Outlier-based initialisation of K-means in colour image quantisation

This paper deals with problems of initialisation of K-means technique (KM) in colour image quantisation. In classic version the KM starts with randomly selected centroids. Authors are more interested in the deterministic initialisations based on the distribution of image pixels in the colour space. Besides initialisations that were proposed earlier (DC and SD), here is considered a new outlier-based initialisation. It is based on the modified Mirkin's algorithm (MM) and puts cluster centroids in peripheral colours of pixels cloud. Such approach taking into account small peripheral clusters allows to obtain a quantised image with perceptually important regions. Tested images were evaluated by means of subjective visual assessment average colour and additionally the loss of colourfulness (ΔM). Pixel clustering was created in the RGB YCbCr and CIELAB colour spaces.