一种倾斜矩形范围框标注方式及遥感目标检测应用分析

An annotation method of arbitrary-oriented rectangle b-box and analysis of its application in remote sensing object detection

为解决机器学习过程中样本标注困难和模型训练遇到的损失函数取值异常和模型回归难的问题,提出了一种新的斜矩形范围框标注方式。采用“倾斜范围框中心点C的坐标、中心点到任意一个顶点D的向量CD、C到D的一个相邻顶点E的向量CE在CD上的投影向量CP与CD的比例系数”来标注倾斜范围框,在给定约束下实现了一个范围框只有一种数值表示,避免了损失异常,有利于模型回归训练。并将该标注方式应用在遥感影像目标检测任务中,通过斜框目标检测公共数据集,与多种其他倾斜范围框标注方式做了效果对比分析,结果表明新提出的标注方式在验证集上得到的平均准确率为0.7752,该斜矩形范围框标注方式对于朝向任意、密集排布的目标检测更具优势。

In order to solve the problem of the abnormal value of the loss function and the difficulty of model regression encountered in the process of model training in the machine learning process,a new oblique rectangular frame labeling method is proposed.Using the center point(C)of the bounding box,the vector(CD)from the center to one of its vertex(D)and the ratio of the vector(CP)to CD.CP is the projection of the vector(CE)to CD.CE is a vector from the center of the bounding box to one of the other vertex(E)that neighbor to D,which avoids loss anomalies and is beneficial to model regression train⁃ing.And this labeling method is applied to the task of remote sensing image target detection,through the oblique frame target detection public data set,and a variety of other oblique range frame annotation meth⁃ods are compared and analyzed.The results show that the average accuracy of the newly proposed labeling method on the verification set is 0.7752,and the oblique rectangular frame labeling method is more advan⁃tageous for the detection of arbitrarily oriented and densely arranged targets.