How many times have you rolled your eyes at those “zoom… enhance… zoom in closer… I SAID ENHANCE!” scenes on television shows? Well, Google is attempting to unroll your eyes with a pair of neural networks that can intelligently “enhance” images using a source image that is just 64 pixels.
The official process and paper, created by the Google Brain team, is called “Pixel Recursive Super Resolution.” In technical speak, the AI “synthesizes realistic details into images while enhancing
their resolution”; translated into TV speak, the AI does a great job of mimicking the impossible “Enhance!” trope.
However you refer to it, the process involves two steps. First, the program is given an 8 x 8 pixel image, which it “maps” to a similar higher resolution sample to produce a workable outline of the portrait or scene. Then, a second network adds resolution to that image, estimating pixel values based on other high res images it finds online with similar pixel locations.
The process is far from perfect, and nowhere near Minority Report or Blade Runner levels, but the results are impressive all the same. In the image above, the leftmost column is the 8×8-pixel source, the next is the computer’s estimation, and the third is the accurate high res version of the 8×8 source image.
To learn more about this incredible research or see more samples, check out the full paper here.
(via The Verge)