An integrated similarity metric for graph-based color image segmentation

摘要

Graph-based method has become one of the major trends in image segmentation. In this paper, we focus on how to build the affinity matrix which is one of the key issues in graph-based color image segmentation. Four different metrics are integrated in order to build an effective affinity matrix for segmentation. First, the quaternion-based color distance is utilized to measure color differences between color pixels and the oversegmented regions (superpixels), which is more accurate than the commonly used Euclidean distance. In order to describe the superpixels well, especially for texture images, we combine the mean and the variance information to represent the superpixels. Then the image boundary information is used to merge the oversegmented regions to preserve the image edge and reduce the computational complexity. An object for recognition may be cut into nonadjacent sub-parts by clutter or shadows, the affinities between adjacent and nonadjacent superpixels are computed in our study. This feature of affinity is not considered in other methods which only consider the similarity of adjacent regions. Experimental results on the Berkeley segmentation dataset (BSDS) and Weizmann segmentation evaluation datasets demonstrate the superiority of the proposed approach compared with some existing popular image segmentation methods.