| The "14th Five-Year Plan" pointed out that it is necessary to accelerate the intelligent transformation of power grid infrastructure and the construction of smart grids.In power transmission and transformation scenarios,it is of great significance to regularly conduct UAV inspections on tower insulators.In the actual inspection process,insulators often have inconsistent scale defects such as pollution,damage,and self-explosion.There are still two problems in the research on insulator image detection after inspection.First,in order to improve the small target detection accuracy of the model,it is necessary to integrate multiple Second,the large amount of edge computing puts forward lightweight requirements for model deployment.In view of the above problems,it is planned to expand the insulator data set through data enhancement algorithm to obtain more insulator sample images.Then,based on the YOLOv5 s model,multiple scale feature maps are fused to different degrees to improve the accuracy of target detection.Finally,the lightweight feature of the PP-LCNet module is used to optimize the backbone network of YOLOv5 s,reduce the amount of model calculation,and improve the speed of model reasoning.The specific work content is as follows:Firstly,the initial insulator image data set is expanded through algorithms such as histogram equalization enhancement,noise transformation and affine transformation to meet the demand of the deep learning model for the number of samples.Secondly,in order to fully integrate multi-scale features,the YOLOv5 s model is optimized from the following three aspects:(1)A new multi-scale detection head is added for the low resolution of small targets,which leads to missed detection of insulator defects.In the existing YOLOv5 model Based on the three scale feature maps of 20×20,40×40,and 80×80,a small target scale of160×160 is added to reduce the high-level semantic information lost by the feature map.(2)In the process of feature extraction,in order to enhance effective features and weaken useless features such as complex backgrounds,the CBAM attention mechanism is introduced into the Neck structure to make the feature information conveyed by the network more accurate.(3)In order to enhance the feature processing capability of the Neck side,on the basis of the original two feature fusion paths,two new horizontal cross-scale connection paths are added to enhance the information transfer between different network layers and enrich the feature information of small objects.Again,in order to reduce the amount of calculation and obtain a lightweight network model,the YOLOv5 model is improved based on the Mobile Net V2,Shuffle Net V2 and PP-LCNet modules respectively.The comparison shows that the obtained P-YOLOv5 model improves the model detection speed while ensuring that the IOU value is within the allowable loss range,and achieves the best balance between performance and speed.Finally,under the sample size of 3166 insulator image datasets obtained from the inspection of dry-shaped transmission towers of 110 kV overhead lines,1900 training sets and 633 test sets were divided into simulation experiments.The results show that in terms of detection accuracy,the average precision m AP of the proposed YOLOv5s-PCB model reaches 95.4%,an increase of 3.1%,and the average recall rate Recall is 95.2%,an increase of 1.9%.In terms of model complexity,the parameter param of the improved lightweight model is only11.32 M,a reduction of 58.58%,and the model inference speed FPS reaches 103,an increase of 58.46%. |
