PPON

Pytorch implemention of "Progressive Perception-Oriented Network for Single Image Super-Resolution"

[arXiv] [Information Sciences]

The schematics of the Progressive Perception-Oriented Network

The example results

Visualization of intermediate feature maps

The inference architecture of our PPON.

Testing

Pytorch 1.1

• Download PIRM_dataset and unzip it into folder Test_Datasets
• Download Checkpoint and put them into folder ckpt
• Run testing:

python test_PPON.py --test_hr_folder Test_Datasets/PIRM_Val/ --test_lr_folder Test_Datasets/PIRM_Val_LR/

Training

• Download DF2K (DIV2K + Flickr2k) training datasets and rename the images to 00xxxx.png (e.g., 003450.png)
• Convert png file to npy file

python scripts/png2npy.py --pathFrom /path/to/DF2K/ --pathTo /path/to/DF2K_decoded/

• Run training x4 model (stage 1, content)

python train.py --nEpochs 1000 --test_every 690 --which_model "content" --lr_steps [200, 400, 600, 800] --save_path 'ckpt_stage1'

(stage 2, structure)

python train.py --nEpochs 300 --test_every 138 --which_model "structure" --pixel_weight 0 --structure_weight 1 --lr_steps [100, 150, 200, 250] --save_path 'ckpt_stage2'

(stage3, perceptual)

python train.py --nEpochs 300 --test_every 138 --which_model "perceptual" --pixel_weight 0 --feature_weight 1 --gan_weight 0.005 --lr_steps [100, 150, 200, 250] --save_path 'ckpt_stage3'

PI VS LPIPS

As illustrated in the above picture, we can obviously see that the PI score of EPSR3 (2.2666) is even better than HR (2.3885), but EPSR3 shows unnatural and lacks proper texture and structure.

Example Results

Citation

If you find PPON useful in your research, please consider citing:

@article{Hui-PPON-2019,
  title={Progressive Perception-Oriented Network for Single Image Super-Resolution},
  author={Hui, Zheng and Li, Jie and Gao, Xinbo and Wang, Xiumei},
  journal={arXiv:1907.10399v1},
  year={2019}
}

@article{Hui-PPON-2021,
  title={Progressive Perception-Oriented Network for Single Image Super-Resolution},
  author={Hui, Zheng and Li, Jie and Gao, Xinbo and Wang, Xiumei},
  journal={Information Sciences},
  volume = {546},
  pages = {769-786},
  year={2021}
}