faster NL-means algo?

[Terka]

http://www.adptools.com/denoisemyimage/description.html
http://www.eurasip.org/Proceedings/Ext/LNLA2008/papers/cr1026.pdf

[MfA]

Indeed it is, I don't think Tritical ever used precomputed squared difference images for speedup ... you want to implement it? :)

[Terka]

no i am not a programmer, want only point theese papers

[tritical]

My implementation is definitely not optimized. It even uses double precision math IIRC. Of course, the version I made uses gaussian weighted euclidean distance to compute neighborhood similarity, so their main speed up (fast computation of the neighborhood distances using sliding window) isn't applicable. Apparently the gaussian weighting is not that important.

[MfA]

It is applicable. Once you have the squared difference image for shift X you can use fast convolution to get any set of weights you want, the sliding sum is just a special case ... you don't even need FFT for Gaussian, an IIR based gaussian convolution would be almost as fast as the sliding sum.

That's the beauty of using the redundancy between different neighbourhoods at the same shift, rather than the redundancy between different shifts for the same neighbourhood.

[tritical]

If you can approximate the weighting window fir convolution with an iir filter that only requires a few operations then the speed up will still be close to that of the no weighting case. I didn't think that gaussian could be approximated well by low-order iir though, and at the neighborhood sizes I'm used to, 5x5-7x7, fft doesn't gain you much over fir.

[*.mp4 guy]

The nlmeans algorithm operates much more effectively at large window sizes, 21X21 for example offers a very nice improvement over 7X7, If large window sizes can be spead up, it would be quite nice.

[MfA]

I didn't think that gaussian could be approximated well by low-order iir though
Well, 2 IIR filters to be exact (a causal and an anti-causal one).

[yup]

Hi all!
See thread at this forum:
http://forum.doom9.org/showthread.php?p=1236283#post1236283
this only work for Nvidia card 8XXX and later. Algo modified not true TNLMeans but work very fast, for my GT8800 GTS at real time.
yup.

[MfA]

Algo modified not true TNLMeans
That's rather generous IMO.

[tritical]

The nlmeans algorithm operates much more effectively at large window sizes, 21X21 for example offers a very nice improvement over 7X7, If large window sizes can be spead up, it would be quite nice.
Do you mean search window size or neighborhood similarity window size? Granted, larger search window size is better. When I mentioned 5x5-7x7 windows I was referring to neighborhood similarity. IMO at some point larger neighborhood similarity window size has to hurt performance (if all pixel differences are treated equally) because the center portion gets too little weight, and the window is too large to capture small details.

The speed-up is related to neighborhood similarity calculation. If you incorporate it, nl-means complexity is reduced close to that of bilateral filtering (and you can use any neighborhood similarity size you like).

[*.mp4 guy]

You are right, I was talking about the search window size. I haven't ever goten any real improvement from similarity windows larger then 7*7, but I 've never noticed them hurting anything either.

[Archimedes]

The nlmeans algorithm operates much more effectively at large window sizes, 21X21 for example offers a very nice improvement over 7X7
That's also my experience (for still images). And a maximum preservation of details can be achieved with a base window of 0 pixel (Bx=By=0).

[markanini]

Sorry to dig up this old post. I found this discussion on parameter tuning a lot more interesting now that we have a faster way of using NL-means via NLMeansCL. I still have some questions though.

IMO at some point larger neighborhood similarity window size has to hurt performance (if all pixel differences are treated equally) because the center portion gets too little weight, and the window is too large to capture small details.
How can you tell by looking at parameter values used that neighborhood window is doing more harm then good? Also, would neighborhood window values need to be adjusted in tandem with search window to ensure efficiency?

[markanini]

Nobody care to comment my question?