Using convolutional neural networks on perceptual and spectral features: Classification experiments on the TREX’13 dataset

Abstract

Distinguishing objects of interest on the seafloor from clutter remains a key problem facing the automatic target recognition (ATR) community. Because scattering at high frequencies relates more to the geometry of the scatterer, traditional ATR on high frequency imaging sonars is known to suffer in areas of high clutter. Recently, there has been interest in low frequency (1–50 kHz) wideband imaging sonars, because the complex combination of external (geometric) and internal (elastic) scattering that occurs in this frequency range is thought to improve classification in areas of high clutter. This work investigates the classification problem of distinguishing unexploded ordnance (UXO) from clutter using image-based Convolutional Neural Networks on the TREX’13 dataset. This dataset consists of experimental and modelled acoustic backscatter for objects interrogated acoustically at 3–30 kHz across a range of aspects. The model data are used exclusively for training, and the experimental data are used exclusively for testing purposes. The generated feature-set is a combination of acoustic colour and perceptual features, which are derived from modelling how humans perceive timbre. Further in an effort to address the differing amplitude modulations between sets of model and experimental data and improve classified performance, several moving window normalizations will be investigated.