Evaluation of neural network and conventional techniques for sonar signal discrimination

The problem of sonar signal discrimination of passive sonar events is addressed. Three generic systems are considered. The first is a conventional system that uses a quadratic Bayesian (QB) classifier. Next is a hybrid approach that uses a neural compound classifier network (CCN) of the type proposed by B.G. Batchelor (1974). Both the conventional and hybrid approaches use a generic automatic detector given by J.J. Wolcin (1984), which is structured to detect signals of arbitrary duration and frequency content. The third system is an all neural network approach which considers neural alternatives to the functions of detection, feature extraction, and feature optimization. The authors discuss a comparison of the first two systems. The third system is addressed by D.W. Cottle and D.J. Hamilton (ibid., this conference, p.13-19, 1991).<>