FRD-Net:输电线路金具锈蚀检测算法OA
FRD-Net:Transmission Line Fittings Corrosion Detection Algorithm
针对输电线路锈蚀金具与正常金具差异细微、锈蚀形态复杂多样以及无人机拍摄的图像存在景深效应导致的特征提取时细节信息丢失严重等检测难题,文章提出了一种输电线路金具锈蚀检测算法(fittings rust detection network,FRD-Net).首先,设计多级融合互补下采样(multi-level fusion complementary downsampling,MFCD)方法并嵌入骨干网络,补偿景深图像下采样过程中细节特征信息的大量丢失;然后,为进一步增强对不同形态金具锈蚀的检测能力,采用独特的特征拓扑重构(feature topology reconstruction,FTR)特征融合方法;接着,引入先验知识指导的特征差分加权(differential weighting of features,DWF)模块表征金具特征信息,精准捕捉并区分锈蚀金具与正常金具间难以具象化的细节差异,该方法通过融合不同尺度的特征信息,实现对金具锈蚀形态的全方位感知与精准检测;最后,提出了细节感知损失函数(detail perception loss function,DP_Loss),通过MFCD生成的空间权重引导模型训练过程聚焦于判别性强的局部特征.在无人机拍摄的输电线路数据集上进行大量实验,FRD-Net交并比阈值为0.5时的平均精度均值(mean average precision at IoU threshold 0.5,mAP50)达到73.7%,参数量仅为1.92 M,计算量为5.2亿次浮点运算(giga floating-point operations,GFLOPs),显著优于主流检测模型,并在跨域锈蚀数据上展现出强泛化能力,为输电线路巡检提供了一种高精度、高效率的解决方案.
In response to the detection challenges posed by the subtle differences between rusty fittings and normal fittings on power transmission lines,the complex and diverse forms of corrosion,and the serious loss of detail information during feature extraction due to depth of field effects in images captured by drones,a corrosion detection algorithm for transmission line fittings has been proposed.First,a multi-level complementary down-sampling method was designed and integrated into the backbone network to compensate for the significant loss of detailed feature information during the down-sampling of depth of field images;then,to further enhance the detection capability for different forms of corroded fittings,a unique feature topology reconstruction feature fusion method was adopted.Next,a prior knowledge-guided feature differential weighting module was introduced to characterize the fitting feature information,accurately capturing and differentiating the difficult-to-visualize detail differences between corroded and normal fittings.This method achieves comprehensive perception and precise detection of fitting corrosion forms by integrating feature information of different scales.Finally,a detail perception loss function was proposed,which uses spatial weights generated by MFCD to guide the model training process to focus on strongly discriminative local features.Extensive experiments on a drone-captured power transmission line dataset showed that FRD-Net achieved an mAP50 of 73.7%,with a parameter count of only 1.92M and a computational load of 5.2 GFLOPs,significantly outperforming mainstream detection models,and demonstrating strong generalization capabilities on cross-domain corrosion data,providing a high accuracy and high efficiency solution for power transmission line inspection.
赵振兵;景群;王津闻;马腾;高凯鹏;李浩鹏
华北电力大学电子与通信工程系,河北省保定市 071003||华北电力大学河北省电力物联网技术重点实验室,河北省保定市 071003||华北电力大学复杂能源系统智能计算教育部工程研究中心,河北省保定市 071003华北电力大学电子与通信工程系,河北省保定市 071003华北电力大学能源动力与机械工程学院,河北省保定市 071003华北电力大学电子与通信工程系,河北省保定市 071003华北电力大学电子与通信工程系,河北省保定市 071003华北电力大学电子与通信工程系,河北省保定市 071003||华北电力大学河北省电力物联网技术重点实验室,河北省保定市 071003
信息技术与安全科学
输电线路金具锈蚀检测景深效应细节感知差分加权
transmission linesfittings corrosion detectiondepth of field effectdetail perceptiondifferential weighting
《电力信息与通信技术》 2026 (8)
1-14,14
国家自然科学基金项目(62571189,62373151,62371188,62303184)中央高校基本科研业务费专项资金资助(2023JC006).
评论