电瓷避雷器Issue(6)：106-116,11.DOI:10.16188/j.isa.1003-8337.2025.06.013
基于Res-CapsNet与改进YOLOv4的绝缘子破损识别与定位
Insulator Damage Identification and Location Based on Res-CapsNet and Improved YOLOv4
摘要
Abstract
Aiming at the problems of poor recognition performance and slow speed of traditional convolu-tional neural network(CNN)in insulator damage detection,the author proposes an algorithm based on Res-CapsNet(residual capsule network)combined with improved YOLOv4,which includes two parts:insulator classification detection and damage location.Firstly,since the residual network can solve the problem of model degradation caused by stacking convolution layers in the traditional classification net-work,ResNet34 is proposed as the pre-training model to extract insulator image features.The extracted convolution features are converted into capsule features,and then the dynamic routing algorithm is used for transmission to ensure the integrity of the feature information.Therefore,not only the direction and angle of the output can be retained,but also the deeper characteristics of the insulator can be extracted,so as to realize the accurate identification of damaged insulators in complex environment.For insulator damage location,Res2Net residual unit is used in CPSDarknet53 to extract insulator fine features,and CBAM attention mechanism is introduced into the location network to pay attention to insulator profile,position and other features to improve the accuracy of the model.At the same time,due to the introduc-tion of Res2Net residual element,the model complexity is increased,so the channel pruning compression network model is adopted to reduce the model computation,which can achieve a high precision while speeding up the model training speed.Finally,compared with SSD,VGG16,Resnet,AlexNet and other networks,the experimental results show that the improved network can achieve 97.98％identification ac-curacy of insulator damage and 96.57％location accuracy,which can quickly and accurately detect the insulator fault state under different weather conditions and distances.The efficiency of intelligent inspec-tion on insulators is greatly improved.关键词
绝缘子破损检测/Res-CapsNet/YOLOV4/Res2Net/CBAM/通道剪枝/智能巡检Key words
insulator damage detection/Res-CapsNet/YOLOv4/Res2Net/CBAM/channel pruning/intelligent inspection引用本文复制引用
卞建鹏,朱泽明,陈璇,安荣廷..基于Res-CapsNet与改进YOLOv4的绝缘子破损识别与定位[J].电瓷避雷器,2025,(6):106-116,11.基金项目
河北省自然科学基金面上项目(编号:E2021210069).Project supported by Hebei Natural Science Foundation General Project(No.E2021210069). （编号:E2021210069）
