Deep Degradation Prior for Real-World Super-Resolution

Kyungdeuk Ko, Bokyeung Lee, Jonghwan Hong, David Han, Hanseok Ko

Research output: Contribution to conferencePaperpeer-review

2 Citations (Scopus)


Real-world Super-Resolution (SR) is a very challenging task to reconstruct a higher resolution image from a real-world image which generally has unexpected artifacts and distortions. Most of the high performance SR methods rely on availability of LR-HR paired datasets, target domain images, or degradation priors. These information, however, are usually not available in real world, thus these methods are not often practical for real world obtained images. Recent studies related to real-world SR mainly focus on constructing the LR-HR paired dataset. The methods estimate the noises and the blur kernels from real-world images to generate a new training set. However, these methods use the degradation only for dataset construction. In this paper, we propose a novel real-world SR method called Deep Degradation Prior-based SR (DDP-SR). Upon completion of training, denoising network and kernel estimation network within DDP-SR becomes capable of extracting degradation representation of any given input image. Thus, the model works regardless of whether the input image from the same or the different domain of the source images. As such, DDP-SR achieves generalization performance on images from different domain while it also outperforms the state-of-the-art methods qualitatively and quantitatively.

Original languageEnglish
Publication statusPublished - 2021
Event32nd British Machine Vision Conference, BMVC 2021 - Virtual, Online
Duration: 2021 Nov 222021 Nov 25


Conference32nd British Machine Vision Conference, BMVC 2021
CityVirtual, Online

Bibliographical note

Publisher Copyright:
© 2021. The copyright of this document resides with its authors.

ASJC Scopus subject areas

  • Artificial Intelligence
  • Computer Vision and Pattern Recognition


Dive into the research topics of 'Deep Degradation Prior for Real-World Super-Resolution'. Together they form a unique fingerprint.

Cite this