Title

Methods of Neural Network Analysis to Identify Weapons in Augmented Reality

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Science and Info Sys

Date of Award

Summer 2021

Abstract

Augmented reality (AR) is a technology being utilized to help give an individual insightwhile doing their job by overlaying information on the real world in real-time. The military is working on using AR to help soldiers on the battlefield to make more informed decisions. The battlefield is a dynamic environment and requires real-time, clear, and effective ways of getting information to the soldier. Convolutional Neural Networks (CNN) are used in this study to classify images of weapons in videos. The YouTube8M video dataset was explored and another hand-picked image dataset was created based on image content and then used in training three ImageNet pre-trained models, Xception, ResNet50, and MobileNet using transfer learning. The effectiveness of this dataset and models were evaluated for speed and accuracy. We found that ResNet50 was more accurate but slightly slower than MobileNet. The Xception pre-trained base was overfitting even though it performed well on the test data. Further research is discussed.

Advisor

Derek Harter

Subject Categories

Computer Sciences | Physical Sciences and Mathematics

Share

COinS