基于LSD与深度学习的轻量化导地线缺陷检测算法  

A Lightweight GroundWire Defect Detection Method Based on LSD and Deep Learning

作　　者：王勇强 周学明 张政[1] 雷波[1] 王晨晟[1] WANG Yong-qiang;ZHOU Xue-ming;ZHANG Zheng;LEI Bo;WANG Chen-sheng(Huazhong Institute of Electro-Optics—Wuhan National Laboratory for Optoelectronics,Wuhan 430233,China;State Grid Hubei Electric Power Research Institute,Wuhan 430077,China)

机构地区：[1]华中光电技术研究所—武汉光电国家研究中心,湖北武汉430223 [2]国网湖北省电力有限公司电力科学研究院,湖北武汉430077

出　　处：《光学与光电技术》2024年第1期60-66,共7页Optics & Optoelectronic Technology

摘　　要：针对无人机航拍图像尺寸过大,且背景环境复杂带来的导地线缺陷检测速度和精度下降的问题,提出了一种基于LSD与深度学习的轻量化导地线缺陷检测方法。首先利用LSD算法提取图像的直线特征,并结合RANSAC拟合出一条分割基线;然后根据分割基线对导地线区域进行粗分割,剔除大部分背景干扰,减小送入检测网络的图片尺寸;最后通过对YOLOv5的主干网络进行修改,减少网络的计算量,使其更适合在边缘计算设备上部署。实验结果表明,该方法使待检区域大大减小的同时抑制了背景的干扰,使检测精度由原来的67.9%提高到71.3%,同时检测速度提升了12%。本算法具有精度高、速度快、适合在边缘计算设备上部署等优点。The detection of defects in ground wires is often slowed down and less accurate due to the large size of UAV aerial images and complex background environment.To address this issue,this paper proposes a lightweight ground wire defect detection method based on LSD and deep learning.First,the LSD algorithm is used to extract linear features from the images.Then,a segmentation baseline is fit by combining it with RANSAC.Based on the segmentation baseline,the region of the ground wire is now clearly segmented,background interferences are eliminated,and image size sent to the detection network reduced.After modifying the YOLOv5 backbone network,the number of parameters is reduced,making it easier to deploy on edge computing equipment.The proposed method reduces the inspected area and suppresses interference.It improves the accuracy from the original 67.9%to 71.3%,at the same time,the detection speed is increased by 12%.It has the advantages of high precision and fast speed,which is suitable for deployment on edge computing equipment.

关 键 词：导地线缺陷检测 航拍图像 直线段检测 直线拟合 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]