This repository is our solution for NTIRE2019 Real Super-Resolution Challenge.
We, team IVIP-LAB, won the 9th PSNR and Top5 SSIM in the final phase of NTIRE2019 Real Super-resolution challenge.
Our paper will be published in CVPR 2019 Workshop. [pdf] [poster]
Encoder-Decoder Residual Network (EDRN)
Quantitative Results
| Method | PSNR (dB) | SSIM | Runtime (s) |
|---|---|---|---|
| Baseline | 26.89 | 0.78 | -- |
| EDRN (ours) | 28.79 | 0.84 | 47.08 |
Qualitative Results
- Python (tested on release 3.5)
- PyTorch (tested on release 0.4.1)
- CUDA9.0
- cuDNN7.1
Our code is tested on Ubuntu 16.04 environment with an NVIDIA GTX 1080Ti GPU.
Please download the dataset from here, and then put the downloaded dataset into ntire file.
$ python main.py --save EDRN --save_results --save_models --model edrn --patch_size 128 --lr_decay 50 --n_GPUs 1 --chop
Quick start (Demo) to reproduce our results. Please download our pretrained model from here.
$ python main.py --save EDRN --save_results --model edrn --n_GPUs 1 --chop --pre_train ../experiment/model/EDRN.pt --self_ensemble --test_only
If you find this work useful in your reseach, please cite our paper.
@InProceedings{Cheng_2019_CVPR_Workshops,
author = {Cheng, Guoan and Matsune, Ai and Li, Qiuyu and Zhu, Leilei and Zang, Huaijuan and Zhan, Shu},
title = {Encoder-Decoder Residual Network for Real Super-Resolution},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2019}
}
This code is built on EDSR (PyTorch). We are grateful to the authors for sharing their codes of EDSR.