高分辨率遥感影像输电杆塔智能检测方法  

Intelligent Detection of Transmission Tower from High Resolution Remote Sensing Imagery

作　　者：宋成根 张正鹏 赵瑞山[1] 卜丽静 SONG Chenggen;ZHANG Zhengpeng;ZHAO Ruishan;BU Lijing(School of Geomatics,Liaoning Technical University,Fuxin,Liaoning 123000,China;College of Automation and Electronic Information,Xiangtan University,Xiangtan,Hunan 411105,China)

机构地区：[1]辽宁工程技术大学测绘与地理科学学院,辽宁阜新123000 [2]湘潭大学自动化与电子信息学院,湖南湘潭411105

出　　处：《遥感信息》2022年第3期65-71,共7页Remote Sensing Information

基　　金：国家自然科学基金青年科学基金项目(41801294)。

摘　　要：针对电网系统中输电杆塔的快速自动检测问题,提出高分辨率遥感影像的输电杆塔智能检测方法。首先,制作高分辨率遥感影像的输电杆塔数据集,针对杆塔目标尺寸特点,通过K-means算法重新聚类先验框;其次,利用迁移学习进行网络参数调整。鉴于原始YOLOv3算法存在IoU不能准确描述目标框的重叠程度,训练过程中损失下降缓慢的问题,提出了引入CIoU并将损失函数中的位置损失替换为CIoU损失。为改善YOLOv3算法对密集目标检测存在漏检的问题,提出将DIoU与NMS结合,从而提高杆塔检测的精度。最后,利用本文数据集实验。实验证明,改进后算法的精确度和召回率分别提高了0.48%、1.53%,精确度达到98%,召回率达到96%以上,与YOLOv3、Faster R-CNN、SSD算法相比,在训练时间、检测精度、定位精度方面更优。Aiming at the problem of fast automatic detection for transmission tower in power grid system,an intelligent detection method of transmission tower from high resolution remote sensing imagery is proposed.Firstly,the high resolution remote sensing images of transmission tower data sets are made.According to the characteristics of tower target size,K-means algorithm is used to re-cluster the prior frame.Then,network parameters are adjusted by using transfer learning.There are two problems in the original YOLOv3 algorithm.One is that IOU cannot accurately describe the overlap degree of the target frame.The other is the loss decreases slowly in the training process.We propose to replace the original loss function with the CIoU function.In order to improve the problem of missing detection in the dense target detection of YOLOv3 algorithm,we combine DIoU with NMS to improve the accuracy of tower detection.Finally,the experimental results show that the accuracy and recall rate of the proposed algorithm are improved by 0.48%and 1.53%respectively,the accuracy is 98%,and the recall rate is more than 96%.Compared with YOLOv3,fast R-CNN and SSD algorithm,the proposed algorithm is better in training time,detection accuracy and positioning accuracy.

关 键 词：输电杆塔 智能检测 迁移学习 YOLOv3 损失函数 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]