Home

Awesome

Toward Real-World Single Image Super-Resolution (RealSR)

Dataset:

Raw images can be downloaded HERE

Captured device: (Canon 5D3 and Nikon D810) + (24∼105mm, f/4.0 zoom lens)

A part of this dataset was used in the RealSR challenge in NTIRE 2019 (in conjunction with CVPR 2019).

Version 1: 234 scenes (204 scenes for training & 30 scenes for testing), as reported in the original paper (HR has the same resolution as LR).

Link: Google Drive, Baidu Drive (code: n77c)

Version 2: 559 scenes (459 scenes for training & 100 scenes for testing), the extended version (HR has the same resolution as LR).

Link: Google Drive, Baidu Drive (byyz)

MethodsPSNR234SSIM234
KPN(K=5)33.4130.4728.800.9130.8600.826
KPN(K=7)33.4230.4928.840.9130.8610.826
KPN(K=13)33.4430.5228.920.9130.8630.829
KPN(K=19)33.4530.5728.990.9140.8640.832
LP-KPN(K=5)33.4930.6029.050.9170.8650.834

Version 3: 559 scenes (459 scenes for training & 100 scenes for testing), the extended version (HR and LR have different resolution).

Link: Google Drive, Baidu Drive(code: 2n93)

Detail for training & testing: Trained on the RGB domain and tested on Y channel (images from Version 3).

MethodsPSNR234SSIM234
Bicubic-28.628427.2378-0.80880.7643
Baseline (Our)-30.600328.6508-0.86300.8206

Visualization (zooming factor: 4) <br> More results can be downloaded here.

<div align="center"> <img src="https://github.com/csjcai/RealSR/blob/master/Sample1.png"/> <img src="https://github.com/csjcai/RealSR/blob/master/Sample2.png"/> </div>

Code:

Model: Pretrained Caffe models

  1. Models for PSNR/SSIM
  2. Models for Visualization

The above provided models (for quantitative metrics and visual quality) are both trained with the loss ratio in 1:1:1. <br> We select different models at different epochs for different purposes.

Caffe: training code & testing code

  1. Download the new layers in folder 'Layer'
  2. Modify the caffe.proto (Path: caffe/src/caffe/proto/)
  3. Compile Caffe and Matcaffe (installation)

-- Training --

  1. Generate the training data
  2. run *solver.prototxt to train the network

-- Testing --

  1. run Test.m

Alignment code:

  1. Put your own image pairs in the folder and modify the path
  2. run Demo.m in folder 'Alignment'
  3. Central region crop
Pipeline:<br>

(1) coarse align the image pairs;<br> (2) central crop the image pairs;<br> (3) finer align the cropped image pairs;<br> (4) discard those misaligned image pairs.

Citation:

If you find this work useful for your research, please cite:

@inproceedings{cai2019toward,
  title={Toward real-world single image super-resolution: A new benchmark and a new model},
  author={Cai, Jianrui and Zeng, Hui and Yong, Hongwei and Cao, Zisheng and Zhang, Lei},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision},
  year={2019}
}
@inproceedings{cai2019ntire,
  title={Ntire 2019 challenge on real image super-resolution: Methods and results},
  author={Cai, Jianrui and Gu, Shuhang and Timofte, Radu and Zhang, Lei},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops},
  year={2019}
}

Contact:

Please contact me if there is any question (Jianrui CAI: csjcai@comp.polyu.edu.hk).