Synthetically-trained neural networks for shape classification from measured acoustic scattering

Abstract

Drawing inspiration from the biological phenomenon of echolocation, ultrasound perception holds immense potential across various engineering domains, spanning from advanced imaging to precise navigation. Despite advances in sensor development and signal processing, current methodologies struggle to match the remarkable perceptual acuity of echolocating animals when deciphering real-world ultrasound echoes. In this study, we bridge this disparity by harnessing Convolutional Neural Networks (CNNs) to discern ultrasound scattering from objects of different shapes. Our novel approach entails training CNNs using exclusively synthetic data, derived from numerical simulations, to process real echoes. We achieve this through (1) sophisticated data augmentation and processing of synthetic echoes that accommodate physical variations and uncertainties inherent in practical scenarios and (2) specialized CNNs (SCNNs) targeted at each shape to compel models to learn features unique to that shape. Rigorous experimentation demonstrates the ability of these synthetically-trained models to accurately classify fundamental geometric shapes of objects based solely on experimentally measured echoes. Furthermore, the intentional selection of the size and shapes of the objects to produce perceptually similar echoes elucidates the efficacy of our approach in handling intricate perception scenarios. By alleviating laborious and costly data acquisition procedures in favor of synthetic data-driven training for real-world perception, our method opens avenues for advancements in diverse fields reliant on ultrasound-based technologies. These advancements bear implications spanning from diagnostics to the realm of autonomous systems and beyond.