Multi-defect Detection Of Insulators In Transmission Lines For UAV Aerial Photography Based On YOLOV5

The country’s economy is growing rapidly and at the same time the demand for electricity is increasing.The power system and transmission lines are constantly expanding and the size of the transmission lines now typically span hundreds of kilometres.Such a large transport system contains many basic components,which must be inspected regularly by the national grid in order to ensure the safe and reliable operation of the entire system.The time and economic costs are very high if manual inspections are carried out,so UAV inspection are introduced to replace manual inspections.The UAV inspection take pictures during the inspection and send them back to the workstation,where the staff then use traditional image processing,machine learning or deep learning algorithms to detect defects in the underlying components.Of all the components of a transmission line,insulators are a very important part of the system,keeping the electrical and mechanical stress unchanged in the complex natural and physical environment,so UAV inspections are required to regularly check and repair insulator components.In this paper,a cascade detection framework is designed based on the YOLOV5 model.The cascade detection framework can realize the function of multiple defect-detection of insulators.The first stage detection model detects all insulator targets in the original image and inputs them into the second stage model.The second stage model does target detection of insulators,detecting all types of defects in the insulator and outputting their location information.The main research in this paper is as follows:(1)The input image is pre-processed using the Mosaic data enhancement algorithm.The input image is first enhanced with three common data enhancements: colour data enhancement,shape data enhancement and contrast data enhancement,and then the four common data enhanced images are stitched together using the Mosaic algorithm.The common data enhancement algorithm and the Mosaic algorithm expand the data set sample and increase data diversity;(2)Design a cascade detection framework.The structure of cascade detection framework is as follows(symbol-> indicates sequential structure): input image-> Mosaic data enhancement algorithm-> first stage insulator detection model(YOLOV5m model:medium size YOLOV5 model)-> insulator target in the input image-> Mosaic data enhancement algorithm-> Second stage defect detection model(YOLOV5s: lightweight YOLOV5 model)-> all types of defects in the insulator target.The cascade detection framework trained and tested on the data set.Finally,the first stage model’s m AP@0.5reaches 93.1% with an inference speed of 32 FP;the second stage model’s m AP@0.5reaches 89.2% with an inference speed of 34 FP;(3)Transform the structure of the cascade detection framework.In this paper,the ECANet channel attention mechanism and Mish activation function are added to the second stage model.The purpose is to strengthen the model’s weighting parameters for defect information and to facilitate the model’s acquisition of deeper information about the network.The ECANet channel attention mechanism is added by adding the ECA structure after the batch normalization layer(BN)in each basic convolution structure,and the Mish activation function is added by removing the Leaky Relu activation function and adding the Mish activation function in each basic convolution structure.After the transformation,the basic convolution structure of the second stage model is changed from Conv＿BN＿LR to Conv＿BN＿ECA＿Mish.The transformed second stage model is trained and tested on the data set.The final result of m AP@0.5 is 91.1%,with an improvement of 1.9% over the model before the transformation,and the model’s inference speed is 30 FPS.