Enhance enhance: Using machine learning to recover lost detail from upscaled photos


Originally published at: https://boingboing.net/2018/05/10/generative-adversarial-network.html


Hmm. /makes a note to not just blur private details in photos.


And it doesn’t just come out all Hounds of Tindalos?


Two things come to mind… Any number of movies where the characters just repeat “enhance” until the image becomes clear.

And I suspect that a good knowledge of the algorithm would enable people that were trying to camouflage something to make it even less obvious after this had been done.


“Give me a hard copy right there”


Yeah, those round blurry features in the original photo were there for a reason – to keep the Hounds at bay. Thank you, data science, for unleashing canine demons.

What’s next? “AI researchers announce breakthrough in summoning Cthulhu via machine learning”?


Cue the inevitable “Enhance!” montage:


How many SAN points does IBM’s Watson have?


I also wish things were that simple.


Be sure to visit that Morning Paper link, which includes a copy of the “ground truth” behind the demo photo.

I’m not so impressed. The up-scaled picture looks more detailed at first blush, but examination of the original shows that those “details” are almost entirely wrong.



It might be worth noting that that paper is from 2016.
Also, the topic of “super-resolution” was worked on before this paper, and continues to be worked on.
Today, it really comes down to “teaching” the computer what “expected” images (i.e. real ones) look like, so that when it upsamples an image, it knows how to plausibly “imagine” the detail that was never there.


Err… the paper was published in 2016…




The video itself could use some by now. Edit: sorted!


I see your enhance, and raise you one un-crop.


This topic was automatically closed after 5 days. New replies are no longer allowed.