CROSS-DOMAIN IDENTIFICATION OF ROAD NETWORKS USING DOMAIN-ADAPTED CONVOLUTIONAL NEURAL NETWORKS
Loading...
Authors
Peterson, Teal A.
Subjects
neural networks
deep learning
convolutional neural network
artificial neural networks
domain adaptation
domain divergence
domain adversarial neural network
gradient descent
gradient reversal layer
H divergence
proxy A distance
back propagation
data
Scan Eagle
ArcGIS
geographic information system
video
imagery
remote sensing
satellite
unmanned aerial vehicle
unmanned aerial system
computer vision
artificial intelligence
machine learning
autonomy
deep learning
convolutional neural network
artificial neural networks
domain adaptation
domain divergence
domain adversarial neural network
gradient descent
gradient reversal layer
H divergence
proxy A distance
back propagation
data
Scan Eagle
ArcGIS
geographic information system
video
imagery
remote sensing
satellite
unmanned aerial vehicle
unmanned aerial system
computer vision
artificial intelligence
machine learning
autonomy
Advisors
Xie, Geoffrey G.
Horner, Douglas P.
Date of Issue
2020-03
Date
Publisher
Monterey, CA; Naval Postgraduate School
Language
Abstract
Convolutional neural networks (CNNs) are a powerful tool for identification of patterns and objects within imagery or video. Training CNNs that can generalize well to their intended target dataset can require large amounts of labeled source data. The characteristics and distribution of this source (training) data must be representative of the target dataset for it to perform well. Labeled source data that fits this requirement is not always readily available. Research published by Ganin et al., in a 2016 paper titled “Domain-Adversarial Training of Neural Networks,” demonstrates that CNNs trained on a labeled source dataset can be adapted to generalize well to a target dataset through a process called domain adaption. In their research, they show that domain-adversarial neural networks (DANNs) improve performance on their target dataset relative to non-adapted CNNs. The purpose of this research is to explore the ability of DANNs to improve unmanned aerial vehicle (UAV) onboard classification of objects by adapting a CNN trained on satellite imagery to UAV aerial imagery. We show that DANNs do improve performance for this use case using several DANN architectures and datasets. This furthers other Naval Postgraduate School research efforts into autonomous UAV navigation and identification of targets of interest.
Type
Thesis
Description
Series/Report No
Department
Computer Science (CS)
Organization
Identifiers
NPS Report Number
Sponsors
SPAWAR
Funder
Format
Citation
Distribution Statement
Approved for public release; distribution is unlimited.
Rights
This publication is a work of the U.S. Government as defined in Title 17, United States Code, Section 101. Copyright protection is not available for this work in the United States.