Google Brain, the neural-networked machine learning arm of Google's experimental side, has created software that can take a grainy, low-resolution image, and produce a higher-resolution image with many missing details filled in. Once the realm of science fiction, the classic "zoom and enhance" is now possible in the real world thanks to two separate methods that Google's software employs equally. One method searches Google's databases for similar images, downscales them to find a match, then either presents the better image if it's an exact match, or uses data from that image to rebuild the less detailed one. This one is called the conditioning network. The other one, called the prior network, follows more traditional machine learning principles by learning about a given subject, person, or object, then trying to identify and reconstruct them in a given image.
In real-world testing, Google Brain managed to fool viewers into thinking that they were looking at an original, high-resolution image about 10% of the time with pictures of celebrities, while participants were fooled 28% of the time looking at upscaled pictures of bedrooms. That may sound awful at a glance, but considering what this technology is, how new it is, and the implications of it, that kind of accuracy this early on is actually pretty amazing. Bicubic scaling, which is the technology currently used for enhancement of grainy images and upscaling older games in emulators, did not manage to goad even a single participant into thinking that they were seeing a real high-resolution image.
The implications here, especially in the future, are pretty obvious, and widespread. Grainy security videos could start showing a perpetrator's face accurately. Grainy night shots taken on a smartphone could be patched up with ease and made to look as good as something shot in optimal lighting. Low-resolution images on the web can be upscaled on the fly to save data. The list goes on and on, but the use cases and implications of the technology behind this breakthrough run even deeper. Machine learning, neural networking, and even artificial intelligence are all improving and advancing at breakneck speeds, and real-world applications like this seem to imply that Google is leading the charge.