Strong VHS cleaning

[AVIL]

Hi all:

I work mainly with VHS footage.,frecuently very noisy (jitter, swimmering, rainbows). Ordinary strong cleaners do wonderful with noise but, alas, destroy also detail. I've combined two denoisers. The frist do a soft denoinsing, suited to clean less and destroy more detail. I've found that feed mvdegrain with vectors obtained with dct=1 leads to the type of denoise I need. After that, all the noise and detail stealed by the first denoiser is added to the original file, enhancing detail an worsing noise. Then the last step. A strong denoiser than clean well and respect detail. This is the script:
#********************************************************************************************
function VHSClean(clip x ,int "ths" , bool "blur_sharp"){
#********************************************************************************************
ths = default( ths , 100 ) #compromise between strong denoising and detail lost
blur_sharp = default( blur_sharp , true )

lambda = 40000
pel = 2
tm = false
srhp = 2
srch = 4
badsad = 2000
chroma = true
thsc = ths * 2
ths1= ths * 4
thsc1= ths1* 2

bs = 8
bblur = 0.6
csharp = 0.6

# Required planar video (tested on YV16)
# Rrequired masktools2, mvtools2. Modern versions
# Only tested on VHS footage
# Video must be progressive, preferiblely using a good bob deinterlacer

sx=x.mSuper(pel=pel, sharp=1)

#phase 1. Soft denoising

f1x=sx.MAnalyse(delta=1,isb=false,truemotion=tm,blksize=16,blksizev=8,overlap=8,overlapv=4,search=srch,searchparam=srhp,badsad=badsad,dct=1,chroma=chroma,lambda=lambda)
b1x=sx.MAnalyse(delta=1,isb= true,truemotion=tm,blksize=16,blksizev=8,overlap=8,overlapv=4,search=srch,searchparam=srhp,badsad=badsad,dct=1,chroma=chroma,lambda=lambda)
f2x=sx.MAnalyse(delta=2,isb=false,truemotion=tm,blksize=16,blksizev=8,overlap=8,overlapv=4,search=srch,searchparam=srhp,badsad=badsad,dct=1,chroma=chroma,lambda=lambda)
b2x=sx.MAnalyse(delta=2,isb= true,truemotion=tm,blksize=16,blksizev=8,overlap=8,overlapv=4,search=srch,searchparam=srhp,badsad=badsad,dct=1,chroma=chroma,lambda=lambda)

x2=x.MDeGrain2(sx, b1x,f1x,b2x,f2x,thSAD=ths,thSADC=thsc)

#phase 2. Reinject denoised over original (like a sharpening using blurred version)

x3=x.mt_lutxy(x2,expr="x 2 * y -",chroma="process")

#phase 3. Strong denoising. Same style as MCDegrainSharp (By Didée and Stainless)


x0 = (blur_sharp) ? x3.sharpen(csharp): x3
x1 = (blur_sharp) ? x3.blur(bblur) : x3

sx0 = x0.MSuper(pel=pel, sharp=1,levels=1) # Only 1 Level required for sharpened Super (not MAnalyse-ing)
sx1 = x1.MSuper(pel=pel, sharp=1)

f1x1=sx1.MAnalyse(delta=1,isb=false,truemotion=tm,blksize=64,blksizev=64,overlap=32,overlapv=32,search=srch,searchparam=srhp,badsad=badsad,dct=0,chroma=chroma,lambda=lambda)
b1x1=sx1.MAnalyse(delta=1,isb= true,truemotion=tm,blksize=64,blksizev=64,overlap=32,overlapv=32,search=srch,searchparam=srhp,badsad=badsad,dct=0,chroma=chroma,lambda=lambda)
f2x1=sx1.MAnalyse(delta=2,isb=false,truemotion=tm,blksize=64,blksizev=64,overlap=32,overlapv=32,search=srch,searchparam=srhp,badsad=badsad,dct=0,chroma=chroma,lambda=lambda)
b2x1=sx1.MAnalyse(delta=2,isb= true,truemotion=tm,blksize=64,blksizev=64,overlap=32,overlapv=32,search=srch,searchparam=srhp,badsad=badsad,dct=0,chroma=chroma,lambda=lambda)
f3x1=sx1.MAnalyse(delta=3,isb=false,truemotion=tm,blksize=64,blksizev=64,overlap=32,overlapv=32,search=srch,searchparam=srhp,badsad=badsad,dct=0,chroma=chroma,lambda=lambda)
b3x1=sx1.MAnalyse(delta=3,isb= true,truemotion=tm,blksize=64,blksizev=64,overlap=32,overlapv=32,search=srch,searchparam=srhp,badsad=badsad,dct=0,chroma=chroma,lambda=lambda)
f4x1=sx1.MAnalyse(delta=4,isb=false,truemotion=tm,blksize=64,blksizev=64,overlap=32,overlapv=32,search=srch,searchparam=srhp,badsad=badsad,dct=0,chroma=chroma,lambda=lambda)
b4x1=sx1.MAnalyse(delta=4,isb= true,truemotion=tm,blksize=64,blksizev=64,overlap=32,overlapv=32,search=srch,searchparam=srhp,badsad=badsad,dct=0,chroma=chroma,lambda=lambda)
f5x1=sx1.MAnalyse(delta=5,isb=false,truemotion=tm,blksize=64,blksizev=64,overlap=32,overlapv=32,search=srch,searchparam=srhp,badsad=badsad,dct=0,chroma=chroma,lambda=lambda)
b5x1=sx1.MAnalyse(delta=5,isb= true,truemotion=tm,blksize=64,blksizev=64,overlap=32,overlapv=32,search=srch,searchparam=srhp,badsad=badsad,dct=0,chroma=chroma,lambda=lambda)
f6x1=sx1.MAnalyse(delta=6,isb=false,truemotion=tm,blksize=64,blksizev=64,overlap=32,overlapv=32,search=srch,searchparam=srhp,badsad=badsad,dct=0,chroma=chroma,lambda=lambda)
b6x1=sx1.MAnalyse(delta=6,isb= true,truemotion=tm,blksize=64,blksizev=64,overlap=32,overlapv=32,search=srch,searchparam=srhp,badsad=badsad,dct=0,chroma=chroma,lambda=lambda)

x4 = x1.MDeGrain6(sx0, b1x1,f1x1,b2x1,f2x1,b3x1,f3x1,b4x1,f4x1,b5x1,f5x1,b6x1,f6x1,thSAD=ths1,thSADC=thsc1)

#phase 4. Recover quick flying objects and water drops

mx=x4.blur(1.5).mt_lutxy(x3.blur(1.5),expr="y x - abs 12 > 255 0 ?",chroma="process")
x4.mt_merge(x3,mx.blur(1.5),luma=true)


}



Very interested in your feedback and ideas to enhance

[anton_foy]

Cool script I like the idea alot! Although when trying this out on a bit noisy UHD footage it will show artifacts where motion is. Kinda like "holes" inthe motion blur or where the motion blur should have been if that makes sense.

[Mounir]

Is it me or your script imply the source to be progressive ? Kinda strange for vhs

edit:
You might want to have a look of my FixRipsP2 mod HERE (https://www.digitalfaq.com/forum/video-restore/12806-fixrips-grain-noise.html)

[AVIL]

@anton_foy
First, thanks for your feedback. I've not proccessrd UHD video. I suppouse that te compromisoe between the. Soft cleaner and te strong one could be different. For videos of this size, the methode blur-sharp calculates a value of 0.75 for parameter bblur instead the actual valué of 0.6. but i doubt this can resolver the problema. I will try on a UHD video to tune other values, like the blksize at obtainig motion vectors.

@Mounir.

Many thanks for your observation. Video must be progressive. Its my bad not to mention It in muy script. I do a good bob deinterlacing bafore start denoising the video. I will edit the scriot and also give your script a try

[AVIL]

@anton_foy
Thinking about UHD video, as distances in pixeles can be four times greater than in VHS could be resasonable also raise the value of searchparam. Change, please the value of parameter srhp from 2 to 8 or greater

[StainlessS]

Maybe try convert to 16 bit before MvTools stuff, and back to 8 bit at end of script.
For 8 bit, vectors in vector clip are limited to signed 8 bit, ie -128 -> +127 for PEL distance (about 64 pixels @ pel=2, 32 for pel=4),
16 bit allows for greater vector length (pixel distance) in vector clips.
Maybe makes a difference, maybe not. [Good for that "Placebo feel good factor"]

[Selur]

Side note: did try a 'quick' port to Vapoursynth (https://github.com/Selur/VapoursynthScriptsInHybrid/blob/master/fromDoom9.py#L233) (not totally sure everything is correct)