On Color Image Quantization by the K-Means Algorithm

In this paper we show the main properties of k-means algorithm as a tool for color image quantization. All experiments have been carried out on color images with different number of unique colors and different colorfulness. We have tested the influence of methods of determination of initial cluster centers, of choice of distance metric, of choice of color space. In our tests we have used two dimensions of palette (256,16) and three different measures for quantization errors. The results of k-means technique have been compared with quantized im- ages from commercial programs.