Denoise without killing one frame content?

[atra dies]

Mdegrain removed fire and smoke unique to single frames and I read in another thread someone had leaves disappearing. I adjusted many parameters including dct instead of spatial (only works with 8x8 blocks) and that only made me have to lower thsad. Reverting to mdegrain1 helps but then there is a lot of noise left.

Anybody have any special techniques they'd like to share?

dfttest works much better but then I am using tbsize 3 and if set too strong one frame content is removed. I am trying to use sst alone but setting sigma to 0 negates it somehow. I think it's a bug.:scared: Eliminating sigma parameter allows it to function. I am using mod 16 dither versions for both.

I have tried smdegrainmod,zzz denoise and tnlmeans and all of them are not as good as:

dfttest(u=false,v=false,ftype=0,f0beta=1,smode=0,sbsize=3,sosize=0,swin=7,tbsize=3,tmode=1,tosize=2,sst="0.0:1 1.0:30",dither=0,lsb=true)

Before I read about ssx ssy sst I had a generic sigma but it blurs of course. I am just playing around with sst values and have to set 1.0 really high to kill a lot of noise. Can sst be set with four points like ssx/y or are 0 and 1 only allowed? I don't know if tritical wrote this 16 bit version or if cretindesalpes did.

[nibus]

Post a sample of the source

[atra dies]

http://www.sendspace.com/file/j5yu1s


dfttest(u=false,v=false,ftype=0,f0beta=1,smode=0,sbsize=3,sosize=0,swin=7,tbsize=3,tmode=1,tosize=2,sigma=10,sst="0.0:1.2 1.0:100.0",dither=0,lsb=true)


I need to read or have someone say how ssx/y/t works. Does sst act solely on the results of ssx and ssy (or sigma=) or does it denoise temporally high frequency video content or the frequency of that content's change between frames?

Another solution may be to limit mdegrain to certain parts of a mask (the noisy grainy part).

ssx="0.0:0.0 1.0:80000.0",ssy="0.0:0.0 1.0:80000.0",sst="0.0:0.0 1.0:1000000.0"

^This has no effect at all but if I use sst alone and have no sigma= something happens. If I have sigma=0 with sst nothing happens but if I put sigma=1 it works. Must be a multiplication thing.

[Didée]

# dgdecode.dll
# fft3dfilter.dll
# fluxsmooth.dll
# mvtools2.dll

mpeg2source("sample_(atra.dies).demuxed.d2v")
o = last
pre = o.fft3dfilter(sigma=3,sigma2=6,sigma3=4,sigma4=2,bw=16,bh=16,ow=8,oh=8,bt=3).fluxsmootht(2)
sup1 = pre.msuper()
sup2 = o.msuper(sharp=2,levels=1)
bv2 = sup1.manalyse(isb=true,delta=2,truemotion=false,global=true,blksize=16,overlap=8,search=5,searchparam=6,DCT=5)
bv1 = sup1.manalyse(isb=true,delta=1,truemotion=false,global=true,blksize=16,overlap=8,search=5,searchparam=6,DCT=5)
fv1 = sup1.manalyse(isb=false,delta=1,truemotion=false,global=true,blksize=16,overlap=8,search=5,searchparam=6,DCT=5)
fv2 = sup1.manalyse(isb=false,delta=2,truemotion=false,global=true,blksize=16,overlap=8,search=5,searchparam=6,DCT=5)
o.mdegrain2(sup2,bv1,fv1,bv2,fv2,thsad=319)

#interleave(o,last)
#stackhorizontal(o,last)

return(last)

[atra dies]

It's better than defaults for mdegrain and the changes I made (except I like search=3) but not better than:

dfttest(u=false,v=false,ftype=0,f0beta=1,smode=0,sbsize=5,sosize=0,swin=7,tbsize=3,tmode=1,tosize=2,sigma=0.1,sst="0.0:0.0 1.0:100000.0",dither=0,lsb=false)

at leaving all of the flame and smoke. sbsize 3 is sharper but removes more smoke. dfttest also leaves some ugly big soft blob grain. I should probably put more points into sst.

I also cobbled together:

a=last
b=last.mt_edge(mode="prewitt", thY1=1,thY2=10000,thC1=0, thC2=0)
c=last.dfttest(u=false,v=false,ftype=0,f0beta=1,smode=0,sbsize=5,sosize=0,swin=7,tbsize=3,tmode=1,tosize=2,sigma=10,dither=0,lsb=false)
d= mt_merge (c, a, b, luma=true)
return d

But I'd have to tweak dfttest more and thY is int when I want float and I don't even know what the units of measurement really are there. Ideally I would be able to denoise below a threshold a lot and then have all the edges necessary remain. There's gotta be a more sophisticated way seeing as how I don't know what I'm doing.

I would use mdegrain on every other scene but I notice it removing some shading from faces in fast motion. Any dfttest fans have some scripts to share?

[Didée]

Well, seems I can't be of any help. Trying your filter proposals (with dfttest v192 from dither package), I see almost no denoising happening at all. (?!?)

After fooling a bit around with dfttest (I use it quite seldom), I wasn't able to achieve anything that - for my taste - comes even remotely close to MDegrain.

Obviously, I have no clue what you want to achieve, or how you want to have things look like.

[Mounir]

Have you tried nlmeansgpu? You should give it a go

separatefields()
NLMeansCL(A=4, S=2, B=1, aa=1.0, h=2.0,HC=1.5, plane=4,cpu=false,buffer=false,sse=true)
weave()

[atra dies]

Well, seems I can't be of any help. Trying your filter proposals (with dfttest v192 from dither package), I see almost no denoising happening at all. (?!?)



Well I may tweak what you posted to try to bring back some noise cause I am changing my mind all the time about this. I am thinking I like how mdegrain smooths the noise more than dfttest and produces a more natural output but I don't want to remove any part of the flame or smoke or make it faded. I see some important principles about how mdegrain works now from your suggestion, ie the calm search clip. I think I might try to make it mdegrain1 and that will help a lot.

(Is search=5 meant to focus the degraining on the horizontal movements of the picture and leave the vertical flame/smoke movement alone?)

I will also try nlmeansgpu and maybe post some comparison shots. This is important dammit cause I thought mdegrain was infallible!

[Didée]

search=5 is not related to horizotal or verrtical motion. It is just another scheme of doing the search cascade: hexagonal pattern with varying distance. It's the "best" available approximation short of doing a full search, but much faster.
(And search=3 "full" search surely is slow, but mostly a nonsense joke when used in combination with the standard "truemotion=true" parameter set of manalyse.)

Could be that the sigma's and thSAD were a little stronger than needed. Going down to

sigma/2/3/4 = 2.5, 5, 3, 2, // and in MDegrain: // thSAD=239

should be sufficient to do its job on the grain, and perform better at keeping the blazing flames.


BTW, nobody will notice if some flame strokes are a bit shortened or weakened in single frames. At playback speed, the changes are too fast to see.

But, if you concentrate too much on single-frame screenshots, and end up with some half-hearted (or overzealous, for that matter) denoising in general - that is what will be noticed.