Research On Insulator Defect Detection In Aerial Images Based On Deep Learning

With the steady and high-speed development of various industries in the country,the demand for electricity continues to increase,and transmission lines are also expanding.Only by the end of 2016,it has reached 1.756 million kilometers,ranking first in the world.The problem of safe operation of the company ha s also become prominent.Insulators are one of the most common and defect-prone components in transmission lines.Traditional manual inspections have gradually withdrawn from the stage of history due to high human resource consumption and poor safety.UAV aerial photography for transmission line inspections has become more and more popular.The more popular.Following this,a large amount of image data brought by drone aerial photography needs to be trouble-shooted,so it is of great practical significance to detect the defects of insulators in aerial images.In this paper,a deep learning-based defect detection method for aerial insulators is carried out.By comparing the existing target detectio n algorithms,the YOLOV4 algorithm with outstanding detection accuracy and detection speed is selected as the basic algorithm of the research.The text collects open s ource images of insulators in many aspects,and uses a variety of methods to expand the data set,and continuously optimizes the insulator defect detection model.The following three aspects are mainly done:(1)An improved YOLOV4 model for aerial insulator detection is proposed,and model compression,SENET attention mechanism,and CBAM at tention mechanism are introduced into the traditional YOLOV4 network.Through sparse training channel pruning to reduce redundant channels,the model p arameters and size are reduced by 75%,and the inference time is reduced by 45%,in the YOLOV4 network.By adding the SE attention module and the CBAM attention module to the three positions of the YOLOV4 backbone network,six improved networks were built.Among them,the YOLOV4-CBAM 3 of the CBAM attention module was added after the last set of residual blocks in the YOLOV4 backbone network.The network performance is the best,and the effect of the feature extraction heat map is the best.The detection accuracy of the insulator detection model t rained by this model is93.15%,which is 2.67% higher than the original YOLOV4 mo del.(2)A cascaded detection network is proposed to solve the problem that the proportion of defects in aerial insulator images is too small,and to more accurately detect defective parts in aerial insulator images.The detection network is composed of a two-stage detection network.The first stage uses a data set composed of the entire aerial image system,and uses YOLOV4S-CBAM network structure training to obtain insulator positioning.The second stage uses the insulator image cropped from the aerial image as the data set.Also use YOLOV4S-CBAM network structure training to obtain the location of the defect in the insulator.Experimental verification shows that the accuracy of the cascade detection network for insulator defect detection is 90.73%,which is 10.07% higher than that of the original YOLOV4 model.(3)Deploy the cascaded detection network model to the Web,adopting the design mode of separation of front and back ends,using t he VUE framework in the front end and the Flask framework in the back end,making the cascading detection network model applicable in engineering,which is convenient for more users It is more convenient to use this model to detect insulator defects in aerial images.