High-Fidelity Generative Image Compression

  • 2020-06-17 16:21:10
  • Fabian Mentzer, George Toderici, Michael Tschannen, Eirikur Agustsson
  • 42

Abstract

We extensively study how to combine Generative Adversarial Networks andlearned compression to obtain a state-of-the-art generative lossy compressionsystem. In particular, we investigate normalization layers, generator anddiscriminator architectures, training strategies, as well as perceptual losses.In contrast to previous work, i) we obtain visually pleasing reconstructionsthat are perceptually similar to the input, ii) we operate in a broad range ofbitrates, and iii) our approach can be applied to high-resolution images. Webridge the gap between rate-distortion-perception theory and practice byevaluating our approach both quantitatively with various perceptual metrics anda user study. The study shows that our method is preferred to previousapproaches even if they use more than 2x the bitrate.

 

Quick Read (beta)

loading the full paper ...