Synthetic Aperture Sonar Motion Compensation using Deep Learning

Abstract:
Synthetic aperture sonar (SAS) systems often rely on redundant phase center (RPC) techniques to determine the ping-to-ping displacements in order to precisely localize the synthetic aperture. This precision is required to maintain coherence across the array, which is necessary to produce high quality data. These measurements are often required to maintain precision in the presence of noise or over complex terrain. Recent advances in deep learning techniques have led to improved performance in the field of computer vision, and deep networks have achieved state-of-the-art results in regres- sion problems. In this paper, we describe the application of deep learning to providing estimates of ping-to-ping displacements for SAS RPC data. The deep network was trained on simulated data and then evaluated against both simulated and in situ SAS data. Results from this analysis and future applications are discussed.