Research On The Identification Method Of Foreign Body Invasion In Overhead Lines Based On Improved YOLOv7

In recent years,the probability of problems such as foreign object invasion represented by engineering vehicles and bird nests has increased sharply,which has seriously affected the safety and operation of overhead lines.To ensure the reliable operation of the power grid,inspection departments mostly use manual,unmanned aerial vehicles,video surveillance,and other methods to inspect the transmission overhead lines.However,compared to video surveillance,both manual and drone inspections cannot achieve real-time monitoring,and there is a possibility of missed inspections.With the vigorous development of deep learning in recent years,the research of overhead line hazard identification based on monitoring images has important scientific significance and application value.Therefore,in order to the current problem of transmission line inspection,this paper proposes a deep learning based hazardous object identification algorithm for the overhead line.(1)To address the problem of data set construction caused by the small number and low quality of overhead line foreign object intrusion images,this paper uses four ways of image scaling,flipping,color transformation,and cropping to achieve data enhancement.Then,the enhanced dataset is annotated to get the final training dataset.Based on the original YOLOv7 algorithm to identify the foreign object intrusion before and after data enhancement respectively,the identification results show that the mAP after data enhancement increases from 85.25%to 87.01%,which proves that data enhancement can effectively improve the identification accuracy.(2)To address the problem that the model accuracy decreases after YOLOv7 comes with anchor box clustering,this paper proposes a method of clustering the sample annotated bounding box size by clustering.The clustering results are used as a guide to fine-tune the default anchor box to improve the model’s localization of targets of different sizes.The method reduces the difficulty of bounding box adjustment during the training process and reduces the training loss,resulting in an increase of mAP to 87.5%.(3)To further improve the recognition accuracy of the model,the attention mechanism is introduced in YOLOv7 in this paper.By comparing the recognition effects under different attention mechanisms and mechanism addition positions,the CA attention mechanism is chosen to be added to the connection part of the feature extraction network and the feature fusion network in YOLOv7.Finally,the mAP of the improved YOLOv7 target detection model is improved.In order to verify the superiority of the improved YOLOv7 algorithm proposed in this paper,it is compared with the other five traditional algorithms,and the results show that the improved YOLOv7 algorithm proposed in this paper has the highest recognition accuracy.