Research On Transmission Line Object Detection Based On Convolutional Neural Network

Transmission line is the main component of the national power network transmission.In order to ensure the continuity,safety and reliability of power supply,it is very important to regularly inspect the transmission line.Traditional transmission line inspection is mainly manual,which has the problems of low efficiency and high danger.With the development of technology in recent years,unmanned aerial vehicle inspections have become mainstream.The combination of image processing technology and unmanned aerial vehicle inspection to the detection of the transmission line operation status not only greatly improves the automation level of transmission line inspection,but also reduces the investment cost.This paper studies the device identification and positioning methods in the transmission line inspection.By using convolutional neural network,UAV aerial images are trained to identify and locate transmission line devices.The main work of the paper is as follows:(1)Establish transmission line data set.At present,there is no standard data set.The video captured by the thermal infrared imager is converted into pictures,and clear images are selected as the original data set.Aiming at the problem that the image data set is small,data enhancement methods such as image flip,rotation,shift,and noise are used to achieve data expansion.Lableimg software is used to manually label the data set,mark the three types of transmission line,tower and insulator string in the transmission line data set,generate corresponding.xml file,and make it into the format of V0C2007 data set.(2)Design and implement the transmission line device identification algorithm based on convolutional neural network.By comparing the advantages and disadvantages of the current deep learning target algorithms,Faster RCNN algorithm and YOLOv3 algorithm are selected to train the transmission line data set.In the training process of Faster RCNN algorithm,three feature extraction networks of VGG16,ResNet101 and MobileNet are used to extract features of the image.For some targets in the image are relatively small,a set of small scale is added in the selection process of candidate box of regional proposal network(RPN)to improve accuracy of object detection,and the final mean average precision(mAP)value could reach 0.9054.Experiments prove that the method has a good effect in the realization of transmission line object recognition.(3)Realize the identification and positioning of the insulator string data set.The visual insulator string data set is obtained from GitHub.The normal insulator string data set in the data set is selected,and the normal insulator data set is expanded to 2,400 pieces by flipping,rotating,shifting,adding noise and so on.Through the Faster RCNN algorithm and YOLOv3 algorithm training,the experimental results are compared with the results of references.The improved Faster RCNN algorithm has better experimental results.The average precision(AP)value of the insulator string can reach 0.9911,which can accurately identify and locate the insulator string.