<p>With the rapid advancement of drone technology, its application in maritime rescue has become increasingly prevalent. Drones offer high flexibility, wide-area coverage, and real-time monitoring, thereby significantly enhancing search efficiency and enabling rapid target localization. However, the aquatic environment poses distinct perception challenges—including specular reflections, severe illumination variations, and substantial object scale changes—which, together with the limited onboard computational resources of UAVs, lead to suboptimal detection accuracy and elevated false-positive rates for small targets. To overcome these limitations, this paper proposes FOE‑YOLO, an improved object detection model built upon YOLO11, specifically tailored for drone‑assisted maritime rescue. First, an ODConv module is incorporated into the backbone to strengthen feature representation under degraded visual conditions while reducing computational overhead. Second, a novel FSEIoU loss function is introduced to address scale variations and improve small‑object localization by adaptively re‑weighting low‑quality predictions. Third, a lightweight neck network, EOA‑YOLO, systematically integrates frequency‑domain enhancement, re‑parameterization, and sequential modeling to boost multi‑scale detection capability without sacrificing efficiency. This lightweight architecture, coupled with a 28.6% reduction in GFLOPs, makes FOE‑YOLO particularly suitable for resource‑constrained UAV platforms, offering clear practical advantages over heavier detectors in real‑world maritime missions. Experimental results demonstrate that, compared with the baseline YOLO11n, FOE‑YOLO achieves increases of 2.4% in mAP50 and 1.5% in mAP50‑95 while simultaneously reducing computational complexity by 28.6%, validating its effectiveness for accurate and efficient small‑object detection in challenging maritime scenarios.</p>