特征优化Faster R-CNN绝缘子检测与定位  

Insulator Detection and Location Based on Faster R-CNN with Optimal Features

作　　者：刘艳琼 张建州[1] LIU Yanqiong;ZHANG Jianzhou(College of Computer Science,Sichuan University,Chengdu 610065)

机构地区：[1]四川大学计算机学院,成都610065

出　　处：《现代计算机》2021年第9期125-130,共6页Modern Computer

摘　　要：绝缘子在高铁架空输电路中起支撑接触网导线和防止电流回地的作用,绝缘子的好坏对高铁线路的安全十分重要。Faster R-CNN因其特征提取网络VGG的网络层数较深,能适应多数场景下的目标检测,但其检测时间相对较长、对小目标和遮挡目标检测精确度低,难以满足绝缘子检测中的高速率、高精确度要求。基于此,借鉴Inception网络中的卷积核组合和非对称卷积核的设计思想,提出一种新的特征提取网络。新的特征提取网络相比VGG网络深度减少而维度增加,缩短了特征提取时间,增加了特征的低维信息和多维性,提高了绝缘子识别速率和识别精确度。实验分别与Faster R-CNN网络、SSD网络和YOLOv3网络对比检测了四种环境下的绝缘子,对比实验结果表明,基于Faster R-CNN网络的改进方法最终在绝缘子检测中达到98.4%的精确度,相比Faster R-CNN网络检测精确度提高了14.3%,而检测时间降至Faster R-CNN网络检测时间的59%,能满足复杂环境中绝缘子的检测要求。Insulators play a role in supporting catenary wires and preventing current from returning to ground in overhead transmission circuits of high-speed railways.The quality of insulators is very important to the safety of high-speed railway.Faster R-CNN can adapt to target de⁃tection in most scenes because of its deep network layers of the feature extraction network VGG,but its detection time is relatively long and the detection ac-curacy for small or occluded targets is low,which makes it difficult to meet the requirements of high speed and high accu⁃racy in insulator detection.In this paper,a new feature extraction network is proposed based on the combination of convolution kernel and asymmetric convolution kernel in the Inception network.Compared with the VGG network,the new feature extraction network has reduced depth and increased dimensions,shortened the feature extraction time,increased the multi-dimensional feature extraction,and improved the rate and accuracy of insulator recognition.Experiments compare our method with conventional Faster R-CNN network,YOLOv3 net⁃work and SSD network,and the result shows that the method of this paper finally achieves 98.4%accuracy on the insulator detection,which is 14.3%higher than that of conventional Faster R-CNN network.While the detection time decreases to 59%of time cost of conventional network,which can meet the requirements of insulator detection in complex environments.

关 键 词：绝缘子检测 Faster R-CNN INCEPTION 非对称卷积核 卷积核组合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术] U225[交通运输工程—道路与铁道工程]