Research On Aerial Insulator Recognition And Defect Detection Method Based On Deep Learning

As an important insulating medium for supporting wires and towers of overhead transmission lines,insulators directly affect the safe and stable operation of the transmission system.Transmission line inspection has gradually changed from shooting by line inspectors to using drone aerial images or video for inspection.The use of human vision for inspection has disadvantages such as low efficiency and large subjective judgment.Although there are certain researches on the detection of insulator defects combined with computers,there are still disadvantages such as slow detection speed and low efficiency.This paper proposes an aerial image insulator identification and defect detection method based on the target detection model in deep learning,focusing on the use of deep learning and computer vision technology to analyze the line inspection image data taken by drones,so as to realize the identification and defect of insulators Testing,specific research content,work done and results are as follows:First,in view of the uneven illumination of the insulator image in the UAV aerial inspection image,the histogram equalization method,the homomorphic filtering method based on the illumination-reflection model,and the retinex image enhancement method with color restoration(MSRCR)are used respectively.Three image enhancement methods were compared and tested,and finally a set of ideal MSRCR parameters were debugged to achieve better preprocessing results.The processed images are used to construct a data set containing 4300 insulator images.Secondly,it analyzes in detail the current mainstream target detection models based on deep convolutional neural networks: Faster R-CNN,SSD and YOLO series models,and analyzes the advantages and disadvantages of the three types of models and the summary of excellent target detection ideas worth learning.A comparative experiment was carried out on the constructed insulator image data set.Finally,in order to improve the detection accuracy of quasi-insulator identification and defect detection,an application-based improvement study of the YOLOv3 target detection model was carried out.Using asymmetric convolutional blocks as structural blocks to build a basic convolutional neural network,a K-means++ algorithm with improved distance indicators is proposed to achieve more effective clustering results for preselected boxes,and a multi-scale detection model is added for the problem of defective small target detection.The improvement scheme of training loss function is given and the parameter adjustment process is introduced.The method of migration learning is adopted during training,and the migration data source is the VOC2007 data set.Aiming at the problem of the readability of the output result,a detection method that increases the logical judgment layer is adopted to improve the readability of the detection result.On the self-labeled data set,the average precision reached 91.9%,and the detection speed reached 29.8FPS,meeting the engineering needs of overhead line inspection.