Deterministic vs. Random Initializations for K-Means Color Image Quantization

We present six methods for initialising the K-means clustering algorithm used for color image quantization. We test these initialization methods on a few quantization levels and on 24 color images contained in the Kodak image dataset. In the vast majority of the examined cases the best results were obtained for the initialization of KM++. The evaluation of the results was carried out using the MSE and several new perceptual quality indices.