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)

Methods	PSNR	2	3	4	SSIM	2	3	4
KPN(K=5)		33.41	30.47	28.80		0.913	0.860	0.826
KPN(K=7)		33.42	30.49	28.84		0.913	0.861	0.826
KPN(K=13)		33.44	30.52	28.92		0.913	0.863	0.829
KPN(K=19)		33.45	30.57	28.99		0.914	0.864	0.832
LP-KPN(K=5)		33.49	30.60	29.05		0.917	0.865	0.834

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).

Methods	PSNR	2	3	4	SSIM	2	3	4
Bicubic		-	28.6284	27.2378		-	0.8088	0.7643
Baseline (Our)		-	30.6003	28.6508		-	0.8630	0.8206

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

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.
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 --

4. Generate the training data
5. run *solver.prototxt to train the network

-- Testing --

6. 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:

(1) coarse align the image pairs;
(2) central crop the image pairs;
(3) finer align the cropped image pairs;
(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).