“Can we enhance this photo?” Sure, but only if you’re in one of those movies or TV shows where the computer expert turns a completely pixelated image of the suspect into a clear, recognizable one. Now, in a case of life imitating art, that capability is actually emerging from the Google Brain research team.
Researchers Ryan Dahl, Mohammad Norouzi and Jonathon Shlens developed the resolution-enhancing software using neural networks. First, their super resolution model checked a set of unidentifiable 8×8-pixel photos of celebrity faces against a bunch of high-res photos that had been reduced to the same small pixel size to find overlap.
Then the model deployed an architecture called PixelCNN — the CNN stands for “convolutional neural network” — to add detail to the 8×8-pixel images. As Sebastian Anthony explained on Ars Technica UK, this network learns how to fill on the missing pixels by learning from a bunch of real high-resolution photos.
“By incorporating the prior knowledge of the faces and their typical variations, an artist is able to paint believable details,” the researchers wrote in their paper published on arXiv. “We show how a fully probabilistic model that is trained end-to-end can play the role of such an artist.”
RELATED: ‘Superhuman’ A.I. Can Locate Any Image
The Google Brain researchers tested the results by flashing pairs of images at human subjects: one 32×32-pixel true image of a celebrity’s face, and one enhanced version created using their computer model. The humans guessed that some of the enhanced images originally came from cameras. As Anthony noted, that happened 10 percent of the time, with 50 percent being a perfect score.
At worst, however, the enhanced photos turned well-known faces into Cubist monsters that had a black hole for an eye or a mostly erased mouth. A separate test using photos of bedrooms also produced hilariously bad images. Mondrian would have loved those.
If the researchers asked the subjects to identify the celebs based solely on the predictions, I wonder what the scores would have been. To me, it looked like a pixelated Gavin Rossdale went in and bizarro pixelated Josh Duhamel came out.
RELATED: Toss Your Glasses: Display Sharpens Blurry Images
Although human subjects thought that some of the enhanced images were real, those images were extremely clever guesses made by a computer model. We still don’t quite have the magical super-photo-enhancing abilities of fictitious crime fighters.
What the Google Brain researchers have achieved is remarkable, though. Combining neural networks was clearly a smart approach, especially when compared to results from other models. The team doesn’t discuss potential applications, but perhaps their computer model could one day help us zoom in on the correct answer to whodunnit.