I've seen two university-type projects for colorizing.
One program was so crude, it required the operator to pick "key points"
and select colors for the key points. Which, as it turns out, is a lot
of work.
The second (also using neural networks), could assign colors itself,
based on training on a large number of photos.
The normal test for these tools is "feathered parrot in rain forest",
and it will select purple or green for the parrot, it gets the green
leaves mostly correct.
But both of them suffer from the same problem. Namely, they cannot
"paint within the lines". There tends to be a bit of "overspray
around the edges" of objects.
This makes the results tantalizingly close, but "not right". It's
just not production grade workmanship. And if you had to fix
the issues by hand, why... you might as well colorize it by hand.
sudo apt install python3-pip
pip install autocolorize
sudo apt install caffe-cpu # not in Ubuntu 2204, but still in Ubuntu 2004.
# The caffe-cuda appears to be out of support,
# which is not a deal-breaker unless you are
# training neural-network designs. This is why it
# takes four minutes - it's CPU based.
/home/bullwinkle/.local/bin/autocolorize -d cpu BW.jpg -o CLR.jpg
"Downloading weights file" [maybe 500MB worth][700KB/sec][cached for next time]
It doesn't use multiple CPU cores, which is... odd.
Start time 11:40AM. RAM usage 17.7GB so far (on a 64GB machine). Finish at 11:44AM.
Considering the amount of time this is taking, she'll have green lipstick.
And that was an amazingly prescient guess on my part,
as the program did NOT colorize the image and just
turned the entire image a tint of green. Bad monkey! Bad!
[Picture]
https://i.postimg.cc/rpKhmQCq/test-and-fix-attempt.jpg
The left image is the original.
The middle image is the colorization output.
The right hand image, is rotating the color dials in GIMP
and "doing the best you can" sorta.
Oh well.
Paul