Detelecine improperly resized NTSC movie

[JClement]

I have a copy of a public domain movie which was captured from an old US SD broadcast, then resized and put on YouTube. I am trying to detelecine it to do good restoration, but the intelaced lines are now scrambled. The original telecine was apparently standard with 3 P frames and 2 I frames. The IVTC in VirtualDub gives me 3 good frames and 1 scrambled one. There are several copies of the same movie on the web, but all are from the same poorly edited copy. I have uploaded 2 frames and have put a short clip at https://1drv.ms/v/c/3107902525652401/IQApWjlxGTAwTJN24TShFlT1AUoq7nT3ERcRM0_CeuSSqFs?e=JDGu9d

I assume they cropped the VHS scrambled lines on the bottom, but the original source may have been the problem, but it was from a good website. There may be dropped frames in the original. The problem is somewhat disguised in the actual drama because of the very static nature of the early 1930s movies. The clip clearly shows the problem.

[Selur]

combed content that was improperly resized and saved as progressive content,..
the usual approach is to:

curse at the one who did it
descale resolution (most often just the height),
resize back,
try to get rid of the ghosting left behind,...

if the source was rather high quality, overlay only replace the parts in motion with the new version. (this will require custom scripting)
Nowadays maybe feed it to discussion based resizer (like SeedVR2 or Topaz Starlight), to see whether that helps 're-imagining' some details,...

Cu Selur

[JClement]

I have already done number one on the list. Of course, it is done by people who have no knowledge of how interlacing/telecining works.

Is there a tip on how to descale the existing frames? How can I readily figure out the scale? I would think the ratio would be somewhere near 480/480 to trim off the messy VHS lines.

I have used an "area" based deinterlacer that removes the ghosting, but messes up other lines. I wonder if there is a possible Avisynth procedure for doing it best?

There is another copy of this video on Archive.org with 720 horizontal where someone increased the vertical and messed that copy. They didn't know that .MP4 videos have a display aspect ratio that fixes the anamorphic 720x480 scan to the correct ratio in the player. Another amateur!

John Clement

[JClement]

The version of the movie on Archive.org yields some interesting results. It is resized from 720x480 to 720x536, and when rescaled back to 720x480 using virtual dub I get similar results to the other video which is on YouTube. So the rescaling back is not quite what is needed. A rescaling back using an inverse transform is needed. Of course I do not know whether the upsizing was bilinear, bicubic, Lanczos, or other filters.

So are there some inverse resize filters that could be used in VirtualDub, or possibly Avisynth. So far I have gotten the best results with the VD internal bicubic downscale to 480 followed by the internal deinterlace Top first-blend fields on the Archive.org Sweet Music MP4. All versions on the web seem to be originally from the same commercial movie site in SD resolution.

[Selur]

When using Avisynth:
http://avisynth.nl/index.php/Descale is the first thing that comes to mind for the down scaling.
In regard of tools to help to figure out the right descale resolution, I know of https://github.com/Asd-g/AviSynthPlus-Scripts/tree/master/getnative, but I have never used it.
Don't know of any plugin for VirtualDub for this.

Cu Selur

[Selur]

crop, descaling height and resizing back, then using seedvr2, then spotless, results in https://www.mediafire.com/file/pjyf7sheu6bvakz/seedvr2_and_spotless.mp4/file

[JClement]

Here is a copy of a clip from the Archive version
https://1drv.ms/v/c/3107902525652401/IQDd8IswS5oqRI76JyIcu0xTAXpcm_zE8-lCr0UIgGZ8r5E?e=Zicl5s

This has 720:536 frame size obvious that the original was 720x480 and was up upscaled vertically to create a 1:1 approximate 4:3 aspect ratio because the person did not understand that there is a tag for display aspect ratio that will make the display program work properly.

The other one I posted has 650:480 frame size which is a different attempt to create the correct aspect ratio which should be 640:480 for 1:1 pixels.

To figure out what is actually going on, I used a bob doubler to see the pattern of interlace. Then I looked at one frame in a photo editor. Every 4th or 5th interlace line is perfect, but between them there is progressive smearing, but with evenly spaced interlace lines. The pattern repeats 4x in 18line or 4.5 line repeats. This is very puzzling. Let us say you change the size from 480 to 481, the interlacing would be sharp at the top and bottom but smeared in the middle. There is no discernable other pattern.

The interlacing has not been disturbed by size changes, but by some other means. I would suspect jitter in a telecine machine which would point to the original source for the captured broadcast. It is possible to detelecine ghost, 3:2, 2:3 and duplicate 5th frame telecines, but this one is a unique problem.

Are there any suggestions for this?

[Selur]

Same problem, same approach,... (https://forum.doom9.org/showpost.php?p=2029608&postcount=2)
crop left=116, right=74, top=44, bottom=24, resize to height 292, resize back, try to get rid of the ghosting, i.e. by decimation
SelectEvery(clip, cycle=5, offsets=[0, 1, 2]) return 17.982=18000/1001 fps 'clean' output

[JClement]

Ah, yes how does one get rid of the ghosting. Would the select every really work? It only works if there are no missing or repeated frames which is not a good bet for videos captured off the air. The proposed solution also reduces the vertical resolution as well as cropping the overall video. I provided some clips, so could a demo of how to do it be created? I am very expert with VirtualDub, but a novice with Avisynth. The raw cliip from the upscale Archive version gives identical results to the YouTube version, but with slightly better horizontal resolution, when it is downscaled to 480x720. Since the even and odd half frames are intermixed by ghosts the decimation will just leave them intact.

[Selur]

Here's what SelectEvery does: https://avisynthplus.readthedocs.io/en/latest/avisynthdoc/corefilters/selectevery.html
So with cycle=5, offsets=[0, 1, 2] it would chunk the clip into pairs of 5 and throw out frame 4 and 5. (which for https://1drv.ms/v/c/3107902525652401/IQDd8IswS5oqRI76JyIcu0xTAXpcm_zE8-lCr0UIgGZ8r5E?e=Zicl5s seem to be always blended frames)

[JClement]

That may be true for the clip I presented, but may not be true for the entire 1 hour movie. The rest of the movie does not have black bars at the edges, so it should not be cropped. I just did the clip from the leadin because that clearly shows the problem. Since it is a 1935 movie it has a lot of scenes which are fairly static so the problem is not as easy to see. The clip from Archive would have to be resized to 480 to recover the original captured video with properly spaced "mice teeth". You said that frames 4 & 5 should be discarded, but then you have a jittery result because there is one frame missing out of 4 of the original video. Frame 4 top contains a duplicate of the half frame 3 plus the half frame 4 of the original movie. frame 4 contains the other half frame of the original frame 4 and a half frame of the original next frame 1 of the original 24fps movie.

To use automatic spot remover programs the original 30fps movie has to be detelecined to 24 fps (actually 23.976) so the 2 interlaced frames must be used to bring back the original frame 4 and not discarded.