Insulator Location Based On Recurring Pattern Recognition

Location of insulators on transmission lines and timely detection of faults are important research contents of smart grid.In recent years,the use of unmanned aerial vehicles(UAVs)for high-altitude transmission line inspection has become more and more popular and has become an important way of power grid inspection.Automatic insulator location in the images captured by UAV can further reduce the workload of manual search.However,the background of UAV inspection is very complicated,including houses,towers and other distractors.UAV fly around the tower during the inspection,which makes the image visual angle and illumination change very much,and the occlusion is frequent.These all bring great challenges to the location of the insulator.This paper uses the computer vision theory and technology to locate the insulators in the aerial images with complex background and solves the problem of target location from two ideas without supervision and supervision.By analyzing the characteristics of the specific mode of insulators,an insulator location algorithm based on recurring pattern detection is proposed.This unsupervised method can detect the insulators in the aerial images of complex background and obtain the equivalent localization effect of the supervised method.On this basis,it extends into a pattern detection framework with a certain shape prior.In the supervised insulator location,we use the deep learning method to locate the insulator and get a higher accuracy.The main work of the paper is:1.An unsupervised method of insulator location based on recurring pattern detection is proposed.For the first time,the method is introduced into insulator location,and it can be used for multi-scale detection conveniently.This method takes MSER as the basic feature and makes full use of the nested characteristics of the group.A visual word aggregation method based on spatial location constraint and similarity measure is proposed.Based on the Helmholtz principle,a scoring model is set up,and a quadratic programming method is designed to further screen the representative features of visual words.Finally,the insulator is located according to the joint optimization framework.It achieves a 77.8%accuracy rate and a 82.3%recall rate on the constructed dataset.Experiments show that the unsupervised method can achieve robust positioning of insulators with different shooting angles,resolutions and different material types under complex background.2.The original recurring pattern detection framework is extended to a certain shape prior detection framework.In the course of the experiment,using the same straight line and equal interval shape priors as examples,by using the SIFT descriptor,experiments are carried out in natural images and insulator images with a priori shape of a visual object in the same line and equal intervals.The experiment proves the feasibility of the method.3.This paper describes several common end to end deep learning target detection networks.The network model SSD under the Caffe framework is used to locate the insulators in aerial images,and 85.778%accuracy and 88.957%recall were achieved.The feasibility of the method was verified by experiments.In addition,since there is no open insulator image data set,an aerial image data set is constructed and manually tagged.The data set contains insulator images with various materials,multiple angles and complex backgrounds.An oblique rectangular box annotation tool was built to manually label the dataset.We also set up an experimental platform for the algorithm studied in this paper,including stepwise debugging,batch identification and evaluation.