面向港口车船目标的航空影像旋转不变检测器

为了解决通用目标检测架构在航空影像下港口场景中的车辆和船只检测任务中性能不足且不能估计目标方向的缺陷。一种单卷积旋转不变检测器被提出以解决这种问题。基于回归网络的旋转不变检测器网络层数更少，边框分配和尺度大小更加公平，在带有角度信息的方向边框策略的配合下能够在特定场景的目标检测任务中实现更鲁棒的检测结果并能评估目标的方向。使用同一数据集与众多先进架构的对比实验和旋转边框消融实验的结果证明了提出架构的优势。

Aerial Images Rotation Invariant Detector for Port Vehicle and Ship Targets

The general purpose target detection architecture is insufficient in the vehicle and ship detection missions in the port scene under aerial imagery and cannot estimate the target direction. A rotation invariant detector has been proposed to solve this problem. Rotation-invariant detector network based on regression network has fewer layers, and border allocation and scale are more fair. In combination with the direction frame strategy with angle information, it can achieve more robust detection results and evaluate the direction of objects in target detection tasks in specific scenes. The results of the comparative experiments with the advanced architecture and the ablation experiments demonstrate the advantages of the proposed architecture.