Research On The Detection Of Porcelain Insulator Drop String Based On Improved YOLOv5

The unmanned aerial vehicle has been developed better and better over last recent years.UAVs are popularly used in insulator detection in power systems.Compared with manual detection,the UAV combined with object detection is more safe,convenient and efficient.You Only Look Once version 5(YOLOv5),as one of the latest representatives of object detection algorithm,has the characteristics of fast detection,good precision and small model,which can increase the advantage of UAV detection.However,the detection precision of YOLOv5 needs to be further improved when the background of porcelain insulator is complex,the object is small and the object is overlapping.Therefore,aiming at this problem,this thesis improved the YOLOv5 model to improve the detection precision.First of all,the complex background environment of porcelain insulators will lead to missed detection of insulators,which will reduce the detection precision.The Coordinate Attention mechanism(CA)introduced in this thesis can encode channel attention as two one-dimensional features aggregated along different directions,and embed location information into channel information to preserve accurate location information.It not only expands the receptive field,but also enhances the attention to the target,and increases the distinction between the insulator and the background.Secondly,the use of Bidirectional Feature Pyramid Network(Bi FPN)is to replace FPN and PANet to solve the problem of small target detection.It can not only fuse multi-scale image features,but also fuse shallow features at the same scale.The detection precision of small targets is improved by repeating stacking to prevent the loss of small targets.Furthermore,for the target overlap(occlusion)caused by missing detection problem;In this thesis,SIoU Loss function is used to replace CIoU Loss function of the original model to train the model.CIoU Loss relies on the aggregate bounding box regression index to train the model,ignoring the mismatch between the real box and the predicted box.This causes the model to get worse.SIoU Loss can solve the deficiency of CIoU Loss.Using SIoU Loss to train the model can effectively improve the detection precision.Finally,the three methods were integrated into the YOLOv5 model at the same time.According to the ablation experiment results,the average precision(m AP@0.5)of the improved model was increased from 84.3%to 89.6% compared with the original model,and the improvement results were significant.The experiment proves that the improvement of YOLOv5 precision in this thesis is feasible,and it is of great practical significance to apply it to UAV to detect porcelain insulator string dropping.