Awesome

PAMS: Quantized Super-Resolution via Parameterized Max Scale

This resposity is the official implementation of our ECCV2020 paper.

Our implementation is based on EDSR(PyTorch).

Dependent

• Python 3.6
• PyTorch == 1.1.0
• coloredlogs >= 14.0
• scikit-image

Datasets

Please download DIV2K datasets from here for training and benchmark datasets for testing. Then, organise the dataset directory as follows:

datasets
  benchmark
  DIV2K

Usage

• train

python main.py --scale 4 --k_bits 8 \
--pre_train ../pretrained/edsr_baseline_x4.pt \
--data_test Set14 --save edsr_x4/8bit/ \
--dir_data ./datasets --model EDSR               

• test

python main.py --scale 4 --k_bits 8 \ 
--pre_train ../pretrained/edsr_x4 --save_results \
--data_test Set5+Set14+B100+Urban100 \
--save edsr_x4/8bit/ --dir_data ./datasets  
--test_only --refine [REFINE] --model EDSR 

set --refine to the saved model path for testing model.

More runnig scripts can be found in run.sh.

• PSNR/SSIM

After saving the images, modify path inmetrics/calculate_PSNR_SSIM.m to generate results.

matlab -nodesktop -nosplash -r "calculate_PSNR_SSIM('$dataset',$scale,$bit);quit"

refer to metrics/run.sh for more details.

Trained Models

We also provide our baseline models below. Enjoy your training and testing! Google Drive.

Citations

If our paper helps your research, please cite it in your publications:

@article{li2020pams,
  title={PAMS: Quantized Super-Resolution via Parameterized Max Scale},
  author={Li, Huixia and Yan, Chenqian and Lin, Shaohui and Zheng, Xiawu and Li, Yuchao and Zhang, Baochang and Yang, Fan and Ji, Rongrong},
  journal={arXiv preprint arXiv:2011.04212},
  year={2020},
  publisher={Springer}
}