MCSN

This repository is an official PyTorch implementation of the paper

@article{li2021lightweight,
  title={A lightweight multi-scale channel attention network for image super-resolution},
  author={Li, Wenbin and Li, Juefei and Li, Jinxin and Huang, Zhiyong and Zhou, Dengwen},
  journal={Neurocomputing},
  year={2021},
  publisher={Elsevier}
}

The code is built on EDSR (PyTorch) and tested on Ubuntu 18.04 environment with TitanX 2080Ti GPU.

Dependencies

• Python 3.6
• PyTorch = 1.2.0
• numpy
• skimage
• imageio
• matplotlib
• tqdm

Data

All scale factor(x2,x3,x4,x8) data:

1. Training data DIV2K(800 training + 100 validtion images)
2. Benchmark data (Set5, Set14, B100, Urban100, Manga109)

Train

1.Cd to './MCSN/src', run the following commands to train models.

python main.py --model MCSN--scale 2 --save mcsn_x2  --n_resblocks 3  --lr 1e-4  --n_feats 64 --res_scale 1 --batch_size 16 --n_threads 6 
python main.py --model MCSN--scale 3 --save mcsn_x3  --n_resblocks 3  --lr 1e-4  --n_feats 64 --res_scale 1 --batch_size 16 --n_threads 6 
python main.py --model MCSN--scale 4 --save mcsn_x4  --n_resblocks 3  --lr 1e-4  --n_feats 64 --res_scale 1 --batch_size 16 --n_threads 6 
python main.py --model MCSN--scale 8 --save mcsn_x8  --n_resblocks 3  --lr 1e-4  --n_feats 64 --res_scale 1 --batch_size 16 --n_threads 6 

Test

python main.py --model MCSN --data_test Set5+Set14+B100+Urban100+Manga109  --scale 4 --pre_train ../experiment/mscn_x4/model/model_best.pt --test_only  --self_ensemble

Results

(1) MCSN Network

(2) Parameters

(3) PSNR&SSIM X2