基于改进Cascade RCNN的输电线路防振锤脱落检测方法  

Detection method of dropped anti-vibration hammer for transmission line based on improved Cascade RCNN

作　　者：阎光伟[1] 刘润泽 焦润海[1] 何慧 YAN Guang-wei;LIU Run-ze;JIAO Run-hai;HE Hui(School of Control and Computer Engineering,North China Electric Power University,Beijing 102200,China)

机构地区：[1]华北电力大学控制与计算机工程学院,北京102200

出　　处：《图学学报》2023年第5期849-860,共12页Journal of Graphics

基　　金：国家自然科学基金项目(62073133)。

摘　　要：无人机巡检输电线路时,因拍摄角度和距离问题,容易出现被输电线遮挡和远距离拍摄的防振锤脱落目标,导致目标特征被遮挡且分辨率较低,且部分防振锤出现滑移现象,导致目标识别准确率降低。针对以上问题,提出一种基于改进Cascade RCNN的防振锤脱落检测网络。第一,设计了对比学习网络,将正负样本与真实样本的特征进行对比学习,利用对比损失函数训练网络,使其能更加关注到被遮挡的防振锤脱落目标,提升其特征提取能力;第二,进行了分类器增强操作,筛选出网络级联结构中回归效果较好的感兴趣区域并送入最后的分类回归队列中,提高了分类器的分类能力,进而提升检测目标的分类分数;第三,设计了并行注意力机制模块,整合网络提取的特征,增大关键特征的权重,使网络关注到图像中更关键的区域;在特征金字塔中,将双线性插值方法代替为反卷积,提升特征还原能力。经交叉验证实验结果表明,改进后的模型召回率、精确率和平均精度达到了97.5%,91.0%和92.0%,相比基线模型分别提高了6.9%,28.4%和8.0%。During the inspection of transmission lines using Unmanned Aerial Vehicle(UAV),there are many dropped anti-vibration hammers that become obstructed by wires or are shot from a distance.This challenge leads to the occlusion of target features and low resolution.In addition,the close proximity of a number of hammers due to sliding poses challenges to the accuracy of target identification.To address the above problems,a deep neural network based on an improved Cascade RCNN was proposed to identify the dropped anti-vibration hammers.The proposed network mainly achieved improvements from the following four aspects.First of all,a contrastive learning network was designed to compare the features of positive and negative samples with real samples.By utilizing a contrastive loss function during network training,the network became more attentive to the blocked dropped anti-vibration hammers and enhanced its feature extraction ability.Secondly,the classifier was enhanced.The selection of interested regions with better regression performance in the cascade structure was filtered and input directly into the final classification regression queue.This improved the classification performance of the classifier,thereby enhancing the classification scores of the detected targets.Thirdly,a parallel attention mechanism module was designed to integrate the extracted features from the network,increasing the weights of key features and directing the network’s attention to more critical features in the image.In addition,in the process of feature fusion of the feature pyramid,the bilinear interpolation method was replaced with deconvolution to enhance the feature restoration capability.The experimental results demonstrated that the improved model achieved a recall rate of 97.5%,precision of 91.0%,and average precision of 92.0%,an improvement of 6.9%,28.4%,and 8.0%,respectively,compared with the baseline model.

关 键 词：输电线路 防振锤脱落 Cascade RCNN 对比学习网络 并行注意力模块 分类器增强 样本相似度 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]