Representing sonar target spectral features using a two-dimensional Gabor wavelet

Abstract

We present an alternate representation of sonar spectral features with popular machine learning techniques for automated target classification. Sonar target recognition suffers from feature uncertainties due to unpredictable environmental parameters: changing sound speed profiles, inhomogeneous propagation of acoustic waves, interference and delayed echoes due to multipath effects, ambient ocean noise and reverberation, and signal deformation due to reflections from the seafloor, sea surface, and the Doppler effect. These parameters are further dependent on target orientation and may combine in a nonlinear fashion, making target classification near impossible. We propose a representation of sonar spectral features using the two-dimensional Gabor wavelet as a kernel filter. This extracts highly informative features specific to a target, regardless of orientation. We validate the robustness of our representation against feature uncertainties by comparing the classification performance of three machine learning techniques—a support vector machine, random forest tree, and a neural network—trained on the raw spectrogram (acoustic color) versus the Gabor-filtered feature space through resulting confusion matrices. We will provide classification results of public domain experimental field data. [This research is funded by the Office of Naval Research under Grant No. N00014-19-1-2436.]