无锚点的遥感图像任意角度密集目标检测方法

Anchor-free remote sensing image detection method for dense objects with rotation

针对基于深度学习的遥感图像目标检测方法密集目标漏检率高、分类不准确的问题,建立了一种基于深度学习的无锚点的遥感图像任意角度的密集目标检测方法。首先采用CenterNet作为基线模型,经过主干网络提取特征,并改造原有检测器结构,即加入角度回归分支进行目标角度回归;然后提出一种基于非对称卷积的特征增强模块,并将主干网络提取到的特征图输入特征增强模块,从而增强目标的旋转不变性特征,消除由于目标的旋转、翻转带来的影响,进一步提升目标中心点、尺寸信息的回归精度。采用HourGlass-101作为主干网络时,该方法在DOTA数据集上的平均精度均值(mAP)比旋转区域候选网络(RRPN)提升了7.80个百分点,每秒处理帧数(FPS)提升了7.5;在自建数据集Ship3上,该方法的mAP比RRPN提升了8.68个百分点,FPS提升了6.5。结果表明,所提方法能获得检测精度和速度的平衡。

Aiming at the problems of high missed rate and inaccurate classification of dense objects in remote sensing image detection methods based on deep learning,an anchor-free deep learning-based detection method for dense objects with rotation was established.Firstly,CenterNet was used as the baseline network,features were extracted through the backbone network,and the original detector structure was improved,which means an angle regression branch was added to perform object angle regression.Then,a feature enhancement module based on asymmetric convolution was proposed,and the feature map extracted by the backbone network was put into the feature enhancement module to enhance the rotation invariant feature of the object,reduce the influence caused by the rotation and turnover of the object,and improve the regression precision of the center point and size information of the object.When using HourGlass-101 as the backbone network,compared with Rotation Region Proposal Network(RRPN),the proposed method achieved a 7.80 percentage point improvement in Mean Average Precision(mAP)and 7.50 improvement in Frames Per Second(FPS)on DOTA dataset.On the self-built dataset Ship3,the proposed method achieved a 8.68 percentage point improvement in mAP and 6.5 improvement vin FPS.The results show that the proposed method can obtain a balance between detection precision and speed.