Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)

First Advisor

Charles A. Harlow


The goal of this research is the recovery of elongated shapes from patterns of local features extracted from images. A generic geometric model-based approach is developed based on general concepts of 2-d form and structure. This is an intermediate-level analysis that is computed from groupings and decompositions of related low-level features. Axial representations are used to describe the shapes of image objects having the property of elongatedness. Curve-fitting is shown to compute axial sequences of the points in an elongated cluster. Script-clustering is performed about a parametric smooth curve to extract elongated partitions of the data incorporating constraints of point connectivity, curve alignment, and strip boundedness. A thresholded version of the Gabriel Graph (GG) is shown to offer most of the information needed from the Minimum Spanning Tree (MST) and Delauney Triangulation (DT), while being easily computable from finite neighborhood operations. An iterative curve-fitting method, that is placed in the general framework of Random Sample Consensus (RANSAC) model-fitting, is used to extract maximal partitions. The method is developed for general parametric curve-fitting over discrete point patterns. A complete structural analysis is presented for the recovery of elongated regions from multispectral classification. A region analysis is shown to be superior to an edge-based analysis in the early stages of recognition. First, the curve-fitting method is used to recover the linear components of complex object regions. The rough locations to start and end a region delineation are then detected by decomposing extracted linear shape clusters with a circular operator. Experimental results are shown for a variety of images, with the main result being an analysis of a high-resolution aerial image of a suburban road network. Analyses of printed circuit board patterns and a LANDSAT river image are also given. The generality of the curve-fitting approach is shown by these results and by its possible applications to other described image analysis problems.