Comparison of conventional and bisecting k-means algorithms on color quantization

Abstract

Color quantization represents an important technique with both many applications, as well as outstanding challenges in the field of image processing. Numerous clustering techniques have been applied to the task of achieving the optimal quantization with varying results. In this paper, we compare the results of the conventional k-means algorithm, indisputably the most celebrated nonhierarchical clustering algorithm, to those of a recent hierarchical variant of it named bisecting k-means. We show that, while bisecting k-means often produces slightly higher quantization distortion than k-means, the former more than makes up for this in stability and computational efficiency.