Filter to correct Colour Recovery quadrant errors on 1970s "Dr. Who"

[alexh110]

2 Entertain recently released "Ambassadors Of Death" on DVD: a "Dr. Who" serial from 1970.
Richard Russell's Colour Recovery software was used to extract and restore colour from black and white film prints of several of the episodes.
However I was disappointed by the number of PAL quadrant errors remaining in the recovered colour content. I think it should be possible to correct many of them with a fully automated AviSynth process.

When the CR process makes a quadrant error, the colours produced correspond to precise reflections of the correct colours in the U and/or V axis. So if the correct colour should be (u,v), the erroneous colours will be either (u,-v), (-u,v) or (-u,-v).
This makes the errors potentially easy to spot, because such precise reflections would be extremely unlikely to occur naturally together in the same image.

The error correction process would have to identify the dozen or so most common colours in each frame (allowing a certain margin for different shades of the same colour).

Then you would check to see if any of these colours are reflections of each other in the U and/or V axis (again allowing some degree of margin for small variations in shade).

If reflections are found, we can assume the erroneous colour(s) will occupy fewer pixels than the correct colour(s), since a fair degree of error correction has already been performed on the material by Richard Russell using his Quadrant Editor software.

The erroneous colours can then be corrected by reflecting them in the appropriate axis (or axes), thus mapping them into the same quadrant as the correct colour.


Unfortunately I don't think my software skills are quite up to writing the code for this process; but wondered if anyone else would like to give it a try? I'm keen to see whether it would produce any improvement.
I'm guessing it would have to be a bespoke plugin, rather than an Avisynth script?
The input would ideally come directly from the vob files of the DVD.

[wonkey_monkey]

As a big Who fan, I'd like to have a go. I haven't got hold of Ambassadors yet, so I'm not too clear on the problem - can you post a short sample cut straight from a VOB?

David

[jmac698]

You should be able to use techniques from colorization to help. Motion tracking could keep once corrected colours stable throughout a scene. This would reduce the manual correction to 3x less or so, as the scene to seconds ratio (even better, scenes to frames ratio)

[alexh110]

I don't think I can really post a clip as I don't want to get into trouble over copyright.
The problem is that there are blotches of erroneous colour all over the place. So for example an area of blue sky will contain blotches of orange where the PAL quadrant has been incorrectly identified. There are more blue pixels than orange; but the orange blotches are still very noticeable and disturbing to the viewer.
What I would like the process to do is identify these blotches as I described, and reverse the sign of the U and/or V signal so they match the quadrant of the correct colour.

jmac the idea is to have a fully automated process with no manual intervention at all. The technique you describe was employed by the "Dr. Who" restoration team for "Planet Of The Daleks". They paid an American company called Legend to manually recolourise one episode of that story. However since then Legend have increased their prices and the process has become unaffordable for "Dr. Who" DVD releases. Instead they're relying entirely on Richard Russell's Colour Recovery technique, described here:
http://colourrecovery.wikispaces.com/

[Asmodian]

If you post a very short clip (10 seconds should be enough) there won't be an issue with copyright.

[IanB]

You should not have trouble with posting the odd frame or two for the purposes of review or comment, which this topic is about.

[alexh110]

OK, I'll post a couple of test frames when I can get round to it, along with their PAL vectorscopes.

Looking at the vectorscope, it becomes clear how much variation in saturation is produced by the Colour Recovery process.
Therefore I think the best way to look for reflections is to convert the U and V values into polar coordinates. That way you could allow for a wide variation along the radial axis; but a much tighter variation on the rotational coordinate. The values of these variations ought to be user-defined parameters that can be passed to the plugin (or script).

[jmac698]

What I don't get is why BBC or other companies don't think of hiring someone from doom9. There are highly skilled people here who are also fans, who would probably give a quote better than a US company. People are already restoring their favorite shows for free, they would jump at the chance to be involved in an actual production.

I'm familiar with Legend, they have a good color tracking technology. IIRC, they got the project before the colour recovery process was mature, so they had no help. What I'm talking about is different, it's much easier. If it has to be fully automatic, I can still modify the idea to use the most frequent colour reflection to track objects.

Btw, they also have a project with a blended telerecording on the backburner, I've finally realized the solution. (I know of 3 now, but the last one should have great quality)

[alexh110]

Here's a couple of test frames.
You may need to increase saturation to get the best idea of the quadrant errors in the image.

More to follow...

[alexh110]

Two more examples.