A Recursive Gaussian Weighted Filter For Impulse Noise Removal

[JohnnyMalaria]

I was rummaging around for an algorithm to remove impulse noise from an image (totally unrelated to video) and came across a paper that describes the subject of this post. I was really astonished at the capability of this method that requires no preprocessing parameter optimization etc. Attached is a before and after - yes, the left is before and the right is after filtering (the original image - the famous Lena - had 90% noise added).

(Compression noise has been added due to the original pdf creation and my grabbing part of it)

The paper is available here (check out the other example images, too):

http://www.icgst.com/gvip/Volume6/Issue3/P1150636003.pdf

I can see this having application to the removal of random shot noise due to low lighting levels etc - I'd have a go at it myself but time doesn't permit it.

I thought this might be of interest to the Avisynth coders etc.

[MfA]

Salt and pepper noise is nice for making pretty pictures ... but you should realise that cleaning it up is just a question of throwing away bad pixels and interpolating from what's left. With shot noise there are no good and bad pixels, just varying levels of bad pixels.

It seems like it's a good filter regardless, but recursive filters and video usually make for very much less than real-time processing.

[halsboss]

Good Grief, that's amazing !!

[krieger2005]

Ah, don't trust those result-pictures. I can't just imagine such results could be ever achieved. Most of such "mirracle" algorithm i saw before presented a super result but in action on general video-artefacts they don't give better results than existing filters.

[scharfis_brain]

it seems they just selected all full-white and full-black areas to be impainted.
That is not a noise filter for normal affected images.