基于典型几何形状精确回归的机场跑道检测方法

在遥感探测领域，实现复杂环境条件下机场跑道类地物目标和轮廓的精确检测具有重要意义。以YOLOv3为代表的主流深度学习算法在目标检测领域取得了显著的成绩，但该方法只能以矩形框给出目标的粗略位置，检测结果具有一定的背景区域且无法准确得到角点位置。针对以上问题，提出一种基于典型几何形状精确回归的机场跑道检测方法。综合利用典型四边形角点回归策略、四边形锚框机制、四边形的非极大值抑制模块以及目标几何拓扑关系，通过网络的轻量化设计和模型压缩，实现对目标在仿射畸变下成像特征的学习，能够快速预测目标的角点坐标，并以目标的四边形轮廓给出其位置。仿真实验结果表明，该算法具备机场跑道目标类型区分和轮廓提取的功能，有效地解决了实际应用中的目标精确定位难题；在不损失精度基础上网络经压缩后较压缩前的检测速度提高了1倍，大幅提升了自动目标检测的准确性和高效性。

Airport Runway Detection Agorithm Based on Accurate Regression of Typical Geometric Shapes

In the field of remote sensing detection, it is of great significance to achieve accurate detection of ground runway targets and contours under complex environmental conditions. The mainstream deep learning algorithm represented by YOLOv3 has achieved remarkable results in the field of target detection, but this algorithm can only give the approximate position of target in a rectangular frame, the detection result has a certain background area and cannot accurately get corner position. For the above problems, an airport runway detection algorithm based on the exact regression of typical geometric shape is proposed. Through the utilization of the typical quadrilateral corner regression strategy, the quadrilateral anchor frame mechanism, the quadrilateral non-maximum suppression module, the target geometric topological relationship, and the lightweight design of the network and model compression, the proposed algorithm can realize to learn the imaging characteristics of target under affine distortion, quickly predict the corner coordinates of target, and finally give its position with the quadrilateral contour of target. Experimental results show that the proposed algorithm has the functions of airport runway target type discrimination and contour extraction, which effectively solves the problem of accurate target positioning in practical applications, and doubles the detection speed without losing accuracy, and greatly improve the accuracy and efficiency of automatic target recognition.