Outlier Detection in Infrared Signatures
Abstract
For a number of years, simulated long wavelength infrared (LWIR) signatures have been used to determine the ability to classify military targets and decoys. Such signatures sometimes exhibit specular behavior, a characteristic displaying a sudden increase in radiant intensity of short duration. This specular behavior is sporadic and is as likely to show up for targets as it is for decoys. Unfortunately, if these outliers (i.e. the specular occurrences) are not removed from the data. the estimated performance of discrimination algorithms can be misleading. Statistical outlier detection provides an useful approach for finding and removing the outliers caused by specular occurrences. This paper considers the statistical properties of the outlier detection algorithms as applied to simulated LWIR signatures. We consider possible statistical models for outliers in order to determine whether or not modifications might minimize the number of outliers left in the signature after editing and minimize the number of good observations deleted from the signature. Ultimately, we are seeking the data editing algorithm which produces the best possible discrimination performance.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1992
- Accession Number
- ADA264353
Entities
People
- Jon A. Magnuson
- Michael Chernick