| The stable operation of the power system is a necessary condition to ensure the normal operation of the country,and the power inspection undertakes the important task of ensuring the normal operation of the power system.The object detection algorithm is applied to the power inspection task to realize intelligent inspection,which can significantly improve the inspection efficiency.Based on the bird’s nest detection operation in power inspection,this paper uses a deep learning-based object detection algorithm to study the bird’s nest target detection method in UAV aerial photography power tower pictures,and designs and develops an intelligent detection system for bird’s nest in power inspection,and the relevant work content is as follows:First,power lines and related facilities belong to the national energy basic resources,it is difficult to obtain enough UAV inspection data from the Internet,in order to smoothly carry out relevant algorithm research,this paper makes a self-made power inspection bird’s nest object dataset.Firstly,the data is collected by the UAV cruise power line,and the pictures with bird’s nest objects are screened out,enhanced and annotated,and finally a power patrol bird’s nest object dataset containing more than 4100 pictures is constructed.Second,aiming at the problems of small size,similar texture features and natural background and partial occlusion by the tower body in the data obtained by UAV aerial photography,this paper conducts research and experiments based on the Sparse R-CNN algorithm with better performance among the existing object detection algorithms,and proposes L-FPNet,an aerial picture bird’s nest detection algorithm based on low-scale feature fusion.Firstly,the "ex-grid" proposal box initialization method is proposed to improve the positioning accuracy.Secondly,the Re LU activation function of the original network dynamic instance interaction head is replaced with the Si LU function,so that the network model has stronger convergence ability.Finally,the feature pyramid structure "LFPN" that integrates low-level features is designed to improve the richness of the feature map and improve the prediction accuracy of the model.The experimental results show that L-FPNet shows high detection accuracy in aerial bird nest detection.Third,in order to realize intelligent inspection,an algorithm with lightweight model and easy integration and deployment is required,and the algorithm should have high detection accuracy and fast detection speed.To solve the above problems,a bird’s nest detection algorithm for aerial pictures based on global attention is proposed,GA-YOLOv5.Based on the YOLOv5 algorithm with relatively light weight and fast detection speed of network model,the K-means++ anchor box clustering algorithm is introduced to improve the object positioning accuracy,and the GAM global attention mechanism is integrated at the neck layer of the network to amplify the global dimension interaction features,reduce information loss,and improve the prediction classification accuracy of the network model.Experiments show that compared with the more advanced algorithms,the GA-YOLOv5 algorithm shows better detection performance in the detection of bird’s nests in aerial pictures,and the weight prediction file generated by training is small,which is easy to integrate in the software system.Fourth,the intelligent detection system of electric power inspection bird’s nest was designed and developed,which realized the functions of intelligent detection of bird’s nest,display of results and generation of test reports.The system is integrated into a lightweight all-in-one machine with graphics computing capabilities,and has been applied and tested in actual scenarios,which shows that the system has the ability to efficiently complete the detection of power inspection birds’ nests. |
