Avisynth woes - ffmpeg hanging, denoising

[Reel.Deel]

I'm glad you like it and improved the overall speed. Here's a more tweaked script, I got rid of the superfluous denoising step and other things like use NNEDI3CL and neo_fft3d. I also incorporated the upscale, I used NNEDI3CL for that, not SuperRes (didn't check the differences) but the source has very low detail so I don't think the difference will be night and day. z_ConvertFormat resizes to the final size, converts it to YUV420, corrects the shift caused by NNEDI3 upscaling and also changes the Colorimetry all in one go . Script now also needs LSFplus.

I don't know the quality of your other tapes but this one here works well for low quality VHS with chroma bands and very noisy. Definitely overkill for higher quality footage. Enjoy, and if you come across another hard to denoise clip, please share

FFVideoSource("tape 37-trim.mkv")
Crop(8,0,-6,-16)
ConvertBits(10)
Levels(20, 1, 1045, 0, 1023, coring=false, dither=true)
Tweak(hue=-3, coring=false, dither=true)
ConvertBits(8)
src = last
NNEDI3CL(3)
o = last

mds().Spline36Resize(o.width(),o.height()/2/4*4)
mds().Spline36Resize(o.width(),o.height()/6/4*4)
mds().Spline36Resize(o.width(),o.height())

o.MergeChroma(last)
nostripes2 = last.MinBlur(1).LSFplus(ss_x=1.0, ss_y=1.0, strength=100, soft=100, overshoot=5)
srchbase   = neo_fft3d(nostripes2,sigma=6, sigma2=4, sigma3=3, sigma4=2, bt=5, bw=16, bh=16, ow=8, oh=8).FluxSmoothT()
pre        = MergeLuma(srchbase, o, weight=.25) 

QTGMCp(src, EdiExt=pre)

NNEDI3CL(1, dh=true, dw=true)
NNEDI3CL(1, dh=true, dw=true)
NNEDI3CL(1, dh=true, dw=true)

z_ConvertFormat(2880, 2160, pixel_type="YV12", colorspace_op="470bg:601:170m:full=>709:709:709:full", \ 
                resample_filter="bicubic", filter_param_a=-0.5, filter_param_b=0.25, src_left=-3.5, src_top=-3.5)



# helper function: vertical-only MinBlur with vertical radius 2

function mds(clip clp)
{
a  = clp.mt_luts(clp,mode="median",pixels="0 -2 0 -1 0 0 0 1 0 2",Y=2,U=3,V=3)
b  = clp.mt_convolution("0 1 0","1 2 1",y=2,U=3,V=3).mt_convolution("0 1 0","1 2 1",Y=2,U=3,V=3)
aD = mt_makediff(clp,a,Y=2,U=3,V=3)
bD = mt_makediff(clp,b,Y=2,U=3,V=3)
DD = mt_lutxy(aD,bD,"x 128 - y 128 - * 0 < 128 x 128 - abs y 128 - abs < x y ? ?",Y=2,U=3,V=3)
clp.mt_makediff(DD,Y=2,U=3,V=3)
return(last)
}

[StainlessS]

RD,
Not sure of your intent there, but
mds().Spline36Resize(o.width(),o.height()/2/4*4)
is same as
mds().Spline36Resize(o.width(),o.height()/2)

as intended ?

maybe something like
mds().Spline36Resize(o.width(),Round(o.height()/2.0/4.0)*4)

[Reel.Deel]

StainlessS, it's a copypasta from a script from Didée, I didn't dare to challenge his logic

Edit: to below - Round() would work, the script is meant for progressive video so mod2 should be enough.

[StainlessS]

I bet it should have been as suggested (or eg Int() instead of round), same for the 6/4/4.

EDIT: Whether or not you should use Round() or Int(), it is intended that result height should be mod4. [suggest Round()]

[StainlessS]

Edit: to below - Round() would work, the script is meant for progressive video so mod2 should be enough.

This lot would seem to require reduction to 1/2 size (mod4),
and then to reduction to 1/6 size (mod4).
Then resize back to original height.
So downsize to 1/12 height, then back up.
I imagine that the big DD had reason for double downsize, and maybe even the modulo stuff between and ending.
(dont think anything to do with Progressive/Interlaced)

mds().Spline36Resize(o.width(),o.height()/2/4*4)  # /2
mds().Spline36Resize(o.width(),o.height()/6/4*4)  # /6 : total /12
mds().Spline36Resize(o.width(),o.height())        # back to original

EDIT:
so maybe

mds().Spline36Resize(o.width(),Round(o.height()/2.0/4)*4)  # /2
mds().Spline36Resize(o.width(),Round(o.height()/6.0/4)*4)  # /6 : total /12
mds().Spline36Resize(o.width(),o.height())                 # back to original

EDIT: And instead of Int() if required, just divide by int instead of float

mds().Spline36Resize(o.width(),(o.height()/2/4)*4)  # /2
mds().Spline36Resize(o.width(),(o.height()/6/4)*4)  # /6 : total /12
mds().Spline36Resize(o.width(),o.height())          # back to original

[Reel.Deel]

Originally Posted by StainlessS

This lot would seem to require reduction to 1/2 size (mod4),
and then to reduction to 1/6 size (mod4).
Then resize back to original height.
So downsize to 1/12 height, then back up.
I imagine that the big DD had reason for double downsize, and maybe even the modulo stuff between and ending.
(dont think anything to do with Progressive/Interlaced)

mds().Spline36Resize(o.width(),o.height()/2/4*4)  # /2
mds().Spline36Resize(o.width(),o.height()/6/4*4)  # /6 : total /12
mds().Spline36Resize(o.width(),o.height())        # back to original

EDIT:
so maybe

mds().Spline36Resize(o.width(),Round(o.height()/2.0/4)*4)  # /2
mds().Spline36Resize(o.width(),Round(o.height()/6.0/4)*4)  # /6 : total /12
mds().Spline36Resize(o.width(),o.height())                 # back to original

EDIT: And instead of Int() if required, just divide by int instead of float

mds().Spline36Resize(o.width(),(o.height()/2/4)*4)  # /2
mds().Spline36Resize(o.width(),(o.height()/6/4)*4)  # /6 : total /12
mds().Spline36Resize(o.width(),o.height())          # back to original

Gotcha. If I decide to make a VHS band reducer script, I will incorporate those changes. Thanks StainlessS

[ENunn]

Thanks for the updated script.

I tried it on one of my typical VHS captures. VCR set to norm. Relatively low noise to begin with but I do want to slightly denoise before scaling. Not sure what I prefer.

Source:


Old script:


New script (with RF's first settings, modified)


New script #2 (with RF's second settings, modified)
]

No scaling applied yet. All look clean though. I'll send a sample once I'm off work. If only I could improve text clarity to make it on-par with the source.

[johnmeyer]

Do you have the edit switch on your VCR turned on? Those captures look pretty soft to me.

If your are unfamiliar with the edit switch, consult your VCR's manual to see the settings they recommend for dubbing to another tape deck. You want to use those settings. Turning that switch ON turns OFF the built-in noise filtering. You do NOT want the VCR to filter noise because it kills details in order to reduce all those little dots. You will never get back that detail once lost. Once you have your capture, the digital noise filters available today are 100x better than the simple analog circuits used in a VCR. The final result will be more detailed and better in every way.

[ENunn]

Originally Posted by johnmeyer

Do you have the edit switch on your VCR turned on? Those captures look pretty soft to me.

I honestly prefer Norm over Edit. I'm not a fan of all that added noise. I find Norm to be more natural to me. I only use Edit if I use my ES15.

I'll take your word for it though :P I'll do a capture with Edit on and try it out.

[johnmeyer]

You are absolutely correct: setting that switch to Edit will make the picture noisier. If you were just watching the video on your TV, that might be what you want, but it is not what you want when digitizing video.

To convince yourself, capture a scene which has a lot of detail (leaves, grass, striped shirt, a shag rug, etc.). Do that with Edit on and then do it again with Edit off. Then do an A/B between your videos while they are running (I do this by putting them above each other in my NLE and then muting/unmuting one track). You can also look at stills captured from each. Ignore the noise, but look at the detail. You should be able to resolve more detail, and even the detail that is already there should reveal more information when Edit is on (i.e., noise reduction is turned off).

There are dozens of AVISynth plugins that can very effectively remove most of the noise without killing the detail. The VCR approach is to just apply a low pass filter which removes the noise and also the details. It has no way to "look" at pixels (because there are no pixels in analog video) and then compare those pixels from one frame to the next, something a temporal denoiser does. It is this frame-to-frame comparison that permits a digital temporal denoiser to get rids of all those dancing dots without removing detail.

BTW, a similar comment applies to removing noise with AVISynth: resist the urge to get rid of ALL the noise because if you do that, you often remove detail.

Remember, "NR" stands for noise reduction, not noise removal.

[ENunn]

Thanks for the help John! I did my first edit mode transfer in a long time, results do look nice, albeit very noisy.

I ran the capture through the script, there's definitely a difference but I do want a little better.

Original:


RealFinder's settings (modified by me):


There's definitely a difference, but I dunno if it's on-par with what my VCR can do. At least detail isn't lost :P

When I try to modify fft3dfilter, MSuper, and mdegrain2, nothing seems to happen. I do want the noise reduced to the point that detail isn't lost.

[kedautinh12]

TemporalDegrain2 can do it but speed very slow

[Reel.Deel]

@ENunn
Post a small sample.

I'm starting to get a little confused. I saw you wrote the name RF earlier and now RealFinder. Are you confusing me with RF?

[kedautinh12]

Maybe he think you're a copy ver of RealFinder

[ENunn]

Originally Posted by kedautinh12

TemporalDegrain2 can do it but speed very slow

Yeah, I used it before I went with SMDegrain. Really slow.

Originally Posted by Reel.Deel

@ENunn
Post a small sample.

Got it.

Originally Posted by Reel.Deel

I'm starting to get a little confused. I saw you wrote the name RF earlier and now RealFinder. Are you confusing me with RF?

Yeah Little bit confusing on my part.

Originally Posted by kedautinh12

Maybe he think you're a copy ver of RealFinder

It's true, they're the exact same person!!!!!!! It's true!!!1one

[Reel.Deel]

RF and I are most definitely not the same person lol

I took a look at your second sample, and in terms of noise it's much better than the first sample. The luma channel has much less noise and no crosshatch patterns. But the chroma is still noisy, not the same as before but enough to use the same script as post #21. I only made one change:

pre = MergeLuma(srchbase, o, weight=.75) # was .45

This way it only does a little bit of denoising on the luma channel and helps retain details.

Here are the results:

• https://imgsli.com/MTM0NTYz
• https://imgsli.com/MTM0NTY0

Images are 2x point resampled and crop, levels, tweak, and upscaling parts of script were commented out.

By the way, I think I recognize some of those commercials. Definitely the Neutrogena one. I used to watch WGN in my younger days

[ENunn]

Sorry for the late response.

Originally Posted by Reel.Deel

I took a look at your second sample, and in terms of noise it's much better than the first sample. The luma channel has much less noise and no crosshatch patterns. But the chroma is still noisy, not the same as before but enough to use the same script as post #21. I only made one change:

pre = MergeLuma(srchbase, o, weight=.75) # was .45

This way it only does a little bit of denoising on the luma channel and helps retain details.

Here are the results:

• https://imgsli.com/MTM0NTYz
• https://imgsli.com/MTM0NTY0

Thanks for this, there's definitely a difference but I do have one issue with it. I don't know how to word it. There's a shift when there's an area of red. I think it's due to the smoothing, but I'd like to alleviate that if possible.

Original:


Script:


I posted another screenshot of the "problem", screenshot 4 on post 27.

Here's a sample.

It's probably not that big of a deal though. I just saw it and it looked a tad off to me.

[Emulgator]

If you consider the brighter, but narrow red stripes weighed into it seems resonable.