Lossy Image Compression with Recurrent Neural Networks: from Human Perceived Visual Quality to Classification Accuracy

  • 2019-10-08 15:43:29
  • Maurice Weber, Cedric Renggli, Helmut Grabner, Ce Zhang
  • 0

Abstract

Deep neural networks have recently advanced the state-of-the-art in imagecompression and surpassed many traditional compression algorithms. The trainingof such networks involves carefully trading off entropy of the latentrepresentation against reconstruction quality. The term quality cruciallydepends on the observer of the images which, in the vast majority ofliterature, is assumed to be human. In this paper, we go beyond this notion ofquality and look at human visual perception and machine perceptionsimultaneously. To that end, we propose a family of loss functions that allowsto optimize deep image compression depending on the observer and to interpolatebetween human perceived visual quality and classification accuracy. Ourexperiments show that our proposed training objectives result in compressionsystems that, when trained with machine friendly loss, preserve accuracy muchbetter than the traditional codecs BPG, WebP and JPEG, without requiringfine-tuning of inference algorithms on decoded images and independent of theclassifier architecture. At the same time, when using the human friendly loss,we achieve competitive performance in terms of MS-SSIM.

 

Quick Read (beta)

loading the full paper ...