Unpaired Real-World Super-Resolution with Pseudo Controllable Restoration

Andrés Romero, Luc Van Gool, Radu TimofteWorkshops @ CVPR 2022 · OpenReview

Abstract

Current super-resolution methods rely on the bicubic down-sampling assumption in order to develop the ill-posed reconstruction of the low-resolution image. Not surprisingly, these approaches fail when using real-world low-resolution images due to the presence of artifacts and intrinsic noise absent in the bicubic setup. Consequently, attention is increasingly paid to techniques that alleviate this problem and super-resolve real-world images. As acquiring paired real-world datasets is a challenging problem, real-world super-resolution solutions are traditionally tackled as a blind problem or as an unpaired data-driven problem. The former makes assumptions about the downsampling operations, the latter uses unpaired training to learn the real distributions. Recently, blind approaches have dominated this problem by assuming a diverse bank of degradations, whereas the unpaired solutions have shown under-performance due to the two-staged training. In this paper, we propose an unpaired real-world super-resolution method that performs on par, or even better than blind paired approaches by introducing a pseudo-controllable restoration module in a fully end-to-end system.
BibTeX
@inproceedings{romero2022unpaired,
  author = {Andrés Romero and Luc Van Gool and Radu Timofte},
  title = {{Unpaired Real-World Super-Resolution with Pseudo Controllable Restoration}},
  booktitle = {Workshops @ CVPR},
  year = {2022},
  url = {https://openreview.net/forum?id=ru5xk7xomedM},
}

Loading related content...

Figure