Skip to main content
SLU publication database (SLUpub)
Doctoral thesis, 2015

Image segmentation using snakes and stochastic watershed : with applications to microscopy images of biological tissue

Selig, Bettina

Abstract

The purpose of computerized image analysis is to extract meaningful information from digital images. To be able to find interesting regions or objects in the image, first, the image needs to be segmented. This thesis concentrates on two concepts that are used for image segmentation: the snake and the stochastic watershed. First, we focus on snakes, which are described by contours moving around on the image to find boundaries of objects. Snakes usually fail when concentric contours with similar appearance are supposed to be found successively, because it is impossible for the snake to push off one boundary and settle at the next. This thesis proposes the two-stage snake to overcome this problem. The two-stage snake introduces an intermediate snake that moves away from the influence region of the first boundary, to be able to be attracted by the second boundary. The two-stage snake approach is illustrated on fluorescence microscopy images of compression wood cross-sections for which previously no automated method existed. Further, we discuss and evolve the idea of stochastic watershed, originally a Monte Carlo approach to determine the most salient contours in the image. This approach has room for improvement concerning runtime and suppression of falsely enhanced boundaries. In this thesis, we propose the exact evaluation of the stochastic watershed (ESW) and the robust stochastic watershed (RSW), which address these two issues separately. With the ESW, we can determine the result without any Monte Carlo simulations, but instead using graph theory. Our algorithm is two orders of magnitude faster than the original approach. The RSW uses noise to disrupt weak boundaries that are consistently found in larger areas. It therefore improves the results for problems where objects differ in size. To benefit from the advantages of both new methods, we merged them in the fast robust stochastic watershed (FRSW). This FRSW uses a few realizations of the ESW, adding noise as in the RSW. Finally, we illustrate the RSW and the FRSW to segment in vivo confocal microscopy images of corneal endothelium. Our methods outperform the automatic segmentation algorithm in the commercial software NAVIS.

Keywords

image segmentation; snakes; active contours; stochastic watershed; minimal spanning tree; corneal endothelium; compression wood

Published in

Acta Universitatis Agriculturae Sueciae
2015, number: 2015:16
ISBN: 978-91-576-8230-7, eISBN: 978-91-576-8231-4
Publisher: Centrum för bildanalys, Sveriges lantbruksuniversitet

      SLU Authors

    • Selig, Bettina

      • Centre for Image Analysis, Swedish University of Agricultural Sciences

    UKÄ Subject classification

    Computer Science
    Computer Vision and Robotics (Autonomous Systems)
    Wood Science

    Permanent link to this page (URI)

    https://res.slu.se/id/publ/64874