Experiment with feature-biased color selection

[gennyble]

Preserving edges would be nice, so here's the idea?

• optionally downscale the image some. 1/2 or 1/4, maybe, to save compute and memory.
• edge deteciton (wikipedia: Canny Edge Detection)
• optionally slightly blur the edge-image so we can get some colours around the edge? experiment here.
• consider colours only on the edge. if we blurred then perhaps we can use the grey value to further bias our selection. we could compute some kind of score where we sum the gray values where the colour is (normalized 0-1).

i don't know how this'll work it's just an idea.

[gennyble]

further, we thought about this while initially developing colorsquash.

some kind of area-bias. influenced by the user almost certainly. us trying to determine the subject of the image is out of scope, but the center might be a good default. perhaps the central third of the image. we'd allocate a specific number of colours, configurable i imagine, to the "subject area". perhaps multiple area? so we can be sure we get necessary colours.

[gennyble]

On that last one, the area-biased selection, this can be a CLI thing. Originally, we didn't think there was a way to make this clean, but there's been some thought on it.

Something like bias-region=center-center or center-left where the first argument is vertical, second horizontal. So it can be top-left, center-right, bottom-center, etc. Instead of the image being divided into thirds, it should be smarter. Taking just the center-left third seems strange. Perhaps we make it half and the arguments just shift the box around?

There can also be like bias-size=50% or something. perhaps 0.5 to mean 50%? To shrink or grow the bias area (be sure not to overflow the bounds of the image.