Comparative Study of YOLOv10, YOLOv11 and YOLOv12 Lightweight Models for Multi-Class Maritime Search and Rescue Using UAV Imagery

Abstract. The maritime Search and Rescue (SAR) operation requires effective and accurate object detection systems capable of identifying various targets in dynamic sea environments and low-light situations. The paper presents a comparative study of the YOLOv10, YOLOv11, and YOLOv12 networks in multi-class marine detection using UAV images. The SeaDronesSee Odv2 dataset has been preprocessed using physics-based augmentation that mimics environmental changes, such as fog, noon, sunset, dawn, and cloudy scenarios. A multi-resolution tiling procedure was implemented to preserve the image consistency of small objects. Results show that YOLOv11s is the model that has the least accuracy-efficiency trade-off, with an mAP@0.5 of 0.888 and an F1-Score of 0.872 at a reasonable inference time. Precision-recall analysis has shown that large maritime objects were detected with high precision, while small objects were detected with average recall. The results show that multi-resolution preprocessing, as well as physicsbased augmentation, enhance the robustness and generalization of the network. Altogether, YOLOv11s is the most stable version to use in real-time maritime SAR missions with UAVs due to its ability to handle a variety of visual conditions.