基于改进Libra-RCNN的输电线路绝缘子识别  

Transmission Line Insulator Identification Based on Improved Libra-RCNN

作　　者：闾海庆 雷远华 王静[2] 邢学敏[3] 杨静[4] LYU Haiqing;LEI Yuanhua;WANG Jing;XING Xuemin;YANG Jing(China Energy Engineering Group Hunan Electric Power Design Institute Co.,Ltd.,Changsha 410007,China;Hunan Third Surveying and Mapping Institute,Changsha 410004,China;School of Traffic and Transportation Engineering,Changsha University of Science&Technology,Changsha 410114,China;China Water Resources and Hydropower Eighth Engineering Bureau Co.,Ltd.,Changsha 410004,China)

机构地区：[1]中国能源建设集团湖南省电力设计院有限公司,湖南长沙410007 [2]湖南省第三测绘院,湖南长沙410004 [3]长沙理工大学交通运输工程学院,湖南长沙410114 [4]中国水利水电第八工程局有限公司,湖南长沙410004

出　　处：《湖南电力》2022年第2期44-49,共6页Hunan Electric Power

基　　金：国家自然科学基金项目(42074033)。

摘　　要：针对无人机航拍输电线路识别绝缘子的定位精度和稳定性较差等问题,提出一种基于ASFF金字塔网络的Libra-RCNN绝缘子检测模型。首先,使用FRN归一化层替代原BN层,消除归一化层对训练批次大小依赖,增加模型学习效率;然后在Libra-RCNN算法金字塔中引入ASFF网络结构,有效解决特征金字塔内部不一致问题;最后借助GIoU交并比替代原IoU交并比,更好精确绝缘子位置。在Insulators_Datasets绝缘子数据集中,改进Libra-RCNN模型平均准确率达94.10%,召回率达97.51%;相较原Libra-RCNN模型分别提高2.23%、2.61%,表明所提算法能稳定、有效地识别绝缘子。To solve the problems of poor positioning accuracy and stability of existing insulators identified by UAV aerial photography transmission lines,a Libra-RCNN insulator detection model based on ASFF pyramid network is proposed.Firstly,the FRN normalized layer is used to replace the original BN layer to eliminate the dependence of the normalized layer on the size of training batch and increase the learning efficiency of the model.Then,ASFF network structure is introduced into Libra-RCNN algorithm pyramid to effectively solve the problem of inconsistency inside the feature pyramid.Finally,GIoU crossover ratio is used to replace the original IoU crossover ratio to better accurate insulator position.In Insulators_Datasets data set,the average accuracy of improved Libra-RCNN model is 94.;0%,and the recall rate is 97.5;%.Compared with the original Libra-RCNN model,the improvement rate is 2.23%and 2.6;%respectively,which indicates that the proposed algorithm can identify insulators stably and effectively.

关 键 词：绝缘子检测 Libra-RCNN模型 FRN归一化层 ASFF网络 GIoU交并比 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TM854[自动化与计算机技术—计算机科学与技术]