Segmentation of vector images by N-level-set-fitting

Abstract

In many applications of segmentation algorithms the number of desired segments is known previously. We present a technique to segment a given vector image (in most cases consisting of three color channels) in a prior known number of segments consisting of connected pixel sets. The main idea is to minimize the Euclidean distance of a vector valued step function to the image, with the step function being constant on a segment. A local minimum of this optimization problem can be obtained by a simple merging algorithm, which starts with a segmentation of the image into a much greater number of segments. The starting segmentation can be computed by using well known histogram based thresholding algorithms.