Deep-Learned Air-Coupled Ultrasonic Sonar Image Enhancement and Object Localization

Abstract

Air-coupled ultrasonic phased arrays are a complement to existing lidar-, camera- and radar-based sensors for object detection and spatial imaging. These in-air sonar systems typically use conventional beamforming (CBF) for high-frame rate image formation. Consequently, in real-world multi-target environments, the unique identification of reflectors is a challenging task due to the array-specific point spread function (PSF). Therefore, we present a neural auto-encoder network based on Xception for removing the PSF characteristics from CBF images and estimating the number of reflectors. Based on this information, the reflector coordinates are extracted by Gaussian mixture model clustering. We train and test the architecture on simulated and randomized multi-target CBF images. The performance is evaluated in terms of the localization precision, reflector count error and the angular resolution obtained. The preliminary results show a low mean error for the localization (-0.61°, -3 mm) and an accuracy of 83% for the reflector count estimation. The angular resolution of the given array can be improved from 14° to 2°. Overall, we highlight the potential of state-of-the-art auto-encoder networks, typically used for optical images, for CBF image enhancement and the combination with clustering for target localization.