Enhancing low-resolution photos can be a pain. The reason is because it's usually impossible to get enough details out of a small number of pixels to make a clearer image. However, researchers from Google have figured out a way to do that.
Using Artificial Intelligence (AI), Google with its Brain division has created a software to extract details out of tiny, pixelated source images. In short, the software is able to "zoom in and enhance", similar to what we usually see in crime movies.
Ryan Dahl, Mohammad Norouzi and Jonathon Shlens came up with a pixel recursive super resolution model that is able to synthesize details in low-resolution images by combining two neural networks.
The first one is the conditioning network. What it does is mapping a low-resolution image source image against other similar but high resolution images to make it understand what the image should at least look like. The network downsizes other high-resolution images to a resolution similar to the source image to to make a match.
The second is the prior network. It uses an implementation of PixelCNN to try and add realistic high-resolution details to image source.
So after the conditioning network uses AI to feed on high-resolution images. And when the source image is upscaled, the prior network tries to match the pixels to match what it "understands."
While the result aren't perfect, but Google has came with an idea to bring fictions in movies to reality.
Using grainy 8x8 pixel (64 source pixels) images of people and bedrooms in the tests, the AI can enhance the practically unusable pictures into something that even humans can now differentiate.
The potential for such technology is huge. As a start, Google can use it to enhance photos on Google+ and on Chrome, for example. Google is already familiar with image compression technology, like RAISR. And this "zoom and enhance" technology can certainly improve the lacks image compression technologies still have.
While the technology is not yet perfect, the system is fairly competent. It can certainly make grainy unusable images more presentable.
Google Brain and DeepMind are two of Alphabet's deep learning research division. The two have published researches that include cryptographic algorithms, AlphaGo AI that defeated professional human Go players, and more.