Google’s neural networks have achieved the dream of CSI viewers everywhere: the company has revealed a new AI system capable of “enhancing” an eight-pixel square image, increasing the resolution 16-fold and effectively restoring lost data.
The neural network could be used to increase the resolution of blurred or pixelated faces, in a way previously thought impossible; a similar system was demonstrated for enhancing images of bedrooms, again creating a 32x32 pixel image from an 8x8 one.
Google’s researchers describe the neural network as “hallucinating” the extra information. The system was trained by being shown innumerable images of faces, so that it learns typical facial features. A second portion of the system, meanwhile, focuses on comparing 8x8 pixel images with all the possible 32x32 pixel images they could be shrunken versions of.
The two networks working in harmony effectively redraw their best guess of what the original facial image would be. The system allows for a huge improvement over old-fashioned methods of up-sampling: where an older system might simply look at a block of red in the middle of a face, make it 16 times bigger and blur the edges, Google’s system is capable of recognizing it is likely to be a pair of lips, and draw the image accordingly.
Of course, the system isn’t capable of magic. While it can make educated guesses based on knowledge of what faces generally look like, it sometimes won’t have enough information to redraw a face that is recognizably the same person as the original image. And sometimes it just plain screws up, creating inhuman monstrosities. Nonetheless, the system works well enough too fool people around 10% of the time, for images of faces.
Running the same system on pictures of bedrooms is even better: test subjects were unable to correctly pick the original image almost 30% of the time. A score of 50% would indicate the system was creating images indistinguishable from reality.
Although this system exists at the extreme end of image manipulation, neural networks have also presented promising results for more conventional compression purposes. In January, Google announced it would use a machine learning-based approach to compress images on Google+ four-fold, saving users bandwidth by limiting the amount of information that needs to be sent. The system then makes the same sort of educated guesses about what information lies “between” the pixels to increase the resolution of the final picture.