Research On Insulator Defect Identification Algorithm For Transmission Lines Based On Improved YOLOv5s

In recent years,with the gradual improvement of China’s Beidou satellite system and the popularization of civil drones,drone inspection technology has also been widely used in the inspection of transmission line insulators.However,there are many types of insulators in aerial images,with complex backgrounds and easy to be obscured,which brings great difficulties to the identification of insulators and their defects.This thesis proposes an insulator defect detection model based on the improved YOLOv5 s algorithm,which is optimized and improved in terms of improving detection performance and model lightweight.Two sets of detection models for insulator defects are trained,and the effectiveness of the improved model detection is verified.The main work of this article is as follows:(1)First of all,in response to the current problems of the insulator dataset,such as a small number of samples,a single type,and an uneven number of positive and negative samples,sample expansion and data enhancement were conducted on the original dataset,enriching the insulator dataset samples,and using the Label Img annotation tool to annotate the dataset.(2)Secondly,the performance optimization of insulator defect detection algorithm for transmission lines based on YOLOv5 s is studied.Aiming at the problem of difficult identification of insulator targets in complex backgrounds,it is proposed to add attention modules to the feature extraction section of the YOLOv5 s network,and add multiple attention modules to the network model for comparative experiments.Finally,the SE attention module is selected.The experimental results show that the network’s feature learning ability for the target region has been effectively improved.Aiming at the problem of small insulator defect targets that are difficult to identify,it is proposed to introduce a context augmentation module into the YOLOv5 s network,adding three fusion methods from the context augmentation module to two locations in the network,and comprehensively comparing the improved core indicators such as detection accuracy,model size,and parameter quantity.Finally,this article chooses to add a context augmentation module to the 10 th layer of the network,and adopts a weighted fusion method to splice it with the 21 st layer to improve the recognition ability of the model for small targets such as insulator defects.In view of the slow convergence speed of YOLOv5 s model,Focal EIOU loss function is proposed to replace the original CIOU loss function to solve the problem of sample quality imbalance in the data set and improve the positioning effect of the prediction box.The three improved methods are combined to obtain the final optimization model.The experimental results show that the detection accuracy for insulator defects is increased from 90.24% to 94.02%,and the average detection accuracy is increased from 93.15% to 95.74%,verifying the effectiveness of the improved algorithm.(3)Finally,lightweight improvements were made to the YOLOv5 s model.In response to the high complexity of the YOLOv5 s network model and its difficulty in deployment on drone devices,this paper proposes to use deep separable convolution and channel shuffling to lightweight improve the YOLOv5 s network model.Comparative experiments are conducted on insulator datasets to obtain the impact of different improvement measures on model performance.Finally,it is found that using the Mobile Net V2 module to replace the original feature extraction network reduces the model parameter count by 47.9%,The computational complexity decreased by 60.1%,while the average detection accuracy decreased by only 4.87 percentage points.The improved method performs best in terms of overall performance and can be better deployed on drone devices with limited memory and computing power.The thesis has 56 figures,13 tables,and 70 references.