DGDenoise

These CUDA filters are packaged into DGDecodeNV, which is part of DGDecNV.
User avatar
admin
Site Admin
Posts: 4415
Joined: Thu Sep 09, 2010 3:08 pm

Re: About DGDenoise

Post by admin » Fri Mar 10, 2017 1:43 pm

Sharc wrote:I have to deinterlace or separate the fields anyway for applying most of the filters, and putting converttoYV12(interlaced=true) early in the script is the way to go.
Yes, Sharc, I agree. Let's keep things as is for now. You're right that 4:2:2 support would be useful and I'd certainly want to support it when CUVID does. People are hopeful for nVidia Video SDK 8.0, but it is still awaited. On the other hand, it is only a limitation of CUVID decoding. I could make CUDA 4:2:2 filters. Let's see what nVidia Video SDK 8.0 brings us.

User avatar
admin
Site Admin
Posts: 4415
Joined: Thu Sep 09, 2010 3:08 pm

Re: About DGDenoise

Post by admin » Fri Mar 10, 2017 5:13 pm

hydra3333 wrote: This wasn't me (it hung on the shack wall for 30 years) but it may as well have been https://drive.google.com/open?id=0B5RV2 ... WhpcjZnY3M
Oh why not, a relative of mine is in this one https://drive.google.com/open?id=0B5RV2 ... HRZYWtFbDQ
The spirit of adventure and discovery shines. So blessed you are to know such people. I wish they were forum members. :)

Aleron Ives
Posts: 113
Joined: Fri May 31, 2013 8:36 pm

Re: About DGDenoise

Post by Aleron Ives » Fri Mar 10, 2017 8:22 pm

This might be a dumb question, but since you're working on adding new filters, are there any prospects for replicating the functionality of Decomb this way? That's my most-used filter, and getting a CUDA speedup for it would be welcome. I'm not sure how closely related this would be to your recent work on DGBobIM or whether CUDA is even appropriate for this purpose... :?

User avatar
admin
Site Admin
Posts: 4415
Joined: Thu Sep 09, 2010 3:08 pm

Re: About DGDenoise

Post by admin » Fri Mar 10, 2017 8:54 pm

Great idea, Aleron! Sure, we could speed up field matching and decimation. Why didn't I think of these obvious things? :scratch: Thank you for the valuable suggestion.

DGBobIM was mostly a proof-of-concept for how screwed up the IMSDK API is. Not exciting to work on that stuff.

BTW, Aussies swim real good. We may have to keep an eye on them.

Sharc
Moose Approved
Posts: 231
Joined: Thu Sep 23, 2010 1:53 pm

Re: About DGDenoise

Post by Sharc » Sat Mar 11, 2017 2:41 am

Aleron Ives wrote:This might be a dumb question, but since you're working on adding new filters, are there any prospects for replicating the functionality of Decomb this way? That's my most-used filter, and getting a CUDA speedup for it would be welcome. I'm not sure how closely related this would be to your recent work on DGBobIM or whether CUDA is even appropriate for this purpose... :?
+1 for implementing Decomb (especially the FieldMatching/IVTC functionality of telecide() )

P.S.
I also like the show= and debug= options of FieldDeinterlace(), however speed is not critical for visual analysis.

gonca
Moose Approved/Curly Approved
Posts: 911
Joined: Sun Apr 08, 2012 6:12 pm

Re: About DGDenoise

Post by gonca » Sat Mar 11, 2017 7:33 am

+1 as well

User avatar
hydra3333
Moose Approved
Posts: 206
Joined: Wed Oct 06, 2010 3:34 am
Contact:

Re: About DGDenoise

Post by hydra3333 » Sun Mar 12, 2017 6:44 am

Thank you, the new gpu based filters work nice.

OK, I did a test or two displaying an interleaved (source,result1,result2) clip in vsedit.
Reasonably clean, if blocky, 15Mb Source: https://drive.google.com/open?id=0B5RV2 ... Hk1UlA5c2c
Card: 750Ti, CPU: i3820 (due for replacement next year; likely a cheapie ryzen 1800X and 1050Ti, not a 1070 though given touted v2's of 1060 and 1080)

Confirming per some other posts eg viewtopic.php?f=8&t=506&start=170#p6074 that the default setting (strength=0.15) for DGDenoise seem to yield a more"softened" visual result compared to KNLMeansCL at defaults (d=1 temporal) at least on my test source. So, easy, use .06 or something.

Sometimes a really old movie can be a bit of a shocker with grain and whatnot and the best thing I've found to date has been the old MDegrain1/2/3 eg
video=MDegrain3i2(video,blksize=8,overlap=4,dct=0)
which I suppose does OK due to temporal processing ?
I haven't such a source to hand to try, so what would be your thoughts re DGDenoise / KNLMeansCL(d=3) / MDegrain for those types of sources ?

I see with interest you considered the possibility of temporal option viewtopic.php?f=8&t=506&start=210#p6137 and then some hesitation viewtopic.php?f=8&t=506&start=220#p6153 and then a possible position of sticking with spatial viewtopic.php?f=8&t=506&start=260#p6253
... I'm hoping the DG density matrix is flagging that gpu temporal denoising still has a significant possibility of a state of existance :) ... no plans for temporal ?
(google gives the words but unfortunately not the intellect to comprehend it).

Just came across the tips from this crowd, although I don't really know what their credentials are
http://www.cineticstudios.com/blog/2015 ... ction.html

vapoursynth x64 shouldn't make any difference, eg

Code: Select all

import vapoursynth as vs 
##import havsfunc as haf # http://forum.doom9.org/showthread.php?t=166582 
import havsfuncTS as haf # this version uses vanilla TemporalSoften instead of TemporalSoften2, as it will be "better" over time 
import mvsfunc as mvs  # http://forum.doom9.org/showthread.php?t=172564 
import finesharp as finesharp # http://forum.doom9.org/showthread.php?p=1777815#post1777815 http://avisynth.nl/index.php/FineSharp 
core = vs.get_core(accept_lowercase=True) # leave off threads=8 so it auto-detects threads 
# the r'' indicates do not treat special characters and accept backslashes 
core.std.LoadPlugin(r'C:\SOFTWARE\Vapoursynth-x64\vapoursynth64\plugins\dll-to-choose-from\KNLMeansCL.dll') # the r'' indicates do not treat special characters and accept backslashes 
# LOAD 64 bit AVISYNTH  plugins  into the avs namespace 
core.avs.LoadPlugin(r'C:\SOFTWARE\Vapoursynth-x64\DGIndex\DGDecodeNV.dll')  
def main(): 
    video = core.avs.DGSource(r'.\z.dgi',deinterlace=1,resize_w=720,resize_h=576) # deinterlace=1 means single rate deinterlacing 
    videooriginal = video 
    video1 = core.knlm.KNLMeansCL(video, device_type="gpu", d=1, a=2) 
    video2 = core.avs.DGDenoise(video,strength=0.06)
    #video = core.avs.DGSharpen(video,strength=0.40) # 
    video = core.std.Interleave([core.text.Text(videooriginal, 'Source clip'), core.text.Text(video1, 'LNKMeansCL clip'), core.text.Text(video2, 'DGDenoise clip')])
    video.set_output() 
    return True 
main() 
[/size]

User avatar
admin
Site Admin
Posts: 4415
Joined: Thu Sep 09, 2010 3:08 pm

Re: About DGDenoise

Post by admin » Sun Mar 12, 2017 7:27 am

hydra3333 wrote:Sometimes a really old movie can be a bit of a shocker with grain and whatnot and the best thing I've found to date has been the old MDegrain1/2/3 eg
video=MDegrain3i2(video,blksize=8,overlap=4,dct=0)
which I suppose does OK due to temporal processing ?
I haven't such a source to hand to try, so what would be your thoughts re DGDenoise / KNLMeansCL(d=3) / MDegrain for those types of sources ?
It's not worth speculating about it in the absence of a clip. When you have one to share I would be happy to discuss comparative results.
I see with interest you considered the possibility of temporal option viewtopic.php?f=8&t=506&start=210#p6137 and then some hesitation viewtopic.php?f=8&t=506&start=220#p6153 and then a possible position of sticking with spatial viewtopic.php?f=8&t=506&start=260#p6253
... I'm hoping the DG density matrix is flagging that gpu temporal denoising still has a significant possibility of a state of existance :) ... no plans for temporal ?
I've asked several times for a clip that clearly shows the benefits of temporal processing, either alone or in conjunction with spatial processing. Do you have one? That is what I would need to find the motivation to do something temporally. Meanwhile, there is always TemporalSoften(). My personal perspective (which can be altered by evidence) is that ghosting is anathema that I would avoid like the plague in the absence of palpable overriding considerations. Yeah, yeah, scene change detection. But not every frame is a scene change while it can still be full of motion, and hence ghosting. Please refute my view with some evidence, i.e., a clip.

If you want to consider motion compensation to combat ghosting, that's perfectly valid. But that has to wait for my CUDA motion estimation stuff.

DG density matrix. That's clever. Not everyone will grok it, though, especially with negative off-diagonal terms.
Just came across the tips from this crowd, although I don't really know what their credentials are
http://www.cineticstudios.com/blog/2015 ... ction.html
It's just warmed-over wives' tales, aka, conventional "wisdom", justifying some commercial product recommendations. I can't take seriously an article that provides a tip on noise reduction that advises to add noise. :facepalm:

Monty Python video enhancement algorithm:

1. Remove noise.
2. Add noise.

That tip is there only to push some commercial tools. The tip to prefer temporal processing is also just a push for a commercial denoiser. They conveniently neglect to mention ghosting. And it's strange that not a single image is given to support anything they say.

Sharc
Moose Approved
Posts: 231
Joined: Thu Sep 23, 2010 1:53 pm

Re: DGDenoise

Post by Sharc » Sat Apr 08, 2017 5:44 am

I am getting ugly artefacts (broad diagonal color stripes) with chroma=true in DGDenoise.
My source is interlaced video. I separate the fields before applying the filter. Unless I am missing something you should be able to duplicate the issue with any source clip.

User avatar
admin
Site Admin
Posts: 4415
Joined: Thu Sep 09, 2010 3:08 pm

Re: DGDenoise

Post by admin » Sat Apr 08, 2017 6:50 am

Yikes! I'll get on it this morning. Thanks for the report.

User avatar
admin
Site Admin
Posts: 4415
Joined: Thu Sep 09, 2010 3:08 pm

Re: DGDenoise

Post by admin » Sat Apr 08, 2017 7:59 am

All fixed. Thanks for reporting it.

Sharc
Moose Approved
Posts: 231
Joined: Thu Sep 23, 2010 1:53 pm

Re: DGDenoise

Post by Sharc » Sat Apr 08, 2017 8:48 am

Confirmed. It's working.

User avatar
admin
Site Admin
Posts: 4415
Joined: Thu Sep 09, 2010 3:08 pm

Re: DGDenoise

Post by admin » Sun Apr 09, 2017 11:29 am

Cool, thanks again Sharc.

gonca
Moose Approved/Curly Approved
Posts: 911
Joined: Sun Apr 08, 2012 6:12 pm

Re: DGDenoise

Post by gonca » Tue Apr 18, 2017 11:45 am

Does DGDenoiseNV handle what is on this sample, film grain or mosquito noise, I believe?
http://www.mediafire.com/file/1djsss779 ... sample.mkv

If not, can you recommend a x64 filter?

User avatar
admin
Site Admin
Posts: 4415
Joined: Thu Sep 09, 2010 3:08 pm

Re: DGDenoise

Post by admin » Tue Apr 18, 2017 1:03 pm

Something is broken. Investigating...

User avatar
admin
Site Admin
Posts: 4415
Joined: Thu Sep 09, 2010 3:08 pm

Re: DGDenoise

Post by admin » Tue Apr 18, 2017 1:57 pm

Oy, that last slipstream was brain-dead. :oops:

OK, I fixed the DGDenoise problem. Now I have one little issue in DGTelecide. Fixes coming soon...

Thanks for pointing this out, gonca.

gonca
Moose Approved/Curly Approved
Posts: 911
Joined: Sun Apr 08, 2012 6:12 pm

Re: DGDenoise

Post by gonca » Tue Apr 18, 2017 2:47 pm

Thanks for creating the filters

User avatar
admin
Site Admin
Posts: 4415
Joined: Thu Sep 09, 2010 3:08 pm

Re: DGDenoise

Post by admin » Tue Apr 18, 2017 3:07 pm

Thanks. Of course, it would be better if they worked properly. ;)

I just slipstreamed the fix. Heading over to the binaries thread to announce it.

User avatar
admin
Site Admin
Posts: 4415
Joined: Thu Sep 09, 2010 3:08 pm

Re: DGDenoise

Post by admin » Tue Apr 18, 2017 3:09 pm

This works pretty good on your sample, gonca:

dgsource("sample.dgi")
dgdenoise(strength=2.0,searchw=9,chroma=true)
dgsharpen()

That's industrial strength. Reduce strength and/or searchw, or turn off chroma, if you like.

gonca
Moose Approved/Curly Approved
Posts: 911
Joined: Sun Apr 08, 2012 6:12 pm

Re: DGDenoise

Post by gonca » Tue Apr 18, 2017 3:27 pm

I have read the documentation but would you recommend searchw=9 or chroma+true over the defaults in general terms?

User avatar
admin
Site Admin
Posts: 4415
Joined: Thu Sep 09, 2010 3:08 pm

Re: DGDenoise

Post by admin » Tue Apr 18, 2017 3:42 pm

Generally, it's a tradeoff of speed against thoroughness (aka, "quality").

Take chroma. Make this script:

dgsource("sample.dgi")
UtoY() # visualize the chroma U plane

and compare it to this:

dgsource("sample.dgi")
grayscale() # visualize the luma plane

You will see that there is lots of noise in the luma but only a relatively small amount in the chroma. That's pretty typical. So adding chroma=true will slow things down while not doing much for the perceived result.

Regarding searchw. The bigger it is the better the smoothing but you lose some detail and it takes longer.

I usually just go with searchw=5 and chroma=false and then set the strength as desired. I don't work with very noisy sources like your sample, however. The knobs are there for you to tweak for noisier sources.

gonca
Moose Approved/Curly Approved
Posts: 911
Joined: Sun Apr 08, 2012 6:12 pm

Re: DGDenoise

Post by gonca » Tue Apr 18, 2017 4:15 pm

Thanks for the explanation and information.
I normally don't work with sources that noisy but I do have some older movies hat seem to be from the era of "Movie Grain is King"

gonca
Moose Approved/Curly Approved
Posts: 911
Joined: Sun Apr 08, 2012 6:12 pm

Re: DGDenoise

Post by gonca » Tue Apr 18, 2017 4:24 pm

Running an encode and it looks like its all working as it should

User avatar
admin
Site Admin
Posts: 4415
Joined: Thu Sep 09, 2010 3:08 pm

Re: DGDenoise

Post by admin » Tue Apr 18, 2017 4:28 pm

Sweet, thanks for the report and testing. I tested this version pretty well too.

Sharc
Moose Approved
Posts: 231
Joined: Thu Sep 23, 2010 1:53 pm

Re: DGDenoise

Post by Sharc » Thu Apr 20, 2017 11:41 am

admin wrote:Generally, it's a tradeoff of speed against thoroughness (aka, "quality").

Take chroma. Make this script:

dgsource("sample.dgi")
UtoY() # visualize the chroma U plane

and compare it to this:

dgsource("sample.dgi")
grayscale() # visualize the luma plane
Why is the chroma U plane view shown at half frame width? Is it because of the YV12 4.2.0 colorspace?

Post Reply