Nasty noise on a bad DVD transfer

[samf88]

Hi, hope someone out there can help me out.

I've ripped a DVD which looks shockingly bad and I was hoping there was something I could do to clean it a little. I fear it's too far gone and I'd end up decimating what little detail there is, but maybe there's a magic filter chain out there that's perfect for making this situation slightly more bearable? I think this will be one of those shows that was edited on video so a lot of the damage may be permanent.

I've been dabbling with MDegrain3 which does clear up small parts of the picture on default settings, but everything else (such as TBilateral, VagueDenoiser, RemoveGrain) has to be ramped up to insane settings to even touch the sides of this awful noise, probably because each speck is massive. Does it have a name? I thought originally it was chroma noise, but now I'm not so sure.

This is a particularly bad frame, just using MPEG2Source and QTGMC's Very Slow preset.
https://ptpimg.me/ge2w50.png

Here's a video excerpt cut from the VOB, incase it's easier to identify frame-by-frame: https://drive.google.com/open?id=1SrRSNdypSfD6ywuGSRfF7Rj8qShxhu3h

If anyone can shed some light - or even recommend a filter/chain that may help in some way - I'd really appreciate it. This show has a very special place in my heart and I wish it was treated better for its (now out of print) DVD release.

I use AVISynth+ 64 bit, if that matters.

[tebasuna51]

For what do you use QTGMC?
It is not interlaced, and you can use the QTGMC denoisers directly to see what is your prefered.

[samf88]

Hi - thanks for replying!

QTGMC is being used to deinterlace the TFF footage which contains visible combing. I did try with EZDenoise early on in my quest but unfortunately it didn't really affect this type of noise on sensible values, and of course high values just turn it to a blur. More advanced settings - such as switching to dfttest - also don't seem to touch it.

I'm guessing I would need to combine a filter with some kind of edge mask to try and retain some detail, but I highly doubt I'll be able to get good results from this.

[hello_hello]

Just throwing some scripts at it, I preferred the TemporalDegrains.

TemporalDegrain removes more noise than TemporalDegrain2, but it blurs more. You'd probably have to encode a bit to judge how it looks when the pictures are moving. MAA is from the AnimeIVTC script, or there's MAA2. It's one of real.finder's functions, as is DeHalo_alpha_mt
https://forum.doom9.org/showthread.php?t=174121

Some SMDegrain experts might come along with some clever denoising settings using it instead.

MAA()
TemporalDeGrain()
DeHalo_alpha_mt()

Source
https://i.postimg.cc/1g6rZ8DN/Source.png (https://postimg.cc/1g6rZ8DN)

TemporalDegrain
https://i.postimg.cc/jDDhZg5c/Temporal-Degrain.png (https://postimg.cc/jDDhZg5c)

TemporalDegrain2
https://i.postimg.cc/WhWnmL2t/Temporal-Degrain2.png (https://postimg.cc/WhWnmL2t)

[samf88]

Thanks! That's wonderful, gives me something to build on. I really appreciate your help, thank you. :-)

If anyone else has any alternative suggestions I'd still be happy to hear them.

[hello_hello]

I had a silly idea. I can't say I've done it before myself (it'd be dog slow), and I don't know if there'd be a better way to go about it, but what if you only denoised the chroma, then denoised the chroma and luma? You could probably get away with less noise removal the second time around because it'd look more like grain. It's probably mainly an improvement for flat areas of color, such as the sky, and not so much everywhere else.

A = MAA()
MergeChroma(A, A.TemporalDegrain())
TemporalDegrain2()
DeHalo_alpha_mt()

https://i.postimg.cc/68pg6KTF/tgft.png (https://postimg.cc/68pg6KTF)

You could do something like this, just to double denoise where it's needed.

A = MAA()
B = A.TemporalDeGrain2()
C = MergeChroma(A, A.TemporalDegrain()).TemporalDegrain2()

B.Trim(0,478) ++ C.Trim(479,683) ++ B.Trim(684,0)
DeHalo_alpha_mt()

Except the logic for running two instances of TemporalDeGrain2 in a script doesn't seem right. To make that work you need to change line 156 in the script to this:

ReplaceGlobals = true

[Mounir]

That should be more than enough to me:

SetMemoryMax(1200)
SetMTMode(3,4)
MPEG2Source("tgft.d2v", cpu=0, info=3)
# Step 1: Remove chroma noise:
assumetff()
ConvertToYV16(interlaced=true)
orig=last
ev=orig.assumetff().separatefields().selecteven()
od=orig.assumetff().separatefields().selectodd()
ev
ue_chroma = UToY(ev).blur(0,1.5).binomialblur(5).ttempsmooth(maxr=2,lthresh=150, strength=1)
ve_chroma = VToY(ev).blur(0,1.5).binomialblur(5).ttempsmooth(maxr=2,lthresh=150, strength=1)
YToUV(ue_chroma, ve_chroma)
MergeLuma(ev)
ev_filtered=last
od
uo_chroma = UToY(od).blur(0,1.5).binomialblur(5).ttempsmooth(maxr=2,lthresh=150, strength=1)
vo_chroma = VToY(od).blur(0,1.5).binomialblur(5).ttempsmooth(maxr=2,lthresh=150, strength=1)
YToUV(uo_chroma, vo_chroma)
MergeLuma(od)
od_filtered=last
interleave(ev_filtered,od_filtered)
assumefieldbased().assumetff().weave()
# Step2: color correction
ConvertToRGB32(interlaced=true)
########
LoadVirtualDubPlugin("C:\Program Files (x86)\VirtualDub\plugins\CMYK_rj.vdf", "CMYK", 0)
CMYK(255,0,266,10,257,-24,255,0,0,42) # Modifying Yellow high tones (-24%) and magenta lows (4%)

https://www.hostpic.org/view.php?filename=2005220003200102.jpg

https://www.hostpic.org/view.php?filename=2005220130260109.jpg

[Boulder]

QTGMC(preset="very slow", inputtype=2) should already help nicely as it will press down the sharp noise and remove some aliasing.
It's hard to remove all the noise without removing all the detail. If you like, you could hit it with something TemporalDegrain2 and apply some light artificial noise with GrainFactory to fake detail.

[samf88]

I had a silly idea. I can't say I've done it before myself (it'd be dog slow), and I don't know if there'd be a better way to go about it, but what if you only denoised the chroma, then denoised the chroma and luma? You could probably get away with less noise removal the second time around because it'd look more like grain. It's probably mainly an improvement for flat areas of color, such as the sky, and not so much everywhere else.


Thank you so much, I do really like those results. Thanks so much for your help :D

That should be more than enough to me:


Thanks so much for that. I've had to modify it slightly for 64 bit, but I do like the results. Still trying to find a 64 bit version of the VDub filter, but I tried it manually in VDub itself to see the results. Thanks for taking the time to work on that, I really appreciate it.

QTGMC(preset="very slow", inputtype=2) should already help nicely as it will press down the sharp noise and remove some aliasing.
It's hard to remove all the noise without removing all the detail. If you like, you could hit it with something TemporalDegrain2 and apply some light artificial noise with GrainFactory to fake detail.

Thanks - definitely thinking I will do pre-treatment with QTGMC before anything else. I will look in to GrainFactory - that's a good idea to reintroduce some "grain" to disguise other things.

Much appreciated everyone! :D

EDIT: Here's what I've been able to achieve so far, using the help provided here: http://www.framecompare.com/screenshotcomparison/99J2JNNU. I'm much happier with this than what I was getting a few days ago.
:thanks:

[Katie Boundary]

I swear by TNLmeans: http://avisynth.nl/index.php/TNLMeans

There's also a KNLmeans but I think it requires an NVIDIA GPU or something.

[samf88]

Thanks! I actually use KNLMeansCL quite often but on its own it wasn't touching the main problem. My current filter chain (with results shown in the link I edited in my post above) does include it as part of a denoise mask. :D

[hello_hello]

Thanks! I actually use KNLMeansCL quite often but on its own it wasn't touching the main problem. My current filter chain (with results shown in the link I edited in my post above) does include it as part of a denoise mask. :D

So don't keep us in suspense. What's the script you're using?

[samf88]

Below is what I've got at the moment. Hopefully it all makes sense and I haven't made any silly mistakes. I'm using AvsPmod to preview/make changes which seems to crash every time I change a parameter, so it wasn't an easy task to adjust things. I haven't yet found an alternative filter to adjust the CMYK levels as suggested by Mounir. I'm also not confident in my grain settings and keep adjusting them. It's possible I've used the masks wrong too - I never was confident using those. I may have switched colour space more than necessary too. I've also taken out all the SetFilterMTMode/Prefetch stuff for the sake of posting here.

MPEG2Source("movie1.d2v", cpu=0)

# Mounir's Chroma Noise Removal
assumetff()
ConvertToYV16(interlaced=true)
orig=last
ev=orig.assumetff().separatefields().selecteven()
od=orig.assumetff().separatefields().selectodd()
ev
ue_chroma = UToY(ev).blur(0,1.5).binomialblur(5).ttempsmooth(maxr=2,lthresh=150, strength=1)
ve_chroma = VToY(ev).blur(0,1.5).binomialblur(5).ttempsmooth(maxr=2,lthresh=150, strength=1)
YToUV(ue_chroma, ve_chroma)
MergeLuma(ev)
ev_filtered=last
od
uo_chroma = UToY(od).blur(0,1.5).binomialblur(5).ttempsmooth(maxr=2,lthresh=150, strength=1)
vo_chroma = VToY(od).blur(0,1.5).binomialblur(5).ttempsmooth(maxr=2,lthresh=150, strength=1)
YToUV(uo_chroma, vo_chroma)
MergeLuma(od)
od_filtered=last
interleave(ev_filtered,od_filtered)
assumefieldbased().assumetff().weave()

## Mounir's Colour Correction
#ConvertToRGB32(interlaced=true)
#LoadVirtualDubPlugin("VirtualDub\x86\plugins32\CMYK_rj.VDF", "CMYK", 0)
#CMYK(255,0,266,10,257,-24,255,0,0,42) # Modifying Yellow high tones (-24%) and magenta lows (4%)

ConvertToYV12()

LoadPlugin("DeGrainMedian64\DeGrainMedian.dll")
Import("kirsch.avsi")
aux=DeGrainMedian(mode=0, limity=7, interlaced=true)
denoised = KNLMeansCL(d=4,a=4,s=6,h=2.0,channels="Y",rclip=aux)
edgemask1 = ConvertToRGB32(interlaced=true).Kirsch().ConvertToYV12(interlaced=true,matrix="PC.601")
edgemask2 = edgemask1.mt_binarize(120).Blur(1.0).Blur(1.0)
mt_merge(last,denoised,edgemask2)

# Deinterlace
QTGMC(preset="Very Slow", FPSDivisor=2)

# End filtering
DeHalo_alpha_mt()
Grainfactory3mod(g1str=2,g2str=2,g3str=2)

# Shifting image up 1px and cropping (I should probably do the shift and crop all in RGB32 colour space actually)
ConvertToRGB32(matrix="Rec601", interlaced=false)
Crop(0,1,0,0)
AddBorders(0,0,0,1)
ConvertToYV12()
Crop(10, 2, -18, -2)
FillMargins(0,1,0,1) # I like to keep to mod4 resolutions, so this just makes the extra 2px less obvious

[Boulder]

Your QTGMC call is incorrect. The sample was not interlaced like tebasuna51 mentioned. Set inputtype=2 instead of FPSDivisor if you want to see how it works on deinterlaced (as that's what your sample looks like) content.

[samf88]

Hi,

I've adjusted it now, thanks.
http://www.framecompare.com/image-compare/screenshotcomparison/77DPWNNX

[Boulder]

Hi,

I've adjusted it now, thanks.
http://www.framecompare.com/image-compare/screenshotcomparison/77DPWNNX

You could also try inputtype=3, I couldn't tell which one looks better but type 2 seems worse in that sample screenshot.

[samf88]

Just gave it a go with inputtype=3 (http://www.framecompare.com/image-compare/screenshotcomparison/99J21NNU). Judging by the pattern on the red blouse alone I'd say it looks best treating it as interlaced (inputtype=0 and discarding the dupe frames using FPSDivisor). The stripey top and pants definitely look better with inputtype=3 though.

[hello_hello]

I use QTGMC in progressive mode quite a lot for denoising, but if you've already denoised and it's just for removing the combed edges...

I ask, as I can't run KNLMeansCL on this PC, so I haven't seen the final denoising, but InputType-2 throws away every second scanline, which is why it's fixing the jaggies. InputType=1 doesn't (I don't think), so it's not fixing the jaggies, but I wonder if MAA().QTGMC(InputType=1) might retain more detail. Just a thought.

A very minor nitpick...
I'm pretty sure the final cropping and resizing can be done this way to avoid the conversion to RGB and back. The resizer is just shifting the pixels by 1.

W = Width()
H = Height()
Spline36Resize(W, H, 0,1,W,H)
Crop(10, 2, -18, -2)
FillMargins(0,1,0,1)

Mind you, without resizing,
Crop(10,2,-18,-2) gets you to 1.32 with a SAR of 12:11
or
Crop(10,2,-18,-2) gets you to 1.29 with a SAR of 16:15

I'm sure that DVD would have a 12:11 (or ITU) SAR.
https://forum.doom9.org/showthread.php?p=1058927#post1058927

I'm a bit OCD about a minimum aspect ratio of 4:3, so I'd probably crop a few more pixels top and bottom to get there. I'd also crop a few more pixels left and right to get rid of the "halo line" on the left and make the right border cleaner.

The extra pixels you'll need to crop top and bottom for 4:3 wouldn't have been seen on a CRT due to over-scanning anyway, and a 4:3 picture fills a bit more of a 16:9 screen. That's just my personal preference though.

If you're interested, you can look at all that fun stuff with my script, whether you resize or not. It'll tell you what the final aspect ratio will be, and you can specify an aspect ratio for it to crop to, to see how much extra it'll crop etc.

CropResize(0,0, 16, 2, -20, -2, InSAR=12.0/11.0, CropDAR=4.0/3.0, NoResize=true, Info=true)
https://i.postimg.cc/PNYCbkx0/Crop-Resize.jpg (https://postimg.cc/PNYCbkx0)

[samf88]

I ask, as I can't run KNLMeansCL on this PC, so I haven't seen the final denoising, but InputType-2 throws away every second scanline, which is why it's fixing the jaggies. InputType=1 doesn't (I don't think), so it's not fixing the jaggies, but I wonder if MAA().QTGMC(InputType=1) might retain more detail. Just a thought.

There seems to be more jaggies doing it that way. :(

https://ptpimg.me/6g29e4.png

A very minor nitpick...
I'm pretty sure the final cropping and resizing can be done this way to avoid the conversion to RGB and back. The resizer is just shifting the pixels by 1.

W = Width()
H = Height()
Spline36Resize(W, H, 0,1,W,H)
Crop(10, 2, -18, -2)
FillMargins(0,1,0,1)

Funny you should suggest that as I was originally using z_Spline36Resize to shift - I can't actually remember the reason why I switched to the colour space conversion method.

I'm a bit OCD about a minimum aspect ratio of 4:3, so I'd probably crop a few more pixels top and bottom to get there. I'd also crop a few more pixels left and right to get rid of the "halo line" on the left and make the right border cleaner.

Annoyingly the video has wildly different cropping values throughout, I suspect due to being edited on video or something. I don't know if there's a filter to automatically detect and rearrange different crop values, but I'm not keen on the idea of overcropping - even though that halo line is really bugging me. I never resize either, unless I'm using nnedi_rpow2 or going from 1080p>720p, etc. I set SAR in the encoder as some devices don't seem to pick up setting a DAR in the container. I get what you mean though, almost like a mini version of when TV channels cropped 4:3 images to 14:9 to help during the AR transition era.

If you're interested, you can look at all that fun stuff with my script, whether you resize or not. It'll tell you what the final aspect ratio will be, and you can specify an aspect ratio for it to crop to, to see how much extra it'll crop etc.

CropResize(0,0, 16, 2, -20, -2, InSAR=12.0/11.0, CropDAR=4.0/3.0, NoResize=true, Info=true)
https://i.postimg.cc/PNYCbkx0/Crop-Resize.jpg (https://postimg.cc/PNYCbkx0)

Very fancy! I will keep a note of that for future use. Thanks so much for all your help.

[hello_hello]

Maybe try MAA or MAA3 before denoising etc. Splitting the fields and denoising before putting them back together might effect it. I'm just theorizing...

[hello_hello]

I get what you mean though, almost like a mini version of when TV channels cropped 4:3 images to 14:9 to help during the AR transition era.

I tend to go with the theory the picture was intended to display as 4:3. You never would've seen the crud at the sides on a CRT due to overscan, and likewise you'd never have seen a few pixels top and bottom.

If you have a 16:9 TV with a 4:3 mode, try playing the DVD with a Bluray/DVD player and the TV in 4:3 mode. It should simulate CRT overscanning and give you a nice clean 4:3 picture without the crud you see on a computer screen.

[hello_hello]

One of the reasons I resize, is because the script adjusts the cropping to prevent aspect error, so you can do something like this to crop extra black for a scene, and the script will crop a bit more as required so you don't have to worry about aspect error.

Trim(0,999).CropResize(640,480, 14,4,-16,-4, InDAR=15.0/11.0) ++ \
Trim(1000,1999).CropResize(640,480, 14,4,-48,-18, InDAR=15.0/11.0) ++ \
Trim(2000,0).CropResize(640,480, 14,4,-16,-4, InDAR=15.0/11.0)

You can go with a "least amount of resizing" resolution for the majority of the cropping, and still know for the sections where the cropping needs to be changed, the picture won't be distorted.

But I invariably resize, especially DVDs, as I always resize to square pixel dimensions.

Edit:

or....

Trim(0,999).Crop(16,4,-16,-4) ++ \
Trim(1000,1999).CropResize(688,568, 14,4,-48,-18, InSAR=12.0/11.0, OutSAR=12.0/11.0) ++ \
Trim(2000,0).Crop(16,4,-16,-4)

would allow you to easily crop some extra and resize a section to match the rest, once again without having to worry about aspect error.

[samf88]

Here is a comparison with what I had originally (InputType=0), and MAA before all denoising with QTGMC InputType=1.
http://www.framecompare.com/image-compare/screenshotcomparison/77DPGNNX

[hello_hello]

Here is a comparison with what I had originally (InputType=0), and MAA before all denoising with QTGMC InputType=1.
http://www.framecompare.com/image-compare/screenshotcomparison/77DPGNNX

I can't argue about what looks best...

[VoodooFX]

Here is a comparison with what I had originally (InputType=0), and MAA before all denoising with QTGMC InputType=1.
http://www.framecompare.com/image-compare/screenshotcomparison/77DPGNNX

Don't deinterlace progressive clip with QTGMC. On some areas it can fix aliasing, when it can create aliasing on other frames and can destroy other things (notice that moire on the butt at the right).

[VoodooFX]

What about something like:

MCDegrainSharp( frames=1, bblur=0.2, csharp=0.0 )
QTGMC( Preset="Medium", InputType=1, Sharpness=0.6 )
Santiag(1,1, nns=4)

QTGMC "Medium" wouldn't blur fast motion areas much like other presets, but sometimes in can create some stabilization issues, when "Slower" stabilization is more precise/better.

I preferred the TemporalDegrains.

Don't you get some strange "effects-bugs" with TemporalDegrain2? Every time I tried it I noticed something strange going on besides noise removal.

[samf88]

What about something like:

MCDegrainSharp( frames=1, bblur=0.2, csharp=0.0 )
QTGMC( Preset="Medium", InputType=1, Sharpness=0.6 )
Santiag(1,1, nns=4)

QTGMC "Medium" wouldn't blur fast motion areas much like other presets, but sometimes in can create some stabilization issues, when "Slower" stabilization is more precise/better.

Many thanks for your contribution. :D

I'm seeing aliasing on the black/white stripey top with InputType=1, but InputType=3 seems to solve that. http://www.framecompare.com/image-compare/screenshotcomparison/999F1NNU

I think I might have to treat the end credits differently to the rest of the script though. There does seem to be 50fps worth of unique frames there in the background graphic, although given that it's only end credits I might just leave it. http://www.framecompare.com/image-compare/screenshotcomparison/999FMNNU InputType=1 does seem to look best for the text, but InputType=3 looks best for the graphics. (Link if anyone wants to investigate: https://drive.google.com/file/d/1ohK7l5jla73WtYxbfPTmG6TXdfmr9s8W/view?usp=sharing)

[hello_hello]

Don't you get some strange "effects-bugs" with TemporalDegrain2? Every time I tried it I noticed something strange going on besides noise removal.

I haven't used it much, but I didn't see anything nasty. The main problem I have with it, is I don't seem to be able to adjust how much noise it removes to any great extent. When there's just the right amount of noise it works, but considering the denoising logic came from QTGMC you'd think it'd have something like the EzDenoise option where you can simply dial in how much noise removal you want.

One day I might pull QTGMC apart to see if I can do better. Probably not, but maybe one day I'll try.