Research On Aerial Insulator Image Recognition Based On Deep Learning

With the development and progress of living standards,people’s demand for electricity grows rapidly,which requires the construction of a larger power system,and the workload of the corresponding inspection task increases day by day.The traditional manual inspection method is inefficient and dangerous,and the construction of an intelligent inspection system is imminent.At the same time,the development of computer hardware technology drives the development of deep learning theory and technology.Deep learning technology promotes the development of different fields and speeds up the transformation from traditional industry to intelligent industry.In the power scenario,transmission lines are exposed for a long time.On the one hand,they will be affected by topography,landforms and seasonal climate,such as lightning and ice.On the other hand,transmission lines run for a long time,and the surface of power equipment will be polluted,worn and cracked.This thesis starts with the insulator on the power transmission line,and achieves the purpose of intelligent patrol inspection by using image processing technology and deep learning algorithm to obtain the insulator image obtained by UAV aerial photography.The main contents of this thesis are as follows:Firstly,the obtained insulator images are preprocessed,and the image enhancement technology is used to enrich the insulator features.For the low-illumination images in the data set,the histogram equalization algorithm and the multi-scale Retinex algorithm are used to enhance the images.Finally,the improved Retinex algorithm is used to better solve the problems of uneven illumination and halos.On this basis,the original data set is expanded by means of rotation and scaling.In terms of detection methods,this thesis first uses the traditional target detection algorithm and combines the HOG feature extraction algorithm with the SVM classifier,but the test results are not ideal.Secondly,it proposes the target detection algorithm YOLOv3 based on the convolutional neural network,and improves YOLOv3 algorithm by using K-means++ and Softer-NMS.The algorithm can achieve high detection accuracy,but the effect of marking box is not very ideal.This thesis uses an insulator detection method based on Gaussian Yolov3.When using Gaussian Yolov3 training,the multi-stage transfer learning method is added,and the original Gaussian distribution strategy of the algorithm is combined.In terms of results,the algorithm can accurately regressive the position of the object.Compared with the Gaussian Yolov3 under the same conditions and the improved Yolov3,The accuracy of insulator detection is increased by 3.0 and 1.0 percentage points respectively,and that of defect detection is increased by 10.5 and 5.5 percentage points respectively.Through the above steps,the identification algorithm based on the detection of insulators in power scene is designed,which can be effectively applied to the inspection task of UAV and help the staff to quickly detect the location of insulator defects.