Progressive Residual Networks for Image Super-Resolution

[Applied Intelligence]

This repository is for PRNet introduced in the following paper

The code is based on EDSR (PyTorch) and tested on Ubuntu 18.04 environment (Python3.6, PyTorch_0.4.0, CUDA8.0, cuDNN5.1) with Titan X/1080Ti/Xp GPUs. .

The architecture of our proposed Progressive Residual Networks (PRNet). The details about our proposed PRNet can be found in our main paper.

If you find our work useful in your research or publications, please consider citing:

@article{Jin2020Progressive,
  title={Progressive residual networks for image super-resolution},
  author={Jin Wan and Hui Yin and Ai-Xin Chong and Zhi-Hao Liu},
  journal={Applied Intelligence},
  volume={50},
  number={7},
  pages={1620–1632},
  year={2020},
}

Train
Test
Results
Acknowledgements

Train

Begin to train

Cd to 'src/', run the following scripts to train models. You can use scripts in file 'demo' to train models for our paper.

# BI, scale 2, 3, 4, 8
# PRNet in the paper (x2)
python main.py --model PRNetx2 --scale 2 --save prnet_x2 --n_resblocks 32 --n_feats 256 --res_scale 0.1 --reset

# PRNet in the paper (x3) - from PRNet (x2)
#python main.py --model PRNetx3 --scale 3 --save prnet_x3 --n_resblocks 32 --n_feats 252 --res_scale 0.1 --reset 

# PRNet in the paper (x4) - from PRNet (x2)
#python main.py --model PRNetx4 --scale 4 --save prnet_x4 --n_resblocks 32 --n_feats 256 --res_scale 0.1 --reset --pre_train ../experiment/prnet_x2/model/model_best.pt

# PRNet in the paper (x8) - from PRNet (x4)
#python main.py --model PRNetx8 --scale 8 --save prnet_x8 --n_resblocks 32 --n_feats 256 --res_scale 0.1 --reset --pre_train ../experiment/prnet_x2/model/model_best.pt

Test

Clone this repository:

git clone https://github.com/jinwan1994/PRNet.git

All the models (BIX2/3/4/8) can be downloaded from BaiduYun(5f61), place the models to ./experiment/model/.

Cd to '/src', run the following scripts.

You can use scripts in file 'demo' to produce results for our paper.

# Standard benchmarks (Ex. PRNetx2)
#python main.py --model PRNetx2 --data_test Set5+Set14+B100+Urban100+Manga109 --scale 2  --n_feats 256 --n_resblocks 32 --res_scale 0.1  --pre_train ../experiment/prnet_x2/model/model_best.pt --test_only # --self_ensemble --save_results --save_gt

# Standard benchmarks (Ex. PRNet_x3)
#python main.py --model PRNetx3 --data_test Set5+Set14+B100+Urban100+Manga109 --scale 3  --n_feats 252 --n_resblocks 32 --res_scale 0.1  --pre_train ../experiment/prnet_x3/model/model_best.pt --test_only # --self_ensemble --save_results --save_gt

# Standard benchmarks (Ex. PRNet_x4)
#python main.py --model PRNetx4 --data_test Set5+Set14+B100+Urban100+Manga109 --scale 4  --n_feats 256 --n_resblocks 32 --res_scale 0.1  --pre_train ../experiment/prnet_x4/model/model_best.pt --test_only # --self_ensemble --save_results --save_gt

# Standard benchmarks (Ex. PRNet_x8)
#python main.py --model PRNetx8 --data_test Set5_x8+Set14_x8+B100_x8+Urban100_x8+Manga109_x8 --scale 8 --res_scale 0.1 --n_feats 256 --n_resblocks 32 --pre_train ../experiment/prnet_x8/model/model_latest_282.pt --test_only # --save_results --save_gt --self_ensemble

Finally, SR results and PSNR/SSIM values for test data are saved to ./experiment/model/*. (PSNR/SSIM values in our paper are obtained using matlab)

Results

Quantitative Results

Benchmark SISR results. Average PSNR/SSIM for scale factor x2, x3 and x4 on datasets Set5,Set14, Manga109, BSD100 and Urban100.

Visual Results

Visual comparison for x8 SR on Manga109 and Urban100. dataset.

Acknowledgements

This code is built on EDSR (PyTorch). We thank the authors for sharing their codes.

jinwan1994/PRNet