Algorithm for sonar-based signal identification

A non-parametric algorithm is described for implementing the optimal multi-dimensional Bayesian classifier, whereby a given feature vector is assigned as belonging to either of two classes on the basis of a log-likelihood ratio test (LLRT). This algorithm addresses the main shortcoming of such classifiers, the determination of the multidimensional distributions essential for the computation of the LLRT, by mapping the sets of learning vectors to a space of orthogonal features approximating statistical independence. The LLRT then simply consists of the sum of LLRTs performed at the level of the individual features, with the determination of the individual (one-dimensional) distributions effected by histogramming the transformed data sets in an off-line “training” process. Results of this procedure are compared against results of conventional classification methods (nearest-neighbor and nearest-mean), conventional methods utilizing orthogonalized data, and procedures using parametric (Gaussian) expressions for the distributions of the features