pics for @ChaosKing, dvd restoration

[feisty2]

following https://forum.doom9.org/showthread.php?p=1917199#post1917199, here are some before and after pics

you will probably notice some detail loss but that's okay, I don't give a darn about preserving details, I only care about perceptual quality (result looks good without reference to the source) and the overall "delicate" and "fine" look that resembles the image quality of a master tape. it's hard to describe what I am trying to say but a pair of trained eyes could tell. I also added some slight chromatic aberration to further enhance the overall "ethereal" and "surreal" look of this clip but that's just my personal preference.

you're welcome to grab the "before" pics and see if you could achieve results similar to the "after" pics with all your favorite avs/vs filters and scripts :devil:

before
https://i.imgur.com/HgYajcv.png
https://i.imgur.com/ZltMP2Y.png
https://i.imgur.com/rJ4WIWj.png
https://i.imgur.com/ulhOIOI.png
https://i.imgur.com/zd6qOwR.png
https://i.imgur.com/UHVi4P7.png
https://i.imgur.com/7WbzW9I.png
https://i.imgur.com/oda4fVW.png







after
https://i.imgur.com/ITDBVhe.png
https://i.imgur.com/fY7Uxbp.png
https://i.imgur.com/B2Jt7hJ.png
https://i.imgur.com/o9WtRE1.png
https://i.imgur.com/MkymoSo.png
https://i.imgur.com/6Wc3cfx.png
https://i.imgur.com/jUJ2XwT.png
https://i.imgur.com/YrJIdGp.png

[Forteen88]

Good job, the colors looks much better in your video-compositing.

[ChaosKing]

you will probably notice some detail loss but that's okay, I don't give a darn about preserving details,

And I thought you would squeeze out every detail to something like 720p quality/res :D

But this looks nice too. I like the new look and the colors. Great job 👍 I mean is does have more of a clean master tape look.
I wonder if something like "contrasharpen" would bring some finer details back.

What are these "markings" in the after pics?

[feisty2]

I wonder if something like "contrasharpen" would bring some finer details back.

What are these "markings" in the after pics?

- maybe, but it would not contribute much to the perceptual quality so I never bothered trying

- it's NTSC closed captioning (line 21), you may also have noticed that the after pics have an unusual 720x512 resolution, it's the resolution of a digital NTSC master defined by SMPTE D10

[FranceBB]

What are these "markings" in the after pics?

Nowadays, timecodes and other ancillary data like subtitles and other stuff are muxed in a container, but this wasn't the case many years ago, in the "good old days". Back then, timecode and subtitles were displayed in non active lines. If you use a VTR, for instance, to play a BetaCAM, the only reason why it knows and displays the timecode in its 7 segments display is because it's reading it in one specific line:

https://cinequipt.com/cms-files/size-867x867/pvw-2800-1.jpg

https://greentreeav.com/images/Videotape%20Formats/Betacam%20SP%203in.jpg

[feisty2]

most NTSC digital masters are either 720x512(SMPTE D10/Sony IMX) or 720x486(betacam), both have a few more lines than the 720x480 DVD resolution and the auxiliary information will be stored in these extra lines, which would then be cropped when the master is transcoded to DVD video. since my goal is to reverse DVD video to the master tape state, of course that includes not just the master tape image quality, but also the master tape video format.

[ChaosKing]

Thank you both for the explanation.
While googling for SMPTE D10 I found a video about the First nonlinear editing system from 1971 https://www.youtube.com/watch?v=9bNmsKBqFPQ&feature=youtu.be&t=87
With "touch" support :eek:

Looks like StainlessS workstation before he discovered xp and avisynth :D

[feisty2]

@ChaosKing
have you successfully reproduced the "after pics" or similar results with "before pics"?

[ChaosKing]

Can't get the colors "right" and it's without the "soft look"

Here's my attempt
https://i.imgur.com/jRgV2ZE.png
https://i.imgur.com/en5387W.png
https://i.imgur.com/TlzZVsv.png
https://i.imgur.com/HhaJId0.png
https://i.imgur.com/HgUznQH.png
https://i.imgur.com/gLuEBxY.png
https://i.imgur.com/BX2CP0d.png
https://i.imgur.com/P88YwDC.png

[feisty2]

it seems you weren't able to reproduce the "delicate" look, it's not "soft", it's actually sharper, but not the kind of "sharpness" that most avs/vs sharpening scripts/filters can produce. take a closer look at the circular pattern on Britney's hat in the "after" pics, it's much more "fine grained".
when people apply sharpening to an image, usually they want this fine grained "delicate" look, but what they get is the poor "bloated" "over-enhanced" look that a pair of trained eyes can instantly recognize. therefore I have said long ago that people should stop using naive sharpening filters/scripts like LSFMod, but no one seems to care anyways. :p

[StainlessS]

I had a light pen like that [CK YT clip], not sure, think was for Sinclair Spectrum, nasty thing, detected when raster scan beam passed the pen 'pick up'
LDR (Light Dependent Restor). Maybe also had one for Commodore Amiga, still nasty, cant remember what I used either one for. [probably just to play with]

EDIT: Here is the exact same light pen I had on spectrum, ghastly thing, and maybe my first printer [or maybe not first one, not sure].
https://www.youtube.com/watch?v=MXST8LaiMHs

[FranceBB]

since my goal is to reverse DVD video to the master tape state, of course that includes not just the master tape image quality, but also the master tape video format.

Well, too bad the original BetaCAM hasn't be stored then, 'cause it would have been of way higher quality, especially because of yv16 planar... nothing is gonna bring back the chroma definition brutally averaged out to yv12 for the DVD... :(
Out of curiosity, how did you put the timecode / subtitles in non active lines? Did you do it with Avisynth or other open source software? I'm asking 'cause the only way I have right now to put subs on a Tape is to convert subtitles to .stl, find a fucking floppy disk that still works, put it into Softel and then route the SDI signal of the file playing from Omneon to Soften and then to the VTR which is recording.
Why we're still making BetaCAM backups (on top of geo backups of RAID6 and LTO), though, remains a mystery...

[feisty2]

Well, too bad the original BetaCAM hasn't be stored then, 'cause it would have been of way higher quality, especially because of yv16 planar... nothing is gonna bring back the chroma definition brutally averaged out to yv12 for the DVD... :(


false, since Y is fully sampled, it could be the guiding reference to restore chroma resolution with very high fidelity. traditional methods (low-rank, sparse representation, etc.) already achieved quite high quality for this kind of task, not to even mention more modern (DNN based) methods. also I've posted something similar long time ago. (https://forum.doom9.org/showthread.php?t=173028)

all the "after" pics have been restored to 4:4:4 using my private neural net model, you can see the 4:2:0 chorma aliasing in the "before" pics is completely gone in the "after" pics

[FranceBB]

false, since Y is fully sampled, it could be the guiding reference to restore chroma resolution with very high fidelity.

perhaps I underestimated the power of neural networks and GPU computation for this kind of tasks. When I have a bit of time, I wanna try to convert a masterfile to yv12, convert it back to yv16 with your filter and compare it with the original masterfile to spot the differences (and I'm pretty sure there will be as it can't be perfect and produce something exactly like the native file). Still, with yv12 sources your plugin really is worth to be used. I'm gonna test its speed and benefits on a few files on Monday at work, 'cause if it ain't too slow on our Xeon/Quadro CPU/GPU server at work, I might (just might) enable it as a conditional for all the yv12 crap we get from mobile phones, facebook, twitter and YouTube for our news channel ('till today I've just used a simple converttoyv16() )...

[feisty2]

chroma reconstructor is no longer supported, as you can see it's posted 5 years ago even before I started college and now I am a graduate student. it's kind of similar to a naïve kernel predicting network (kernel predicted on Y, then applies cross-correlation accordingly on CbCr). These days I would just use a "real" neural network for this kind of things (multichannel super resolution/guided filtering).

[FranceBB]

I see... that explains why it was talking about 16bit stacked, for instance... Anyway, I'm not really expert in neural networks; as a matter of fact, professors have been making examples about how things like gradient descent and back propagation were used in NN but that was pretty much it as we were far more interested in the math at the time...
That said, I've never implemented a neural network myself and I haven't really used any other networks other than NNEDI; as a matter of fact, I'm begging to appreciate the potentialities of neural networks only in recent times, like when there was the topic about modern upscaling methods.