Wavelet feature extraction of high-range resolution radar profiles using generalized Gaussian distributions for automatic target recognition.

摘要

This dissertation provides a new technique for improving future extraction of high range resolution (HRR) radar profiles for automatic target recognition (ATR) systems. Although not new, HRR radar is an important sensor for ATR. This sensor collects data which are a range profile of a target. Targets in this study are aircraft. HRR radar target identification is a challenge, because it requires representing a three-dimensional object as a one-dimensional signal. Typically, these radar signals are modeled as complex exponentials which, when combined during the dimensional reduction process, add constructively and destructively depending on their relative phases. Thus, a slight change in the relative phases in the radar returns can have significant effect on the HRR signature. Hence, the goal of the ATR system is to identify the target on the basis of its HRR profile and to properly classify it amongst a set of target classes.; The problem studied in this dissertation is one of identifying six classes of aircraft on the basis of their HRR range profile. This dissertation solves this HRR ATR problem using a "best bases" algorithm approach that relies on wavelets and principal component analysis for extracting the features and for reducing the overall dimension of the original feature space, respectively. A statistical-classification supervised learning approach is used to construct and train the classifier. The algorithm employs the statistical distribution of the target class in wavelet feature space to obtain six independent classifiers, one for each target class. To ensure separation amongst these target classes and to simplify classification, Bayesian Classification discriminants and maximum likelihood analyses were used.; The classifiers were then used against the training and test set, respectively, with and without noise. The classifiers resulted in 100% and 98.1132% correct classification against the training and test set, respectively.