Masters Theses

Amol G. Bokil

Date of Award

12-1988

Degree Type

Thesis

Degree Name

Master of Science

Major

Electrical Engineering

Major Professor

Mohan M. Trivedi

Committee Members

D. W. Bouldin, M. O. Pace

Abstract

Accurate and reliable detection of objects is an important component of an image interpretation system. Objects might appear against a wide variety of background regions and as such characteristic properties of the objects are sought. Texture analysis based approaches are powerful in such applications since texture is a property inherent to a region. A gray level co-occurrence (GLC) based texture analysis package useful for object detection and image segmentation is developed. Most of the techniques found in literature employ a supervised classification scheme, where the classifier is trained with a set of training samples. Unavailability of the training samples and the expense associated with it has forced researchers to explore other strategies. Cluster analysis is the technique suggested in this study. It provides a logical means to perform unsupervised classification in a d-dimensional space. Forward sequential search based measurement selection strategy and Mahalanobis distance classifier are the elements of the supervised classification scheme. K-means cluster analysis algorithm is the unsupervised counterpart. The main focus of this study is the application of these two techniques to aerial images. A high resolution multispectral urban scene image and a high resolution thermal infra-red channel image is used to study the performance of these two techniques. Resubstitution testing procedure is applied generally, while in the analysis of the thermal infra-red image jackknifing is also employed. Supervised statistical classification scheme IV seemed to perform better than the cluster analysis based classification scheme. Investigation of multispectral texture analysis is the other issue addressed in this thesis. A gray level difference (GLD) based technique for multispectral texture analysis is studied. Alternate approaches based on the GLC and the GLD methods are proposed to perform texture analysis in multispectral images. Ap pending the measures from each of the channels and using them for the object detection task is the underlying idea of the approaches suggested here. In the analysis of multispectral imagery the techniques suggested in this study, per formed better in object detection as well as image segmentation tasks. Also, in general the proposed GLC based technique yields better results as compared to the gray level difference based method. The results obtained in the case of the analysis of multispectral imagery are better than those obtained when analyzing a single channel of the image indicating the superiority of multispectral analysis approach over the single channel analysis strategy.

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

Share

COinS