Applying Deep Learning Methods to Identify Targets in Synthetic Aperture Radar Images
Abstract
Synthetic aperture radar (SAR) provides high-resolution imagery and can operate in the day and at night andin every weather condition. SAR has been used for military reconnaissance and surveillance. Examining SARimages manually, however, is challenging even for a specialist, since it is difficult to find high-value targets in awide area of SAR images. This is especially true when time is critical for operations. Thus, an efficient, reliablemethod to analyze SAR images automatically is needed. To solve this problem, deep learning (DL) methods aredeveloped for automatic target recognition (ATR). A convolutional neural network (CNN) is a deep-learningalgorithm made up of several processing layers for target recognition and classification. One of the challenges indeveloping and testing a CNN algorithm is to find relevant datasets. The dataset used in this thesis comes from theMoving and Stationary Target Acquisition and Recognition program (MSTAR).In this research, the SAR ATR concept and performance are analyzed using several CNN DL architectures.Specifically, this investigation examines the effects of a few variable parameters within CNN DL architectures togain insight into optimal strategies for using these architectures. Using CNN structures with different numbers oflayers, it was possible to classify SAR targets successfully and automatically with state-of-the-art accuracy. Thismethod proved useful for classification and recognition of military targets.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 2020
- Accession Number
- AD1126746
Entities
People
- Serkan Aktas
Organizations
- Naval Postgraduate School