A comparison of color coordinate systems using color classification and image segmentation results of humans

Abstract

Here we present a comparison of color coordinate systems using two types of image segmentation data created by human subjects. The first type was created based on color classification of pixels using color information alone as perceived by human subjects, while the second type was created by humans directly segmenting images. We define four criteria for comparative purposes: uniformity of the color distribution; color concentration; color separability; and local color distribution in regions of images. Copyright © 2004 Wiley Periodicals, Inc.Our results show that the Laaab, Lauav, and Joblove models closely match human perception of color from the viewpoint of color classification and region segmentation. However, it is also clear that in areas of low brightness these models' distributions of color are spread inappropriately wide. On the other hand, the RGB color coordinate system, and models based on linear transformations such as the opponent color axes, YES and YIQ models, differ rather significantly from human perception of color. The recently proposed HSV models 1 and 2 are seen to be close to human perceptions of color with no areas in which color uniformity is poor. On average these models received high evaluations. This research has shown that there does not exist one single color coordinate system that is appropriate for all image processing tasks and that the most suitable color coordinate system will be dependent on the particular processing task at hand. © 2005 Wiley Periodicals, Inc. Syst Comp Jpn, 36(10): 93–105, 2005; Published online in Wiley InterScience (). DOI 10.1002sscj.10095