Single Image Super-Resolution, An AI Driven Image Enhancement Technique

Published on August 31, 2020

You know how in the movies, when the police or other authorities are looking at grainy surveillance footage and someone requests a nameless technician to ‘enhance’ it? Yeah, well before now that was just movie magic, but AI is making that possible. Some recent work in the field shows a lot of promise after achieving some pretty incredible results.

Image Credit: [Pixabay/Geralt]

The results are found in a paper titled ‘Single Image Super-Resolution via a Holistic Attention Network’ by researchers at the State Key Laboratory of Information Security of the Chinese Academy of Sciences, Northwestern University, among others.

Their method involves using a CNN (Convolutional Neural Network) to map the exact differences between a low and high resolution version of the same image. Convolutional Neural Networks are made up of many layers of artificial neurons that are all interconnected in ways that enable a machine to simulate learning. Through this means, the system is able to ‘understand’ what the image is based on a low resolution version, and can the recreate a clearer picture.

Each layer, however, can only respond to the layer above it, so detail can get a bit lost. The layers toward the bottom make an imprint upward, so there is some detail that you won’t find throughout.

This paper explains how overcome this loss of history within these type of neural networks.

The researchers uses two new modules that examine correlations between multiple layers so that they can detect the feature interdependencies of each layer. The new modules provide summation of all the missing data while amplifying the most significant details.

Check out the example below. It shows the kind of quality that SISR is able to produce vs legacy methods.

As you can see, this technology is a pretty significant leap forward. This software can be used for all kinds of things, from increasing the quality of surveillance footage to making your selfies come out a little sharper.

A vertical component added to the horizontal stack of layers within a CNN turned out to be critical component of getting image enhancing to work a lot better.

Featured Image Credit: [Pixabay/Geralt]

Enjoyed this video?
"No Thanks. Please Close This Box!"