Target recognition in diverse synthetic aperture radar image datasets with low size weight and power processing hardware

Abstract

AbstractThis paper studies the performance of target detection and classification algorithms applied to synthetic aperture radar (SAR) data. We describe a process to merge measured environmental SAR scene images with target image chips to produce a large dataset for training deep learning algorithms. Three algorithms, RetinaNet, EfficientDet, and YOLOv5, were trained using a powerful cloud server. Performance at inference time, in terms of speed and accuracy, was tested on both the cloud server and a low size weight and power (SWAP) single board computer. YOLOv5 was found to be the most accurate and fastest algorithm on the cloud server but the slowest on the low‐SWAP device. RetinaNet and EfficientDet produced operationally useful throughput on the low‐SWAP device for surveillance applications, with RetinaNet having the higher accuracy. Further qualitative analysis of algorithm performance on additional data with different characteristics highlighted the importance of gathering relevant training data and carrying out suitable pre‐processing steps.