Fixing bad color joining in Technicolor films

[billw6560]

I realize that this is probably a pie in the sky request, but I recently (and finally) got my first 4K TV, an 77" LG B4. I previously had a Samsung 61" DLP HDTV that has served me well over the last 9 years, but it finally started failing. With that said, I always knew that several Technicolor films in my library were poorly done, specifically the joining of the resultant Red, Green, and Blue sections to a single print. On the 61" TV, the films that have poor joining, such as "The Glenn Miller Story", in many scenes the Red and Green are not well aligned up with the Blue, displaying Red and Green edge shadows. I have two different sources of this film, OTA and BluRay, and both have the same conditions.

I mentioned my TV's in that with the 61" DLP HD set, these anomalies were present, but mostly watchable. However, with the 4K 77" set with it's very good video processing, these anomalies are very noticeable and annoying.

So my question is, is there any video processing software, whether free or commercial, that would take the video file, separate the RGB data, then have the ability to recombine/realign these three separate streams into a new single file.

Again, I'm guessing that this is s hard no, but there are no dumb questions, only those that were not asked. Thanks!

[poisondeathray]

Dedicated R,G,B color channel align would be NukeX's f_colour_align . Other options would be photoshop image stack R,G,B and autoalign , or maybe avisynth autooverlay using R,G,B

Potential issues

1) Sometimes what is mistaken for R,G,B registration/alignment issue, is actually a grading issue affecting 1 channel (or a combination) . These cases won't be "magically" fixed, you would have to do a bunch of others things at minimum

2) Normally fixes are done upstream because a consumer chroma subsampled 4:2:0 source like Blu-ray is going to blur and deteriorate the color edges, making a clean "fix" more difficult

If you see the problem on a retail BD, either they couldn't fix it easily , or were too cheap to fix it properly

[billw6560]

I agree, it's probably an issue with the source of the mastering of the three color reels into one. The original process was to take all three reels, and project them onto a single master/copy print. In HD times, I'd imagine that each individual color reel was telecined into a 4K capture camera, then digitally merged in post processing.

However, in this example, the separation is not consistent. Most scenes are OK, but others (shown below) are quite out of alignment. I will look into NukeX's f_colour_align. With the proliferation of AI, this would be a good project for it, to analyze each frame and make the appropriate adjustments.

I think that these Technicolor releases were probably a bit rushed and/or the daughter prints were not in the best condition.


[Attachment 84465 - Click to enlarge]

[poisondeathray]

It's not just a "simple" 3 strip alignment issue - typical alignment/registration techniques wont work for your case

Red channel has a diffuse blur and bloom (with quantization artifacts - maybe those were partially from jpeg) - the effective resolution is lower than the other channels - this is the biggest problem

With the proliferation of AI, this would be a good project for it, to analyze each frame and make the appropriate adjustments.

Yes, there could be some machine learning projects that could help , but the Red channel needs special attention

[poisondeathray]

I tried of the few earlier suggestions and they don't work for your problem, but they work fairly well for "normal" misalignment cases

Here is one for matlab that is for "normal" misaligment, but won't work well for problem
https://github.com/AhmedIsmail007/Aligning-RGB-Channels-using-SSD

Your case is more like a lower resolution red channel , and that results in a sort of "chroma bleeding"

There are some avisynth filters that attempt to sharpen (awarpsharp2) and align the chroma based on the luma, and they can improve some areas, but also worsen some areas . The filters are applied at a fixed value - they do not automatically calculate some ideal values for you. If you have varying degrees of severity in different scenes, you would need to apply the filter(s) with different strengths


[Attachment 84478 - Click to enlarge]

ImageSource("GMS Color 1.jpg")
ConvertToPlanarRGB()
z_convertformat(pixel_type="YUV444P16", colorspace_op="rgb:709:709:f=>709:709:709:l")
mergechroma(awarpsharp2(depth=70))
mergechroma(awarpsharp2(depth=50))

Machine learning might be suited for this, but I haven't seen any already developed with pretrained models for this specific problem yet

[salini]

This is really interesting! I think AI could help fix the color issues in these films. Maybe using machine learning to adjust each frame could work better for different kinds of misalignment or color problems. Has anyone tried AI tools like Topaz Video AI for this? I'd love to know how it worked

[billw6560]

All good points. This issue is that only parts of the "The Glenn Miller Story" are affected, most of the film is relatively OK, but not perfect. Another example is the BD re-release of "White Christmas", which was done very well. However, there is one scene at the beginning, when Bing enters the MASH tent to thank Danny for saving his life. Here too, there's noticeable color separation. However, this is the only scene this is occurs. The rest of the film is near perfect.

The other points are also well taken. If the entire film was misaligned, there's a much better chance it could be fixed. With only some scenes affected, this AI would need to be very smart to differentiate what is a real misalignment, and what was intended (i.e. artistic effects). I guess if I was a studio that did this remastering, it would be worth the time.

As seen below, the Top Left closeup shot is near perfect, but when the scene went to a long shot (Top Right), it got fuzzy, probably from this misalignment. The bottom two shots are blown-up of Danny's scarf, where the separation is very noticeable. Interestingly, all this only occurs in the long shot (10:46 to 11:05).


[Attachment 84547 - Click to enlarge]