Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data

  • 2021-07-22 17:43:24
  • Xintao Wang, Liangbin Xie, Chao Dong, Ying Shan
  • 48

Abstract

Though many attempts have been made in blind super-resolution to restorelow-resolution images with unknown and complex degradations, they are still farfrom addressing general real-world degraded images. In this work, we extend thepowerful ESRGAN to a practical restoration application (namely, Real-ESRGAN),which is trained with pure synthetic data. Specifically, a high-orderdegradation modeling process is introduced to better simulate complexreal-world degradations. We also consider the common ringing and overshootartifacts in the synthesis process. In addition, we employ a U-Netdiscriminator with spectral normalization to increase discriminator capabilityand stabilize the training dynamics. Extensive comparisons have shown itssuperior visual performance than prior works on various real datasets. We alsoprovide efficient implementations to synthesize training pairs on the fly.

 

Quick Read (beta)

loading the full paper ...