Hierarchical Segmentation and Clustering Using Similarity Analysis

189 p.Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 1997.This dissertation addresses the problem of unsupervised low-level classification of point patterns. Low-level classification is understood as the problem of unsupervised detection of structure in multidimensional images and point patterns. Images and point patterns consist of multiple features or measurements of physical phenomena. The features are either measurements within a set of regularly distributed samples (e.g., brightness of digitized photographs) or independent measurements of phenomena within any set of irregularly distributed samples (e.g., height and weight of people in the USA). These two types of features differ according to the regular or irregular samples at which features are obtained. A set of features is viewed as a point in high-dimensional space. Repetitive measurements of the same set of features give rise to multiple points denoted as a point pattern. The resulting structure is in a form of image regions (a segmentation problem) or pattern clusters (a clustering problem). Thus a typical low-level classification problem breaks down to segmentation and clustering problems. Features from one structure are similar to each other and dissimilar to all surrounding features. The similarity and dissimilarity of features is modeled on local or global properties of structures. Classification based on local feature properties leads to the problems of nontexture image segmentation and clustering of points, while global feature properties give rise to problems of texture image segmentation and clustering of point aggregations. Images and point patterns contain information about phenomena that is present at various levels of detail. This dissertation addresses the problem of information detail by detecting structures at multiple levels, called multiscale classification. Multiple results are then organized into a tree or hierarchy of detected structure. The first part of the dissertation presents a new framework for segmentation of nontexture and texture images. The hierarchical segmentation methods partition a regular multidimensional grid of sample points based on similarities of features at the sample points. The second part of this work presents a new clustering approach applied to point patterns. A point pattern is partitioned into clusters based on local or global similarity of features.U of I OnlyRestricted to the U of I community idenfinitely during batch ingest of legacy ETD