OT: Noise reduction and image enhancement

farss wrote on 1/25/2008, 2:57 AM
Just found this:
http://www.ee.ucsc.edu/~milanfar/software/superresolution.html.
I don't know if it's commercially available or if the spooks are keeping it to themselves but I know a few here have devoted a lot of time to noise reduction software and this seems to almost validate some of what we see on CSI.

Some examples of what this software can do here:
http://www.ee.ucsc.edu/~milanfar/software/SR-Examples.pdf


Bob.

Comments

johnmeyer wrote on 1/25/2008, 8:40 AM
Just before I ran Live Picture back in 1998, the founders of one of the companies we had purchased (Eric Chen) had left to start a company that was going to use this same technique (integration over multiple frames) to increase resolution in each frame. The underlying concept is simple enough to understand, namely that the information you miss in one frame (because your camera's resolution can't differentiate the detail) will be partially picked up in a new location in the next frame as the camera or the object moves. The devil is in the details, and I am not aware of any vendor that has been able to make it work reliably.

I had always thought that the technique would do nothing if the camera is stationary and most of the scene is static, but the last of the examples in the PDF you link to shows books on a bookshelf, and the mass of pixels in the original is replaced by extremely clear and legible text in the recovered image (although it took over 300 frames to achieve this).

Very neat stuff. It will be interesting to see if they can commercialize it. I'd certainly buy it.
john-beale wrote on 1/25/2008, 10:29 AM
If the subject is motionless and the camera is absolutely perfectly still, the technique can do nothing.

If the camera is doing a slow pan, or moving in some way, AND the camera (and subject) motion can be accurately tracked to a small fraction of a pixel, then the technique can do impressive stuff as you see in the examples (given enough frames). For fast object and/or camera motion, or objects visible for only a few frames, I don't think it can do anything useful.
Shergar wrote on 1/25/2008, 11:41 AM
I hunted for "super-resolution" a while ago and found this...



which led me to the page Bob mentions plus some others

I think we're talking about several techniques, not just one, and from what I've seen results on fixed camera sequences are definitely useful...

The QE labs product for example does seem to deliver a real resolution increase for a video frame ... which gave me hope that someone will be along soon with a plugin to up-res HDV to 2k or even 4k !!! Strangely QE labs has a dead website, with no links whatsoever.

The software Bob mentioned is being used at http://www.FixMyMovie.com by MotionDSP, but only for low res source... I assume they've got plans for pro applications but so far nothing public, it seems (except that they are working with the intelligence guys :-)

Definitely a space to watch, IMO

DataMeister wrote on 1/26/2008, 3:47 PM
A couple years ago I ran across the products from Ocean Systems which used a similar type concept of gaining information from multiple frames of video. Looks like they have expanded and fine tuned things a little since.

http://www.oceansystems.com

dTective (for video)
ClearID (for still photos)
QEAS (for audio)

GlennChan wrote on 1/28/2008, 1:52 PM
It's highly likely that the human visual system does something like this.