| The modern society relies more and more on electricity,and it is necessary to guarantee the stable supply of electricity through regular inspection of power lines.When inspecting power lines through UAVs,the corresponding detection algorithm needs to minimize complexity and improve real-time on the basis of high accuracy because of the special detection environment,many types of foreign objects,inconspicuous defects and complex algorithm model,and high real-time requirements.In order to meet the requirements of UAV power line inspection,this paper optimizes the foreign object detection and insulator defect detection algorithms for power lines,and the main research contents include the following three points:1.The initial limited data set of positive samples(foreign objects and insulator defects)is expanded using various data enhancement methods.First,the positive samples are expanded uniformly using spatial transformation,additive noise,and image blurring methods.Poisson fusion is used for data enhancement of foreign object images,which makes the foreign object types more diverse and complex.For insulator defect images,Laplace sharpening is used for data enhancement,which makes the insulator edges,contours and other detailed areas richer.The enhanced data set has higher sample richness and more balanced positive and negative samples,which makes the subsequent model have better accuracy and robustness.2.To address the problem of various types of foreign objects and high real-time requirements in power line foreign object detection scenarios,this paper proposes an improved algorithm based on YOLOv4,replacing the original backbone network with the lighter Mobile Netv3,which significantly improves the model’s operational efficiency and adds a new layer of feature output to improve the accuracy of transmission line foreign object detection by using four-channel fusion.The new layer of feature output and fourchannel fusion are used to improve the accuracy of transmission line foreign body detection.And the average pooling is introduced in the SPP module for better adaptation to the detection scenario.Then,by optimizing the loss function and collecting more information after pooling,we can effectively avoid the situation of missed and false detection and improve the detection capability of small foreign objects.Through comparative experimental analysis with the classical algorithm,it is verified that the improved algorithm in this chapter has good performance and robustness with an average accuracy of 93.43% and a frame rate of 38 FPS in a variety of foreign object detection scenarios,and ablation experiments are designed to verify the impact of adding different optimization measures on the foreign object detection algorithm in power line foreign object detection scenarios.3.To address the problems of lack of images and low defect resolution in power line insulator defect detection,Mobile Netv3 network structure is used to replace the backbone network of YOLOv4 to improve the detection efficiency,and rotating frame object detection method is introduced to solve the detection problem that traditional detection algorithms are difficult to eliminate a large amount of background when insulators are tilted too much in the image.In addition,a dual attention mechanism of spatial attention and channel attention is embedded to enhance the feature information.Through the analysis of fused features,the detection layer can obtain more accurate insulator defect information,thus greatly improving the efficiency of insulator defect detection.The experimental results show that the improved algorithm has an average accuracy of 91.03%and a frame rate of 33 FPS in insulator rotation and small target scenarios,as well as a good leakage detection rate.Ablation experiments are designed to verify the impact of adding different optimization measures to the defect detection algorithm in power line insulator defect detection scenarios. |
