基于CenterNet的改进遥感旋转目标检测

Improved Remote Sensing Rotating Object Detection based on CenterNet

遥感影像由于目标角度各异且普遍密集、小目标占比高、背景复杂等特点,检测精度低。针对水平框算法不再适用于遥感旋转目标,以及主流五参数法存在角度回归的周期性与边缘互换性问题,提出VR-CenterNet,采用向量表示法来进行旋转目标的检测与损失设计,规避角度回归的根本性问题,优化细长目标的偏移高敏问题;针对浅层特征融合的高冗余问题,引入自适应通道激活过滤杂质信息,为强化关键点信息,在主干输出部分引入改进后的全局上下文自适应层激活注意力块。首先在HRSC2016与UCAS-AOU数据集上进行不同算法的性能比较;再在两数据集上进行方法消融实验,以验证各改进方法的有效性。实验结果表明:在HRSC2016与UCAS-AOU数据集上分别取得的了88.48%与90.35%的精度。改进算法能够提升遥感旋转目标的检测精度,为遥感旋转目标的准确检测提供了另外一种解题思路。

Remote sensing images have low detection accuracy due to the characteristics of different object angles,generally arranged densely,high proportion of small objects and complex background.In view of the inapplicability of the horizontal detection algorithm for remote sensing rotating object detection,and the periodicity and edge interchangeability of angle in the mainstream five-parameter method,a VR-CenterNet is proposed,which used the vector representation to detect the rotating box and design the loss function to avoid the problem of angle regression,and to optimize the high displacement sensitive problem of slender objects.For the high redundancy problem of shallow feature fusion,self-adaptive channel activation is introduced to automatically filter impurity information.In order to strengthen the key point information,an improved global contextual self-adaptive layer activation attention block is introduced in the output of backbone.First,the performance of different algorithms is compared on HRSC2016 and UCAS-AOD data sets.Then,the module ablation experiment is conducted on the two data sets to verify the effectiveness of each improved method.Experimental results show that:88.48%and 90.35%accuracy are obtained on HRSC2016 and UCAS-AOD data sets respectively.The improved algorithm can improve the detection accuracy of remote sensing rotating objects,and provide another problem-solving idea for the accurate detection of remote sensing rotating objects.