New approach for initialization of K-means technique applied to color quantization

Color quantization of images is still an important auxiliary operation in the rapidly developing field of color image processing. Color quantization methods include fast divisive techniques, e.g. median-cut (MC), and slower adapted clustering techniques e.g. the most popular K-means (KM) technique. The results obtained by KM strongly depend on a method of initialization, i.e. the method of determining the initial cluster center. The classic version of the KM uses a random selection of the initial centers. The aim of research is to find a fast initialization method that leads to high performance clustering and does not allow for formation of empty clusters. The new approach involves using one of splitting methods, e.g. MC or Wu's algorithms as a method for the KM initialization. In the paper this approach is successfully compared with other methods and tested for different numbers of clusters k.