pouët.net

I want to be able to enlarge an image liek in these videos

http://www.youtube.com/watch?v=Vxq9yj2pVWk
http://www.youtube.com/watch?v=3uoM5kfZIQ0

This seems to be related to the image "eigenvalues".

My secret idea is to render to a one pixel large target and let the algorithm create the missing cubes and ribbons as I enlarge the texture. Ideas ?

s-spline

Best one evar:

http://www.youtube.com/watch?v=KUFkb0d1kbU

start with a very very very big image :)

cubes and ribbons won't work - tried it! ..but shadebobs gave me something almost perfect.

There's more actual results in that area than one would think :)
http://www.ifp.illinois.edu/~jyang29/papers/CVPR08-SR.pdf
http://www.wisdom.weizmann.ac.il/~vision/single_image_SR/files/single_image_SR.pdf

"Turn the camera by 30 degrees":
http://www-cvpr.iai.uni-bonn.de/pub/pub/OTKC_dagm09.pdf (Actually, it extracts meshes too. And textures. And does automatic segmentation. In a few seconds).

This is all starting from a single 2D input image btw, if you have video the real fun starts.

Link to the third paper for the lazy.

xeron, nice one :)

GREYCstoration

a tool.

hmm.. nice too maytz. i also recall that some of the dutch guys from aardbei was involved in a company that created an image scaling tool. i've forgotten the tool name ...

s-spline, as mentioned above.

I guess we're just talking about enlarging images? and the s-spline mentioned is the one used in tools like the PhotoZoom software?
It's been a few years since i last used it but it had pretty good results, then i switched to Genuine Fractals and more recently tried Blow Up from Alien Skin.
A relatively new free tool came out recently also, SmillaEnlarger which also gives decent results. So if i need to resize something i'll use one of those 3 and pick the best looking one.
Reshade gives excellent results, but it ignores Colour Profiles in images, i reported this to them and they said "Thank you for bringing this to our attention. We'll do everything possible to include color profile handling in the next release. We will contact you when this happens." but last i tried this software(around 1.41) it still hadn't be implemented, 2.0 beta is out now though so maybe it does now.
As for SR stuff from video, i tried out Topaz Moment well over a year ago and it was a horrible buggy piece of crap and the results were only so-so.

Useful! I have a use for some of this stuff - I need to extract a high res still image from a low res video source. Anyone seen anything on doing that in realtime by any chance?

Avisynth has a handy lanzcos4 filter, but it will not "invent" details like the resampling software cited above (which are better perceptually).
If your source is crap, nothing can really be done.

This is something i'm writing myself, so avisynth is out :) I've tried lanzcos before, it's better than bicubic but no use for upscaling. I was thinking more of one of the super resolution methods (the sources I'll use will often have sub-pixel misalignments and the like that are useful).

If you are not happy with lanzcos while upscaling, try to raise the number of taps... else you can look at how SmillaEnlarger work.

Well, to be fair, any upsampling algorithm will make up any pixels that don't correspond to exact integer coordinates in the source image.

As far as correctness is concerned, anything that will downsample to the original image is fair game. And there's no particular reason to prefer separable linear filters (other than their simplicity). In fact, over most representatitve samples of "real-world" (photographic) images, linear filters will probably be worse (in terms of relevant error metrics) than the more advanced methods.

If you have lots of very similar images, you have a lot of that information a regular upscaler would have to invent, so some of the more advanced techniques look a lot more promising for my particular case :)

paper on image compression/reconstruction/processing with JPG image compression in the PDF ---> automatic defeats the purpose of the quality comparison charts ---> me can't tell which method is best ---> me drops the paper to the recyclebin :(

hmm.. would be fun to see, if one could remove the censorship from a tv-documentary, when someone is interviewed :)
The blurry/blocky areas should in theory provide enough data even with only very little motion.

I like how everyone compares its algorithm against Photoshop ugly bicubic one.

Restoration of a Single Superresolution Image from Several Blurred, Noisy, and Undersampled Measured Images? :)

*swirl* !!!

very interesting thread, this!