Best quality and speed video denoisers 2011

[Zabar12]

[QUOTE=mammo1789;2127815]

mt_masktools-25.dll (12/31/2010)
Another thing i noticed (which I sad before that with each pass there is a tape degradation and someone here sad not ) evidence by 4 and 5 pass bytes increase almost 2 mb from the original, for comparison it continue to grow and pass 9 =369.437.191 ( that's almost 10 mb difference) maybe the passes add dirt to the head each time but probably additional noise due to degradation is put inside the material also . Am I right?

I measured degradation

[Gavino]

Thanks G! I see you are avoiding the chroma resizer, I guess the reason is the implicit chroma offset. What is the assumed chroma placement and what style (mpg2, dv etc.) does it correspond to?
I still don't find this obvious to find in the documentation.

http://avisynth.org/mediawiki/Sampling

Avisynth uses MPEG2 chroma placement for YV12 (see sect 2.5 of linked wiki page), with chroma samples vertically situated halfway between luma samples.
Consequently, when converting YV12 to YUY2, all chroma pixels are interpolated and original chroma is lost (see sect 2.3).

The same applies when going from YUY2 to YV12 of course (sect 2.2), but because of your PointResize before the conversion, the interpolation is done between two equal pixels, preserving the original values, so that bit works as desired.

[jagabo]

I measured degradation

Interesting test but I bet that's mostly just temperature related noise. Otherwise, why would the tape get better after sitting for a day? And measuring noise by testing the compressibility is probably not a reliable measure of tape signal condition.

[Zabar12]

Can anyone explain how Median handles sound?

[Gavino]

The median script functions posted previously don't touch the sound - they just return the audio from the first source and ignore the others.

I suppose in principle something similar could be done with the audio - a simple mean (average) is probably sufficient, and could be constructed using MixAudio().

[Zabar12]

Gavino
Thank you for clarification. Median is wonderful tool!

[Gavino]

I suppose in principle something similar could be done with the audio - a simple mean (average) is probably sufficient, and could be constructed using MixAudio().

In practice, this might not be so easy to do since it would require the audio in all the sources to be exactly in sync at the sample level.

[jmac698]

oh it's not the same, the way we perceive audio is not the same as how we perceive video. If you made equivalent signals, our sensing of them would be different - I mean sharpness in a picture would be like treble in a speaker. But I would say the ear is most sensitive to frequency effects than our eyes. So combining two sounds would cause a phasing effect that's quite audible. I think it would also get noisy or distorted, cause you would be making jumps in the waveform. I guess trading noise for distortion. But you'd have to try it.
Actually I could try that easily just be listeing to a picture as sound, even though it's not music you can still hear what the relative difference would be.

Anyhow, the linear audio track wouldn't have the dropouts effect in the same way, but hi fi audio could have the same "comets" effect. Hi fi audio is recorded as FM the same way as video, and my theory for the dropouts is that a low carrier signal causes the FM circuit to drift high which causes the white, and the circuit returns to normal as it locks onto the carrier again. This would make a clicking noise in hi fi audio, and in fact I do remember hearing a clicking noise before on VHS. I think it would work if you lined up the audio and did a careful splice. In audio you do a mxing overlap between two ends of an inserted sound, otherwise you hear a click. I think a direct median could help but there's probably a more effective way to repair the audio. The difference here is that audio wants a smooth waveform and DC doesn't matter, in video DC (brightness) would matter.
It's hard to explain, sorry
I'm sure Jagabo can follow
Also the 'gamma' of audio is different, not 2.2 like video (our sense of weight etc. is different too).

[Zabar12]

Can using a mode deliver a better result than a median?

[jagabo]

Can using a mode deliver a better result than a median?

Maybe. But you will need a lot of caps to get any duplicate values.

[jagabo]

But I would say the ear is most sensitive to frequency effects than our eyes. So combining two sounds would cause a phasing effect that's quite audible... I'm sure Jagabo can follow

I know what you mean. And, yes our ears are very sensitive to phase issues, it's how we determine where a sound is coming from. But given high speed and relatively high precision of the video sync signals (we're talking KHz for the audio vs MHz for the video) I suspect successive audio samplings will be fairly well phase aligned -- at least in the short term (seconds? minutes?) with clean recordings. Over the long term (minutes? hours?) I suspect the audio caps would get far enough apart that you would get noticeable phase differences.

[jmac698]

That's a good point, there's only 12 samples at the highest sampling frequency, per line of video. So the video spikes are practically inaudible anyhow. So what is causing the clicks?
I've recorded video with an audio sampler, just a tad blurry though

[Zabar12]

Guys, please post a Median script for 9 files.

[jmac698]

temporal medianblur:


medianblurt(clip,int radiusy,int radiusu,int radiusv,int temporalradius, bool mc, bool calcborder,bool markscenechange,int blksize,int pel,int lambda, int thSCD1, int thSCD2,bool spfull,bool bsfull)
http://avisynth.org/tsp/

And

interleave(clip1,clip2,clip3,clip4,clip5,clip6,clip7,clip8,clip9)
medianblurt(0,0,0,4,false,false)
selectevery(4,9)

Should do it.

[Zabar12]

Thanks. Is there a faster Median (9) based on technology developed by G-force?

[Gavino]

interleave(clip1,clip2,clip3,clip4,clip5,clip6,clip7,clip8,clip9)
medianblurt(0,0,0,4,false,false)
selectevery(4,9)

Should be selectevery(9,4)

[jmac698]

The g-force script relies on sorting through the use of min, max. This is an extremely complicated way to sort, and I wouldn't want to work it out by hand. You can call a median of 3 a total of four times;

v1=median2(clip1,clip2,clip3)
v2=median2(clip4,clip5,clip6)
v2=median2(clip7,clip8,clip9)
median2(v1,v2,v3)

[jagabo]

The g-force script relies on sorting through the use of min, max. This is an extremely complicated way to sort, and I wouldn't want to work it out by hand. You can call a median of 3 a total of four times;

v1=median2(clip1,clip2,clip3)
v2=median2(clip4,clip5,clip6)
v2=median2(clip7,clip8,clip9)
median2(v1,v2,v3)

That was my first thought too. And it may look ok with some, or even most, material. But it doesn't return a true median. Consider the following list:

1 1 2 1 5 5 1 9 9

reordered:

1 1 1 1 2 5 5 9 9
        2

using your algorithm:

(1 1 2) (1 5 5) (1 9 9)
   1       5       9
           5

[Gavino]

You can call a median of 3 a total of four times;

v1=median2(clip1,clip2,clip3)
v2=median2(clip4,clip5,clip6)
v2=median2(clip7,clip8,clip9)
median2(v1,v2,v3)

That won't work, as a simple counterexample shows.
Suppose (for a given pixel) the clips are ordered 1 2 4 5 3 6 7 8 9 (so the median is 3).
v1 returns clip2, so 3 is eliminated.

I don't think you can get away without doing a full sort (or the equivalent of it).

Ha! jagabo beat me to it!

[jmac698]

What about
med(v1,v2,v3)
med(last,v4,v5)
med(last,v6,v7)
med(last,v8,v9)
This won't work either, as the order of operations could matter.
Is it even possible to use median of medians?

[Zabar12]

Can anyone ask G-force on this matter? Back in 2008 he was working on Median (7)
http://forum.doom9.org/showthread.php?t=140282

[poisondeathray]

Can anyone ask G-force on this matter? Back in 2008 he was working on Median (7)
http://forum.doom9.org/showthread.php?t=140282

why don't you pm him ?

[jmac698]

What's wrong with median blur? And it could be done, but I don't want to work out the sequence of min/max to do sorting.

[Zabar12]

nothing

[Zabar12]

why don't you pm him ?

I'm a new kid on the block, while some participants of this thread have already had discussions with G-force on median

What's wrong with median blur?

Of course, nothing is wrong. The only concern is it's speed. I haven't tried it yet, though

[Zabar12]

Unfortunately, Avisynth reported that Medianblur is incompatible with YUY2 and can process only yv12

[jmac698]

just resize it beforehand
double height

[Zabar12]

just resize it beforehand
double height

I'm sorry, but can you elaborate on this a bit?

[jagabo]

YV12 has half the vertical chroma resolution of YUY2. You can convert YUY2 to YV12 and back to YUY2 losslessly* by doubling the frame height with a PointResize(), converting to YV12, Converting back to YUY2 with a PointResize() again, and halving the frame height:

PointResize(width, height*2)
ConvertToYV12()
...other filters here...
ConvertToYUY2()
PointResize(width, height/2)

So you would do the upscale/ConvertToYV12() to each of your source, run medianBlur(), then downscale/ConvertToYUY2() the result.

* In theory. In practice I still see some changes in the chroma channels.

[Zabar12]

Please help me to merge Resize into Median (9) script:

input_1=AVISource("1.avi")
input_2=AVISource("2.avi")
input_3=AVISource("3.avi")
input_4=AVISource("4.avi")
input_5=AVISource("5.avi")
input_6=AVISource("6.avi")
input_7=AVISource("7.avi")
input_8=AVISource("8.avi")
input_9=AVISource("9.avi")
interleave(input_1,input_2,input_3,input_4,input_5,input_6,input_7,input_8,input_9)
medianblurt(0,0,0,4,false,false)
selectevery(9,4)