Salient Feature Identification and Analysis using Kernel-Based Classification Techniques for Synthetic Aperture Radar Automatic Target Recognition
Abstract
An investigation into feature saliency for application to synthetic aperture radar (SAR) automatic target recognition (ATR) is presented. Speci cally, research is focused on improving the SAR binary classi cation performance aspect of ATR, or the ability to accurately determine the class of a SAR target. The key to improving ATR classi cation performance lies in characterizing the salient target features. Salient features may be loosely de ned as the most consistently impactful parts of a SAR target contributing to e ective SAR ATR classi cation. To better understand the notion of salience, an investigation is conducted into the nature of saliency as applied to Air Force Research Lab's (AFRL) civilian vehicle (CV) data domes simulated phase history data set. After separating vehicles into two SAR data classes, sedan and SUV, frequency and polarization features are extracted from SAR data and formed into either 1D high range resolution (HRR) or 2D spectrum parted linked image test (SPLIT) feature vectors. A series of experiments comparing vehicle classes are designed and conducted to focus speci cally on the saliency e ects of various SAR collection parameters including azimuth angle, aperture size, elevation angle, and bandwidth. The popular kernel-based Bayesian Relevance Vector Machine (RVM) classi er is utilized for sparse identi cation of relevant vectors contributing most to the creation of a hyperplane decision boundary. Analysis of experimental results ultimately leads to recommendations of the salient feature vectors and SAR collection parameters which provide the most potential impact to improving vehicle classi cation. Demonstrating the proposed saliency characterization algorithm with simulated civilian vehicle data provides a road map for salient feature identi cation and analysis of other SAR data classes in future operational scenarios. ATR practitioners may use saliency results to focus more attention on the identi ed salient features of a ta
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 27, 2014
- Accession Number
- ADA598707
Entities
People
- Matthew S. Flynn
Organizations
- Air Force Institute of Technology