Research On Insulator And Damage Detection Algorithm Of Improved YOLOv5

Insulators are key electrical components commonly used in high-voltage transmission lines to insulate overhead lines from towers and the earth.It is exposed to the complex natural environment and electromagnetic environment all year round,and is highly susceptible to breakage,which in turn affects the safe and stable operation of transmission lines.Traditional insulator breakage detection adopts manual patrol,which consumes a lot of manpower,is inefficient and ineffective.In recent years,the intelligent inspection method based on UAVs（Unmanned Aerial Vehicle）has become the development trend of insulator breakage detection.However,the aerial images of UAVs have the characteristics of complex background,small and overlapping targets,which cause difficulties for the subsequent intelligent analysis of insulators and their damage.To this end,this dissertation purposefully proposes an insulator and its breakage detection algorithm based on improved YOLOv5,considering both detection accuracy and speed,providing a new idea for UAV intelligent inspection.First,for the complex and diverse background of aerial images,this dissertation introduces the attention mechanism module ECA-Net in the algorithm of YOLOv5 backbone feature extraction layer.The algorithm achieves both weighting the different amounts of information contained in the feature channels,while taking into account the degree of interdependence between neighboring feature channels,And the k neighboring channels are used to interact and pass the information to the next layer of the network,and finally to enhance the weight of target features to reduce the weight of background features,so as to distinguish the background from the target,and the module does not increase the computational complexity of the model.Second,for the characteristics of small targets,this dissertation combines the Bidirectional Feature Pyramid Network in the enhanced feature extraction layer part of YOLOv5.The algorithm combines features at different scales as well as more shallow features at the same scale,and repeats the stacking with a Bi-FPN as a cyclic unit,so as to effectively compensate for the loss of targets in small feature maps and the loss of target location information in large feature maps,thus improving the detection capability of small targets.The algorithm also eliminates network layers that are less useful for the feature fusion layer,reducing the amount of model computation.Finally,for the characteristics of insulator target overlap,this dissertation combines the new candidate frame selection algorithm Soft-NMS（Soft Non-Maximum Suppression）in the YOLOv5 model.The algorithm considers both the confidence score of the candidate frame and combines the overlap score of the candidate frame with the real frame.The main implementation process is to take the Gaussian index of the overlap score and weight it with the confidence score,and then perform a threshold-based rejection of the weighted score.This process reduces the probability of rejection of candidate frames with high overlap,improves the selection ability of the model in candidate regions,and ultimately achieves improved detection accuracy of adjacent overlapping insulators.In this dissertation,we construct a dataset containing ceramic and glass insulators by designing histogram equalization,affine transformation and noise enhancement digital image processing techniques.There are 1059 insulator targets in the data set and 808 broken insulators.Finally,by designing a deep learning-based target detection algorithm for comparison experiments,it is concluded that the improved YOLOv5 model improves the insulator localization accuracy from 93.30% to 94.68%,and the insulator breakage localization accuracy from 90.63% to 95.04%.The real-time frame rate reaches 53.4 FPS,which verifies the effectiveness of the proposed method in this dissertation.