Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.771181
Title: Selective image segmentation models and fast multigrid methods
Author: Roberts, M. T.
Awarding Body: University of Liverpool
Current Institution: University of Liverpool
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
This thesis is concerned with developing robust and accurate variational selective image segmentation models along with fast multigrid methods to solve non-linear partial differential equations (PDEs). The first two major contributions are the development of new distance terms and new intensity fitting terms for selective image segmentation models. These give state-of-the-art segmentation results, with high robustness to the main parameters and to the user input. Therefore, these models are highly applicable to real-world applications such as segmenting single organs from medical scans. The final major contribution is to develop new novel non-standard smoothers for the non-linear full approximation scheme multigrid framework. Multigrid is an optimal O(N) iterative scheme when it converges. However, typically if we directly apply a multigrid solver to a non-linear problem, it will not converge. This is principally due to the ineffectiveness of the standard smoothing schemes such as Jacobi or Gauss-Seidel. We review the true reason that these smoothers are ineffective using local Fourier analysis and develop a smoother which is guaranteed to be effective. Experiments show that the smoother is effective and the algorithm converges as desired. These new non-standard smoothing schemes can be used to solve a whole class of non-linear PDEs quickly. This work also lays the groundwork in the development of a "black-box" non-linear multigrid solver which doesn't require the degree of tuning that current multigrid algorithms do.
Supervisor: Chen, Ke ; Irion, Klaus L. Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.771181  DOI:
Share: