Skip to Main Content
In this paper, two unsupervised color image segmentation methods based on color clustering are explored: k-means (KM) and mixture of principal components (MPC). KM and MPC use respectively the Euclidean distance and the vector angle as color similarly measures. It is shown that the vector angle is an intensity-invariant measure in RGB based on the dichromatic reflectance model. Results are given for various color spaces: RGB, XYZ, rgb (normalized RGB), CIELAB, CIELUV, h1h2h3 (a new space), and l1l2l3. Quantitative and qualitative results show the effectiveness of the MPC algorithm on the RGB, rgb, and XYZ color spaces whereas the KM combination seems most effective in the CIELAB, h1h2h3, and l1l2l3 color spaces. Finally, poor color clustering results with MPC in h1h2h3 and with KM in rgb suggest that some assumptions in deriving a simplified version of Shafer's model for matte surfaces might have been violated.