ChaoXiang661 / LBNet

This repository is an official PyTorch implementation of our paper "Lightweight Bimodal Network for Single-Image Super-Resolution via Symmetric CNN and Recursive Transformer". (IJCAI 2022)

LBNet-Pytorch: Lightweight Bimodal Network for Single-Image Super-Resolution via Symmetric CNN and Recursive Transformer

This repository is an official PyTorch implementation of the paper "Lightweight Bimodal Network for Single-Image Super-Resolution via Symmetric CNN and Recursive Transformer".

Paper | Project

Dependencies

Python>=3.7 
PyTorch>=1.1
numpy 
skimage 
imageio 
matplotlib 
tqdm

For more informaiton, please refer to EDSR

Dataset

We used DIV2K dataset to train our model. Please download it from here or SNU_CVLab.

You can evaluate our models on several widely used benchmark datasets, including Set5, Set14, B100, Urban100, Manga109. Note that using an old PyTorch version (earlier than 1.1) would yield wrong results.

Results

All our SR images can be downloaded from Results.[百度网盘][提取码:xpuh]

All pretrained model can be found in IJCAI2022_LBNet.

The following PSNR/SSIMs are evaluated on Matlab R2017a and the code can be referred to Evaluate_PSNR_SSIM.m.

Training

  LBNet: num_heads = 8
  
# LBNet x4
python main.py --scale 4 --model LBNet --save experiments/LBNet_X4

# LBNet x3
python main.py --scale 3 --model LBNet --save experiments/LBNet_X3

# LBNet x2
python main.py --scale 2 --model LBNet --save experiments/LBNet_X2

  LBNet-T：num_heads = 6, 'dim//2' in util/rlutrans.py/EffAttention is changed to 'dim'

# LBNet-T x4
python main.py --scale 4 --model LBNet-T --save experiments/LBNet-T_X4

# LBNet-T x3
python main.py --scale 3 --model LBNet-T --save experiments/LBNet-T_X3

# LBNet-T x2
python main.py --scale 2 --model LBNet-T --save experiments/LBNet-T_X2

Testing

  LBNet: num_heads = 8
  
# LBNet x4
python main.py --scale 4 --model LBNet --pre_train test_model/LBNet/LBNet-X4.pt --test_only --save_results --data_test Set5

  LBNet-T：num_heads = 6, 'dim//2' in util/rlutrans.py/EffAttention is changed to 'dim'

# LBNet-T x4
python main.py --scale 4 --model LBNet-T --pre_train test_model/LBNet-T/LBNet-T_X4.pt --test_only --save_results --data_test Set5

Performance

Our LBNet is trained on RGB, but as in previous work, we only reported PSNR/SSIM on the Y channel.

Model	Scale	Params	Multi-adds	Set5	Set14	B100	Urban100	Manga109
LBNet-T	x2	404K	49.0G	37.95/0.9602	33.53/0.9168	32.07/0.8983	31.91/0.9253	38.59/0.9768
LBNet	x2	731K	153.2G	38.05/0.9607	33.65/0.9177	32.16/0.8994	32.30/0.9291	38.88/0.9775
LBNet-T	x3	407K	22.0G	34.33/0.9264	30.25/0.8402	29.05/0.8042	28.06/0.8485	33.48/0.9433
LBNet	x3	736K	68.4G	34.47/0.9277	30.38/0.8417	29.13/0.8061	28.42/0.8559	33.82/0.9460
LBNet-T	x4	410K	12.6G	32.08/0.8933	28.54/0.7802	27.54/0.7358	26.00/0.7819	30.37/0.9059
LBNet	x4	742K	38.9G	32.29/0.8960	28.68/0.7832	27.62/0.7382	26.27/0.7906	30.76/0.9111

Visual comparison

SR images reconstructed by our LBNet have richer detailed textures with better visual effects.

Model complexity

LBNet gains a better trade-off between model size, performance, inference speed, and multi-adds.

Acknowledgements

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

Citation

If you use any part of this code in your research, please cite our paper:

@article{gao2022lightweight ,
  title={Lightweight Bimodal Network for Single-Image Super-Resolution via Symmetric CNN and Recursive Transformer},
  author={Gao, Guangwei and Wang, Zhengxue and Li, Juncheng and Li, Wenjie and Yu, Yi and Zeng, Tieyong},
  journal={arXiv preprint arXiv:2204.13286},
  year={2022}
}