基于旋转中心点估计的遥感目标精确检测算法  

Remote sensing target accurate detection algorithm based on rotation center point estimation

作　　者：蒋光峰[1] 胡鹏程 叶桦[1] 仰燕兰[1] Jiang Guangfeng;Hu Pengcheng;Ye Hua;Yang Yanlan(School of Automation,Southeast University,Nanjing 210096,China)

机构地区：[1]东南大学自动化学院,南京210096

出　　处：《计算机应用研究》2021年第9期2866-2870,共5页Application Research of Computers

基　　金：中央高校基本科研业务费专项资金项目(2242020K40244)。

摘　　要：由于遥感图像背景复杂、目标密集分布以及目标尺度、形状差异巨大,给检测带来挑战。当前基于R-CNN的两阶段算法在水平框(HBB)检测上取得了良好效果,然而在定向框(OBB)检测上效果有限。基于点估计的HBB目标检测框架,提出用于定向遥感目标检测的旋转中心点估计网络(RCNet),大幅提升一阶段anchor-free算法在倾斜目标检测上的性能,同时保持较高的检测速度。RCNet通过添加一个用于方向预测的分支,实现旋转中心点估计。提出新的角度表示方式,解决回归角度参数loss不连续以及宽高交换导致训练过程不稳定的问题。所提方法在DOTA数据集上取得66.68 mAP的检测精度以及29.4 fps的检测速度,实现了最佳的速度和精度平衡。Due to the complex background of remote sensing images,the dense distribution of targets,and the huge differences in target size and shape,it brings challenges to detection.The current two-stage algorithm based on R-CNN works well in horizontal box(HBB)detection,but its effect in directional box(OBB)detection is limited.Based on the HBB target detection framework of point estimation,this paper proposed a rotation center point estimation network(RCNet)for directional remote sensing target detection,which greatly improved the performance of the one-stage anchor-free algorithm on tilted target detection while maintaining a high detection speed.RCNet realized the estimation of the center of rotation by adding a branch for direction prediction.It proposed a new angle representation method to solve the problem of instability in the training process caused by the discontinuous loss of the angle parameter regression and the exchange of width and height.The proposed method achieves a detection accuracy of 66.68 mAP and a detection speed of 29.4 fps on the DOTA data set,achieving the best speed and accuracy balance.

关 键 词：定向框检测 遥感图像 点估计 anchor-free 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]