Study On Image Recognition And Processing Used By Drones For High Voltage Transmission Lines Inspection

Regular inspection on transmission lines and accessary equipment is a foundamental work of ensuring their operation safety.However,the traditional manual inspection methods have met lots of limits during practical application in nowadays.Recently using helicopter and drones is becoming an important way during transmission lines inspection because of their advantages like high efficiency,high speed,reliability,low cost and all terrain operation mode.The research on drone’s patrol has begun for years.Especially the image processing and recognition is becoming more and more important in the research.For the fault detection on the targets during transmission line inspection,traditional methods have weak performance under complex enviroment.And methods based on deep learning usually try to detect the fault of the targets directly in an object detection manner.But because of the diversity of fault such methods have their limits.In this paper we try to combine the advantages of both traditional methods and deep learning to detect particular fault of the targets.For such purpose we propose a method using semantic segmentation to separate the targets from background before fault detection.This method is convenient for subsequent processing like fault detection.Furthermore we have made some research on detecting the defect and fault of the insulator.For the insulator damage we use two methods which combine dictionary learning,feature extracting by CNN and FFT.For the smudginess detection on insulator we propose a method based on abnormal detection and multilayer perceptrons(mlp).The abnormal detection is an unsupervised learning method which is to extract the abnormal areas on the insulator’s surface as step 1.And mlp is used to filter the result further from step 1.We also give the test results of all the methods proposed.Moreover,in order to implement CNN on high resolution pictures we use BING as the region proposal method before deep neural network.Such region proposal method is to resolve the frame and hardware limits problem in the deep learning.