Slashdot Mirror
Google Brain Creates Technology That Can Zoom In, Enhance Pixelated Images (softpedia.com)
Google Brain has created new software that can create detailed images from tiny, pixelated images. If you've ever tried zooming in on an image, you know that it generally becomes more blurry. You'd just get larger pixels and not a clear image. Google's new software effectively extracts details from a few source pixels to enhance pixelated images. Softpedia reports: For instance, Google Brain presented some 8x8 pixel images which it then turned into some pretty clear photos where you can actually tell facial features apart. What is this sorcery, you ask? Well, it's Google combining two neural networks. The first one, the conditioning network, works to map the 8x8 pixel source image against other high-resolution images. Basically, it downsizes other high-res images to the same 8x8 size and tries to make a match on the features. Then, the second network comes into play, called the prior network. This one uses an implementation of PixelCNN to add realistic, high-res details to that 8x8 source image. If the networks know that one particular pixel could be an eye, when you zoom in, you'll see the shape of an eye there. Or an eyebrow, or a mouth, for instance. The technology was put to the test and it was quite successful against humans. Human observers were shown a high-resolution celebrity face vs. the upscaled image resulted from Google Brain. Ten percent of the time, they were fooled. When it comes to the bedroom images used by Google for the testing, 28 percent of humans were fooled by the computed image.
I don't care how fancy the algorithm is, the original data was lost. This is still just a guess about the original content. It's just a better guess than was possible before.
I just hope law enforcement doesn't think they can use this to solve any crimes.
One of our competitors trademarked the term "hypothesis". From now on, we will call them "boneheaded ideas".
A while ago, someone made the nnedi upsampler that uses neural networks to upsample. It's still one of the best image upsamplers available.
Google's approach is quite a bit different. Where nnedi worked to better extract detail out of what was already in the image, Google seems to literally fill in detail that was probably in the source but maybe not. Much, I guess, like how our own memories work. It's an interesting approach and the results look quite fantastic. My only question is how well it will work on a random sampling.
I see you still don't actually know what a Fourier transform is.
The summary's explanation of what this does isn't correct. It says:
Google's new software effectively extracts details from a few source pixels to enhance pixelated images.
It doesn't extract details from a few source pixels. It invents details to add to those source pixels, based on the knowledge that the pixelated image is of a face, and of what faces look like. It produces something that plausibly fits the input data. How close this is to the original image, pre-pixelation, depends on what images were in its training set.
This is an interesting piece of work, but it doesn't mean that you can recover data that has been discarded.
Note to ACs: I usually delete AC replies without reading them. If you want to talk to me, log in.