Re: [Gimp-developer] News of a new image enhancement algorithm ...

[Nicolas Robidoux <nicolas robidoux gmail com>]

Joao:

Read the article!

My guess is that the SR scheme combines two things:

1) Look for "similar things" at the same scale within the image to
reconstruct a "platonic ideal" which you then replicate (this is
super-resolution using data from the same image; classical
super-resolution uses multiple shots of the same scene).
2) Do the same at multiple scales, in the hope that there is
self-similarity to pick up on.

But I've only read the summary.

-----

Prequel: I am very biased because I develop and program competing
methods which could be described as belonging to "older" generation
approaches. So, I'm fighting obsolescence.

Now: If I was to suggest a "state-of-the-art cool" method to someone,
I would suggest NNEDI3 (which is programmed in Avisynth).

Why?

a) NNEDI3 not an academic scheme. It's well tested in real world
situations, I'm pretty sure that the code is FLOSS, and there is an
open source community around it. It was put together so that it does
pretty well all the time, not only when the data fits the "ideal
input".

b) No matter how wonderful the "discovery" of how common
self-similarity is in this world (a crush with fractals actually
contributed to getting me back into grad school a few years back), the
world is not universally self-similar.

Take out the test images from the Weinzmann site in which you do not
find repetitions of the same pattern, either at the same scale or at
multiple scales. Does the SR method really create a faithful or better
enlargement?

(Bias warning.)

If you look for self-similarity, you'll find it...where it's not.

Look at the skin of the baby. Do you see that the woolen hat was
"woven" into it?

And I really dislike what SR does to eye pupils.

Also: Some of the sharpness comes from applying something which in the
end is a lot like a variation diminishing limiter (Jensen like?). Do
you really want this "waxy look"?

Not that these things could not be fixed (it's most likely a matter of
setting thresholds).

This being said: Do my methods do better? Maybe not. But they are
local (SR requires an analysis of the whole image), fairly cheap, and
adapt to large enlargements or reductions robustly.

-----

Again, in the same ballpark, it is my opinion that NNEDI3 (which I
have absolutely no connection with) is likely to disappoint less.