ArbSR

Pytorch implementation of "Learning A Single Network for Scale-Arbitrary Super-Resolution", ICCV 2021

[Project] [arXiv]

Highlights

• A plug-in module to extend a baseline SR network (e.g., EDSR and RCAN) to a scale-arbitrary SR network with small additional computational and memory cost.
• Promising results for scale-arbitrary SR (both non-integer and asymmetric scale factors) while maintaining the state-of-the-art performance for SR with integer scale factors.

Demo

Motivation

Although recent CNN-based single image SR networks (e.g., EDSR, RDN and RCAN) have achieved promising performance, they are developed for image SR with a single specific integer scale (e.g., x2, x3, x4). In real-world applications, non-integer SR (e.g., from 100x100 to 220x220) and asymmetric SR (e.g., from 100x100 to 220x420) are also necessary such that customers can zoom in an image arbitrarily for better view of details.

Overview

Requirements

• Python 3.6
• PyTorch == 1.1.0
• numpy
• skimage
• imageio
• cv2

Train

1. Prepare training data

1.1 Download DIV2K training data (800 training images) from DIV2K dataset or SNU_CVLab.

1.2 Cd to ./utils and run gen_training_data.m in Matlab to prepare HR/LR images in your_data_path as belows:

your_data_path
└── DIV2K
	├── HR
		├── 0001.png
		├── ...
		└── 0800.png
	└── LR_bicubic
		├── X1.10
			├── 0001.png
			├── ...
			└── 0800.png
		├── ...
		└── X4.00_X3.50
			├── 0001.png
			├── ...
			└── 0800.png

2. Begin to train

Run ./main.sh to train on the DIV2K dataset. Please update dir_data in the bash file as your_data_path.

Test

1. Prepare test data

1.1 Download benchmark datasets (e.g., Set5, Set14 and other test sets).

1.2 Cd to ./utils and run gen_test_data.m in Matlab to prepare HR/LR images in your_data_path as belows:

your_data_path
└── benchmark
	├── Set5
		├── HR
			├── baby.png
			├── ...
			└── woman.png
		└── LR_bicubic
			├── X1.10
				├── baby.png
				├── ...
				└── woman.png
			├── ...
			└── X4.00_X3.50
				├── baby.png
				├── ...
				└── woman.png
	├── Set14
	├── B100
	├── Urban100
	└── Manga109
		├── HR
			├── AisazuNihalrarenai.png
			├── ...
			└── YouchienBoueigumi.png
		└── LR_bicubic
			├── X1.10
				├── AisazuNihalrarenai.png
				├── ...
				└── YouchienBoueigumi.png
			├── ...
			└── X4.00_X3.50
				├── AisazuNihalrarenai.png
				├── ...
				└── YouchienBoueigumi.png

2. Begin to test

Run ./test.sh to test on benchmark datasets. Please update dir_data in the bash file as your_data_path.

Quick Test on An LR Image

Run ./quick_test.sh to enlarge an LR image to an arbitrary size. Please update dir_img in the bash file as your_img_path.

Visual Results

1. SR with Symmetric Scale Factors

2. SR with Asymmetric Scale Factors

3. SR with Continuous Scale Factors

Please try our interactive viewer.

Citation

@InProceedings{Wang2020Learning,
  title={Learning A Single Network for Scale-Arbitrary Super-Resolution},
  author={Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo},
  booktitle={ICCV},
  year={2021}
}

Acknowledgements

This code is built on EDSR (PyTorch) and Meta-SR. We thank the authors for sharing the codes.