[placebo]

It definitely makes a significant qualitative improvement, making the picture appear more in sync with what our brain interprets as a higher resolution picture, but my first thought is whether this particular example goes beyond aesthetics. Is there really any instance where this method could for instance turn an unintelligible picture of a license plate to something in which the characters can be recognised? More generally, I wonder whether there has been any research on the limits - i.e, what needs to be the combined minimal size of the information stored in the neural network plus the information on its inputs before the output can be said to be true to the source with probability x ?

[semi-extrinsic]

I imagine that if you trained this on a set of license plate photos, it would be able to enhance license plates illegible to an untrained human such that they're readable. However, I doubt it would be better than a human specifically trained at this task.

I've seen some videos from Cold War satellite photo analysts, and the way they can look at some tiny gray blobs and go "That's a T-64 tank, that's a T-62 tank, that's an SA2 launcher" etc.

[PeterisP]

Well, it doesn't create any information that wasn't in the original data (nothing can do that, you can only lose information in processing) so if e.g. the characters can be recognized in the processed image of a licence plate, then by definition they could have been recognized from the original data as well in some manner.

However, they can make things more easily interpretable by humans. A rough analogy is turning up the contrast - given a very dark image of licence plate where the black parts are totally black (#000000) and white parts are just very dark (#010101), the characters definitely can be recognized even while human in normal conditions would just see it as totally black, and processing would help.

[placebo]

> Well, it doesn't create any information that wasn't in the original data (nothing can do that, you can only lose information in processing)

I'm not sure this is correct. In a sense, it does contain information that wasn't in the original inputs - i.e information added by the weights in the neural network which itself was obtained by information extracted from an enormous amount of previous samples. Of course, the largest and best trained neural network won't be able to tell the license number given 2 pixels of information, but I am curious as to the theoretical limits of what can be achieved in extreme cases of with very little information as input and a neural network that has almost limitless resources.