Exploiting high dimensional data for signal characterization and classification in feature space

Abstract

The challenge of target classification is addressed in this work with both feature extraction and classifier hyperparameter optimization investigations. Simulated and measured high-range resolution radar data is processed, features are selected, and the resulting features are given to a classifier. For feature extraction, we examine two techniques. The first is a supervised method requiring an "expert" to identify and construct features. The performance of this approach served as motivation for the second technique, an automated wavelet packet basis approach. For this approach, we develop the Kolmogorov-Smirnov best-basis technique that utilizes empirical cumulative distribution functions and results in improved classification performance at low dimensionality. To measure classification efficacy, we use a quadratic Bayesian classifier, which assumes a Gaussian distribution as well as a support vector machine. The support vector machine is a classifier, which has generated excitement and interest in the pattern recognition community due to its generalization, performance, and ability to operate in high dimensional feature spaces. Although support vector machines are generated without the use of user-specified models, required hyperparameters, such as kernel width, are usually user-specified or experimentally derived. We develop techniques to optimize selection of these hyperparameters. These approaches allow us to characterize the problem, ultimately resulting in an automated approach for optimization, semi-alignment .