Combining Color, Texture and Contour Cues for Image Segmentation

We present a method for measuring similarity among any two image pixels, based on the distribution of image chromaticities near each pixel. This measure is designed for algorithms that use pairwise pixel similarity measures to segment images. We describe a way to combine the proposed color-based measure with previously developed grayscale similarity measures. We show that adding the colorbased similarity measure to segmentation algorithms that only rely on gray-scale information improves their performance in breaking up color images. We demonstrate performance improvement due to the addition of a color cue by comparing segmentations produced by two versions of the Normalized Cuts algorithm: one using only gray-scale similarity measures (contour and texture) and a second version that adds in the proposed color similarity measure. We quantify the performance of either algorithm by measuring its consistency with human segmentations.