We’ll all seen it in various Police and Espionage dramas: some horribly blurry or pixelated image is taking from an equally horrible surveillance camera and after “enhancement” a face come out of it. For years, many of us have laughed at these movie scenes, and rightly so, because there is simply not enough original information to reconstruct an image that is clear and sharp from a blobby bunch of pixels. Each group of four images A, B, C, D, represent (from left to right):
- A:16×16 input
- B: a blurred version of A
- C: computer generated image
- D: clear photo of the person blurred in A
However, the image presented here shows something very interesting: how an AI using deep learning and a lot of face photos (130,000) can “guess” what a very pixelated face may end up looking like. You can find the sources from David Garcia on github. This is somewhat to having AI create art from what the computer has seen before, which has been demonstrated before.
Based on all the face that is has seen, the Deep Learning AI will generate a clear image of what the face would “probably” look like, based on the blurry 16×16 input image. The output result is a 64×64 image with clearly defined face features.
First of all, you will see that the likeness of each computer-generated photo does not exactly match the actual, real photo. However, the quality is surprisingly good, especially since the training dataset of images is relatively small (a large one would be hundreds of millions or billions).
The computer-generated images are the equivalent of an artistic rendering which is based on really geeky science, which is something like “from all the humans I’ve seen, this blob would resemble this computer-generated one”.
Although these images cannot be considered and be used as “proof” (they’re not!) they can prove useful as the equivalent of a super-fancy sketch artist. The important part is to remember that the generated image is a loosely close rendering and not a person.
Of course, I’m being sarcastic about apologizing about how laughable some of the movie depiction of “photo image enhancement” are sometimes, but this is yet the closest thing to what Hollywood has been showing for years, and it will get only much better with time.