NavShip-5V: A Multi-view Dataset and Lightweight Framework for Autonomous IMO Inspection

Autonomous maritime inspection by Unmanned Aerial Vehicles (UAVs) requires precise relative viewpoint estimation to locate identity markers (specifically, the International Maritime Organization (IMO) number) on vessel sterns. However, existing datasets lack navigational semantics, and visual ambiguity arising from vessel symmetry poses significant challenges for robust control. To address this, we propose a lightweight, hierarchical perception framework tailored for embedded edge deployment. First, we introduce NavShip-5V (Navigation-oriented Ship 5-Viewpoints), a novel multi-view dataset annotated with a function-oriented protocol. Second, we design a viewpoint estimation network based on MobileNetV3-Small. By optimizing the classification topology (merging symmetric side views) and employing a class-balanced weighted loss (20:1), we effectively resolve the visual ambiguity issue. Counter-intuitively, our lightweight model achieves a Top-1 accuracy of 90.9%, significantly outperforming the heavier ResNet-18 (84.1%) and the unstable ShuffleNetV2 in both accuracy and convergence stability. Finally, integrated with YOLOv8-nano (90.8% mAP@0.5) for local ROI detection, the proposed system demonstrates a complete, safe, and efficient Find-Navigate-Read pipeline, paving the way for fully autonomous maritime law enforcement.