基于深度可分离卷积及SVD输电线路缺陷检测  

Methods of Defect Detection in Transmission Line Based on Depthwise Separable Convolution and SVD

作　　者：杨可军 杨建旭 陈思宝[2] 吕苗苗 李程启 杨波 郑文杰 YANG Kejun;YANG Jianxu;CHENG Sibao;LYU Miaomiao;LI Chengqi;YANG Bo;ZHENG Wenjie(Anhui Nanrui Jiyuan Electric Power System Technology Co.,Ltd.,Hefei 230088,China;School of Computer Science and Technology,Anhui University,Hefei 230601,China;State Grid Shandong Electric Power Research Institute,Jinan 250003,China;State Grid Anhui Maintenance Company,Hefei 230022,China)

机构地区：[1]安徽南瑞继远电网技术有限公司,安徽合肥230088 [2]安徽大学计算机科学与技术学院,安徽合肥230601 [3]国网山东电力科学研究院,山东济南250003 [4]国网安徽检修分公司,安徽合肥230022

出　　处：《智慧电力》2020年第10期64-69,77,共7页Smart Power

基　　金：国家自然科学基金资助项目(61976004);国家电网有限公司科技资助项目(SGTYHT/17-JS-199)。

摘　　要：针对输电线路实时巡检中无人机嵌入式移动端的存储和计算受限问题,在Faster R-CNN目标检测框架及VGG特征提取网络基础上,融合深度可分离卷积及SVD分解,构建了一种轻量级的输电线路缺陷检测方法。在公共的PASCAL VOC 2007与COCO数据集上的实验结果验证了所提方法的优越性。在输电线路缺陷数据集上实验,在保证检测精度不下降的同时降低了模型存储空间,提升了模型的检测速度,实现了检测的实时性。In order to solve the problem that the storage and calculation of UAV embedded mobile terminal is limited in the real-time inspection of transmission line,based on Faster R-CNN object detection framework and VGG feature extraction network,this paper combines depthwise separable convolution and SVD to build a lightweight transmission line defect detection method.The experimental results on the public PASCAL VOC 2007 and COCO dataset show the superiority of the proposed method.Furthermore,the experiment on a transmission line defect dataset shows that small storage space for the model and rapid and real-time detection is implemented while ensuring the detection accuracy.

关 键 词：输电线路缺陷检测 深度可分离卷积 奇异值分解 模型压缩 

分 类 号：TM755[电气工程—电力系统及自动化]