FOE-YOLO targets faster drone detection for maritime rescue
A new YOLO11-based model aims to sharpen drone detection during maritime rescue missions.
FOE-YOLO is designed for one of the hardest perception problems in UAV search and rescue: finding small targets over water. Maritime scenes create glare, sharp illumination changes and large shifts in object scale. Those conditions can cut detection accuracy and raise false positives, especially when drones must process images with limited onboard computing power.
The model builds on YOLO11 and changes several parts of the detection pipeline. An ODConv module is added to the backbone to strengthen feature extraction under degraded visual conditions while reducing computational overhead. The goal is to give a lightweight detector more reliable image understanding when cameras face reflections, weak contrast or changing light at sea.
FOE-YOLO also introduces an FSEIoU loss function to improve small-object localization and handle scale variation. The method adaptively re-weights low-quality predictions, helping the model focus on difficult targets that are easy to miss or misplace. A lightweight neck network, called EOA-YOLO, combines frequency-domain enhancement, re-parameterization and sequential modeling to improve multi-scale detection without adding heavy compute demand.
In experiments, FOE-YOLO raised mAP50 by 2.4% and mAP50-95 by 1.5% compared with the baseline YOLO11n model. It also cut computational complexity by 28.6% in GFLOPs, a key measure for UAV platforms that cannot carry large processors. The result points to more practical real-time object detection for drone-assisted maritime rescue, where faster and more accurate target localization can directly affect mission outcomes.