Comparison of a multilayered perceptron with standard classification techniques in the presense of noise

摘要

The problem of classifying radar pulses in the presence of noise is described. Using simulated pulsed signals and noise, the authors compare a multilayer perceptron neural network against a parametric template-based technique, the Mahalanobis distance classifier, and the k-nearest neighbor classifier. The multilayer perceptron using error backpropagation was found to have better performance than the k-nearest neighbor classifier for training data having a low signal- to-noise ratio (SNR). Template matching and Mahalanobis distance classification both gave much lower accuracy than the former two classifiers for high testing SNRs regardless of the SNR of the training data.