Algorithm for color quantization/reduced image col

问题:

I'm writing a web app that takes a user-submitted image, gets the pixel data via a canvas element, does some processing, and then renders the image using vector shapes (using Protovis). It's working well, but I end up with several thousand colors, and I'd like to let the user pick a target palette size and reduce the color palette to that size.

At the point where I want to reduce the color space, I'm working with an array of RGB pixel data, like this:

[[190,197,190], [202,204,200], [207,214,210], [211,214,211], [205,207,207], ...]

I tried the naive option of just removing least-significant bits from the colors, but the results were pretty bad. I've done some research on color quantization algorithms, but have yet to find a clear description of how to implement one. I could probably work out a cludgy way to send this to the server, run it though an image processing program, and send the resulting palette back, but I'd prefer to do it in JavaScript on the client side.

Does anyone have an example of a clearly explained algorithm that would work here? The goal is to reduce a palette of several thousand colors to a smaller palette optimized for this specific image.

Edit (7/25/11): I took @Pointy's suggestion and implemented (most of) Leptonica's MMCQ (modified median cut quantization) in JavaScript. If you're interested, you can see the code here.

Edit (8/5/11): The clusterfck library looks like another great option for this (though I think it's a bit slower than my implementation).

回答1:

With the caveat that I don't claim any expertise at all in any field of image processing: I read over the Wikipedia article you linked, and from there found Dan Bloomberg's Leptonica. From there you can download the sources for the algorithms discussed and explained.

The source code is in C, which hopefully is close enough to JavaScript (at least in the core "formula" parts) to be understandable. The basic ideas behind the "MMCQ" algorithm don't seem super-complicated. It's really just some heuristic tricks for splitting up the 3-dimensional color space into sub-cubes based on the way colors in an image clump together.

回答2:

I wrote a web app that extracts a color palette from an image. It allows you to load an image, then play around with three different algorithms/approaches for extracting a color palette from it:

1. Simple histogramming
2. Median Cut
3. k-means

You can find a copy of it running here

You can find the code for it on github

It's written 100% in Javascript, and uses Plotly.js for the example plots.

I also wrote a post describing the three approaches/algorithms in more detail - you can find that here