Color quantization for image processing using self information

A digital picture generally contains tens of thousands of colors. Therefore, most image processing applications first need to apply a color reduction scheme before performing further sophisticated analysis operations such as segmentation. While a lot of color reduction techniques exist in the literature, they are mainly designed for image compression and are unfortunately not suited for many image processing operations (e.g. segmentation) as they tend to alter image color structure and distribution. In this paper, we propose a new color reduction scheme (SICR), using probabilities and information theory elements to balance between the information provided by the selected colors and the necessity to accurately represent the selected colors. We also advocate for the use of perceptually accurate metrics for evaluation. Experimental results on a diversified dataset of images selected from the internet show that our technique performs well compared to other color reduction schemes.