Insulator Defects Detection Based On UAV Inspection Images

Insulators is the key component of transmission lines,and regular inspections of insulators is an important preventive measure to ensure the normal operation and stability of the power system.At present,UAV inspection is gradually replacing inefficient manual inspection as the main inspection method in the inspection of transmission lines in our country.Furthermore,With the rise and development of 5G technology and computer vision technology,intelligent inspection has become a key part of the construction of the smart grid.This thesis mainly studies the instance segmentation algorithm of the insulators in UAV inspection images of the transmission lines,the detection algorithm of the overheating defect of composite insulators and the self-explosion defect of glass insulators.The shooting quality of inspection images is directly related to the accuracy of image segmentation and defects detection.In this thesis,Lucy-Richardson algorithm is used to perform deblurring and restoration of inspection images with motion blur,and local adaptive gamma correction is used for image enhancement of low-contrast images.At the same time,a variety of online random data enhancement methods are used to expand the size of datasets.Through analyses and experimental verifications,we choose the first-stage Yolact++as the instance segmentation network of insulators and optimize it properly as follows:Modifying the setting of anchor’s sizes and ratios according to the elongated characteristic of the insulators;Proposing a post-processing algorithm to optimize the output masks;Improving the Feature Pyramid Networks by using path enhancement idea because of the lack of bounding box positioning deviation and imprecise edge segmentation.In the detection of overheating defect of composite insulators,we use the method of skeletonization and filling to optimize the result of the instance segmentation,obtain the position information covering only the core rod part,and then use the orthogonal regression method to fit the relationship between the temperature data of infrared images and the true temperature values to obtain temperature information.Finally,we integrate the position and temperature information to determine the specific level of the overheating defect according to the infrared diagnosis guidance document.In the detection of self-explosion defect of glass insulators,we propose a lightweight key point detection network named IKpNet to end-to-end position the glass insulators.At the same time,we propose a new resize method to avoid image distortion because of the characteristics of different input image specifications.The method of linear fitting and distance threshold calculation for key point coordinates is adopted to realize the division of different strings of insulator slices and the statistical work of the number of self-explosive slices.Finally,all the work in this thesis is integrated into the visual interface for intelligent inspection of insulator defects in UAV inspection images.