU	Inter Xeon E5-2620 v3 2.40GHz
RAM	32GB DDR4 2133MHz
GPU	NVIDIA GTX1080Ti
Learning_rate	0.007
Batch_size	64
Epoch	200

Datasets

You can download from , 'https://pan.baidu.com/s/1Fdmc5Ua2su9o7rNv0R8UCQ 提取码：2333'(BaiDuYun)

Make these dataset to 'data/train/groundtruth' and 'data/train/rain'

The datasets come from 'https://github.com/hotndy/SPAC-SupplementaryMaterials' , 'Rain12600' and 'Rain1400'(https://xmu-smartdsp.github.io/), and make some process get mine datasets.

Demo

Demo.py

Make the model in to path 'checkpoints/model.pth'

python demo.py

Train

Train.py

If use pre-trained model, make the true path of model.

python train.py

Val

val.py

Make the val datasets into true path

python val.py

Reference

[1] Ren D, Zuo W, Hu Q, et al. Progressive image deraining networks: A better and simpler baseline[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019: 3937-3946.

[2] Chen L C, Zhu Y, Papandreou G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 801-818.