[IPAG Home Page] [Publications] [Dissertation Archive] [Personnel]

IPAG Dissertation Archive

Feature and Module Integration for Image Segmentation

A Dissertation

Presented to the Faculty of the Graduate School


Yale University

in Candidacy for the Degree of

Doctor of Philosophy


Amit Chakraborty

Dissertation Director: James Scott Duncan

May 1996


A systematic approach towards the problem of designing integrated methods for image segmentation has been developed in this thesis. This is aimed towards the analysis of underlying structures in an image which is crucial for a variety of image analysis and computer vision applications. However, a robust identification and measurement of such structure is not always achievable by using a single technique that depends on a single image feature. Thus, it is necessary to make use of various image features, such as gradients, curvatures, homogeneity of intensity values, textures, etc. as well as model-based information (such as shape). Integration provides a way to make use of the rich information provided by the various information sources, whereby consistent information from the different sources are reinforced while noise and errors are attenuated. As a first step, integration is achieved in this work by using region information in addition to gradient information within the deformable boundary finding framework. This considerably increases the robustness of the final boundary output to noise and initial estimate. This feature integration paradigm for deformable boundary finding is then further developed through the addition of curvature information which makes the boundary solution better localized. Next, a more general integration framework is considered, whereby computational modules are associated with the boundary and region processes which are simultaneously updated. This is achieved through the use of a new game-theoretic procedure where the modularity of the underlying objectives are retained. The integration problem is framed as a family of coupled and coexisting objectives using a Bayesian strategy whereby the output of one module depends upon the previous outputs of the other modules. This mode of information sharing, where only the final decisions of the different modules (decision makers) are broadcast to the other decision makers, is not only technically more general than other single objective function approaches, but is also computationally less burdensome especially in cases like the present one where incommensurate objectives are involved. This further improves both the region and boundary estimates.

BibTeX Entry

author =  "Amit Chakraborty",
title =   "Feature and Module Integration for Image Segmentation",
school =  "Yale University",
month =   "May",
year =    "1996")

The complete text of the thesis is available as a .pdf file. (228 pages, 2.9 MB)

Questions regarding the content of this page should be directed to the group director.

Questions or bug reports regarding this interface should go to webmaster@noodle.med.yale.edu.

Home Page, webmaster@noodle.med.yale.edu