In a significant advancement in radar technology, the IET Radar, Sonar & Navigation journal published an insightful study examining deep learning algorithms designed for target detection and classification using synthetic aperture radar (SAR) data. With the integration of environmental SAR scene images and target image chips, the research presents a comprehensive methodology for creating an extensive training dataset. The study’s evaluation spans the performance of three leading algorithms on both high-powered cloud servers and low size weight and power (SWAP) single-board computers, offering profound insights into their efficacy and applications.
Researchers investigated the effectiveness of target detection and classification algorithms applied to SAR data by combining environmental images with target image chips to develop a substantial dataset for deep learning training. The deep learning algorithms RetinaNet, EfficientDet, and YOLOv5 were examined. These algorithms were trained on a powerful cloud server and their speed and accuracy were put to the test on both the cloud server and a low-SWAP single-board computer.
Performance Analysis
Notably, YOLOv5 demonstrated the highest accuracy and speed on the cloud server, making it the top performer in high-resource environments. However, its performance on the low-SWAP device was less impressive, ranking as the slowest among the three algorithms. In contrast, RetinaNet and EfficientDet showed operationally useful throughput on the low-SWAP device, suitable for surveillance tasks, with RetinaNet achieving higher accuracy.
Qualitative Insights
Further qualitative analysis on additional datasets with varying characteristics underscored the critical role of acquiring relevant training data and performing appropriate pre-processing steps. This analysis highlights that algorithm performance can significantly vary based on data characteristics, which emphasizes the need for tailored approaches in different operational scenarios.
Earlier reports on similar topics also discussed the utilization of SAR data for target detection but often lacked the comparative analysis of different deep learning algorithms. Previous studies primarily focused on single algorithm performance and did not emphasize the importance of dataset creation and pre-processing steps as profoundly as the current study. Moreover, earlier findings often did not address the practical applications on low-SWAP devices, which is a key aspect explored in this research.
Comparatively, recent articles have begun to explore the application of deep learning algorithms on SAR data but still seldom delve into the detailed performance metrics across different hardware configurations. This study stands out by providing a thorough examination of how these algorithms perform under varying conditions, offering valuable insights into their potential operational use cases.
Given the findings, it becomes evident that while YOLOv5 excels in high-resource environments, its utility on low-SWAP devices is limited. RetinaNet and EfficientDet, however, offer a balance between speed and accuracy that makes them viable for embedded systems. For practitioners in the field, it’s crucial to consider the specific requirements of their applications when choosing an algorithm. Additionally, the study’s emphasis on the importance of data relevance and pre-processing provides a clear directive for further research and implementation strategies to enhance algorithm performance in diverse operational conditions.
