On-Board Ship Detection in SAR Images Based on L-YOLO
Xiaowo Xu, Xiaoling Zhang, Tianwen Zhang, Jun Shi, Shunjun Wei, Jianwei Li · 2022 IEEE Radar Conference (RadarConf22) · 2022
Ship detection in Synthetic Aperture Radar (SAR) images is an essential but challenging task. Considering the high-delay between spaceborne SAR system and ground system communication, it is difficult for ground system to achieve realtime SAR data processing. It brings the difficulty for obtaining real-time SAR ship detection. To solve this problem, a novel onboard SAR ship detection method based on the Lightweight You Only Look Once (L-YOLO) algorithm is proposed. Different from the original You Only Look Once version 5 (YOLOv5) algorithm, L-YOLO utilizes lite convolution block instead of original convolution block to compact the network by the means of introducing low-cost linear operations. To further compensate the accuracy, we apply a cost-free K-means algorithm to re-cluster the priori boxes of the dataset. For the following algorithm validation, we also transplant L-YOLO to Nvidia Jetson TX2 to validate the practicability of the scheme. Experimental results on Large-Scale SAR Ship Detection Dataset-v1.0 (LS-SSDD-v1.0) show that our method can achieve ~ 73% mean Average Precision (mAP) with a compact network architecture, where the Floating point Operations (FLOPs) are half (from 16.3G to 8.1G) of YOLOv5.