| High voltage transmission line is the essential medium for transmitting electric power today.The scale of high voltage transmission lines is likewise expanding yearly in tandem with the steadily rising demand for electric power.As a result,regular inspection of high voltage transmission lines is critical to ensuring the safe and steady transmission of power.At the moment,it has become a common inspection mode to obtain images of high voltage transmission lines by UAV inspection.However,if these massive inspection images are analyzed and judged only by manual,the detection efficiency is low,and the situation of missing detection will inevitably occur.Therefore,an intelligent detection technology is urgently needed to detect the images of high voltage transmission lines accurately and quickly.Based on deep learning technology,this thesis realizes the intelligent detection of multiple inspection objects in high voltage transmission line images which using improved target detection algorithms,further improves the efficiency of UAV inspection and promotes the development process of intelligent inspection for high voltage transmission lines.The main works are as follows:(1)Aiming at the problem that there is no public image data set of high voltage transmission lines,the image data set of high voltage transmission lines is constructed by ourselves.The data was enhanced by flipping,rotating,zooming,translating and changing brightness of the original image,so as to increase the number of the original sample data set and mark the images.Finally,3154 images of two types including insulator and bird’s nest were selected as the experimental data set.(2)Aiming at problems in high voltage transmission line images target detection,such as low recognition rate of small target,comliex background and target overlap,this thesis presented a target detection network model based on improved YOLOv5 algorithm.Firstly,the structure of multi-scale detection network is optimized,and 160×160 detection feature map is added to improve the recognition rate of small target bird nests.At the same time,the weight value of the model on the vital features is increased by introducing the the CBAM attention mechanism in the backbone feature extraction network.In addition,the original non-maximum suppression algorithm is improved,which to increase the recognition rate of overlapping target objects.Through experimental analysis,the method is able to achieve better detection results for high voltage transmission line images with complex backgrounds,achieving a detection performance of 94.92% average accuracy mean for insulators and bird nests.(3)A cascaded network-based insulator defection method is proposed for the problem of low detection accuracy due to the small insulator defect location area.Firstly,the improved YOLOv5 model was used to detect the insulator location region in the image and cut it down.Then,the improved Faster R-CNN model was used to detect the defect location region in the clipped image.This method effectively improved the detection accuracy of insulator defects.By comparing the experimental results of insulator defect detection by single network structure and cascade network structure,the feasibility and effectiveness of the cascade network proposed in this thesis for insulator defect detection is verified,and the average accuracy of insulator defect detection performance of 91.86% is realized,which is 13.18% higher than that directly detected by the improved YOLOv5 algorithm. |
