Employing geometric pattern discovery techniques for sonar target recognition

Abstract

Automated sonar target recognition suffers from linear superposition of acoustic color features, which need to be precisely disambiguated to correctly identify a target of the specific composition and geometry. Most target acoustic color images exhibit feature overlap between environmental reverberation as well as reflections due to target geometry, and elastic waves that are representative of the material composition of the target. From a signal processing perspective, this is essentially mitigating and disentangling interference from two highly competing yet hard-to-predict sources: reflections from the environment and reflections due to the target geometry, which represent different physical phenomena with distinct geometric patterns. We will present recent results in feature selection and interference mitigation employing Gabor wavelets to separate target acoustic color features due to elastic wave orbits from interfering features due to the target geometry and environmental effects. We will provide results from simulations as well as experimental field data and discuss the trade-offs in feature representations employing acoustic color and those that also include the phase information. Finally, we will discuss how this type of feature extraction can be used to discover and create robust priors for Bayesian inference networks for learning salient target features.