Espozo wrote:
Quote:
so this is not as implausible as it might first seem.
It's entirely impossible here. Unless you're going to say that it used an object detection algorithm and guestimated what the original image looked like, but disregarding how absurd that is, I couldn't even tell what the heck the "before" image was.
"Guesstimation" is essentially what those super-resolution algorithms do.
Here's a quick example I found:
https://www.slrlounge.com/zoom-and-enha ... a-reality/You can see that the NN (neural network) is able to reproduce the most important features from the really low resolution input. Of course it doesn't match the ground truth perfectly -- there's simply not enough information in the input. Faces similar to the ones in the NN results were probably present in the training set.
(BTW, the article is wrong in claiming this makes the TV trope a reality -- these algorithms wouldn't allow you to, say, extract a license plate number from a low resolution image, they would simply dream up some random license plate number if the information in the source image is not sufficient to reproduce it.)
That said, temporal stability (= results should appear consistent from frame to frame in videos) is a huge problem with these algorithms, so I'm very skeptical how their "UHD upscaler" actually performs in practice.