Image Color Extractor

Pulling the main colors out of a photograph is a clustering problem rather than a lookup. A modest image contains hundreds of thousands of pixels and often tens of thousands of distinct color values, almost none of which repeat exactly. There's no single color sitting in the file waiting to be read. So the colors you see here are computed by grouping similar pixels together and choosing a representative for each group.

Why averaging doesn't work

The obvious approach, averaging every pixel, produces a muddy result almost every time. Average a photograph of red flowers against green foliage and the answer is a brown that appears nowhere in the picture. Averaging finds the center of the colors present, and the center of two opposing hues is a neutral. This is why proper extraction partitions the colors first and only averages within each partition.

The usual algorithms

Median cut works by treating all the pixels as a box in color space, repeatedly splitting the box along its longest axis at the median, and stopping when it has as many boxes as colors requested. It's fast and predictable. Octree quantization builds a tree that subdivides color space and merges the least populated branches, which keeps memory bounded. Clustering methods such as k-means iterate toward the groupings that minimize total distance, which often looks best but costs the most to compute.

Why the result sometimes ignores the obvious subject

Extraction counts pixels, and pixel count isn't the same as visual importance. A photograph of a small bright bird against a large expanse of sky yields sky, because the sky occupies most of the frame. A logo on a white background yields white for the same reason. Tools compensate by discarding near-white and near-black pixels, by weighting saturated colors more heavily, or by sampling the center of the frame more than the edges, all of which are heuristics rather than a correct answer.

Transparency is a related trap. Fully transparent pixels still carry color values in the file, and an extractor that doesn't check the alpha channel will happily average in color data from regions that are invisible, which is how a logo with a transparent background sometimes produces a palette that seems unrelated to anything on screen.