Date of Award

12-1992

Document Type

Thesis

Degree Name

Master of Science in Electrical Engineering

Department

Department of Electrical and Computer Engineering

First Advisor

Steven K. Rogers, PhD

Abstract

This study investigated neural networks for face verification and classification. The research concentrated on developing a neural network based feature extractor and/or classifier to perform authorized user verification in a realistic work environment. Recognition accuracy, system assumptions, training time, and execution time were analyzed to determine the feasibility of a neural network approach. Data was collected using a camcorder and two segmentation schemes: manual segmentation and motion-based, automatic segmentation. Data consisted of over 2000. 32x32 pixel, 8 bit gray scale images of 52 subjects; each subject had two to ten days worth of images collected. Several training and test sets were created and then used to train and test the following networks: a back propagation network using the raw data as inputs, a back propagation network using Karhunen-Loeve Transform coefficients, computed from the raw, data, as inputs; and a back propagation network using features extracted by an identity network as inputs. The classification networks performed well on constrained, single day captured, data bases but performed poorly on data gathered over multiple days . For multiple days, a verification network using a single hidden layer with back propagation obtained 95% verification accuracy and is suitable for use in a face verification system.

AFIT Designator

AFIT-GE-ENG-92D-23

DTIC Accession Number

ADA259587

Comments

The author's Vita page is omitted.

Download

Share

COinS