Masters Theses

Date of Award

12-1992

Degree Type

Thesis

Degree Name

Master of Science

Major

Electrical Engineering

Major Professor

Dragana Brzakovic

Committee Members

I. Alexeff, D. B. Koch

Abstract

This thesis presents a study of object recognition and classification using the Forward Only Counterpropagation neural network (CPN). The structure of the network combines a Kohonen layer, a self-organizing map functioning as a feature extractor, and the outstar structure of a Grossberg layer functioning as a classifier to the extracted features. Stemming from the fact that the network has the capability to recognize and classify objects, the CPN is used to estimate the orientation angle of a given object. The object is oriented in a two dimensional plane with an angle varying between 0 and 180 degrees. Using an appropriate number of nodes in the Kohonen layer, the network is expected to perform with acceptable accuracy. The study also covers the effects of increasing the number of nodes in the processing layers to obtain better accuracy, and network behavior in the presence of randomly generated normally distributed noise added to the tested patterns. For more accurate computation of the orientation angle of the presented objects, the network structure is modified by adding a third layer to the original CPN structure. The new layer identifies the correct pattern from the set of patterns that can be recognized and classified by the winning nodes in the first and second layers of the modified network. The final step of the presented study involves a method of defect detection in images. The CPN is used for image recognition and classification and detects the sections in the image with possible defects. Defects include shift, size change, and absence of details in the processed image. The network is trained on defect free images, and images with possible defects can be inputted as testing patterns.

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

Share

COinS