Using machine learning to colourise grayscale images, is it being done? Options

[Habukaz]

I am just wondering whether machine learning is being used (by people here or elsewhere) to colourise space images that lack colour data (like MER Navcam images, LORRI images and so on).

I have a few ideas on how to do it; but if it is already routinely being done (and with good results), it wouldn't be as cool to try to implement them.

[scalbers]

I found this via an online search, though it looks to be just for Earth ground-based images: http://cs231n.stanford.edu/reports2016/219_Report.pdf

[Habukaz]

Thanks for the link (here's another). I'd expect colourising images from Earth to be more tricky than many celestial targets (including Mars) due to the greater number of common colours as well as types of objects, at least in many common settings.

[algorithm]

Never mind machine learning, how about human learning!
I've been trying to figure out how it's done in Photoshop for ages now. Maybe PS isn't up to the job!

[JRehling]

Funny, I have been thinking recently on a related problem, to add texture to images of a given resolution after upscaling them so that resolution could be increased.

The basic truth, though, is: You can't get something for nothing. A BW photo of Uranus, Titan, and Venus might be exactly the same. You can color them by knowing the right answer, but it's absolutely impossible to extract color from the image itself. And the same principle applies on all levels.

I have colored a BW image that I took of Mars – that's quite predictable, on a global scale, but only because I know what the colors of Mars are.

If color is easily predictable from brightness alone, then this can be done. If not, it can't. And this will have to be world by world, nebula by nebula, etc.

I'd say that Mars and Europa, to choose a pair, make this quite possible at global resolution. Jupiter, for example, does not (the GRS is the same brightness as many non-red portions of it).

[Gerald]

If you have fractal patterns or some kind of implicite rule, you may be able to fill in a plausible solution. But when interpreting such images, you take the risk to interprete plausible artifacts. For any confidence estimates, you need to multiply the probability of your solution being correct for the ill-posed problem with the probability, that your interpretation is correct under the condition that your solution is correct.

[scalbers]

Extrapolating a Fourier spectrum can help fill in more resolution. An algorithm sort of along these lines used on Earth clouds is here:

https://www.google.com/url?sa=t&rct=j&a...147448319,d.cGw