ADVANCES You can't create detail...can you?

30 January 2018

SIG: Imaging Science

Artificial intelligence upscales low resolution to high resolution in new software.

EnhanceNet-PAT takes a low resolution image (right) and adds texture to an upscaled high resolution version (centre) that compares very well with the original high resolution image (right). Image © MPI-IS

We all know what happens. A friend gives you a file shot at low resolution on a smart phone and asks if there is anything you can do with it to make it bigger. You know that you can make it bigger, but if there is no detail there it can't be created.

Well, researchers at the Max Planck Institute for Intelligent Systems in Tübingen, Germany, have developed software that might just change all that. Called EnhanceNet-PAT, it uses machine learning and artificial intelligence to recognise image content and apply texture as you enlarge the image.

The software learns by trial and error, much as humans do. "The algorithm is given the task of sampling millions of low-resolution images to a high-resolution one, and is then shown the original. Notice the difference? OK, then learn from your mistake'" says Mehdi M.S. Sajjdi, who with Dr Michael Hersch and Prof Dr Bernhard Schölkopf created the new software. Once EnhanceNet-PAT has been trained, they continue, it no longer needs the original images.

EnhanceNet-PAT is more efficient than any comparable software, the team claims. The difference is than EnhanceNet-PAT doesn't try to be pixel-perfect. Instead it creates what it believes are faithful textures. In the sample above, where it sees the bird it has applied feather textures, where it thinks there are twigs it applies a wood surface texture.

A ready-trained Creative Commons version of the software is available to those familiar with Linux or commands in OS X. However, we do wonder how long it will be before EnhanceNet-PAT is snapped up by a major software company such as Adobe or even appears as an app on your smartphone.

The original paper, with links to the software, is available by clicking here.

Gary Evans ASIS FRPS

Chair, Imaging Science Group