Active sonar classification using perceptual signal features from musical acoustics

Abstract

Because of the relatively low frequencies at which they operate, navy active sonars are often plagued by false-alarm returns resulting from geological structures. In the lexicon of sonar operators these false returns are referred to as geoclutter or simply clutter. Despite mounting evidence that human operators can aurally discriminate target returns from clutter, attempts to develop robust automatic classification algorithms have, as yet, met with limited success. This paper investigates the possibility of improving the performance of automatic classifiers by exploiting the signal processing employed in the human auditory system. This amounts to replacing the statistical signal features used by conventional automatic classifiers with perceptual signal features that reflect the way a human listener would perceive a given return. Drawing an analogy between active sonar returns and percussive musical timbre, this paper considers perceptual signal features—like loudness centroid, sub-band attack/decay, and sub-band correlation—that have previously been identified as underlying the perception of timbre. Values for each of these timbre features are measured for a series of target and clutter returns recorded during an experimental sea trial. The effectiveness of these perceptual features as target-clutter discriminators is then evaluated using a Gaussian linear discriminant classifier.