Experimental comparison between neural networks and classical techniques of classification applied to natural underwater transients identification

Abstract

The authors present an application of the joint use of signal processing techniques and neural networks to identify transient natural underwater sounds. The work focused on sounds of very short duration (typically 5 to 50 ms). Each context-free click is described by a reduced set of 31 input parameters, by the use of the autoregressive modeling and the Daubechies wavelets transform. The performances obtained by the Adaline-like-network (ALN) and the multilayered perceptron (MLP), and those obtained by classical techniques of classification (factorial discriminant analysis, and a clustering algorithm) are compared. A dichotomic approach and a multiclass approach were used.<>