Dissertation/Thesis Abstract

Variational image segmentation based on pixel pairwise similarities
by Bertelli, Luca, Ph.D., University of California, Santa Barbara, 2009, 192; 3350379
Abstract (Summary)

The main goal of this thesis is to develop robust computational methods to address some of the open problems arising in the field of variational image segmentation. In particular, we focus our attention on a specific sub-class of variational methods, proposing several novel variational frameworks based on pixel pairwise (dis)similarities. The starting point for most of the main contributions can be considered Graph Partitioning Active Contours, the framework for binary (i.e. foreground/background) segmentation based on pixel similarities introduced in [122]. The motivation for the work of this thesis stems from the fact that a general binary segmentation framework is not adequate in the case of natural images, which usually require segmentation into multiple regions. Furthermore, in the presence of occlusion or high levels of noise, prior information about objects of interest or domain knowledge would be needed for a robust segmentation. This research addresses these issues in two main directions.

First, we introduce novel variational frameworks based on pairwise pixel similarities for the segmentation in multiple regions. This can be considered a principled solution to the multi-region segmentation problem, that in previous work on pairwise similarity based cost functions have been solved mainly by recursive bi-partition. In addition, we explicitly address the problem of multiphase curve regularization. In fact, when extending curve evolution frameworks to segment images in multiple regions, traditional regularization techniques, aimed at increasing robustness to noise and artifacts, cease to be adequate. We therefore design novel length and area regularization terms, whose minimization yields evolution equations more suitable to eliminate spurious regions and other kind of noisy artifacts. To the best of our knowledge, this is the first attempt in this direction.

Secondly, we address the problem of introducing prior knowledge within the segmentation framework. Prior information in the form of multiple views of the same object/scene is incorporated by reformulating the cost function in [122] in such a way that pixel pairwise dissimilarities are computed across different views, granting robustness to occlusions and high level of noise. Minimization of these cost functions is carried out, after imposing a warping constraint between the views. By doing this, the need of introducing a specific shape term in the cost function is avoided and, at the same time, the shape prior is exploited in a more complete way, taking into account also intensity or color information of the emerging regions. We also introduce non-rigid registration models (based on thin plate splines) within the level set framework to cope with non-rigid deformation of the object shape, which, to the best of our knowledge, is the first attempt in this direction. We finally demonstrate that this model is effective in segmenting bio-medical images, in which one of the views is represented by a reference image (or atlas) containing information about the structures of interest.

Indexing (document details)
Advisor: Manjunath, B. S.
Commitee: Chandrasekaran, Shivkumar, Gibou, Frederic, Rose, Kenneth
School: University of California, Santa Barbara
Department: Electrical & Computer Engineering
School Location: United States -- California
Source: DAI-B 70/03, Dissertation Abstracts International
Subjects: Electrical engineering
Keywords: Image segmentation, Level set methods, Multiphase curve evolution, Pixel pairwise similarities, Variational segmentation
Publication Number: 3350379
ISBN: 9781109082357
Copyright © 2019 ProQuest LLC. All rights reserved. Terms and Conditions Privacy Policy Cookie Policy