Date of Award

1990

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

First Advisor

Charles Harlow

Abstract

This dissertation proposes a new approach for analyzing high resolution aerial photographs of urban areas. Analyzing aerial photographs is the process of constructing an overall description of a scene. It involves knowledge of visual sensors, computing systems, artificial intelligence, software engineering, and perceptual psychology. Researchers have had only limited success in this area. This dissertation considers a high level analysis approach. Most aerial photograph interpretation systems concentrate on analyzing an airport, roadway, or urban scene. Those systems, however, do not explain how they know they were examining such a scene. This dissertation concentrates on how to reach that point. It begins with "this is an aerial photograph" and works its way down through a hierarchy of labels until it reaches the point of "this is an urban area--find and label the objects." The new analysis approach introduces a unique use of three basic ideas. These ideas are (1) the use of context, expectations, selective attention, and the perceptual cycle, (2) analyzing the image through a hierarchy of increasingly specific labels, and (3) the interplay between the segmentation and interpretation processes. These are developed in a computer vision system for analyzing aerial photographs. The system comprises (1) a control mechanism, (2) a knowledge base, (3) a belief maintenance system, and (4) an image processing interface. In general, the system uses the knowledge stored in frames to investigate areas in the image. The control mechanism calls low level routines in the image processing interface. They report the results back to the control mechanism which invokes the belief maintenance system. The belief maintenance system reports which frame is the most probable label for the area under investigation. To demonstrate the system, this dissertation presents the results of analyzing a high resolution, multi-spectral, aerial image of an urban area. It also presents the results of analyzing three different housing areas taken from a single channel, gray scale image of a color aerial photograph. These show the validity of the new approach and the power and portability of the system.

Pages

217

Share

COinS