Abstract

A tool that quickly calculates the dominant colors of an image can be very useful in image processing. The k-means clustering algorithm has this potential since it partitions a set of data into n clusters and returns a representative data point from each cluster. We discuss k-means with sampling for images, which applies k-means clustering to a random sample of image pixels. We found that even with a small random sample of pixels from the image, k-means with sampling exhibits no significant loss of correctness. We examine the usefulness and limitations of k-means clustering in determining the prominent colors of an image and identifying trends in large sets of image data.