EM/MPM-based segmentation techniques with improved boundary accuracy

Joel M Dumke, Purdue University

Abstract

The segmentation problem essentially involves partitioning the domain of an image into homogeneous regions. This is a significant problem in areas ranging from computer vision to medical imaging to materials science. The Expectation Maximization/Maximization of Posterior Marginals (EM/MPM) algorithm has been applied to this problem with some success. This dissertation explores modifications to the algorithm which are primarily intended to improve the accuracy of the locations of boundaries in the resulting segmentations. The first modification presented is an adaptive version of the algorithm which focuses on texture-based segmentation. To extract texture features, the input images are filtered at various window sizes and the modified algorithm progresses from the largest windows sizes, which have the advantage of high selectivity, to the smallest window sizes, which allow for the greatest boundary accuracy. Another modification involves developing a new probabilistic image model which is inspired by parametric active contours. A novel application of morphological filtering techniques allows us to influence boundary curvature. This modification allows us to gain some of the benefits of active contours while retaining the benefits of the EM/MPM algorithm. Finally, this dissertation presents a modification based on selective application of Maximum Likelihood (ML) estimators to the results of the EM/MPM algorithm. Since the inaccuracies in boundary locations are sometimes the result of prior models that are more suited to interior regions, the ML estimators allow more accurate boundary locations by focusing only on the image data.

Degree

Ph.D.

Advisors

Comer, Purdue University.

Subject Area

Electrical engineering

COinS