New very slow FFT denoiser

[DarkFoon]

@morsa
what kind of sourse do you use these settings on?
analog capture? DV capture? DVD? VHS? Broadcast?
Also, what sigma values do you actually use?

[Leo 69]

Actually I'd be very grateful to some guys here who'd post some good settings for this filter.For example "light denoising" "heavy denoising" etc. And "High Quality Standard denoising" would be great!

Сheers

[Fizick]

So, people want 2D

Got it!

Version 0.3, 21 December 2004 - added block temporal size parameter bt (was fixed =3 internally).

bt - block temporal size, number of frames (1, 2, or 3, default = 3):

1 - only current frame (spatial 2D filter - fastest);
2 - previous and current frame;
3 - previous, current and next frame (slowest)


IMHO, for big noise we must use big sigma, and more large block 32x32 to prevent grid artifactes.

Probably for 2D block 32x32 is better too.

[morsa]

for More oe less noisy DV I use plane 0 sigma=2, plane 1 sigma=3 ,plane 2 sigma =4.
I don't remember the block size right now but I guess it was 16....

[krieger2005]

great filter, ass always... and you are right: it's very slow.

To prevent "bluring" of edges with high sigma-values one can use "FineEdges"-Function from "Didée" i guess. Here my try:

function CleanUp(clip c, int "eEnd", int "eStart", int "bw", int "bh", float "sigma", int "sharp", bool "mask"){c

eEnd=default(eEnd,8)
Flaechen" markiert
eStart=default(eStart,68)
bw=default(bw,32)
bh=default(bh,32)
sigma=default(sigma,3.0)
sharp=default(sharp,180)
mask=default(mask,false)

ed=c.FineEdge(16)
SurfaceMask=ed.Levels(eEnd,1,eStart,255,0,false)
ret=mask ? MaskedMerge(c,c.BlankClip(pixel_type="YV12", color=$66FF66),SurfaceMask,Y=3,U=3,V=3,useMMX=true) : NOP()
cleanSurface=c.AddBorders(0,0,bw+20,bh+20).FFT3DFilter(bw=bw, bh=bh, sigma=sigma).crop(0,0,-(bw+20),-(bh+20)).xsharpen(sharp)

ret= (!mask) ? MaskedMerge(c,cleanSurface,SurfaceMask,Y=3,U=3,V=3,useMMX=true) : ret

return ret
}

great job.

[DarkFoon]

@Fizick
How does the "previous frame" part of the filter work?
does it take the unprocessed previous frame, and then process it, or does it take the processed previous frame and process it again?
Hmm, an example would be good, that doesn't seem to make sense...

For example:
The filter is set "bt=2".
I randomly access frame 10. The filter processes frame 9, and frame 10 in the FFT, and then gives the filtered output frame 10.
Then I move over to frame 11. When the filter processes frame 11, it also processes frame 10, but does it take filtered output frame 10, or does it take the original frame 10?

Gosh, I hope that makes sense. Anybody else know what I'm saying?

[State of Mind]

Ok, I also have a noisy DV source, what are the commands to use this plugin? I am VERY eager to see how well this denoises my DV.

[State of Mind]

???

[Fizick]

DarkFoon,
current versions has no memory, and procesed (spatially) all frames at every time step. It is not optimal, but easy coded. My filter is still experimental. I had some think about optimizing, and may be another temporal algo (Kalman) for 2D frequency "frames", but not rigth now. Sorry.
But (IMHO) the speed factor will not =2 for bt=2. Overlapped frame is large.

State of Mind,
do you have your own eyes? Try!

[DarkFoon]

@State of Mind
I tested this filter on my noisy DV capture and *angelic choir* it worked a miracle!!!
It would appear that the Y channel (plane=0, in the filter) is the noisiest. You can turn the filtering on so that it processes the other channels, but you wont get as dramatic of an improvement as you would with just the luma (Y, or whatever) being filtered.

Supposing that your capture is still 720x480 when you run this filter:

FFT3dFilter(sigma=2.0, bt=3, plane=0, bh=48, bw=48, measure=false)

That is what I used on my capture. Experiment with filtering the other planes in addition to this one.

@Fizick
Does the filter need square block sizes? Or could I get away with, for example, bh=48, bw=90? [for this example, assume that the block sizes fit exactly into the video frame.]

[Fizick]

DarkFoon, do you see any limitation or constrain on block size in my filter "manual"
But some sizes will result in very slow procesing due to FFT method implementation.
See FFTW doc for more info.
48 and 90 is OK.

[DarkFoon]

where in the docs does it say what sizes might slow it down? Or what should I look for? Sorry, I'm not a programmer, so I am lost in those kinds of things.

[Fizick]

FFTW works most efficiently for arrays whose size can be factored into small primes (2, 3, 5, and 7), and otherwise it uses a slower general-purpose routine.

48 = 2*2*2*2*3 is OK
90 = 2*3*3*5 is OK

[DarkFoon]

Thanks!

Keep up the good work!
This filter has great potential, don't let it fall between the cracks.

[Leak]

Originally posted by Fizick
FFTW works most efficiently for arrays whose size can be factored into small primes (2, 3, 5, and 7), and otherwise it uses a slower general-purpose routine.

48 = 2*2*2*2*3 is OK
90 = 2*3*3*5 is OK

Shouldn't this info be in your filter's documentation, under "Performance considerations" or something like that?

(NB: I haven't tried your filter as I don't have sources that are noisy enough, so apologies if it's already in there... )

np: Donnacha Costello - Dry Retch (Pop Ambient 2002)

[gst]

This filter is really powerfull. Maybe the best denoiser out there, and one of the slowest to.
I've got interesting results with this combination on a noisy dvd source :

DRemoveGrain(mode=2).DRemoveGrain(mode=2)
crop()
Lanczos4Resize()
Unfilter(-5,-5).FFT3DFilter(sigma=3,bh=32,bw=32).Unfilter(-5,-5).DRemoveGrain(mode=1)

I've also made a few test with the block size. You can see results here :
http://x.spirit.free.fr/FFT3D.zip (4.69Mo)

Block size tested :
8x8;16x16;32x32;64x64

Codec : xvid
Matrix : eqm_v3ulr_rev2
Bitrate: 1.050mb/s

32x32 give me the best quality, but there's no big difference between 16x, 32x and 64x. 8x is the worst value for my source.

Good job Fizick

[DarkFoon]

I usually work on anime, so my filter combination is much stronger than for video, for obvious reasons. But here it is, for those interested:

fft3dfilter(sigma=2.0, plane=0, bt=2, bh=48, bw=48, measure=false) #<- Y plane
fft3dfilter(sigma=3.0, plane=1, bt=2, bh=48, bw=48, measure=false) #<- U plane
fft3dfilter(sigma=4.0, plane=2, bt=2, bh=48, bw=48, measure=false) #<- V plane
msmooth(threshold=5, strength=3, mask=false)

Note: I use 48 because it is the largest number the divides into 720 and 480.

This is incredibly slow on my PIII, but it works excellently.
An explanation: the eye is most sensitive to luma, so that is filtered the least. But the difference between not filtering the luma and filtering it is quite visible; this can make or break the quality.
The reds (V plane, whatever you call it) are the most subsampled with my source (MPEG2 TV capture) so they can be filtered more powerfully.
Msmooth is to blend out the 'ghosting' that results (sometimes) from the temporal part of the filter. It's most noticeable when there are fast motions across large solid color areas.

To illustrate this, I have some pictures for comparison.

Original source frame (after IVTC, but no denoising):
http://www.freewebs.com/cassumnox/6343-original.png

Frame after FFT3D, but no Msmooth:
http://www.freewebs.com/cassumnox/6343-unsmoothed.png

Frame after FFT3D and Msmooth:
http://www.freewebs.com/cassumnox/6343-smoothed.png

The 'ghosting' I am talking about can be seen on the suitcase.

Note: compare the file sizes of the pictures too

Sorry the host is so slow, my usual server is down. :'(

[Fizick]

...When the filter processes frame 11, it also processes frame 10, but does it take filtered output frame 10, or does it take the original frame 10? Gosh, I hope that makes sense. Anybody else know what I'm saying?

DarkFoon, seems now I understand your question

All normal temporal filters (fft3d too) always take only original frames on input, not prefiltered by previous steps.


Other comment: V plane is not quite "red". Ask Wilbert (or see his guide) for detail

[DarkFoon]

V is not red?
Crap! Curse the internet's wealth of misinformation!

[Wilbert]

V is the red minus the luma. Highering V will turn your clip to red and lowering it to green (you can check that with ColorYUV(off_v=x)).

btw, V and U are sampled "equally", so you can apply the same amount of filtering on them.