Best quality and speed video denoisers 2011

[2Bdecided]

I've always found this multi-capture thing doubtful. If I still-pause the tape on a decent VCR (with NO digital frame-store), the noise is almost all fixed. This, to me, proves that it's all on the tape, not in the playback.

Still, if people can experience something different, that's probably useful for them, but I can't quite believe there would be a simpler solution (e.g. VCR and capture chain with less noise to start with). I suspect a lot of it is the slight (sub-pixel) timing mis-match causing a slight horizontal blur.

Where it does work is for comet trails. They certainly change from one play to another, so do benefit from multi-capture processing.

Cheers,
David.

[pandy]

I've always found this multi-capture thing doubtful. If I still-pause the tape on a decent VCR (with NO digital frame-store), the noise is almost all fixed. This, to me, proves that it's all on the tape, not in the playback.

Uncorrelated noise is reduced approx 3dB for each doubling of passes and this work quite well and it is practically used for example in low noise electronics.

[Gavino]

The theory is not disputed, but I think what 2Bdecided is saying is that in practice there is no (or very little) uncorrelated noise.

[pandy]

So his assumption is perfect head to tape matching, no noise in amplifiers, no noise in ADC etc - even capturing mono recording with 2 channels (stereo) then adding two channels in digital domain will improve SNR.
There is lot of uncorrelated noise... it doesn't work only for digital world but for analog it works quite well and we talk about converting analog to digital.

[jagabo]

OK, here's some empirical evidence. I dug an an old VHS deck out of the closet and captured a section of a tape (a commercial release, from the bargain bin many years ago) four times. I used a Panasonic ES15 in passthrough mode as a TBC. I used the following script:

v1=AviSource("cap2.avi").Trim(16,0)
v2=AviSource("cap3.avi").Trim(20,0)
v3=AviSource("cap4.avi").Trim(24,0)
v4=AviSource("cap5.avi").Trim(20,0)

v12=Overlay(v1,v2,opacity=0.5)
v34=Overlay(v3,v4,opacity=0.5)
Overlay(v12,v34,opacity=0.5)

Here's a frame from cap2.avi:


Here's the same frame averaged over the four caps (the output of the above script):


I picked this particular frame because it's easily identified by the white spot. Otherwise it's representative of the rest of the video. There is definitely less noise after averaging the four videos together. As expected, the white spot, being "on the tape" (and probably on the film it was made from), did not disappear.

[pandy]

uncorrelated vs correlated - drop is a signal correlated - however - from my perspective averaged seems to have less chroma cross talk on white circle, also seems to be less noisy - it is just pure average or additional processing (median filtering?)

recorded by VCR signal which already is SNR limited and this is real limit that can be achieved by multiple capture (assuming good quality of source maybe up to 46 - 48dB)

[jmac698]

@2bedecided
Both views are correct, it's just that in your experience, the noise originally on the tape was more visible than the additional noise added by the capture chain. However, as in the example given by jagabo, if some part of the original at least is clean, you can clearly see the difference in noise reduction.
Also it's a great observation that there's some horizontal blurring from horizontal itter; this is essentially a spatial denoise as well. That's why I recommend using a TBC in the process.

I agree with Pandy's observations, in fact earlier in the thread I listed types of electronic noise like thermal, magnetic domain alignment (tape hiss) etc. I've read specs on VCR's of >45dB video SNR.

Further if you do your multiple passes with multiple copies (like buy 3 copies of the same tape), you eliminate the *mastering* noise as well.

I've done the same with analog TV broadcasts recorded with repeats; I've even done something equivalent with digital broadcasts by combinng the best frames of each copy (say, the B follow an I frame).

I have a technique for doing these experiments perfectly, I made a test signal which lets me create a perfect TBC. Just some lines as bookends to the picture that I can align in software to subpixel precision. I can test multigenerational tapes and measure the exact noise per generation or other artefacts I see. I think I could come up with a perfect Neat Video noise profile for example that could work for any tape (on my tested machine, at least), so I don't have to rely on finding a smooth area in the video itself (normal profile procedure).

[mammo1789]

Hi to all i didn't had time past 2 days to get involved in the tread.
I see clearly that in jagabo sample there is some evidence of noise reduction, but even if you capture one copy of the material and run it trough noise reduction filter ( avysinth or vdub including neat video) you will cancel(reduce) that noise ( from the equipment and ac/dc and so on ) and much faster and easy than capturing 4 different copies. Another problem in this approach I see ( i didn't tried yet my self) is the perfect timing of 4 copies or align of frames in time ( for example on first copy you press "rec" in time, and on second copy you press "rec" 1/2 second later or maybe 1 second later and so on and so on, and you get 4 differently in time copies meaning on time x ( lets say frame number 1000th you have head on frame 1001 you have mouth opened this is because of interlacing frame by frame so 1000th frame on copy 2 will be mouth and 1000th frame on copy 3 maybe something else which makes the task even harder to align only by eye?)

One more problem is that if understand it correctly fundamentally every time you play copy of vhs tape and that tape is passing trough head and vcr path it losses quality permanently ( quality of recorded material of tape not the quality of vcr).

Maybe this approach is better suited for signal drops ( that occur each time on different place ) or comets as someone mentioned but for type of noise that jagabo canceled in the example I think is not worth it.

Some of the tapes that I have problem with are these kind of nature I will try these weekend to solve ( if i can ) where i had problem with https://forum.videohelp.com/threads/340068-sort-of-black-comets-trails-in-hi8-video ( what do you think jmac possible ?) and this https://forum.videohelp.com/threads/339642-Tbc-makes-bad-picture-worse

If not i sad it its not worth the time and effort for only that type of noise reduction seen on picture

[jagabo]

I aligned my 4 caps with Trim() -- adjusting by eye.

Quality loss from a few plays of a VHS tape are negligible.

I don't know what the cause of those defects in your examples were. But if they're different each time you play the tape a median of 3 filter will get rid of many of them.

[jmac698]

I know, some people say this technique is not worth it. It's up to you. I also said you can try Neat Video to TNLMeans, a few others are listed in a similiar thread, just look below where it says "Similiar Threads" and click on "denoisers comparisons".
Yes it's best use is on comets, but for me also I found a very noisy tape where it worked miracles. Before this technqiue I had used removedirt as the most powerful technique but it ruined the video. This was extremely bad video.

Yes I believe the tape loses a bit of signal every time it's played, but it doesn't matter as the first copy will still have the full signal, but also it doesn't wear that quickly - think of kids tapes that are played for years. I have one and it still looks normal. Anyhow have experience with this? I've bought old rental tapes that are still watchable.

[pandy]

If not i sad it its not worth the time and effort for only that type of noise reduction seen on picture

This is very complex and noise reduction is only one of benefits for multiple pass. Ordinary noise reduction is always lossy - with multiple pass there is no loss - 4 pass equal 6dB ie 1 additional bit of resolution - only at a cost of time.

[jagabo]

If not i sad it its not worth the time and effort for only that type of noise reduction seen on picture

This is very complex and noise reduction is only one of benefits for multiple pass. Ordinary noise reduction is always lossy - with multiple pass there is no loss - 4 pass equal 6dB ie 1 additional bit of resolution - only at a cost of time.

Yes, noise reduction by filtering will reduce picture elements with properties similar to the noise. Noise reduction by averaging multiple caps will reduce noise without reducing the picture elements. For example, a fuzzy sweater may look like noise to a filter resulting in the fuzz being smoothed out. But with multiple caps the fuzz will remain and only real noise will be reduced.

[mammo1789]

Quality loss from a few plays of a VHS tape are negligible.

Thanks jagabo i read many times that with every playing of the tape it losses particles maybe I'm wrong but i heard it many times , unlike cd or dvd which doesn't have physical contact with the media, that's why i don't tend to play many times mine precious vhs footage.

but for me also I found a very noisy tape where it worked miracles. Before this technqiue I had used removedirt as the most powerful technique but it ruined the video. This was extremely bad video.

As i understand you correctly before you sad that this technic is only for noise from equipment and not from the tape or material it self.

I know about similar treads i saw them but i was thinking about personal settings of experience guys like you and jagabo and ls
For example for chroma noise you usualy use that and this( setting x) for grain this (setting x) for other type this in form of plugins or some kind of software

Yes, noise reduction by filtering will reduce picture elements with properties similar to the noise. Noise reduction by averaging multiple caps will reduce noise without reducing the picture elements

this will only imply for noise from the equipment right? not the noise from the tape itself.
I'm saying these because the most of my bad tapes or tapes that have bad signal or lots of noise i tried them in 3 or four different combination of equipment and its probably the same ( to my eyes at least) are material problems not equipment ( although I agree that equipment is adding some anomalies like noise for instance)

jagabo you sad you align them by eye ( can i do the same or could i screw something and made it worse )

[Gavino]

For example, a fuzzy sweater may look like noise to a filter resulting in the fuzz being smoothed out. But with multiple caps the fuzz will remain and only real noise will be reduced.

Although some of the fuzz can be still be smoothed out because of the slight horizontal blurring effect:

You have to be careful though; some multiple playback averaging is actually doing a spatial averaging as well, due to the copies not being lined up, it could be something like a 4x1 weighted convolution.

[pandy]

OK, few facts - luminance bandwidth from typical VCR is around 2 - 3MHz (at best case this is around 320 pixels - chroma bandwidth is way lower - around 75 pixels - we capture 720 pixels for luma and 360 for chroma), video can be stabilized by TBC (even in software after capturing and usually with half-pel accuracy that can be extended). Tapes are noisy as storage medium - each pass is different, noise in equipment is also each time different - this is very complex - at same point we can isolate like few tens of potential noise sources from tape to final picture.

[jagabo]

this will only imply for noise from the equipment right? not the noise from the tape itself.

Yes. Noise from a noisy broadcast, for example, would not be removed.

jagabo you sad you align them by eye ( can i do the same or could i screw something and made it worse )

Yes. I used StackHorizontal() to view two caps side by side, then Trim() to sync the two. Then repeated the process for the other two caps until I had all four caps synchronized.

Although some of the fuzz can be still be smoothed out because of the slight horizontal blurring effect:

Yes.

[jmac698]

I measured 2.1MHz

[jmac698]

I found out the standard used in vcr spec sheets, it's not the same as audio for example. It was set at 6db down or so, if I recall. Maybe even more.

[jagabo]

I use the point where the output signal is half the amplitude of the input signal, about -3db. Just because it's easy to see. VHS chroma resolution calculation:

https://forum.videohelp.com/threads/319420-Who-uses-a-DVD-recorder-as-a-line-TBC-and-wh...=1#post1981589

from the caps in this post:

https://forum.videohelp.com/threads/319420-Who-uses-a-DVD-recorder-as-a-line-TBC-and-wh...=1#post1980652

[jmac698]

I made that chart the easy way, just by recording sinewaves and measuring the peak to peak luma amplitude. The y scale in luma values, which translates to 5 mV I believe. Btw SVHS is off the scale, I didn't record fine enough sinewaves to find the limit. Probalby 4-5mHz. IN pixels I can see spaces of 3 pixels clearly, but pure hceckerboard if fairly grey.
Nice links.

[2Bdecided]

@2bedecided
Both views are correct, it's just that in your experience, the noise originally on the tape was more visible than the additional noise added by the capture chain.

Yes, that's the thing. It's a camcorder original taken in bright light, but still all the noticeable noise is on the tape itself. old example. It's much noisier than the example jagabo posted.

I haven't looked carefully at anything I recorded from TV, which should be far cleaner (excellent reception). I'll have to try.


With analogue cassette tape, it's very noticeable that the tape noise is usually similar to the playback deck noise. i.e. you can hear the noise where the tape starts, but you can hear almost as much noise by pressing play with no tape in the machine. Some machines are better than others, but it would sound better without the playback machine noise. I suppose at best VHS is a similar beast - but with VHS you have frames+lines to aid synchronisation, and just a little softening if it goes wrong. With cassette tape, you have nothing to aid synchronisation except cross-correlation, and horrible phasing and very audio high frequency loss if it goes even one sample wrong. People have often talking about doing this processing with multiple copies of scratchy old records, but I've never seen it done in practice.

Cheers,
David.

[mammo1789]

I had spare time these weekend and i tried capturing from hi8 trough Panasonic dmr es15, i captured 4 copies made trim in virtual dub and virtual dub mod and 4 copies are now perfect match. The problem is when i load the script its says no overlay ( is it some dll or other involved)

[mammo1789]

I also tried jmac version of median it says syntax error.

[poisondeathray]

overlay is spelled incorrectly (missing a "l")

v34=overay

[mammo1789]

errors

[mammo1789]

overlay is spelled incorrectly (missing a "l")

v34=overay

Thanks i didn't see that silly me i just copied from jagabo without thinking

[mammo1789]

I must say that there is definitely some noise reduction happening with this method( I was really skeptic). So if i understand it correctly i don't loose any information of the original picture with these ( unlike noise reduction plugins) right? Here is some examples

[Gavino]

I also tried jmac version of median it says syntax error.

You need to comment out (or remove) the lines marked in red:

                                               Median1-function that gives median of 3 clips:
 Function Median1(clip input_1, clip input_2, clip input_3, string "chroma")
{# median of 3 clips from Helpers.avs by G-force

...

}
 
  Median2-function that gives median of 5 clips:
  
 Function Median2(clip "input_1", clip "input_2", clip "input_3", clip "input_4", clip "input_5", string "chroma")
...

[mammo1789]

In the 'median' approach, there is no blending - for each pixel, the one from the clip which has the median value for that pixel (from the set of input clips) is used.

So which approach will gain better picture assuming all else equal

By the way i tried to comment the red lines and delete them now i got different error

[Gavino]

So which approach will gain better picture assuming all else equal

Depends what sort of noise you have. The median approach works best to filter out 'rogue' pixels like spots, etc, while the average will tend to be better with Gaussian types of noise.

By the way i tried to comment the red lines and delete them now i got different error

What was the error?
In case it's not clear, what jmac698 posted was just the definitions of the median functions. You still need to write the part that loads your sources and calls either median1 (for 3 sources) or median2 (for 5). You will also need to have the necessary dependent plugins loaded.