Defect Detection And Information Quantification Of UAV Power Inspection Images

In order to meet the development needs of the national industry and science and technology industry,my country’s power scale continues to expand,and the resulting operation and maintenance costs have also increased year by year.At the same time,the rapid development of 5G technology,drone technology and artificial intelligence technology in recent years has spawned the concept of smart grid and promoted its rapid development,which has enabled the effective application of drone patrol transmission lines The inevitable trend of development.This thesis mainly studies the instance segmentation algorithm of key components and defects in the transmission line images of UAV inspection and the method of quantifying the information data of some components.This thesis first introduces the application situation of UAV power inspection and the development status of deep learning instance segmentation algorithm.Afterwards,for poor imaging and low-contrast UAV images,this thesis proposes a fast-estimated dark channel defogging and adaptive gamma correction image enhancement method.At the same time,a variety of online random data enhancement methods are used to expand the training data set,which enriches the location diversity and background diversity of the small sample data set.In the detection and segmentation tasks of power components and defects,the Mask R-CNN algorithm is selected as the basic instance segmentation algorithm according to the characteristics of the object.At the same time,in order to solve the problem of poor positioning accuracy of the long target in the high-resolution UAV image and low detection accuracy of the small target,the path enhancement structure and the adaptive pooling structure are used to improve the utilization of feature maps at different levels Design a feature extraction network that includes grouping ideas and channel attention mechanisms.Compared with the original algorithm,the average detection accuracy of the optimized algorithm is improved by 9.3%.In the quantitative research on the information of glass insulators,this thesis proposes to use the case segmentation algorithm combined with the invariant moment feature to make statistics on the types,number of pieces and number of self-explosive pieces of glass insulators.At the same time,in order to reduce the workload of labeling,an auxiliary labeling method for glass insulator sheets based on clustering and watershed algorithm is proposed.For the information quantification of the fittings(shock-proof hammers,overhanging clamps),the de-lighting process of color maintenance is first used to reduce the interference of shadows on the quantification of the fittings information.Then the super red algorithmcombined with the HS component statistical algorithm is used to quantify the percentage of corrosion.Drawing on the idea of human skeleton point detection,the deep learning key point detection algorithm is used to locate the key points to quantify the deflection angle of the pendant and to correct the angle in combination with perspective transformation.This thesis experimentally tests the segmentation effect and defect detection accuracy of the components in the UAV inspection image,and evaluates the accuracy of information quantification,verifies that the algorithm has good accuracy and strong anti-interference ability,and implements the algorithm On the visual interface.