Doctoral Dissertations

Date of Award

12-1993

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Major

Electrical Engineering

Major Professor

Mohan M. Trivedi

Committee Members

Don Bouldin, Marshall Pace, Michael Thomason

Abstract

The analysis of reconnaissance images acquired from satellites or aircraft is one of the important application areas of computer vision. In this dissertation, the focus is on the detection of objects in high-resolution aerial images acquired with thermal infrared sensors. The images are characterized by; small target signature, high background clutter, continuous variations in background, large amounts of image data, and high processing speed requirements. As a result, present detection techniques suffer from the following drawbacks: (i) Low processing speed due to computationally intensive algorithms, (ii) Detection results which could be improved, and (iii) Lack of a formal evaluation procedure for performance analysis. A new framework for aerial image analysis is developed to combat these drawbacks. The analysis procedure attempts to draw ideas from the human visual system at each stage. It consists of texture-based segmentation to guide the detector, neural network "filter" for detection, and image clutter measures to provide confidence levels in the results. The multiresolution texture-based segmentation approach is based on the following observations: (i) Many regions in a large image can be logically eliminated from further analysis, for purposes of object detection, (ii) The human visual system utilizes relatively "global" information about an image, in order to narrow the "focus of attention" prior to detection, and (iii) Image texture plays a vital role in the segmentation of images into perceptually meaningful parts. A novel unsupervised segmentation approach based on gray level cooccurence (GLC) matrix texture models was developed and tested extensively. The concept of Multiresolution Associated Region (MAR) was developed along with normalized match distances, region homogeneity criteria and the MAR Aggregation Rule, in a systematic fashion. The segmentation strategy is based on multiresolution cooperative texture model computation. It considerably reduces the detection stage computational load and improves the accuracy. The results of the segmentation algorithm will be used to direct the "focus of attention" of the detection stage. The detector will operate only on those image regions deemed to be of interest. A neural network "filter" is designed and trained to detect targets in thermal infrared images. New concepts incorporated in the design are: (i) Operation ol the neural network like a spatial domain filter, (ii) Training methodology based on model-based samples, (iii) Modified neural transfer function and backpropagation algorithm. The detection and false alarm rates were determined for the neural network filters, and Receiver Operating Characteristic (ROC) curve analysis performed. It showed that the performance of the model-trained neural network filter, was much superior to that of the size-matched contrastbox filter, especially in the images with higher amounts of visual clutter. The final major contribution of this research deals with the quantitative characterization of image clutter. In the past, object detection approaches have typically been tested on application-specific image data sets. There is a need for standards to be developed to judge the relative merits of such approaches. The accomplishments of research toward this goal were: (i) Development of a Texture-based Image Clutter (TIC) measure, (ii) Evaluation of existing clutter measures and identification of their limitations, and (iii) Development of clutter measure evaluation methodology to cover a wide spectrum of images. The TIC measure will be used to determine the confidence level in the object detection results. A new framework for image analysis was developed. It incorporates (i) A novel procedure for unsupervised multiresolution segmentation based on texture, (ii) Neural network "filters" for object detection, and (iii) "Image Clutter" measures for confidence levels in the detection results. Performance evaluation tests for each aspect of the above framework consisted of extensive experimental studies. The techniques developed are expected to have an impact on a broad range of image understanding applications.

Download
Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS