Auto-pick colours from image for use in the palette

[Shrinks99]

Would be cool to auto-choose x number of colours from the input image using colour quantization as defined by a single integer set as the --palette value. For example --palette 5 would choose the 5 best colours from the source image and dither accordingly.

Additionally, don't limit users to a minimum of 2 values! By allowing people to select a single average colour the application now has a bonus feature of spitting out average coloured images. It's not a bug, it's a feature!

[makew0rld]

Other ways of doing this:

--palette mmcq --palette num 4

--palette mmcq:5

These methods allow the color quantization algorithm to be chosen, in case more are added in the future.

https://github.com/joshdk/quantize can be used for this.

[makew0rld]

I've now got this working on the quantize branch! But I want to investigate whether implementing the same algorithm but with linearizing the image first would work better. I think it would, so I'll have to create my own library for that and test that before merging anything.

There also seem to be options available to how I implement this. Some (most?) implementations split every bucket every time, resulting in power-of-two palettes. Some implementations choose a single bucket to split at each iteration (the one with the largest range in any channel), resulting in arbitrary palette sizes. The latter sounds like it would give better results but I should test both if possible.

• https://github.com/joshdk/quantize/blob/65999d3a4c76/mmcq.go
• https://en.wikipedia.org/wiki/Median_cut
• http://leptonica.org/papers/mediancut.pdf
• https://rosettacode.org/wiki/Color_quantization#Go
• https://muthu.co/reducing-the-number-of-colors-of-an-image-using-median-cut-algorithm/
• https://joelcarlson.github.io/2016/01/15/median-cut/

[makew0rld]

I made my own lib that does median cut in linearized sRGB. But it performs worse! Claude explained why: https://claude.ai/share/e8473098-8669-4fd2-9e11-eefde6658157

Essentially, this is a perceptual task (pick the top colours according to humans) and gamma compression helps match human perception. This was obviously visible with shadows and highlights. My linear code would keep highlight detail but blow out shadows, while the original library did the reverse. Overall the latter was much better.

The results with the original lib (joshdk/quantize) are not always as good as you might expect, but it's not worth worrying about that right now.

[makew0rld]

Added in 870ed78. Please test!

[makew0rld]

Still testing with other color spaces... I think this can be better. Also it shouldn't be called mmcq.

[makew0rld]

Hoping for a response on this: https://computergraphics.stackexchange.com/questions/14530/how-to-automatically-pick-palette-colours-for-dithering