Research On Real-time Identification And Location Method Of Insulators In Aerial Image

Power insulators are resistant to voltage and mechanical stress and are indispensable controls in overhead transmission lines.Since insulators are usually made of glass and ceramics,they are prone to cracks,breakages,and the like,which affect the normal operation of overhead transmission lines.Therefore,it is necessary to carry out periodic state detection of the insulators in the transmission line.The real-time identification and location of the insulators in the aerial image by using equipment such as UAV and then taking photos of insulators are necessary prerequisites for state detection.Therefore,this paper studies the method of real-time identification and localization of insulators in aerial imagery for devices such as UAV without GPU.The main research work is as follows:(1)After studying the current state-of-the-art identification and localization methods,this paper proposes an object identification and location method based on deep learning to perform real-time identification and localization of insulators.This paper conducts a comparison test of speed and accuracy under the same conditions.Finally,because the YOLOv2 algorithm can balance speed and accuracy and is superior to other models,this paper chooses YOLOv2 algorithm as the basis of this research and impro ves it under the framework.(2)Based on the basic network structure of YOLO-LITE algorithm,this paper proposes three lightweight full convolution blocks that can be used for stacking neural networks,and uses the k-means algorithm to cluster the bounding boxes in the training set to obtain priori boxes.Then,a multiple feature maps fusion method to improve the accuracy of the model and a method to reduce the model input size to speed up the model are proposed.The optimal real-time object identification and localization model is obtained through various stacking experiments.In comparison with other object identification and localization algorithms,the model proposed in this paper achieved 41.83%mAP accuracy on the PASCAL VOC2007 test set,and reached 25 FPS speed on the Intel 15 7th generation processor,and is superior to YOLO-LITE in speed and accuracy,and is far superior in speed to other object identification and localization algorithms.(3)In this paper,the insulators in the 1200 images of UAV aerial photography are marked and the training set and test set are constructed in a ratio of 2:1.Then the object identification and localization algorithm and other algorithms proposed in this paper are trained and tested on the above training set and test set,and compared and tested from different angles.Finally,it is found that the proposed algorithm can balance the accuracy and speed,reaches 40.65%AP accuracy on the test set and 25 FPS speed and is superior to other algorithms compared.