A Comparative Study of the Use of a Robust Color Image Segmentation Method

Abstract

In this paper, a comparative study of some basic close related color image segmentation methods is presented. It is focused in the evaluation of two segmentation methods based on a recently published adaptive color similarity function making use of: 1) pixel samples of both figure and background and classifying by maximum similarity, and 2) pixel samples of only figure and classifying by automatic thresholding thus employing only half of input information. It is also presented for comparison, the results of classification using the Euclidean metric of a* and b* channels rejecting L* in the L*a*b* color space and with the Euclidian metric of the R, G, and B channels in the RGB color space. From the results it was obtained that the segmentation technique using the adaptive color similarity function and classifying by automatic thresholding (employing only half of the information supplied to the other methods) had better performance than those implemented in the L*a*b* and RGB color spaces in all cases of study. The performance is equivalent to that using pixel sample of both figure and background and classifying by maximum similarity. The improvement in quality of the segmentation techniques using the color similarity function is substantially significant.