Image colour quantization using competitive learning

Abstract

The display of full colour images on devices with limited colour capabilities requires the mapping of the true colour information of an image on to a restricted colour palette. This palette, or colour table, generally consists of a limited number of elements which must be used to represent all colours within an image. Selecting an appropriate subset of colours which accurately represent the distribution of colour in the original image is not a trivial task. We explore the use of competitive learning for the selection of an image's colour table. Our results demonstrate that frequency sensitive competitive learning is capable of selecting an appropriate colour table for a given image. Also, when compared with the colour tables produced by traditional techniques, it was found that the competitive learning approach produced tables with improved performance in terms of reduced overall quantization error. The results reported examine colour tables of various sizes from modest tables of 256 entries down to highly restricted tables of 8 colours.