基于Gromov-Wasserstein最优传输的输电线路小目标检测方法

Transmission line small object detection based on Gromov-Wassertein optimal transport

针对输电线路无人机线路巡检场景中目标检测算法在处理线路缺陷、零部件缺失等小目标时性能严重下降的问题,从标签分配角度提出新的损失函数,提高小目标检测的准确性和效果。区别于传统目标检测方法,将每个目标预测框视为高斯感受野,将真实值视为高斯热图,通过计算2个高斯分布之间的距离进行标签分配;提出利用Gromov-Wassertein最优传输引导模型学习,该方法可以建立在现有的检测模型之上。对多个输电线路目标检测数据集进行试验,结果表明,采用高斯感受野和最优传输的标签分配方案在输电线路巡检中的小目标检测方面具有良好的效果。

In order to address the issue of severe performance degradation of target detection algorithms in the scenario of unmanned aerial vehicle(UAV)line inspection for power transmission lines,specifically when dealing with small targets such as line defects and missing components,a new loss function was proposed from the perspective of label assignment to improve the accuracy and effectiveness of small target detection.Different from traditional target detection methods,each predicted bounding box was treated as a Gaussian receptive field,and the ground truth value was treated as a Gaussian heat map.Label assignment was performed by calculating the distance between two Gaussian distributions.A Gromov-Wassertein optimal transport-guided model learning method was introduced,which could be built upon existing detection models.Experimental results on multiple power transmission line target detection datasets demonstrated that the label assignment scheme using Gaussian receptive fields and optimal transport had achieved good performance in small target detection during power transmission line inspection.