Dealing with "localized" noise

[martinfrombern]

I have this source (tv cap) I'm currently dealing with. The problem with it is that there is an awful noise on a lot of frames which is not present in each part of frame, but instead is somehow localized (maybe even on a block-level?). Anybody can suggest a filter (chain) to deal with it? I'm thinking about some sort of "block-adaptive denoiser", cause the noise I'd like to remove is only this one (to avoid destroying the details and/or making it plastic).
Maybe I'm not perfect clear, but I believe you'll understand what I'm talking about when you see the screenshot (try zooming it a little bit to see more clearly).
Screenshot (thumbnail)
http://thumbnails2.imagebam.com/5154/4aa53251539824.gif (http://www.imagebam.com/image/4aa53251539824)
Sample for those who'd like to test their scripts
http://www.mediafire.com/?2ztwftkzzlz

Thanks

[Dark Shikari]

Those are badly quantized DCT blocks.

[martinfrombern]

Those are badly quantized DCT blocks.any filter/trick to correct them, without touching the rest of frame?

[Didée]

Hardly. Not only that all those DCT-artifacts (blocking and hi-freq noise) are all over the frame. Worse, actually those artifacts are contributing a major amount of the signal energy that you "see" in the source. After removal of the artifacts, the result will look dull, no matter what.

Perhaps an acceptable starting point:
fft3dfilter(sigma=4.5,sigma2=2.25,sigma3=0.49,sigma4=0.24,bw=24,bh=24,ow=12,oh=12,bt=1,plane=4)

[Archimedes]

FFT3DFilter looks surprisingly good (tested on the still frame). Never played with the different sigma values so much.

TNLMeans(Ax=10, Ay=10, Sx=3, Sy=3, h=1.4), if you have (a lot of) time, or dfttest() with default parameters may also candidates to do the job.

[Didée]

The result of TNLMeans seems not bad at all, but also not that much better than FFT3D. It's different, but not necessarily better.

When I compare the speed, then (on my ancient PC) TNL renders at 0.067 fps, while FFT3D renders at 4.7 fps.

From the results, I don't see much justification for TNL being 80 times slower. Waiting one day for a result may be hard (or not), but waiting 80 days is nothing I would accept ...

[Archimedes]

The result of TNLMeans seems not bad at all, but also not that much better than FFT3D. It's different, but not necessarily better.
May be it's a personal taste. Hard to say. Regarding the still image, TNLMeans (settings can be improved) looks better to me than FFT3DFilter (settings can also be improved, in know :)). It preserves more details (on most areas) while removing more artefacts (and noise) than FFT3DFilter.

TNLMeans:
http://img33.imageshack.us/img33/5425/ss0720x0574tnlmeans.th.png (http://img33.imageshack.us/i/ss0720x0574tnlmeans.png/)

FFT3DFilter (settings from Didée):
http://img180.imageshack.us/img180/9494/ss0720x0574fft3dfilter.th.png (http://img180.imageshack.us/i/ss0720x0574fft3dfilter.png/)

dfttest:
http://img210.imageshack.us/img210/7942/ss0720x0574dfttest.th.png (http://img210.imageshack.us/i/ss0720x0574dfttest.png/)

[Didée]

It's always about personal taste, isn't it?

TNLMeans ... preserves more details (on most areas) while removing more artefacts (and noise) than FFT3DFilter.

An interesting finding. The screen in front of me is showing something different:

Source:
http://thumbnails22.imagebam.com/5160/be3ea151593295.gif (http://www.imagebam.com/image/be3ea151593295)

FFT3DFilter:
http://thumbnails21.imagebam.com/5160/0d889351593300.gif (http://www.imagebam.com/image/0d889351593300)

TNLMeans:
http://thumbnails21.imagebam.com/5160/dcc5bb51593303.gif (http://www.imagebam.com/image/dcc5bb51593303)

I'd say that - with given settings - TNLMeans is less effective in artifact removal.

Now, of course one can argue that TNL also seems "sharper" than FFT3D. That's true, and partly a side-effect of the lesser filtering it does.

On the other hand - throw a little (contra-) sharpening at FFT3D, and it'll look same sharp. Speed is no concern. If FFt3D is ~80 times faster than TNL, then FFT3D+controsharp will be ~79 times faster than TNL.

I know you (and many others) like TNL very much. The problem is that it's only practical for still image processing. With *effective* settings (i.e. such settings that really put TNL in front of other spatial(!) filters), it is perfectly inpractical to have it process thousands 'n thausands of frames. You want to get a result before they've put you six feet under.

[Archimedes]

I'd say that - with given settings - TNLMeans is less effective in artifact removal.
But that's not the screen capture above. The screen capture of TNLMeans is totally blocky. :eek: Normally, on still images, TNLMeans makes a good job on deblocking.

I know you (and many others) like TNL very much.
I'm not the marketing chef of it. I see what i see.

The problem is that it's only practical for still image processing. With *effective* settings (i.e. such settings that really put TNL in front of other spatial(!) filters), it is perfectly inpractical to have it process thousands 'n thausands of frames. You want to get a result before they've put you six feet under.
That's the point. On still images i have no doubts about the quality of TNLMeans. What it can do and what not. On noisy images with many details i go, for example, with Bx=By=0 which increase again the calculation time enormous. We talk about 1 Frame per hour (may be more). :eek: Regarding quality i never reached such results with FFT3DFilter or dfftest. May be, i used the wrong parameters. dfttest seems to be the closest to TNLMeans. TNLMeans is not a all in one solution. There are thoroughly situations, where it completely fails (as described in the technical paper) and, for example, where FFT3DFilter is the way to go. For example, very bad sources where you need some “wiping effect”. This is, where TNLMeans fails.

[g-force]

The screen in front of me is showing something different:


Never seen THAT Morrissey video before... ;)

-G