DGPC-YOLO: an enhanced YOLOv8 model for SBMSR target detection

In Space-based Maritime Surveillance Radar (SBMSR), factors such as the high-speed motion of the satellite platform, significant ground/sea clutter, and the broad distribution of physical size among observed maritime targets lead to the degradation of target detection performance in conventional Constant False Alarm Rate (CFAR) detection algorithms. Deep learning-based methods offer a promising alternative for target detection in two-dimensional images due to their powerful capability for automatic feature extraction and learning. However, the lack of real measured data from SBMSR has hindered the development of high-performance models in this domain. To address this issue, this paper applies down-sampling to the high-resolution SAR images dataset (HRSID) to emulate the characteristics of maritime targets in SBMSR and proposes an enhanced YOLOv8n detection model named DGPC-YOLO. First, the neck and detection head of YOLOv8 are redesigned by adding a dedicated tiny-object detection head while removing the large one. To improve the utilization of local detailed features in small targets, a novel feature extraction module, C2f_GLVSS, is proposed to replace part of the C2f modules in the backbone network. Then, a composite loss function combining Normalized Wasserstein Distance (NWD) and PIoUv2 is proposed to replace the original CIoU loss. Finally, the Coordinate Attention (CA) mechanism is incorporated after the SPPF layer to enhance the model's focus on small targets and challenging samples while suppressing background interference in complex scenarios. To evaluate the detection performance of DGPC-YOLO, comprehensive ablation and comparison experiments are conducted. Compared to the baseline, the proposed model achieves improvements of 2.4% in Precision, 6.5% in Recall, and 4.0% in mAP50:95. DGPC-YOLO effectively reduces both false alarms and missed detections of marine targets, demonstrating strong potential for practical application in SBMSR systems.