# Generative Adversarial Networks for Extreme Learned Image Compression

**Authors:** Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, Luc Van Gool  
**arXiv:** [1804.02958](https://arxiv.org/abs/1804.02958)  
**Date:** 2018-04-09  
**Summary page:** https://www.summarizepaper.com/en/paper/1804.02958v1/

## Key points

- Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool propose a framework for extreme learned image compression using Generative Adversarial Networks (GANs).
- Their approach combines a GAN formulation of learned compression with a generator/decoder that operates on the full-resolution image and is trained in combination with a multi-scale discriminator.
- This enables visually pleasing images to be produced at significantly lower bitrates than previous methods.
- The method can fully synthesize unimportant regions in the decoded image such as streets and trees from a semantic label map extracted from the original image.
- Only the preserved region and the semantic label map need to be stored, resulting in significant savings in storage space.
- A user study confirms that for low bitrates, their approach outperforms state-of-the-art methods by up to 67%, including the next-best method BPG.
- The authors' project website provides additional information about their research.

## Summary

In their paper "Generative Adversarial Networks for Extreme Learned Image Compression," Eirikur Agustsson, Michael Tschannen, Fabian Mentzer, Radu Timofte, and Luc Van Gool propose a framework for extreme learned image compression using Generative Adversarial Networks (GANs). Their approach combines a GAN formulation of learned compression with a generator/decoder that operates on the full-resolution image and is trained in combination with a multi-scale discriminator. This enables visually pleasing images to be produced at significantly lower bitrates than previous methods. Additionally, their method can fully synthesize unimportant regions in the decoded image such as streets and trees from a semantic label map extracted from the original image. This means that only the preserved region and the semantic label map need to be stored, resulting in significant savings in storage space. A user study confirms that for low bitrates, their approach outperforms state-of-the-art methods by up to 67%, including the next-best method BPG. The authors' project website provides additional information about their research.