mClean spatio/temporal denoiser v3.2 (01 March 2018)

[Taurus]

@ burfadel:
1.:I really like your script!
2.:Bad news.
Your first version (1.1) of the function is working flawless on my side.
The new versions (1.2+1.3) getting stucked at line 54 (outbits).
------------------------------------
Avisynth 2.60 MT
All necessary plugins uptodate.
YV12 8bit video.
-----------------------------------
Can you give me a hint for troubleshooting?

[MysteryX]

His script is designed for AVS+ only.

If you want to use with Avisynth 2.6, you need to remove ConvertBits, and you will only process in 8-bit, which will considerably degrade the results because of applying a series of actions each having a subtle effect. The rounding errors will add up.

[lansing]

Originally Posted by MysteryX

I did some tests on a noisier source, only mClean to start with.

Source: KARA - Jumping

Here I get better results with higher thSAD such as 550. Enh=20 is unwatcheable. I again get best results with enh=12, rn=10. Because of the renoise, I think it's safer to go with higher thSAD.

Screenshots:
- Original
- RemoveGrain(21)
- KnlMeansCL(D=2, A=2, h=1.4, channels="YUV")
- mClean(550, enh=12, rn=10)
- mClean(450, enh=20, rn=12)

I'm not seeing any difference in sharpness between the two mclean screenshots, heck I don't even see any difference between them as a whole, except that tiny bit of artifact variation around the fingers on the 2nd comparison. Saying that one is unwatchable compares to the other is just absurd.

[MysteryX]

I guess I developed a sense for the details.

[MysteryX]

Burdafel, you did the bitrate conversion wrong. FF3DFilter and SMDegrain were unnecessarily processing in HBD resulting in poor performance. Here I fixed your script.

Testing back my first video, thSAD is good at 450 (difference between 450 and 550 is extremely minimal). Here I've set enh=13 and rn=10, but it will take more tests on various sources. But so far, I wouldn't go higher on low-noise source, and sources with high noise can't take any more sharpening. The only case I can see where I'd set those higher are for videos that are a little blurry.

# mClean spatio/temporal denoiser
# Version: 1.3 (27 August 2017)
# By burfadel


#  +++ Description +++
# This script is intended to remove noise whilst retaining as much detail as possible.
# Typical spatial filters work by removing large variations in the image on a small scale, reducing noise but also making the image less
# sharp or temporally stable.

# mClean works primarily in the temporal domain, although there is some spatial limiting.
# Chroma is processed via a different method to luma for optimal results.
# Input must be 

#  +++ Artifacts +++
# Spatial picture artifacts may remain as removing them is a fine balance between removing the unwanted artifact whilst not removing detail
# Additional dering/dehalo/deblock filters may be required, but should ONLY be uses if required due the detail lose/artifact removal balance
# A deblocking filter such as Deblock_QED can be applied after, as MClean currently does not deblock; this may be provided as an option later

#  +++ Sharpening +++
# Additional sharpening filters may not be required, mClean does some light detail enhancement. Any additional sharpening filters may require
# a little less strength. Alternatively use a higher 'enh' setting. Range of normal sharpening is 0-100. There are two extra options, 101 and
# 102. These are for 'overboost' sharpening, suitable only for high quality, high resolution sources. Overboost sharpening requires the modplus
# plugin, this is only required if overboost is used.

# +++ ReNoise +++
# ReNoise adds back some of the removed luma noise. Re-adding original noise would be counterproductive, therefore ReNoise modifies this noise
# both spatially and temporally. The result of this modification is the noise becomes much nicer and it's impact on compressibility is greatly
# reduced. It is not applied on areas where the sharpening occurs as that would be counterproductive. Settings range from 1 to 20, default
# value is 12. The strength of renoise is affected by the the amount of original noise removed and how this noise varies between frames.

# +++ Outbits +++
# Specifies the bits per component (bpc) for the output for processing by additional filters. It will also be the bpc that mClean will process.
# By default, mClean processes as 12 bits if the input is 8 bit, and converts back to 8 bit. If the input is 10 bits or higher no conversion is
# done unless outbits is specified and is different to the input bpc. If you output at a higher bpc keep in mind that there may be limitations
# to what subsequent filters and the encoder may support.

#  +++ Required plugins +++
# Latest RGTools, MVTools2, Masktools2, FFT3DFilter
# Latest Modplus, only required if using sharpening overboost (enh settings 101 and 102)
# Requires latest fftw.dll to be installed as instructed on the website - http://www.fftw.org


function mClean(clip c, int "thSAD", int "blksize", int "blksizeV", int "overlap", int "overlapV", int "enh", int "rn", int "outbits", int "cpu")
{
    defH        = Max (C.Height, C.Width/4*3)     # Resolution calculation for auto blksize settings
    thSAD       = Default (thSAD, 450)     # Denoising threshold
    blksize     = Default (blksize, defH<360 ? 8 : defH<750 ? 12 : defH<1200 ? 16 : defH<1600 ? 24 : 32)     # Horizontal block size for MDegrain2
    blksizeV    = Default (blksizeV, blksize)     # Vertical block size for MDegrain2, default same as horizontal
    overlap     = Default (overlap, blksize>4?(blksize/4+1)/2*2:0)     # Horizontal block overlap
    overlapV    = Default (overlapV, blksize>4?(blksizeV/4+1)/2*2:0)     # Vertical block overlap
    enh         = Default (enh, 13)     # Detail enhancement (detail orientated sharpen) strength
    rn          = Default (rn, 10)     # ReNoise strength from 0 (disabled) to 20
    outbits     = Default (outbits, c.BitsPerComponent)     # Output bits, default input depth
    calcbits    = c.BitsPerComponent == 8 ? 12 : c.BitsPerComponent
    calcbits    = outbits > calcbits ? calcbits : outbits
    cpu         = Default (cpu, 4)     # Threads for fft3dfilter


    Assert(isYUV(c)==true, """mClean: Supports only YUV formats (YV12, YV16, YV24)""")
    Assert(isYUY2(c)==false, """mClean: Supports only YUV formats (YV12, YV16, YV24)""")
    Assert(isYV411(c)==false, """mClean: Supports only YUV formats (YV12, YV16, YV24)""")
    Assert(enh>=0 && enh<=102, """mClean: "enh" ranges from 0 to 102""")
    Assert(rn>=0 && rn<=20, """mClean: "rn" ranges from 0 to 20""")
    Assert(outbits>=8 && outbits<=16, """mClean: "outbits" ranges from 8 to 16""")

# Spatio/temporal chroma noise filter
filt_chroma  =  fft3dfilter (c, bw=blksize*2, bh=blksizeV*2, ow=overlap*2, oh=overlapV*2, sharpen=0.12, bt=3, ncpu=cpu, dehalo=0.3, sigma=2.35, plane=3)

# Temporal luma noise filter
super      =  c.MSuper (chroma=false,hpad=16, vpad=16)
bvec2      =  MAnalyse (super, chroma=false, isb = true, delta = 2, blksize=blksize, blksizeV=blksizeV, overlap=overlap, overlapV=overlapV, search=5, searchparam=5)
bvec1      =  MAnalyse (super, chroma=false, isb = true, delta = 1, blksize=blksize, blksizeV=blksizeV, overlap=overlap, overlapV=overlapV, search=5, searchparam=3)
fvec1      =  MAnalyse (super, chroma=false, isb = false, delta = 1, blksize=blksize, blksizeV=blksizeV, overlap=overlap, overlapV=overlapV, search=5, searchparam=3)
fvec2      =  MAnalyse (super, chroma=false, isb = false, delta = 2, blksize=blksize, blksizeV=blksizeV, overlap=overlap, overlapV=overlapV, search=5, searchparam=5)
clean      =  c.MDegrain2 (super, bvec1, fvec1, bvec2, fvec2, thSAD=thSAD, plane = 0)
clean      =  calcbits != clean.BitsPerComponent ? clean.ConvertBits(calcbits) : clean
c          =  calcbits != c.BitsPerComponent ? c.ConvertBits(calcbits) : c

# Masks for spatial noise reduction and noise independent detail enhancement
noised     =  mt_makediff (clean, c, u=1, v=1)
noise      =  mt_binarize (clense(mt_makediff(mt_binarize(noised, u=1, v=1), mt_edge(sharpen(clean, 0.85), "prewitt", u=1, v=1), u=1, v=1), grey=true), u=1, v=1)

# Spatial luma denoising
clean2     =  mt_merge (clean, removegrain(clean, 18, modeU=-1, modeV=-1), noise, u=1, v=1)

# Unsharp filter for spatial detail enhancement
clsharp    =  (enh>0<=100) ? mt_adddiff (mt_makediff(clean, blur(clean, 0.80*(enh/100), 0.50*(enh/100)), u=1, v=1), clean2, u=1, v=1) : clean
clsharp    =  (enh>=101<=102) ? mt_adddiff (mt_makediff(clean, gblur(clean, enh-100), u=1, v=1), clean2, u=1, v=1) : clsharp

# If selected, combining ReNoise
renoise    =  (rn==0)    ? nop : tweak(temporalsoften (noised, 4, 128, 0, scenechange=0, mode=2), cont=1.010+(0.020*(rn/20)), bright=1.01+(0.04*(rn/20)))
clean2     =  (rn>0<=20) ? mergeluma (clean2, mt_adddiff(clean2, renoise, u=1, v=1), 0.3+(rn*0.035)) : clean2

# Combining spatial detail enhancement with spatial noise reduction using prepared mask
filt_luma  =  mt_merge (clean2, clsharp, mt_invert(mt_convolution(noise, u=1, v=1), u=1, v=1), u=1, v=1)

# Converting bits per channel
filt_luma  =  outbits < filt_luma.BitsPerComponent ? ConvertBits(filt_luma, outbits, 1) : filt_luma

# Combining result of luma and chroma cleaning
return mergechroma (filt_luma, filt_chroma)
}

[lansing]

Originally Posted by MysteryX

I guess I developed a sense for the details.

I literally grab the two mclean screenshots from the first comparison, stacked them into photoshop, set layer overlay to "difference", and I got a complete black image. You need an eye checkup.

[MysteryX]

Originally Posted by manolito

The "super" param does not have any prefiltered content in it. Only "bak" and "fwd" are based on the prefiltered clip, and these two only have motion vector data in them.

You have a good point. It produces more clearly-defined motion vectors ... but how does it result in clearer hair textures?

The added grain won't make it to the output to benefit the encoder. This explains why I saw improvement in quality but no difference in file size.

[MysteryX]

Originally Posted by lansing

I literally grab the two mclean screenshots from the first comparison, stacked them into photoshop, set layer overlay to "difference", and I got a complete black image. You need an eye checkup.

I checked back with the script to make sure I have the right images.

Oddly enough, first screenshot has 0 difference between the 2 settings.

2nd screenshot however has considerable difference and doesn't look good with enh=20

I was about to upload 3 screenshots but the 3rd has files missing after upload, perhaps because postimage detected some images as being exactly the same.

[MysteryX]

Here are some tests on another source that definitely has more noise

KARA - Dazzling Red

- Original
- RemoveGrain(21)
- KnlMeansCL(D=2, A=2, h=1.8, channels="YUV") # raised from 1.4 to 1.8 here
- mClean(450, enh=12, rn=10)
- mClean(450, enh=20, rn=12)





Here we see more clearly the characteristics of each method. RemoveGrain does little, KNLMeans blurs out details, and mClean does a pretty good job in both cases.

[manolito]

Originally Posted by MysteryX

You have a good point. It produces more clearly-defined motion vectors ... but how does it result in clearer hair textures?

My next question would be:
If you do get better results with mClean instead of RemoveGrain, then what is the major difference between the two?

mClean works primarily in the temporal domain, although there is some spatial limiting.

RemoveGrain is purely spatial, so the next test I would do is replacing RemoveGrain with a (fast) purely temporal denoiser like TTempSmooth (or the even faster TTempSmoothF) as the prefilter in FRC. This should be way faster than using mClean, and maybe you will get more precise motion vectors compared to RemoveGrain, too.


Cheers
manolito

[lansing]

Originally Posted by MysteryX

Here are some tests on another source that definitely has more noise

KARA - Dazzling Red

- Original
- RemoveGrain(21)
- KnlMeansCL(D=2, A=2, h=1.8, channels="YUV") # raised from 1.4 to 1.8 here
- mClean(450, enh=12, rn=10)
- mClean(450, enh=20, rn=12)


Here we see more clearly the characteristics of each method. RemoveGrain does little, KNLMeans blurs out details, and mClean does a pretty good job in both cases.

What? How on earth did you come up with that conclusion?

First of all, your video has no grain, so the smdegrain inside mclean is basically doing nothing except creating artifacts on motion objects by its nature, you can see the artifacts in the fog which doesn't exist in the original. Then what you have left in the script is a simple removegrain(18). So right now you're basically comparing a removegrain(21) against KNLMeansCL and a removegrain(18), and you're saying removegrain(18) wins, I am speechless.

Oddly enough, first screenshot has 0 difference between the 2 settings.

There's nothing odd about it because we can visually SEE it ourselves that there're no difference. The only one that insisted that one is better than the other is you.

I have a feeling that you're just trolling the OP with all these nonsense.

[burfadel]

Originally Posted by MysteryX

Burdafel, you did the bitrate conversion wrong. FF3DFilter and SMDegrain were unnecessarily processing in HBD resulting in poor performance. Here I fixed your script.

Testing back my first video, thSAD is good at 450 (difference between 450 and 550 is extremely minimal). Here I've set enh=13 and rn=10, but it will take more tests on various sources. But so far, I wouldn't go higher on low-noise source, and sources with high noise can't take any more sharpening. The only case I can see where I'd set those higher are for videos that are a little blurry.

Thanks, I've updated the first post . I also added in a line for filt_chroma, seeing as it needs to be the same bit depth for merging. There will be adjustments with that later as I update the chroma filtering.

Do you think there should be an auto adjustment for default enh and rn values based on resolution like there is for blocksize? Hopefully people won't have too many low quality 720P and 1080P videos.

[MysteryX]

Originally Posted by lansing

First of all, your video has no grain

There's a big difference between a denoiser and a degrainer. I'm testing this as a denoiser. Noise is generally encoding artifacts. Grain is something else.

Originally Posted by lansing

Do you think there should be an auto adjustment for default enh and rn values based on resolution like there is for blocksize? Hopefully people won't have too many low quality 720P and 1080P videos.

What kind of auto-adjustment are you thinking about? So far I tested enh=13, rn=0 for 288p content and enh=13, rn=10 for 1080p content.

Here's my feedback/observations so far

1. I like the results of this script

2. When it comes to using it as a prefilter for FRC, it does bring some benefits: better motion vectors

3. The re-added grain to help the encoder isn't one of those benefits

When it comes to FRC with preset="slower", the performance of mClean isn't really an issue, but perhaps not all parts of it are necessary in that case, it has to be tested. Does renoise contribute to better motion vectors at all? My suspicion is that motion vectors are better defined for having more clearly-defined edges.

The main issue is that it's hard to do each test because it can't yet be run in the same script.

[burfadel]

Yes, renoise probably would affect further analysis. Ideally in that case you would interpolate the renoise and add it back after framerateconverter.

[MysteryX]

Does it really make a difference on textures or are my eyes bad? Had to test for sure. I took an image of a full head and compared RemoveGrain(21) with mClean as a FRC prefilter. Then I put it in Photoshop, do a diff, use magic wand to select all with tolerence set to 0, 1 and 2.

There *IS* a difference. Not sure I can explain it though. This is a mostly static scene, differences are larger on high-animation scenes.

[MysteryX]

This test shows clear differences

- No prefilter
- RemoveGrain(18)
- RemoveGrain(21)
- TTempSmooth
- mClean(rn=10)
- mClean(rn=0)



TempSmooth gives no gain and in fact degrades picture. mClean wins. RemoveGrain on its own... 21 is more balance than 18, but within mClean 18 works.

Yes, renoise helps.

You can't "add things back" after interpolation because source and destination frames count don't match

Originally Posted by lansing

I have a feeling that you're just trolling the OP with all these nonsense.

I think I've settled the case.

[burfadel]

Originally Posted by MysteryX

You can't "add things back" after interpolation because source and destination frames count don't match.

That's true, I didn't explain what I meant very well, but thinking of what I wrote it woudn't make sense anyway .

[lansing]

Originally Posted by MysteryX

Does it really make a difference on textures or are my eyes bad? Had to test for sure. I took an image of a full head and compared RemoveGrain(21) with mClean as a FRC prefilter. Then I put it in Photoshop, do a diff, use magic wand to select all with tolerence set to 0, 1 and 2.

There *IS* a difference. Not sure I can explain it though. This is a mostly static scene, differences are larger on high-animation scenes.

Dude, you'll have to pull that FRC thing of yours out of your arse, because you're confusing the f out of everybody.

with mClean as a FRC prefilter

What does this even mean?

This test shows clear differences

- Original
- RemoveGrain(18)
- RemoveGrain(21)
- TTempSmooth
- mClean(rn=10)
- mClean(rn=0)

...

I think I've settled the case.

Your test is a fake, the first screenshot is not the original, that's the result of you putting the script inside your FRC thing, which does NOTHING on denoising. You faked the results.

You're not helping the OP in any way, such that you tested his denoising script on videos that have no noise at all, you gave him false feedback on setting comparison, you exaggerated invisible differences that photoshop can't even see, and you hijacked his thread and forced his script onto your own little filter, which in itself is a big confusing mess that doesn't make sense.

[MysteryX]

Watch your language. If you don't like what we're doing, just get the heck out.

Originally Posted by lansing

that's the result of you putting the script inside your FRC thing

Yes that's what I meant by original, perhaps I should have called it "no prefilter", but I think anyone with a brain understood that. I can't compare with original because generated frames don't match source frames.

Please, if you got nothing constructive to say, shut up.

Edit: I renamed it to "no prefilter" to avoid confusing anyone else

If someone else views any screenshot other than mClean(rn=10) to give better results, I'm listening, but please have something to back up what you say. This test is pretty simple: comparing 6 versions of a script to see which gives better results. I don't care about how "things should be done", I only care about concrete results.

[StainlessS]

Guys, please be nice.
Both of you are valued members of the community and both totally respected, if something aint working quite right, then
maybe in time it will, it dont need no negs.
Consider this a good slap on the arse for both

(Soon be Christmas )