Lightweight and Accurate Video Synthetic Aperture Radar Target Detection Network

Video synthetic aperture radar (SAR) target detection has seen rapid development in recent years, but current methods generally suffer from high computational complexity as well as frequent false alarms and missed alarms. Additionally, in video SAR scenarios, small targets become more difficult to identify due to low resolution and complex backgrounds. Recently, leveraging shadows for video SAR moving target detection has proven advantageous as it provides accurate location information and additional boundary details. To address the challenges of target detection in video SAR caused by low resolution and complex backgrounds, as well as the high computational cost of current methods, this paper proposes a lightweight algorithm for video SAR shadow detection based on the YOLOv5 architecture, named Lavs-DeNet. First, a lightweight feature extraction backbone network combining both global and local information called CGLLNet is proposed. This backbone improves the detection performance by mitigating false alarms and missed alarms caused by defocus in video SAR images. Next, a slanted ladder bidirectional feature pyramid network (SL‐BiFPN) is designed. By using tilted upsampling with multiple stacked layers, this network efficiently extracts multi-scale features, further reducing missed alarms. Finally, a lightweight information interaction module (LII‐C3) has been developed, significantly reducing computational complexity. To validate the effectiveness of the proposed method, experiments were conducted using the data set released by Sandia National Laboratories. The experimental results show that the proposed Lavs-DeNet can achieve 97.88% detection accuracy, requiring only 21.02G floating points and 0.88M parameters, which is superior to the current classical video SAR target detection networks.