This is to extend the SR-GAN and ESR-GAN to do blind super-resolution. The problem statement is to reconstruct the high-resolution image from low resolution (a.k.a. super-resolution), but without knowing how the low resolution is derived from the original high-resolution image, i.e., blind super-resolution.

The contribution of this paper: (1) A process to create high-quality synthetic dataset for super-resolution, (2) a network for SR, especially using U-Net in the discriminator with spectral normalization to increase discriminator capability.

The classical degradation model includes blur, downsampling, noise, and JPEG compression:

\[\mathbf{x} = D(\mathbf{y}) = [(\mathbf{y} \oast \mathbf{k})\]

first-order vs high-order degradation modeling for real-world degradation sinc filter for ringing and overshoot artifacts discriminator of more powerful capability gradient feedback from discriminator needs to be more accurate for local detail nehancement U-net design with spectral normalization (SN) regularization

Further Reading

  • First SR network: SRCNN, 9, 10
  • Blind SR survey: 28

Bibliographic data

   title = "Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data",
   author = "Xintao Wang and Liangbin Xie and Chao Dong and Ying Shen",
   booktitle = "Proc. ICCV",
   year = "2021",
   arXiv = "2107.10833",