Awful vertical noise Stripes

[Mr. Brown]

http://img329.imageshack.us/img329/656/stripes3mh.png (http://imageshack.us)

These Vertical Noise Stripes i doesnt get out
The other "normal" noise is no problem with fft3d(slow) or hqdn3d (speed)
but the noise stripes keep alive

i tried

horizontalreduceby2() is good but i loose much details

turnleft().DeGrainMedian(limitY=6,limitUV=10,mode=0,norow=true).turnright() helps a little bit but not much

defreq(fx=0) doesnt help too

here is a sample (phase shifted & fielblended PAL)
http://rapidshare.de/files/12986763/stripes.rar.html

excuse me for my bad english

[iantri]

forgive me for being a bit obtuse, but .. what stripes? :|

Maybe the JPEG compression artefacts are killing them in the image you posted?

[Didée]

I tried:

Preparing with FFT3D, improving with limited horizontal medianfiltering, then re-merging the original on what still appears edgy.

Before - after of frame 210 (bobbed) :

http://img134.imageshack.us/img134/9769/frame210orig3un.th.png (http://img134.imageshack.us/my.php?image=frame210orig3un.png) http://img135.imageshack.us/img135/134/frame210after5qf.th.png (http://img135.imageshack.us/my.php?image=frame210after5qf.png)


It's not exactly fast :D - had no other idea how to get a reasonable grip to the problem.
The script needs MaskTools v2.0a26.

mpeg2source("J:\stripes.demuxed_cut.d2v")

o = last
before = o.bob()

fft = o.fft3dfilter(sigma=2,sigma2=5,sigma3=2,sigma4=1,bt=3,degrid=1,interlaced=true)
sep = fft.separatefields()
flat = mt_luts( sep, sep, mode="med", pixels=mt_rectangle(2,1,false), expr="y", U=2,V=2 )
mt_lutxy(sep,flat,yexpr="x 4 + y < x 3 + x 4 - y > x 3 - y ? ?",U=2,V=2)
weave()
bobbed = bob()
edge = bobbed.mt_edge("min/max",0,3,0,3)
final = mt_merge(bobbed,before,edge,U=4,V=4)

#interleave(before,final) # for switch-frame comparison
final

return( last )

[Oline 61]

Out of curiosity, have you tried Pixie Dust?

[Didée]

No, we're waiting for you to report the results :p

MoComp'ed denoising is very difficult with this: in areas without real detail (sky), the motion search engine most likely will manage to "snap on" the noise, finding some best matches in such a way that the noise persists.

[BruceL]

@Mr Brown
Doesn't look like you had much in your script for Defreq.
Can you post an image with Defreq set to "Show=True"?

[Boulder]

Next time, could you please post the thumbnails only for faster page loading. Many people still use 56k modems (not me though).

[Mr. Brown]

@ Didée
you can not use fft3d(interlaced=true) & Separatefields().filter().weave() in your script because some frames are progressive and some frames are interlaced in my source this will produce bad artifacts

your script is good to remove the vertical noise stripes but it also destroy much details

Out of curiosity, have you tried Pixie Dust?

Yes of course PixieDust(5) but it doesnt help much

@Mr Brown
Doesn't look like you had much in your script for Defreq.
Can you post an image with Defreq set to "Show=True"?

defreq is a frequency remover and not a noise remover i tested it only because someone told it was a good vertikal filter

[Didée]

@ Didée
you can not use fft3d(interlaced=true) & Separatefields().filter().weave() in your script because some frames are progressive and some frames are interlaced in my source this will produce bad artifacts
Sure I can :)
Most frames *are* interlaced, and on the few progressive ones, fft3d with interlaced=true will not do very much harm.
Then, what above is happening inbetween separatefields()/weave(), actually is a pure spatial, horizontal-only filtering, so it won't do any harm either. (It's not that easy to see from the script, I agree ;) )

However, some detail gets lost, that's true. Compare with
pixie = o.bob().converttoyuy2().pixiedust(3).converttoyv12()
pixie .fft3dfilter(sigma=2,sigma2=4,sigma3=2,sigma4=0,bt=3,degrid=1,interlaced=false)
this removes the noise-stripes not fully as good, but also removes detail.

You'll have to make some compromises here, no way out.
Mo'comped filters definetly will have problems with this source, (a) because of what was noted in previous post and (b) for the simple fact that the source is fieldblended ... motion compensation *can not* work fully successful on blended fields/frames.

And with traditional spatio/temporal filters, you will necessarily loose the detail that's lying below the noise level.

Frequency filtering could be the best way to go ... but it will also remove those (vertical) details that look similar to the noise.

Choose your poison.