Research On Insulator Object Detection Based On Deep Learning

Power engineering is the fundamental industry related to the development of the national economy.Insulator detection is an important assignment in the field of the operation and maintenance of power system.In practical applications,the insulator images collected by drone aerial photography have complex backgrounds,incomplete objects,and occlusions,which brings difficulties to the subsequent detection.The thesis takes the identification requirements of insulators and other components in the field of power inspection as the starting point,and combines the actual project to carry out insulator object detection research based on deep learning.Through comparing the network principles and model characteristics of the common deep learning algorithms,the thesis aims to find algorithm models that match the needs of engineering projects.The insulator detection algorithm based on YOLOv3 and the insulator detection algorithm based on CornerNet-Lite are respectively proposed,and the loss function of the algorithm is improved.The main work of this thesis is as follows:According to the principle development of object detection technology,due to the two-stage and single-stage in the deep learning algorithm,the visual analysis of the structure of the network model and the reproduction of the process steps are classified,as well as combing with the application requirements of the actual project.By comparing the performance characteristics of the algorithm,the thesis selects the applicable algorithms YOLOv3 and CornerNet-Lite.Aiming at the performance effects of the existing deep learning algorithms and the research situation of insulator object detection,an insulator object detection algorithm based on YOLOv3 is proposed.Based on the data processing requirements of aerial images in the project,the thesis applied the YOLOv3 algorithm to insulator object detection in the field of power inspection,which improves the efficiency and safety of transmission line safety supervision work,and solves the problem of using machines to replace manual identification technology.By using the self-built insulator database to train the model,it can be seen that the algorithm can meet the real-time and accuracy requirements in engineering practice.Aiming at the common problems of missed detection and error detection in the field of insulator identification,an insulator object detection algorithm based on CornerNet-Lite is proposed.To think of the phenomenon of staggered frames and small objects insensitive in the experimental results of the YOLOv3 algorithm model,the thesis constructs a corner classification model based on CornerNet-Lite,and based on the performance of the algorithm model,an improved loss function algorithm that expands the vector distance between different types of corner points is proposed.By comparing the detection results before and after the improvement,it can be seen that the detection ability of the CornerNet-Lite model is improved.The experimental results show that the single-stage algorithm can meet the requirements of real-time detection in practical applications under the premise of ensuring accuracy.Comparing the detection results of the network model on the selfbuilt insulator dataset and the COCO dataset,the insulator object detection algorithm based on YOLOv3 and CornerNet-Lite is suitable for practical engineering projects,and the CornerNet-Lite algorithm based on the improved loss function can effectively improve the accuracy of the object recognition and solve the problems of inaccurate object positioning and wrong recognition,At the same time,it can also solve the phenomenon of the staggered frame recognition between the same object categories due to the high overlap of insulator strings.