U-Matic Capture Experiments with Non-Traditional Capture Setup (SDI/ProRes)

[aramkolt]

I've been working on seeing what the best Umatic capture configurations are for my specific player and am going a little non-traditional with SDI and/or Component into an AJA KiPro compressed to ProRes422.

I tried several captures of the same section of a U-Matic tape which is probably a second generation Dub at least due to the original editing. I do not know if the original local TV company that made this would have used dub cables during editing or composite.

Here is a link to a folder that has the raw capture samples:
https://www.dropbox.com/scl/fo/on1mllei82t501lybac58/AGkVSHCjI13eFRdQ16FlAeM?rlkey=hyt...ikfjl001q&dl=0

I apologize in advance for the annoying audio that is at the beginning, but it did serve as a good cue that the video was about to start.

You'll also want to download the files to get the full experience as they are the raw files and have not been deinterlaced. Seems the dropbox player actually does do some form of deinterlacing which isn't the worst I've seen, but it probably is also at a reduced bitrate as well.



Here's the file descriptions for the different configurations:

CVR900_Composite - Regular composite out of the player, to the CVR900 to digitize, and then via SDI into the KiPro. There's not a ton of info on the CVR900, but it should act as a frame TBC during conversion.

CVR900_SVideo - Uses the same noise reduction circuit as composite output, just tapped chroma and luma before they are mixed.

CVR900_SVideo_PreNR - Eliminates one noise reduction circuit on luma, otherwise same as above. I suspect that just grabbing luma off of the Dub output would look virtually identical, but will have to test that later.

DPS_DUB_NNR - DPS TBC connected via DUB cable, then into the AJA KiPro via Component - I believe this setup uses no noise reduction at all - Color saturation is quite a bit higher just due to how the DPS unit happened to be set, so try to ignore that part. If the tape had color bars on it at the beginning, I could have probably had the levels roughly the same, but it didn't. The on screen graphics do show some interesting color noise that seems to confuse the other setups, but I can't say which is actually more accurate to what is on the tape. Of note, this particular DPS grabs chroma from the 688KHz DUB signal and not from composite like most TBCs of that era as well as the Keystrobe dub optimizer (which I don't have for testing).

Elgato Video Capture via Composite - Just put that one in there for fun to show what some of the YouTube capture companies consider adequate quality. The couple videos I've seen use older generation machines (5xx0 series vs 9xx0 here) so I don't imagine that helps the quality either. The above setups really don't take more time to do (other than some post capture deinterlacing for a delivery copy), but they certainly take up a lot more disk space and some more initial expense. I'd summarize that the elgato is watchable given the file size and as long as the title isn't on screen, but for some reason it really doesn't like whatever character generator made the title characters. In a vacuum, it's passable, but knowing the quality could have been like some of the other captures is hard to un-see.




I'm curious if others think that one method is obviously superior? The noise reduced S-Video to me appears to have less ringing (which Umatic is known for), but at the cost of sharpness (or at least perceived sharpness). Not sure if Avisynth would do a better job of reducing the ringing on the more detailed capture, or if in some cases it would be better to use the normally present noise reduction for most captures? I think the DPS and the pre-noise reduction S-Video look the best, but it could be due to the amount of grain making me think there's more detail there than there actually is.


I'd also like to inquire as to what kind of Avisynth scripts would be recommended on these samples for QTGMC as well as to help get rid of some of the occasional line dropouts and ringing while trying to preserve some of the sharpness?

[lollo]

Missed your interesting post, sorry for the late reply. Here my quick analysis.

You are loosing details in the composite capture versus the s-video capture, because whites clipping (the shoulder of the man)





comparison:


with slider: https://imgsli.com/MzY1OTg0

Overall, i prefer s-video capture because sharpness and colors:

comparison:


with slider: https://imgsli.com/MzY1OTg4

Comparing SVideo versus SVideo_PreNR is more difficult. I think the second is more detailed while having more noise (the vertical pattern in the middle of the van (assuming is not an artifact), the logo in the front of the van, etc.):



with slider: https://imgsli.com/MzY1OTkw

If a restoration with AviSynth is planned, this capture is a better candidate, because we can soften the noise while (hopefully) mantain the higher number of details present compared to the SVideo capture.

Introducing now the 4th capture (DPS_DUB_NNR.mov) the bad colors do not allow directly a fair comparison, and the image is loosing a portion of the image (left shift by 4 pixels); in addition, there is a distortion of the top of the picture, clockwise.



Reducing a little bit the saturation and the V channel to equalize and compare, at first impression I prefer the CVR900_SVideo_PreNR capture:



with slider: https://imgsli.com/MzY1OTk3

Mind that I did not check all parameters for the whole sequences, just few shots to have an overview. So YMMV.

Concerning the appropriate AviSynth flow, the source is interlaced BFF, so a basic QTGMC with a setting to do not smooth too much and denoise/sharpening may fit. For dropouts, a frame replacement with interpolation is quick and save, otherwise intraframe correction.

edit: looking back to SVideo_PreNR maybe the noise is too much to be removed without side effects. Many tests need to be done

[Sharc]

Using the waveform monitor:
Apart from the noise I would prefer the CVR900_SVideo_PreNR because it is the only one which does neither crush nor clip the brights
- The CVR900_Composite produces out of gamut RGB and clipping at Y=255, loosing all details in the brights
- The CVR900_SVideo is within Y(16....235) but still clips/crushes the brights at Y=209 and is therefore still loosing details in the brights

Edit: All 3 with chroma ghosts

(DUB and Elgato are out)

[dellsam34]

Looking at Lollo's findings, I see ringing increases going from composite all the way to pre-dub as well as chroma artifacts, If composite did not have overblown whites I would opt for it. Hoepfully there are post filters that can mitigate all those problems.

[Sharc]

Hoepfully there are post filters that can mitigate all those problems.

Clipped signals (luma or RGB) cannot be recovered in post. What is lost is lost. One has to improve the capture process to avoid clipping upfront. In this respect the noisy SVideo_PreNR is the best.

[aramkolt]

Thanks for all of the replies! Very interesting analysis to be sure - I am pretty confident that I can get rid of the clipping on all samples with appropriate proc amp adjustments. Both the DPS and CVR do have proc amp adjustments, but it might be the case that the clipping occurs at the input if the CVR900 is clipping if it's showing that at Y=226 or Y=209 in the one capture.

I know some TBCs will clip right at the input if incoming signal is too hot and then adjusting with their proc amps will just adjust the "already clipped" levels lower. This is a well known limitation of the DPS475/575 for example. I'll have to run an amplified ramp pattern into it and see if it clips at a certain input level on input.

The CVR900 interested me in particular because I believe it has line-TBC-like effects (which Umatic otherwise lacks). If the CVR900 clips hot signals at the input, that makes it a less "all in one device" than I was hoping, or I may have some setting for automatic gain control set incorrectly. Automatic color control produces strange artifacts occasionally at least with U-Matic, so that I'll be leaving off.


If clipping was not an issue on any of the captures, which is preferable for restoration in terms of detail? If the DPS/DUB levels are fixed, are there further issues with that?

I'd be curious if someone could post an Avisynth converted samples (and the scripts that made them) that does the best they can with the ringing in particular while hopefully not losing too much of the other detail.

[Sharc]

I'd be curious if someone could post an Avisynth converted samples (and the scripts that made them) that does the best they can with the ringing in particular while hopefully not losing too much of the other detail.

In addition to revisit the levels/clipping: Before throwing scripts at it make sure that the Y/C tapping does not pick up noise via unshielded wires and does not cause signal reflections (possibly interpreted as "ringing") due to introducing an impedance mismatch for example.
Hardware modifications are quite delicate.

[dellsam34]

Clipped signals (luma or RGB) cannot be recovered in post. What is lost is lost. One has to improve the capture process to avoid clipping upfront. In this respect the noisy SVideo_PreNR is the best.

No, not the composite, I've eliminated it already due to clipping, I was refering to ringing and chroma artifacts in the S-Video and Dub.

[lollo]

I'd be curious if someone could post an Avisynth converted samples (and the scripts that made them) that does the best they can with the ringing in particular while hopefully not losing too much of the other detail.

Minimum processing and deringing on capture CVR900_SVideo.mov to upscale for youtube:

QTGMC(preset="slow", matchpreset="slow", matchpreset2="slow", sourcematch=3, tr1=2, tr2=0, NoiseTR=2, sharpness=0.1)
dering(A=0.3, W=3.5, RelVisibleTh=0.01, AbsVisibleTh=0.005, PostFilter="median2", PreFilter="FIR2", SkipAlonePeaks=true)
TemporalDegrain2(degrainTR=1)
LSFmod(defaults="slow")
nnedi3_rpow2(rfactor=2, nns=4, qual=2, cshift="Spline36Resize", fwidth=1440, fheight=1080)

deringing effect: https://imgsli.com/MzY2MTk3

org vs restored: https://imgsli.com/MzY2MTk4

ouuput: https://www.youtube.com/watch?v=GQMVtFD0s5g (uploaded to my channel because large size, then compressed by YT)

This was just a quick attempt, try to experiment changing the parameters of the filters.

The capture CVR900_SVideo_PreNR requires more aggressive denoise (among other stuff) so, for instance, try to increase temporal radius of TD2 from 1 to 3.

For lines dropout you can use interpolation with MVtools or RIFE (better):

function interpolate_frame_GMa_1frame(clip c, int "frame_number")
{
	# interpolate bad frame
sup = c.MSuper()\
bv = MAnalyse(sup, isb=true, delta=2)
fv = MAnalyse(sup, isb=false, delta=2)
interpolated = MFlowInter(c, sup, bv, fv)

#stackhorizontal(c, interpolated)
#return(interpolated)

	# replace bad frame with interpolated frame
video_repaired = \
	c.trim(0,frame_number-1)++\
	interpolated.trim(frame_number-1,-1)++\
	c.trim(frame_number+1,0)

	# video repaired
return(video_repaired)
}

As Sharc and dellsam34 pointed out, for optimal results try to have the best capture possible, fixing all issues before AviSynth