Improvement of Automatic Target Recognition Through Synthetic Data Augmentation

Data sets of well­ labelled and diverse acoustic imagery of the seabed are scarce. However, a recent breakthrough in synthetic aperture sonar (SAS) image simulation has facilitated the rapid generation of realistic echo data. The synthetic data include important aspects of the acoustic wave physics, such as aspect­ dependence, layover, diffraction, speckle, focusing errors, and artefacts. Moreover, it provides high­ fidelity label information. This combination of speed, realism, and detail has enabled the use of synthetic data to improve the volume and diversity of training data for deep learning algorithms in automatic target recognition (ATR). We present an overview of the rapid simulation model, alongside an existing SAS simulation model, and demonstrate its application to ATR training for the detection and classification of underwater munitions and unexploded ordnance.