BibTeX entry

@PHDTHESIS{200109Mike_Rogers,
  AUTHOR={Mike Rogers},
  TITLE={Exploiting Weak Constraints on Object Structure and
    Appearance for Segmentation of 2-D Images},
  SCHOOL={University of Manchester},
  MONTH=Sep,
  YEAR=2001,
  URL={http://www.bmva.org/theses/2001/2001-rogers.pdf},
}

Abstract

Prior knowledge can be used to constrain complex image interpretation tasks. When objects are highly variable, the constraints derived from the prior knowledge will be weak. This thesis investigates how best to maximise and utilise weak constraints, using as an example the location of boundaries (segmentation) in nerve capillary images. Capillaries imaged by electron microscopy show a complex textured appearance, making segmentation difficult. Considerable variation occurs among boundaries manually positioned by human experts. Previous attempts to detect boundaries using a data-driven approach have proved unsuccessful due to the existence of confusing image evidence in the vicinity of these boundaries. Point Distribution Models (PDMs) are found to be the most appropriate modelling method, despite the fact that the capillaries have no identifying landmarks. An extension to PDMs called Structured PDMs is proposed, which is designed to represent the intermittently present lumen structure. Including this feature results in a more specific model. A comparison of texture analysis schemes is carried out. Results show that wavelet texture analysis gives the best classification of capillary structures. Finally, a quantitative evaluation shows that model driven methods improve on the accuracy and robustness of data-driven segmentation performance. More specifically, SPDMs are shown to significantly improve segmentation results compared to PDMs. Genetic search provides a promising method of fully automatic segmentation.