YAVHSR (Yet Another VHS Restoral): "LILI"

[jagabo]

The UV vector graph is useful in determining color shifts. For example, if you increase U and decrease V, ie move the colors towards the lower right corner, you shift the colors towards blue:

[jagabo]

I just read vectorscope's doc and searched a bit on doom9, but nothing told me what a slant or slope means.

It means the video has faded (or otherwise changed) unevenly across the width (if in the bottom graph) or height (if in the right graph) of the frame. Maybe the reel of film was left sitting on a flat surface and one side of the film faded more than the other. Or somehow a VCR introduced the slant (though I don't really see how that would happen). Note how the sky is very blue on the right, magenta on the left.

[sanlyn]

The cable versions are dim, unsaturated and green in comparison. But if you pump up the gamma and saturation, you see the same effect. On one hand, a big problem. On the other, I'm even happier now that I dug deep to buy After Effects, where this can be fixed (if only I can get time to learn AE in detail!).

Jagabo, you're correct about this imbalance. And I certainly learned something about vectorscopes. In many scenes you'd never notice it, but in some it becomes more obvious when you start working with them. In the 1-minute MPEG problem scenes posted earlier, the video begins with a carnival scene at night. You can see blue fogging in shadows, more on the right than on the left. And besides that, a cyan bias in that shot that doesn't help.

Well . . .I can partially correct that in a hurry. But in version #2 the work's cut out for me. Well, splendid. I love a challenge (don't advertise that statement, please). Anyway, I got bright blue fixed in YUV thanks to your tips on that, and tweaked in RGB. Fortunately the next 35 minutes in version #1 are satisfactory. There are only one or two problems left to go, and then the last 4 minutes.

[jagabo]

I see one way you could repair the chroma in AviSynth. Make an alpha mask that a gradient from black to white across the frame. Adjust the frame for the left edge of the video (less magenta), and adjust another for the right edge (less blue). Then Overlay the two frames together using the alpha mask:

left=ColorYUV(less magenta)
right=ColorYUV(less blue)
Overlay(left, right, 0, 0, alpha)

[sanlyn]

I didn't know you could use alpha masks in YUV, except for grayscale (luma only). I know one can modify overlay affects via opacity. Lemme look that up in the morning.

[jagabo]

And in case Gavino is reading this, you can use mt_merge() instead of Overlay():

left=ColorYUV(less magenta)
right=ColorYUV(less blue)
mt_merge(left, right, mask, true)

LOL

[lordsmurf]

Interesting.
Subscribed.

[Gavino]

I'm able to fiddle with ColorYUV luma OK and reign it in to 16-235, fixing some crushed blacks and fried highlights. But this does nothing for u and v colliding into the bright side of the histogram. Gain and offset just screw up everything across the board. Or does ColorYUV chroma contrast do something I'm not aware of?

"Gain" multiplies the signal (pixel value) by a given factor, preserving the zero point.
As the ColorYUV() documentation says, "it doesn't make sense to apply this setting to the chroma of the signal" (since it's neutral point is at 128). By contrast (pun intended ), the 'cont' parameters modify the signal preserving the centre point, so cont_u and cont_v are more appropriate when adjusting chroma (and cont_y is not normally useful for luma).

And in case Gavino is reading this, you can use mt_merge() instead of Overlay():

left=ColorYUV(less magenta)
right=ColorYUV(less blue)
mt_merge(left, right, mask, true)

LOL

Good to see you're picking up on Masktools, jagabo.
Note the mask here would not be a (RGB32) alpha mask, it has to be a YV12 clip - you could use mt_lutspa to create a horizontal gradient mask: mt_lutspa(expr="x 255 *").

[jagabo]

Here's an example. I start with the sample image (to have something interesting too look at) and make it grayscale (so it will be very obvious what happens later). I make a copy of that grayscale image with the U plane shifted up, making it purple. Then make another copy of the image with the V plane shifted up, making it magenta. A horizontal gradient alpha mask was made using Gavino's suggested mt_lutspa(). Then the purple and magenta images were blended together using that mask:

ImageSource("tooblue.jpg")
BilinearResize(320,240)
ConvertToYV12()
Greyscale()

purple=ColorYUV(last, off_u=50)
magenta=ColorYUV(last, off_v=50)
Mask=mt_lutspa(expr="x 255 *")
mt_merge(purple, magenta, mask, true)

StackVertical(StackHorizontal(purple, magenta), StackHorizontal(mask, last))

The result is an image that's purple at the left edge, magenta at the right edge, with a smooth gradient between the two extremes:

[sanlyn]

OK, jagabo, I'll pursue that idea. Been reading up on masktools all morning and will research more to learn what your script is doing.

Going to what gavino posted above: The problem with blue is in the brights from RGB 180 up. Midtones are about right, then darks get too green, losing magenta. So the 0 and 128 points with blue are about OK, it's above RGB 128 that I was trying to reduce. For some reason cont_u at extreme negative values got blue out of the 155-plus range. But Blue was really shooting up there, I added a sharp VirtualDub grad curve from 180 on up!

The image I posted is early in the scene. It's 13.5 seconds, during which it displays the same hue changes as the titles on all versions. I thought this was Macrovision effects on the tape, but it's on cable too. Hue changes last longer as the movie plays, then gradually dissipate until 30 minutes later they're almost gone.

So, the opening minutes are a mess in all versions. Vectorscopes of following scenes show no slope, but slower hue drifts between cyan and magenta (more subtle, too). Then you suddenly hit something like the carnival scene, with problems similar to this one. It's the same on VHS and cable.

Frame 194 gives this vectorscope, for 1 frame only. But I don't see any color changes at all:


Frames 58 and 59 have a small but bright Sunkist orange to the left of the larger red vehicle -- exactly like both cable versions (No bobbing this time. I'll use SeparateFields. In the IVTC version I'll run RemoveSpots)

150 frames after the frame I posted, it's too green. 80 frames later, too red. Just as blue starts taking over again, the scene ends and the next scene actually looks good with scant color work needed. After spending an hour reading up on masktools, looking at the scene again it appears the left side street and buildings are also too blue, just as bad, all above RGB 180 - and no red/green color difference in bright building facades between left and right. Effectively, then, the entire scene has too much bright blue and too much green in shadows. So the mask colors reduce browns and yellows in the buildings. Green in the darks is another problem. The suit worn by the man walking left to right would change color. So I think I know why the cable versions are so green: apparently it masks many of these hue changes thru the movie.

Vectorscope for the next scene has less of a slope (mixed light with bluish shadows, and lots of greenery here). By frame 1210 there's no slope at all.


OK, I'll keep at it. Gotta go rig up my other video PC now and let it run with another Christmas project while I work on this one.

Fortunately(?), the camera doesn't move in this shot.

[sanlyn]

These are RGB histograms of the unmodifed capture. The problem street scene is on the left. You can see bright blue stretched beyond the other colors.



I just looked at all 4 VHS captures. This scene looks like the same doodoo on all of them. But all the other scenes, including 4 more shots of this same location, don't have the blue problem or the red offset seen in the 'gram on the left. The other scenes (all of them) look similar to the histograms on the right. The cable recordings have a different problem with this shot (red).

Any mask that adds blue to any part of the scene makes blue look worse and harder to handle. After trying a few mask ideas in Photoshop and Avisynth, no one mask resolves blue or other left-right color imbalance. The trick with this scene involves green shadows and too much red on the left and middle. I'm getting my proc amp and stuff and capturing this scene again. I hate VHS.

[jagabo]

These are RGB histograms of the unmodifed capture. The problem street scene is on the left. You can see bright blue stretched beyond the other colors.

I'm not arguing that the street scene doesn't have too much blue, but there's no reason the RGB histograms should look similar. The colors in this image are all close to what the creator intended and the histograms look nothing alike:



Similar RGB histograms usually means there's lots of grays in the image.

[sanlyn]

Yes, you're right, not every shot should look exactly "alike". None of them do. I refer to overall patterns. This sequence of scenes in the street has the same locale, lighting, etc., but none of them have a color stretched the way blue is, jerked back and peaking almost backwards like red, etc. The other scenes still need tweaking, even the other street shots very nearly like this one, but none of them are this whacko. It's just a maverick piece of source, even in the cable videos.

I just recaptured it. The PA-1 proc amp gave me something more workable -- but in the IVTC version, I'll still have to take time for serious masking on that scene. As you said, there is a left/right imbalance (it's red-vs-green) that's even more obvious once blue got tamed. I discovered I've been trying a mask to correct 3 colors in 2 ranges. Will post later after I repack all this gear, and you'll see what I mean.

Well, at least I learned something about masktools.

[lordsmurf]

What are you using to show histograms in Vdub?

[sanlyn]

What are you using to show histograms in Vdub?

Copy VirtualDub output window to clipboard, paste in new image in ancient Photoshop or ancient PaintShopPro. Then crop, etc.

[jagabo]

What are you using to show histograms in Vdub?

ColorTools. http://www.trevlac.us/colorCorrection/colorTools.html

My cap was an Alt+PrintScrn cap of the VirtualDub window, pasted into an image editor, saved as JPG. The forum software displays it downscaled. Click on it a few times to get full size.

[sanlyn]

Ed: Ooops. I hit Submit accidentally. Isn't there a way to delete these booboo's?

[sanlyn]

For now, I'm saving levels and color fixes of the street scene until next year. The VHS scene is too corrupt to spend another 2 weeks on masking and other tricks. I'll use the PBS broadcast version of this shot, which is more workable. Anyone who wants to play with the original 13-second AVI can use this copy (640x480, Lagarith YUY2, 103MB):
http://dc400.4shared.com/download/Eo7kiX1d/EB3_LIL2B_L.avi
But I don't think it's worth the effort.

The VHS production lab decided that instead of warmish morning light, this scene should look like a brilliant, blue-blasted high noon with a color temp approaching 12000K (you don't even get that much blue under heavy overcast!). You expect to see Gary Cooper on the street under a blazing noon sun. The cable version has the more correct Fall morning light. Reduce bright blue on the VHS, and you find data mostly destroyed in the RGB 30 to 90 range. A new capture with a proc amp didn't fix blue.

unmodified VHS:



unmodified analog cable print (buildings have correct colors for French Alpine villages):



For the upcoming IVTC version I'll recapture, plus another capture with new proc amp settings for shots with serious level and chroma glitches.

Found user comments on Amazon: some complaints about horrible color and tint drift. So a new tape won't help. From reviews, it appears the DVD looks worse than the tape.

Later, all versions of the movie have crap like below. Nothing like a huge reel change signal to liven things up.

[2Bdecided]

There are clips on YouTube. They seem to have come from TCM, rather than DVD.

I mention this in case they're any help, but sadly I doubt they are.

Cheers,
David.

[sanlyn]

2BDecided: yes, I've seen those for some time. It's questionable that I'd be able to use them. Only place anyone could get clips is from old or new VHS (which all look alike and get ruined on UTube anyway) or the DVD (NTSC or PAL, which look worse) or the TCM print (bitrate so low it looks blurred). Where the PBS print came from I don't know; later it was broadcast thru digital box and looked as bad as TCM's. Anyway, the PBS version has worse problems later in the movie. Tough to denoise, too, as it's already gone thru 2 or more "digitizing" processes before it reached my DVD recorder. There are some night and dim-light scenes where the PBS version just falls apart, and long stretches of cable transmission noise.

A new capture will help. I'm learning After Effects, which has more advanced color controls. But the big first step is fixing as much as possible during capture and in YUY2.

[sanlyn]

IVTC/Decimate: I note on a recent capture that VirtualDub capture says it dropped 7 frames during a 90-minute video (tape damage, no surprise) and inserted 2 frames to maintain audio sync. What happens when IVTC hits these dropped or added frames, and how would I correct it? I haven't found anything on Doom9 about this with IVTC/Decimate.

[jagabo]

Something like TFM().TDecimate() tries to adapt to breaks in the telecine pattern. Otherwise you would have to manually remove, duplicate, or add motion interpolated fields or frames.

[sanlyn]

Hm. It figures. I haven't encountered that problem yet, jagabo, but I'll get there eventually if VirtualDub's capture messages are anything to go by. I'll have to come up with something specific and play it by ear (what else?). Thank you.

I captured this whole video again, this time with capture filter settings more appropriate for the problem scenes. Far more workable results now. Still a pain in the neck, though. I hate VHS (did I mention that?)

[sanlyn]

This is the problem I mentioned with the SignVideo PA-100 proc amp and some Macrovision tapes. At bthe bottom of the image is a 20-pixel high discoloration ("gray bar"). I've had to exagerrate the brightness/contrast to make it more visible a PC monitor (it's far more visible on TV). Arrows in the images indicate the problem area. The affected area is visible as lighter gray in dark areas, darker gray on bright shots, around RGB 24. It doesn't appear on home-made or unprotected tapes. About 30 to 45 minutes into a tape, the affect dissipates and can't be seen.





I know the images show other noise problems; I fixed those, but the gray bar persists. In the "Home Video" logo, the gray on the right-hand border isn't in the movie. This happens with every player, capture card and OS. My usual setup for VHS capture with VirtualDub is:

VCR > DVD recorder (TBC pass-thru) > PA-100 > AVT 8710 (if needed) > ATI AIW 9600XT or 7500 > VirtualDub > AVI

These images were made with:

VCR > PA-100 > ATI card > VirtualDub > AVI

How I arrange components in the circuit makes no difference. As soon as I remove the PA-100, the gray area is gone. Same problem if I capture directly to a DVD recorder or play directly to TV with the PA-100 hooked up, so the capture card or TBC isn't a factor.

[jmac698]

Just a reminder, Lili is playing Feb 19 on TCM. If you look here, http://www.tcm.com/mediaroom/video/1466/Lili-A-TCM-Original-Featurette-.html, the clip they used looks restored. You can see the big splotches on the trailer in the hi lili song scene, but it looks much better in the featurette version, so hopefully the restored version is playing. You can also use that as a color reference.

[sanlyn]

jmac698 I saw it listed on TCM's schedule and will record it. TCM broadcasts a low-bitrate version, probably similar to other recordings I made, but I'll record it anyway to see if it's any better.

[sanlyn]

Look, folks, this ivtc/deinterlace stuff is really cool for running Avisynth utils, and encoding with 3:2 pulldown. Removed lots of noise, got great color. Problem: THE DVD MOVIE JERKS AROUND LIKE A SONOVABITCH WHEN PLAYED ON TV!!!

Sorry, people., I'm going back to the telecined/interlaced version. Interesting experiment, though. But this ain't anime or UTube.

[jagabo]

Look, folks, this ivtc/deinterlace stuff is really cool for running Avisynth utils, and encoding with 3:2 pulldown. Removed lots of noise, got great color. Problem: THE DVD MOVIE JERKS AROUND LIKE A SONOVABITCH WHEN PLAYED ON TV!!!

Then you did something wrong. Did you encode at 23.976 fps progressive with 3:2 pulldown flags? Or perform hard 3:2 pulldown and encode at 29.97 fps interlaced? Either of those should have worked. If you step through the IVTC'd AVS output and verify motions were smooth?

The majority of movies on DVD are encoded 23.976 fps progressive with pulldown flags. So you can be sure that is not a problem.

[manono]

Sorry, people., I'm going back to the telecined/interlaced version.

That's a mistake, if for no other reason than encoder efficiency (and there are others). Much better would be to figure out where you messed up. Samples of before and after and the script used might be helpful.

[sanlyn]

Did you encode at 23.976 fps progressive with 3:2 pulldown flags? Or perform hard 3:2 pulldown and encode at 29.97 fps interlaced? Either of those should have worked. If you step through the IVTC'd AVS output and verify motions were smooth?

Yes to all three.

OK, the panic button is disabled. I been looking at frames, frames, frames, fields, fields, fields. We know the original source (VHS and cable broadcast as well) are telecined. If the originals are also interlaced, they don't "play" that way with SeparateFields: that is, Top and Bottom fields viewed in succession are "the same moment in time", not two different instants.

All clips encoded after these statements have the problem:

AssumeTFF()
TFM(order=1)
TDecimate()

None of the clips encoded before those statements have the problem. Also played with mode, order, etc., etc. No go. Starting tomorrow I have to back up to frame 0 and redo all of it. Filtering with Bob, yadif, SeparateFields, etc., never gave me this problem.

The source doesn't look interlaced to me. Using separate fields on the telecined source, you see:
aa bbb cc ddd ee fff gg hhh ii jjj kk (looks like progressive/telecine to me)

the ivtc'd version of the same scene:
aa bb cc dd ee ff gg hh ii jj kk (looks like pure progressive to me)

But there's something screwy with the 3:2 pulldown encodes, from all 3 encoders, because it now looks interlaced:
ab cd de ef fg hi ij jk kl lm no (looks like interlaced/telecined)

Hmm.