Analysis of merge criteria within a watershed based segmentation algorithm

The watershed transform is a very powerful segmentation tool which guarantees closed contours. In this paper the watershed transform is used for the segmentation of a very simple image consisting of a circle, a rectangle and a background region. The ability of different merge criteria to find these major structures based on the highly over-segmented watershed transform for different signal to noise ratios (SNR) is analyzed. Special focus is given to the compensation of prior merge probabilities induced by the topology of the over-segmented watershed images. Herby a relative performance increase of 5.1% to 23.5% is achieved for the different merge criteria.