What's that technology where they find hidden things from a 2D photo? Using radiosity, shadows, etc.

[demagogue]

You could probably count a few of the symbols on it, noting their colour and position and using this information to take an informed guess at which card it was. (Though differentiating clubs and spades might be tricky.) You couldn't use the same technique to "unblur" a face though.

 

You might if, just like the card, you already had features of the face (or common features of faces) in memory, and were just confirming if it was the same face, and morphing those features onto the blur.

 

I've seen neural-net setups for face recognition (where a set of hidden-layer nodes are weighted over many trials to conform with some input, like a set of photographs of the face), where they distort an input face quite severely and the computer can still identify it with a high probability.

 

And then, when asked to reconstruct the facial features, a programmer can simply run the distorted information in reverse through the same net that correctly identified it, and it returns this image that is very strikingly similar to the actual, original face. In the example of this I remember best, they occluded like 40% of the face around the eyes, and after correct identification, they ran the occluded picture in reverse through the set of weights and it returned this image that had these really haunting ghost eyes that completely do not exist in the original input ... as if the computer actually had those eyes in its "mind's eye" when making the id.

Edited June 10, 2007 by demagogue

[sparhawk]

Which of course is what I said before. The neural net was trained to recognice that face, so the information was already there. So of course it might be possible to recognize even the distorted version of that face. Humans do this all the time as well.

[OrbWeaver]

Exactly, if you have sufficient prior knowledge of what you are looking for, constructing an algorithm to fill in the blanks may be possible. This is very different from the generic "magic detail enhancer" that you see on the movies.

[Crispy]

You might if, just like the card, you already had features of the face (or common features of faces) in memory, and were just confirming if it was the same face, and morphing those features onto the blur.

Indeed, but I'm assuming that we don't already know what the face looks like, as in the serial-killer-reflected-in-someone's-pupil-which-was-caught-on-a-security-camera example.

[Nyarlathotep]

Exactly. Computers are no better than us at figuring out what blurry photos are.

[sparhawk]

Computers are more patient and stubborn.