Filter to increase video detail using Super-Resolution?

[cweb]

First of all, I hope this is the right forum to post this. If not, I
apologise beforehand...

There's an interesting posting on Slashdot today asking whether it is possible to increase video detail using super-resolution algorithms. Poptones did mention Avisynth too...

I thought that this might be of interest to some filter developers or somebody at any rate... perhaps someone might come up, someday, with a new Avisynth filter based upon a super-resolution algorithm.

Here's the link to the article (http://ask.slashdot.org/article.pl?sid=03/06/19/2216246)...

[trbarry]

It looks both interesting and very complicated. Personally I'd never heard of that before and don't know of any Avisynth filters.

- Tom

[Blight]

My brother actually wrote a research paper on super-resolution. It sort of works as reversed anti-aliasing. But... it's only good for stills reallly. At least until some computer is able to actually interpolate 3D objects in a scene and calculate motion to a serious sub-pixel element.

So... right now, forget it.

[sh0dan]

.. Furthermore applying this kind of filter to video will create unbearable amounts of "warping". Having even slighly noisy material will make the algorithm propose very different solutions for each frame, making the final result unuseable.

[CAFxX]

i don't know if any of you read out the entire paper attached to the article.
at its end there's the source code of the filter and some tests

[MfA]

The MERL algorithm isnt really what I would call super-resolution, it is more a smart upscaling based on clustering ... it doesnt look very natural to my eyes though. I wouldnt be surprised if less complex edge based interpolation methods look better (some examples here (http://www.ee.princeton.edu/%7Elixin/interpolation1.htm) and here (http://www.general-cathexis.com/interpolation.html)). Ive said before, it would be nice to have edge directed interpolation (http://www.ee.princeton.edu/%7Elixin/papers/nedi.pdf) in Avisynth :) (Some implementation tips are here (http://chiranjivi.tripod.com/EDITut.html) if anyone is interested.)

[Kurosu]

Originally posted by MfA
it would be nice to have edge directed interpolation (http://www.ee.princeton.edu/%7Elixin/papers/nedi.pdf) in Avisynth :) (Some implementation tips are here (http://chiranjivi.tripod.com/EDITut.html) if anyone is interested.)

NOW that sounds really interesting and... new. The math doesn't look that complex (though matrix operations will most probably be very costly in means of calculation).

*bookmarks*

[edit]
I had overlooked the calculation of the inverse of a 4x4 matrix, and that for each pixel of the output...

[trbarry]

I guess trbarry have gone close to the EDI feature. I have achieved a good aproach using the following function

Bach -

I didn't realize anyone was trying to use TomsMoComp to double in both dimensions. I never announced it because I didn't do much testing but if you use SearchEffort=-2 TMC should already double the size in both directions. But this currently works for YUY2 only (if at all). I know I did finish the code once, though I'd planned a better and faster version and then got side tracked with the YV12 stuff. But it should work.

However this is using Edge Based Line Averaging which I don't think is the same as Edge Directed Interpolation.

- Tom