Efficient Single Image Super-Resolution Using Dual Path Connections with Multiple Scale Learning
This repository is for EMSRDPN introduced in the following paper
Bin-Cheng Yang and Gangshan Wu, "Efficient Single Image Super-Resolution Using Dual Path Connections with Multiple Scale Learning", [arxiv]
It's an extension to a conference paper
Bin-Cheng Yang. 2019. Super Resolution Using Dual Path Connections. In Proceedings of the 27th ACM International Conference on Multimedia (MM ’19), October 21–25, 2019, Nice, France. ACM, NewYork, NY, USA, 9 pages. https://doi.org/10.1145/3343031.3350878
The code is built on EDSR (PyTorch) and tested on Ubuntu 16.04 environment (Python3.7, PyTorch_1.1.0, CUDA9.0) with Titan X/Xp/V100 GPUs.
Contents
Introduction
Deep convolutional neural networks have been demonstrated to be effective for SISR in recent years. On the one hand, residual connections and dense connections have been used widely to ease forward information and backward gradient flows to boost performance. However, current methods use residual connections and dense connections separately in most network layers in a sub-optimal way. On the other hand, although various networks and methods have been designed to improve computation efficiency, save parameters, or utilize training data of multiple scale factors for each other to boost performance, it either do super-resolution in HR space to have a high computation cost or can not share parameters between models of different scale factors to save parameters and inference time. To tackle these challenges, we propose an efficient single image super-resolution network using dual path connections with multiple scale learning named as EMSRDPN. By introducing dual path connections inspired by Dual Path Networks into EMSRDPN, it uses residual connections and dense connections in an integrated way in most network layers. Dual path connections have the benefits of both reusing common features of residual connections and exploring new features of dense connections to learn a good representation for SISR. To utilize the feature correlation of multiple scale factors, EMSRDPN shares all network units in LR space between different scale factors to learn shared features and only uses a separate reconstruction unit for each scale factor, which can utilize training data of multiple scale factors to help each other to boost performance, meanwhile which can save parameters and support shared inference for multiple scale factors to improve efficiency. Experiments show EMSRDPN achieves better performance and comparable or even better parameter and inference efficiency over SOTA methods.
Train
Prepare training data
-
Download DIV2K training data (800 training images for x2, x3, x4 and x8) from DIV2K dataset and Flickr2K training data (2650 training images) from Flickr2K dataset.
-
Untar the download files.
-
Using src/generate_LR_x8.m to generate x8 LR data for Flickr2K dataset, you need to modify 'folder' in src/generate_LR_x8.m to your directory to place Flickr2K dataset.
-
Specify '--dir_data' in src/option.py to your directory to place DIV2K and Flickr2K datasets.
For more informaiton, please refer to EDSR(PyTorch).
Begin to train
-
Cd to 'src', run the following scripts to train models.
You can use scripts in file 'demo.sh' to train models for our paper.
To train a fresh model using DIV2K dataset
CUDA_VISIBLE_DEVICES=0,1 python3.7 main.py --scale 2+3+4+8 --test_scale 2+3+4+8 --save EMSRDPN_BIx2348 --model EMSRDPN --epochs 5000 --batch_size 16 --patch_size 48 --n_GPUs 2 --n_threads 16 --SRDPNconfig A --ext sep --data_test Set5 --reset --decay 1000-2000-3000-4000-5000 --lr_patch_size --data_range 1-3450 --data_train DIV2K
To train a fresh model using Flickr2K dataset
CUDA_VISIBLE_DEVICES=0,1 python3.7 main.py --scale 2+3+4+8 --test_scale 2+3+4+8 --save EMSRDPN_BIx2348 --model EMSRDPN --epochs 5000 --batch_size 16 --patch_size 48 --n_GPUs 2 --n_threads 16 --SRDPNconfig A --ext sep --data_test Set5 --reset --decay 1000-2000-3000-4000-5000 --lr_patch_size --data_range 1-3450 --data_train Flickr2K
To train a fresh model using both DIV2K and Flickr2K datasets to reproduce results in the paper, you need copy all the files in DIV2K_HR/ to Flickr2K_HR/, copy all the directories in DIV2K_LR_bicubic/ to Flickr2K_LR_bicubic/, then using the following script
CUDA_VISIBLE_DEVICES=0,1 python3.7 main.py --scale 2+3+4+8 --test_scale 2+3+4+8 --save EMSRDPN_BIx2348 --model EMSRDPN --epochs 5000 --batch_size 16 --patch_size 48 --n_GPUs 2 --n_threads 16 --SRDPNconfig A --ext sep --data_test Set5 --reset --decay 1000-2000-3000-4000-5000 --lr_patch_size --data_range 1-3450 --data_train Flickr2K
To continue a unfinished model using DIV2K dataset, the processes for other datasets are similiar
CUDA_VISIBLE_DEVICES=0,1 python3.7 main.py --scale 2+3+4+8 --test_scale 2+3+4+8 --save EMSRDPN_BIx2348 --model EMSRDPN --epochs 5000 --batch_size 16 --patch_size 48 --n_GPUs 2 --n_threads 16 --SRDPNconfig A --ext sep --data_test Set5 --resume -1 --decay 1000-2000-3000-4000-5000 --lr_patch_size --data_range 1-3450 --data_train DIV2K --load EMSRDPN_BIx2348
Test
Quick start
-
Download benchmark dataset from BaiduYun (access code: 20v5), place them in directory specified by '--dir_data' in src/option.py, untar it.
-
Download EMSRDPN model for our paper from BaiduYun (access code: d2ov) and place them in 'experiment/'. Other multiple scale models can be downloaded from BaiduYun (access code: z5ey).
-
Cd to 'src', run the following scripts to test downloaded EMSRDPN model.
You can use scripts in file 'demo.sh' to produce results for our paper.
To test a trained model
CUDA_VISIBLE_DEVICES=0 python3.7 main.py --scale 2+3+4+8 --test_scale 2+3+4+8 --save EMSRDPN_BIx2348_test --model EMSRDPN --epochs 5000 --batch_size 16 --patch_size 48 --n_GPUs 1 --n_threads 16 --SRDPNconfig A --ext sep --data_test Set5+Set14+B100+Urban100+Manga109 --reset --decay 1000-2000-3000-4000-5000 --lr_patch_size --data_range 1-3450 --data_train DIV2K --pre_train ../experiment/EMSRDPN_BIx2348.pt --test_only --save_results
To test a trained model using self ensemble
CUDA_VISIBLE_DEVICES=0 python3.7 main.py --scale 2+3+4+8 --test_scale 2+3+4+8 --save EMSRDPN_BIx2348_test+ --model EMSRDPN --epochs 5000 --batch_size 16 --patch_size 48 --n_GPUs 1 --n_threads 16 --SRDPNconfig A --ext sep --data_test Set5+Set14+B100+Urban100+Manga109 --reset --decay 1000-2000-3000-4000-5000 --lr_patch_size --data_range 1-3450 --data_train DIV2K --pre_train ../experiment/EMSRDPN_BIx2348.pt --test_only --save_results --self_ensemble
To test a trained model using multiple scale infer
CUDA_VISIBLE_DEVICES=0 python3.7 main.py --scale 2+3+4+8 --test_scale 2+3+4+8 --save EMSRDPN_BIx2348_test_multi_scale_infer --model EMSRDPN --epochs 5000 --batch_size 16 --patch_size 48 --n_GPUs 1 --n_threads 16 --SRDPNconfig A --ext sep --data_test Set5 --reset --decay 1000-2000-3000-4000-5000 --lr_patch_size --data_range 1-3450 --data_train DIV2K --pre_train ../experiment/EMSRDPN_BIx2348.pt --test_only --save_results --multi_scale_infer
Results
All the test results can be download from BaiduYun (access code: oawz).
Citation
If you find the code helpful in your resarch or work, please cite the following papers.
@InProceedings{Lim_2017_CVPR_Workshops,
author = {Lim, Bee and Son, Sanghyun and Kim, Heewon and Nah, Seungjun and Lee, Kyoung Mu},
title = {Enhanced Deep Residual Networks for Single Image Super-Resolution},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {July},
year = {2017}
}
@inproceedings{2019Super,
title={Super Resolution Using Dual Path Connections},
author={ Yang, Bin Cheng },
booktitle={the 27th ACM International Conference},
year={2019},
}
@misc{yang2021efficient,
title={Efficient Single Image Super-Resolution Using Dual Path Connections with Multiple Scale Learning},
author={Bin-Cheng Yang and Gangshan Wu},
year={2021},
eprint={2112.15386},
archivePrefix={arXiv},
primaryClass={eess.IV}
}
Acknowledgements
This code is built on EDSR (PyTorch). We thank the authors for sharing their code.