Dissimilarity based perceptual categories of underwater targets and related auditory features

摘要

Automatic classification of underwater targets has a variety of applications in many fields, but state of the art systems still perform worse than skilled operators. To design a system that follows auditory principles, this study investigates how human beings classify various underwater targets and identify the auditory features they use via a pair-wise comparison experiment. The stimulus set contains 20 underwater sounds from 5 typical categories. The dissimilarity ratings were analyzed by a sophisticated version of the multidimensional scaling algorithm CLASCAL, yielding an optimal spatial model with three dimensions. Each dimension separates one category of targets, and the five clusters in the space are in accordance with the five categories chosen for the experiment. For the purpose of automatic target classification, the quantitative expressions interpreting three perceptual dimensions were derived by exploiting an auditory filterbank. It was found that these three dimensions corresponded to the measure of maximum subband envelope variance, spectral roll-off, and pitch duration, respectively.