| With the rapid development of transmission and transformation technology in our country,the transmission distance of overhead lines continues to increase,and the voltage level continues to increase,which makes the scale of overhead lines continue to expand,but at the same time,it will face the problem of safe and stable operation of overhead lines.As one of the most important and fault-prone components in overhead lines,insulators mainly play an important role in providing mechanical support,increasing creepage distance and preventing current from returning to ground.Since the insulator is exposed to the harsh outdoor environment for a long time,a series of problems such as breakage and string drop will occur,and it is a fault-prone component.On the basis of in-depth analysis of related research at home and abroad,this thesis focuses on the detection of power insulator defects in aerial images,mainly including target recognition,image segmentation and defect detection of insulators.The research work of this thesis is as follows:(1)Aiming at the problem of insufficient recognition of insulator targets in aerial images,this thesis selects the YOLOv4 target detection algorithm,which is more balanced in terms of recognition speed and detection accuracy,as the basic algorithm for insulator target recognition.Based on the YOLOv4 algorithm,the K-means++clustering algorithm is used for the insulator target size distribution,and the recognition effect in the insulator identification task is improved by selecting the bounding box regression loss function and attention mechanism.The experimental results show that the improved algorithm can better solve the problem of missed detection in the YOLOv4 algorithm,and improve the accuracy of the insulator identification model.(2)An improved DeepLabV3+model for insulator string segmentation in aerial images is proposed.On the basis of the traditional DeepLabV3+model,the MobileNetV2 network is used instead of Xception as the backbone network,and a selfattention mechanism is added to the network in series with the ASPP module in the original network structure.The improved DeeplabV3+algorithm is used to compare with the commonly used segmentation algorithms,and the experimental comparison proves that the improved algorithm has more advantages in segmentation performance compared with the commonly used algorithms,and achieves the expected performance of the experiment.(3)According to the regularity of the distribution of insulator strings in the segmented image,the least squares fitting method is used to fit the insulator strings in the binarized image to build a mathematical model.Then use the straight line to scan the insulator string one by one,according to the shape rule of the insulator sheet,sum the number of white pixels in each cycle of the scan,and compare the summed value with the discrimination threshold to determine whether the insulator in the image has defects or not.Experiments show that this method has certain reference value in the task of insulator defect detection. |
