liuguoyou / physics_sr

This repository is an official PyTorch implementation of the paper "Image Formation Model Guided Deep Image Super-Resolution" from AAAI 2020.

PHYSICS_SR

This repository is an official PyTorch implementation of the paper "Image Formation Model Guided Deep Image Super-Resolution" from AAAI 2020.
The code is built on EDSR (PyTorch) and tested on Ubuntu 16.04 environment (Python3.6, PyTorch_0.4.1, CUDA8.0, cuDNN5.1) with Tesla V100/1080Ti GPUs.

Dependencies

• ubuntu16.04
• Python 3.6(Recommend to use Anaconda)
• PyTorch0.4.1
• numpy
• skimage
• imageio
• matplotlib
• tqdm
• cv2

Get started

Trainset:

We use the DIV2K dataset to train our models. You can download it from here

Benechmarks:

You can evaluate our models with widely-used benchmark datasets:

Set5 - Bevilacqua et al. BMVC 2012,

Set14 - Zeyde et al. LNCS 2010,

B100 - Martin et al. ICCV 2001,

Urban100 - Huang et al. CVPR 2015.

Models

All the models(X2, X3, X4) can be downloaded from GoogleDrive.

Quicktest with benchmark

You can test our super-resolution algorithm with benchmarks. Please organize the testset in testset folder like this:

|--testset  
    |--Set5  
        |--LR
            |--X2
                |--babyx2.png  
                     ：   
                     ： 
            |--X3
            |--X4
        |--HR
            |--baby.png  
                 ：   
                 ： 

Then, run the following commands:

cd code
python main.py --dir_data ../testset --data_test Set5 --model physics_sr --pre_train ../models/X2/model_best.pt --scale 2 --save physics_sr_x2 --save_results --test_only

And generated results can be found in ./experiment/physics_sr_x2/results/

• To test all scales, you can modify the options(pre_train, scale) of the command above.
• To test other benchmarks, you can modify the option(data_test) of the command above.
• To change the save root, you can modify the option(save) of the command above.

How to train

If you have downloaded the trainset, please make sure that the trainset has been organized as follows:

|--DIV2K
    |--train  
        |--DIV2K_train_LR_bicubic
            |--X2
                |--0001x2.png  
                     ：   
                     ： 
            |--X3
            |--X4
        |--DIV2K_train_HR
            |--0001.png  
                 ：   
                 ： 

The command for training is as follow:

cd code
python main.py --dir_data <your root> --data_test DIV2K --model physics_sr --scale 2 --save physics_sr_x2 --save_results

The trained model can be found in ./experiment/physics_sr_x2/model

• To train all scales, you can modify the option(scale) of the command above.
• To test other benchmarks, you can modify the option(data_test) of the command above.
• To change the save root, you can modify the option(save) of the command above.

Citation

If our work is useful in your research or publication, please cite our work:

@inproceedings{pan2020physics_sr,
    title={Image Formation Model Guided Deep Image Super-Resolution},
    author={Jinshan Pan, Yang Liu, Deqing Sun, Jimmy Ren, Ming-Ming Cheng, Jian Yang, and Jinhui Tang},
    booktitle={AAAI},
    year={2020}
}

Acknowledgements

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