文档名：基于ERYOLO算法的跨环境输电线路缺陷识别方法
摘要：为提高输电线路缺陷智能检测算法在不同环境条件下的鲁棒性,克服现有智能检测算法在不同环境下识别性能下降的问题,该文提出了跨环境鲁棒YOLO(ER-YOLO)算法.首先,基于广义注意力理论在YOLOv8的骨干网络中引入Transformer注意力机制;其次,使用大卷积核和通道注意力模块优化特征提取;最后,应用多重注意力机制检测头网络强化算法多尺度、空间位置和多任务感知能力.为获得测试数据,该文探索生成了模拟暗光、雾霾、模糊环境的虚拟数据集.经消融实验和对比分析,跨环境鲁棒YOLO算法在跨环境测试中展现了更高的缺陷识别精度和鲁棒性,各测试数据集下mAP平均值为0.726,相对改进前提升0.069,同时在实际环境下进行了验证,证明了该算法的有效性.该文提出的跨拍摄环境的输电线路缺陷识别方法,在跨环境识别中表现出卓越的性能.跨环境图像生成方法可为后续虚拟数据集生成技术提供借鉴.
Abstract：Thedefectdetectionalgorithmforpowertransmissionlinesbasedondeeplearningheavilyreliesonthedatasetused.Thelearnedfeaturesoflinedefectsandimagefeaturesarehighlycorrelatedwiththefeaturedistributionofthedatasetused.Currently,mostdatasetsusedtotraindefectdetectionalgorithmsarecapturedinasingleenvironment,leadingtothespecificdefectrecognitionalgorithmperformingwellonlyinaparticularenvironment.However,itsgeneralizationabilityisinadequateindifferentshootingenvironments,suchaslowlight,haze,andblur,resultinginpooralgorithmrobustness.Toimprovetherobustnessoftheintelligentdefectdetectionalgorithmforpowertransmissionlinesindifferentenvironmentalconditionsandovercometheissueofdecreasedrecognitionperformanceindifferentenvironments,thispaperproposesthecross-environmentrobustYOLOalgorithm(ER-YOLO).Firstly,basedontheyouonlylookoncev8(YOLOv8)algorithm,ER-YOLOenhancesthealgorithm'slong-distancemodelingcapabilitybyemployingtheTransformerattentionmechanismbasedonthegeneralizedattentiontheory.Secondly,ER-YOLOimprovesthepathaggregationnetwork(PANet)intheYOLOv8algorithmbyincreasingtheconvolutionalkernelsizeandintroducinganefficientattentionmechanisminthecrossstagepartiallayer(CSPLayer),enhancingthenetwork'sobjectdetectioncapability.Finally,ER-YOLOusesamultipleattentionmechanismdetectionheadnetworktostrengthenthealgorithm'smulti-scale,spatiallocation,andmulti-taskperceptioncapabilities,enablingthenetworktofocusoncrucialtargetinformation.Additionally,tovalidatethecross-environmentdetectionperformanceofthedefectrecognitionalgorithm,acertainnumberofcorrespondingdatasetsofpowertransmissionlinedefectsindifferentenvironmentsarerequired.However,duetoenvironmentalandpracticalengineeringconstraints,collectingalargenumberofdatasetsinspecificenvironmentsischallenging.Toaddresstheissueofthetestdataset,thispaperproposescommonmethodsforgeneratingadverseenvironmentdataforpowertransmissionlinedefectrecognition.High-fidelitytestdatasetsweregeneratedbasedonnormalenvironmentdata.Thepaperexploresmethodsforsimulatingadverseenvironments,includingdarklightenvironmentsimulationusingexposurefusionalgorithmsandbrightnessreductionmethods,hazeenvironmentsimulationusingCycleGANnetworks,andimagingblurenvironmentsimulationusingmeanfilteringmethods.Theeffectivenessofeachmethodwasevaluatedandcomparedwithothermethods,providingconditionsfortestingtherobustnessofdefectrecognitionalgorithmsacrossdifferentenvironments.Throughablationexperimentsandcomparativeanalysis,ER-YOLOdemonstratedhigherdefectrecognitionaccuracyandrobustnessincross-environmenttesting.TheaveragemAPvalueundervarioustestdatasetswas0.726,showinganimprovementof0.069comparedtothepreviousversion.Thealgorithm'seffectivenesswasfurthervalidatedinrealenvironments.Thisstudyproposesadefectrecognitionmethodforpowertransmissionlinesacrossshootingenvironments,exhibitingexcellentperformanceincross-environmentrecognition.Italsoexploresmethodsforgeneratingcross-environmentimages,providinginsightsforfuturevirtualdatasetgenerationtechniques.Futureresearchdirectionsmayfocusoncross-environmentdefectrecognitionstudiesforothertypesofdefectsandexploreothereffectivemethodsforgeneratingmulti-environmentvirtualdatasets.
作者：裴少通 张行远 胡晨龙 杨文杰 刘云鹏Author：PeiShaotong ZhangHangyuan HuChenlong YangWenjie LiuYunpeng
作者单位：河北省输变电设备安全防御重点实验室(华北电力大学)保定071003
刊名：电工技术学报
Journal：TransactionsofChinaElectrotechnicalSociety
年，卷(期)：2024, 39(9)
分类号：TM755TP39
关键词：输电线路 缺陷检测 深度学习 数据集生成
Keywords：Powertransmissionlines defectdetection deeplearning datasetgeneration
机标分类号：TP391TP273.4TN98
在线出版日期：2024年5月13日
基金项目：中央高校基本科研业务费资助项目基于ER-YOLO算法的跨环境输电线路缺陷识别方法[
期刊论文] 电工技术学报--2024, 39(9)裴少通 张行远 胡晨龙 杨文杰 刘云鹏为提高输电线路缺陷智能检测算法在不同环境条件下的鲁棒性,克服现有智能检测算法在不同环境下识别性能下降的问题,该文提出了跨环境鲁棒YOLO(ER-YOLO)算法.首先,基于广义注意力理论在YOLOv8的骨干网络中引入Transfo...参考文献和引证文献
参考文献
引证文献
本文读者也读过
相似文献
相关博文
基于ER-YOLO算法的跨环境输电线路缺陷识别方法 The Defect Detection Method for Cross-Environment Power Transmission Line Based on ER-YOLO Algorithm
基于ER-YOLO算法的跨环境输电线路缺陷识别方法.pdf
- 文件大小:
- 5.26 MB
- 下载次数:
- 60
-
高速下载
|
|
