It’s interesting if not useful. Compare the “truth” image in the second to the last column to the guess in the far right column. Considering how bad the input is (far left column) the guesses may have some value. They can’t always tell a male from a female face at this point, however.
You see it all the time in movies and TV shows: A security camera records footage of an intruder, but the image is too blurry or pixelated to make out who it is.
Some nerdy-looking “hacker” then clacks at his keyboard and — boom — seconds later, pixelated image turns into a crisp one revealing the person’s face in glorious detail.
“Oh, come on!” we all say while rolling our eyes. Well, you might have to break that habit because Google has figured out a way to turn movie magic into reality (sort of).
According to ArsTechnica, researchers at Google’s deep learning research project, Google Brain, have created software that attempts to “sharpen” images made up of 8 x 8 pixels.
Of course, Google Brain’s software can’t actually enhance the original block of pixels. Instead, what it’s doing is using machine learning to try to guess what the original image might be if it had been downsized to 64 pixels.
Google Brain’s software does this with two stages of neural network training. The first stage involves a “conditioning network” that cross references the 8 x 8 pixelated image with similar-looking images that are higher resolution and then downsized, checking for patterns and colors, as you can see below:
The second stage, called the “prior network” then uses details from high-resolution images to try to fill out the low-resolution images.
Finally, the images produced from both neural network training sessions are then composited together to create the best approximation of what the original image might be.
Google Brains’ software isn’t technically “zoom and enhance” magic, but according to the researchers’ findings, it comes damn close and the “enhanced” images are good enough to fool most people.
Squint hard and you might think the “hallucinations” (Google Brain generated images based on the training) are enhanced versions of the low-res images, too. They could have fooled me.