PAGE11基于卷積神經(jīng)網(wǎng)絡(luò)的高鐵沿線隔音屏障缺陷檢測方法研究【摘要】：隨著高速鐵路網(wǎng)的迅猛發(fā)展，噪音污染問題日益突出，隔音屏障作為降低噪音的關(guān)鍵設(shè)施得到廣泛應(yīng)用。然而，隔音屏障在運(yùn)營中受環(huán)境因素影響，可能產(chǎn)生螺栓松動(dòng)、單元板破損等結(jié)構(gòu)缺陷，影響隔音效果并帶來安全隱患。傳統(tǒng)的人工巡檢和目視檢查方法效率低、成本高、主觀性強(qiáng)，難以滿足大規(guī)模隔音屏障的檢測需求，迫切需要一種安全有效、高精度、高速的方法來檢測隔音屏障結(jié)構(gòu)中存在的缺陷。近年來，基于卷積神經(jīng)網(wǎng)絡(luò)的圖像識(shí)別方法已廣泛應(yīng)用于各個(gè)領(lǐng)域。卷積神經(jīng)網(wǎng)絡(luò)強(qiáng)大的特征提取和學(xué)習(xí)能力，可以實(shí)現(xiàn)對(duì)目標(biāo)物體的自動(dòng)識(shí)別和定位。因此，研究基于卷積神經(jīng)網(wǎng)絡(luò)的高鐵隔音屏障缺陷檢測方法具有重要的理論意義和應(yīng)用價(jià)值。對(duì)此，本文深入研究了基于卷積神經(jīng)網(wǎng)絡(luò)的高鐵沿線隔音屏障缺陷檢測方法，特別關(guān)注了基于Yolov5模型的目標(biāo)檢測技術(shù)在該領(lǐng)域的優(yōu)化與應(yīng)用。由于現(xiàn)有的隔音屏障缺陷數(shù)據(jù)集不足，本文首先通過無人機(jī)采集高鐵沿線隔音屏障的圖像數(shù)據(jù)，并進(jìn)行了數(shù)據(jù)清洗、標(biāo)注、劃分處理，自制了一個(gè)隔音屏障缺陷檢測（NBDD）數(shù)據(jù)集。隨后，基于Yolov5模型進(jìn)行了優(yōu)化，在自制數(shù)據(jù)集上進(jìn)行訓(xùn)練，建立了一個(gè)適用于隔音屏障缺陷檢測的高效目標(biāo)檢測模型，該模型在隔音屏障缺陷檢測任務(wù)中展現(xiàn)出了較高的性能。在此基礎(chǔ)上，為了進(jìn)一步提升檢測精度和泛化能力，進(jìn)一步嘗試了添加注意力機(jī)制對(duì)模型性能的影響，并通過對(duì)比實(shí)驗(yàn)分析了添加注意力機(jī)制的效果，選擇實(shí)驗(yàn)結(jié)果最好的CBAM模塊對(duì)yolov5模型進(jìn)行改進(jìn)。此外，考慮到數(shù)據(jù)集規(guī)模及各類別缺陷樣本數(shù)量不平衡問題對(duì)模型性能的影響，還實(shí)施了數(shù)據(jù)增強(qiáng)處理，通過擴(kuò)充數(shù)據(jù)集規(guī)模來提升模型的泛化能力，并對(duì)實(shí)驗(yàn)結(jié)果進(jìn)行了對(duì)比分析，實(shí)驗(yàn)結(jié)果表明，數(shù)據(jù)增強(qiáng)處理顯著提高了模型的檢測精度和穩(wěn)定性，改進(jìn)的yolov5-CBAM模型對(duì)比原模型在檢測精度上提升0.5%，yolov5-CBAM模型在隔音屏障缺陷檢測上具有良好的檢測性能。最后，本文總結(jié)了研究的主要成果和局限性，并對(duì)未來的研究方向進(jìn)行了展望?！娟P(guān)鍵詞】：隔音屏障；缺陷檢測；目標(biāo)檢測；卷積神經(jīng)網(wǎng)絡(luò)1緒論1.1研究背景及意義高速鐵路是我國重要的交通運(yùn)輸基礎(chǔ)設(shè)施，截至2023年底，全國高速鐵路營業(yè)里程達(dá)到4.5萬公里，而高鐵沿線的噪音污染問題日益凸顯，尤其是在人口密集區(qū)域，噪音對(duì)居民生活和工作環(huán)境造成干擾。為了降低噪音對(duì)沿線居民的影響，隔音屏障被廣泛應(yīng)用。圖SEQ圖\*ARABIC1高鐵線路上的隔音屏障如圖1所示，隔音屏障主要由鋼結(jié)構(gòu)立柱和吸隔聲屏板兩部分結(jié)構(gòu)組成，立柱是隔音屏障的主要受力構(gòu)件，通過螺栓固定在軌道邊的預(yù)埋鋼板上；吸隔聲板是主要的隔聲吸聲構(gòu)件，它通過高強(qiáng)彈簧卡子固定在H型立柱槽內(nèi)，形成隔音屏障。高鐵隔音屏障是高速鐵路沿線的關(guān)鍵設(shè)備，用于減少高鐵列車噪音傳播的設(shè)施，可以有效隔絕噪音，保護(hù)周圍居民免受干擾。隔音屏障不僅在降低鐵路噪聲方面表現(xiàn)出色，同時(shí)，它也是防止外來物侵入的堅(jiān)固屏障，對(duì)于確保高速鐵路的安全運(yùn)營至關(guān)重要。然而，在高鐵運(yùn)營過程中，隔音屏障不僅受到自然環(huán)境中的風(fēng)雨侵蝕、溫度變化等影響，還要承受高速列車運(yùn)行產(chǎn)生的脈動(dòng)力反復(fù)沖擊。這些因素共同作用，可能導(dǎo)致隔音屏障的隔音效果下降，甚至可能導(dǎo)致隔音屏障有脫落、倒塌至線路上的安全隱患，進(jìn)而對(duì)高鐵的正常運(yùn)行構(gòu)成威脅。2013年，在我國京滬高鐵線路上發(fā)生過隔音屏障脫落事故，不僅造成了巨大的經(jīng)濟(jì)損失，也給人們的生命安全帶來了嚴(yán)重威脅。因此，對(duì)隔音屏障的結(jié)構(gòu)缺陷進(jìn)行定期檢測并及時(shí)發(fā)現(xiàn)故障隱患，以確保高鐵安全運(yùn)營具有重要意義。高鐵隔音屏障有六種構(gòu)件和常見缺陷：螺栓、正常柱、正常砂漿層、表面損壞、生銹柱和劣化砂漿層，如圖2所示。隔音屏障長期受紫外線、雨水等作用，會(huì)導(dǎo)致聲屏障柱子生銹和砂漿層變質(zhì)；隔音屏障長期在高速列車運(yùn)行產(chǎn)生脈動(dòng)力的反復(fù)沖擊作用下，可能造成立柱螺栓松動(dòng)和單元板破損。圖2隔音屏障結(jié)構(gòu)和缺陷細(xì)節(jié)圖（a）螺栓；（b）正常柱；（c）正常砂漿層；（d）表面損壞；（e）生銹柱；（f）劣化砂漿層。隨著高速鐵路網(wǎng)的擴(kuò)展，目前已安裝的隔音屏障總長度超過4000km，約占高鐵運(yùn)營里程的10%，鑒于高鐵線路中隔音屏障的安裝數(shù)量龐大，相應(yīng)的檢測工作也變得異常繁重。更為復(fù)雜的是，這些隔音屏障主要被安裝在高鐵線路的高架橋上。由于橋面并未預(yù)先設(shè)計(jì)檢修通道，對(duì)隔音屏障外側(cè)進(jìn)行人工檢測變得異常困難。檢查和維護(hù)面臨著很大的挑戰(zhàn)。在過去的幾十年里，隔音屏障的檢查完全依靠檢查員和他們的經(jīng)驗(yàn)。檢查維護(hù)工作人員的工作強(qiáng)度大，效率低下，成本高，缺乏客觀性且主觀性強(qiáng)，難以保證檢測的準(zhǔn)確性和一致性。傳統(tǒng)的隔音屏障缺陷檢測方法主要依賴人工巡檢和目視檢查。人工巡檢主要采用夜間入網(wǎng)巡視、人工摸排觀測的方法進(jìn)行，人工檢查效率較低，缺乏客觀性，且由于光照條件不足，夜間誤報(bào)率很高。而對(duì)于聲屏障外側(cè)，無法直接觀察到到其結(jié)構(gòu)情況，目前缺乏有效的檢查作業(yè)方法和手段對(duì)聲屏障外側(cè)的檢查維護(hù)，主要采用撬杠撬起后再進(jìn)行檢查維護(hù)等方式進(jìn)行，檢查維護(hù)難度很大，工作人員的工作強(qiáng)度大且效率低下。目視檢查主要由工人們觀看安裝在復(fù)合檢查列車上的攝像頭對(duì)鐵路線進(jìn)行重新編碼的視頻，以了解隔音屏障的損壞情況，這種方法效率低下且不準(zhǔn)確?，F(xiàn)有的隔音屏障檢測方法仍然嚴(yán)重依賴人工檢測，效率低，成本高，主觀性強(qiáng)，并且難以檢測到隔音屏障的外部結(jié)構(gòu)。并且，高鐵隔音屏障缺陷檢測的挑戰(zhàn)是在非常有限的時(shí)間內(nèi)完成工作，要求相關(guān)技術(shù)方法具有更高的檢測精度和更快的識(shí)別速度。因此，迫切需要一種安全有效、高速、高精度的方法對(duì)隔音屏障進(jìn)行檢測維護(hù)。近年來，隨著深度學(xué)習(xí)技術(shù)的迅猛進(jìn)步，卷積神經(jīng)網(wǎng)絡(luò)（CNN）在計(jì)算機(jī)視覺領(lǐng)域的應(yīng)用已日益廣泛，特別是在工業(yè)缺陷檢測任務(wù)中展現(xiàn)出強(qiáng)大的特征提取和識(shí)別能力。因此，越來越多的學(xué)者開始致力于利用CNN進(jìn)行目標(biāo)檢測方法的深入研究。在缺陷檢測領(lǐng)域，利用深度學(xué)習(xí)算法進(jìn)行表面缺陷檢測有了大量的研究和應(yīng)用案例。因此，將卷積神經(jīng)網(wǎng)絡(luò)應(yīng)用于高鐵沿線隔音屏障的缺陷檢測問題，有望提高檢測的效率和準(zhǔn)確性，降低人工巡檢的成本和風(fēng)險(xiǎn)，為高鐵運(yùn)行安全提供有力保障。同時(shí)，開展這一研究對(duì)于推動(dòng)高鐵沿線隔音屏障維護(hù)管理的現(xiàn)代化、智能化具有重要意義。1.2國內(nèi)外研究現(xiàn)狀目前，基于卷積神經(jīng)網(wǎng)絡(luò)(CNN)的目標(biāo)檢測算法在缺陷檢測任務(wù)中展現(xiàn)出強(qiáng)大的特征提取能力，具有大量的研究和應(yīng)用案例。基于卷積神經(jīng)網(wǎng)絡(luò)(CNN)的目標(biāo)檢測算法在土木工程基礎(chǔ)設(shè)施缺陷檢測中得到了廣泛的應(yīng)用。例如混凝土裂縫檢測、橋臺(tái)檢測、路面裂縫檢測、污水管道缺陷檢測、鋼結(jié)構(gòu)螺栓松動(dòng)檢測等。ChenADDINZOTERO_ITEMCSL_CITATION{"citationID":"654d7uJy","properties":{"formattedCitation":"\\super[1]\\nosupersub{}","plainCitation":"[1]","noteIndex":0},"citationItems":[{"id":723,"uris":["/users/local/Mc1qNPH1/items/ZRJ9JFH4"],"itemData":{"id":723,"type":"article-journal","abstract":"Thefastnetworkingofhigh-speedrail(HSR)maycausein-servicefatigueandultimateloaddamagetobridges.Thispaperinvestigatestheapplicationofdeepconvolutionalneuralnetworks(CNNs)formulti-categorydamageimageclassificationrecognitionofHSR-reinforcedconcrete(RC)bridges.Thepresentstudyestablishesadeeplearning(DL)systembasedonalargeamountofHSRbridgetestdata.Whentobegin,thedamagedonetoHSRbridgepiersmaybebrokendownintothreeprimarycategories:concretecracks,concretespalling,andreinforcementexposure.ThesecategoriesaredeterminedbythestatisticsofHSRbridgepiertesting.Secondly,inordertodevelopanautomatedrecognitionmodelforthedamageofHSRpiers,AlexNetCNNsweretaughtatransferlearningapproachandthenusedtotrainthemselves.Thecorrectrecognitionrateofthethreedamagedpicturesintheactualapplicationofthemodelis86%forcracks,82%forreinforcementexposure,and70%forconcretespalling,allofwhichhavegoodrecognitionrates.Thestudy'saccuracyandprecisionenhancedetectionefficiencyandmaybeutilizedtoidentifyHSRpierdeteriorationquickly.Theresearchcomprisesarandomselectionfromthetrainingandvalidationsets.Italsoassessesthetrainingmodel'sgeneralizationtoout-of-samplepicturesforengineeringapplications.Thisaspectoftheworksetsitapartfrompreviousresearchinthesamestudyarea.","container-title":"EngineeringStructures","DOI":"10.1016/j.engstruct.2022.115306","ISSN":"0141-0296","journalAbbreviation":"EngineeringStructures","page":"115306","source":"ScienceDirect","title":"Convolutionalneuralnetworks(CNNs)-basedmulti-categorydamagedetectionandrecognitionofhigh-speedrail(HSR)reinforcedconcrete(RC)bridgesusingtestimages","volume":"276","author":[{"family":"Chen","given":"Lingkun"},{"family":"Chen","given":"Wenxin"},{"family":"Wang","given":"Lu"},{"family":"Zhai","given":"Chencheng"},{"family":"Hu","given":"Xiaolun"},{"family":"Sun","given":"Linlin"},{"family":"Tian","given":"Yuan"},{"family":"Huang","given":"Xiaoming"},{"family":"Jiang","given":"Lizhong"}],"issued":{"date-parts":[["2023",2,1]]}}}],"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[1]等研究了深度卷積神經(jīng)網(wǎng)絡(luò)(CNN)在橋梁結(jié)構(gòu)損傷圖像分類識(shí)別中的應(yīng)用，利用遷移學(xué)習(xí)技術(shù)構(gòu)建了3種橋梁結(jié)構(gòu)損傷的自動(dòng)識(shí)別模型，建立了一個(gè)深度學(xué)習(xí)(DL)系統(tǒng)，具有較高的識(shí)別率。Cha等ADDINZOTERO_ITEMCSL_CITATION{"citationID":"NRLH6NCo","properties":{"formattedCitation":"\\super[2]\\nosupersub{}","plainCitation":"[2]","noteIndex":0},"citationItems":[{"id":742,"uris":["/users/local/Mc1qNPH1/items/KLER8U6A"],"itemData":{"id":742,"type":"article-journal","abstract":"Computervision-basedtechniquesweredevelopedtoovercomethelimitationsofvisualinspectionbytrainedhumanresourcesandtodetectstructuraldamageinimagesremotely,butmostmethodsdetectonlyspecifictypesofdamage,suchasconcreteorsteelcracks.Toprovidequasireal-timesimultaneousdetectionofmultipletypesofdamages,aFasterRegion-basedConvolutionalNeuralNetwork(FasterR-CNN)-basedstructuralvisualinspectionmethodisproposed.Torealizethis,adatabaseincluding2,366images(with500×375pixels)labeledforfivetypesofdamages—concretecrack,steelcorrosionwithtwolevels(mediumandhigh),boltcorrosion,andsteeldelamination—isdeveloped.Then,thearchitectureoftheFasterR-CNNismodified,trained,validated,andtestedusingthisdatabase.Resultsshow90.6%,83.4%,82.1%,98.1%,and84.7%averageprecision(AP)ratingsforthefivedamagetypes,respectively,withameanAPof87.8%.TherobustnessofthetrainedFasterR-CNNisevaluatedanddemonstratedusing11new6,000×4,000-pixelimagestakenofdifferentstructures.ItsperformanceisalsocomparedtothatofthetraditionalCNN-basedmethod.Consideringthattheproposedmethodprovidesaremarkablyfasttestspeed(0.03secondsperimagewith500×375resolution),aframeworkforquasireal-timedamagedetectiononvideousingthetrainednetworksisdeveloped.","container-title":"Computer-AidedCivilandInfrastructureEngineering","DOI":"10.1111/mice.12334","ISSN":"1467-8667","issue":"9","language":"en","license":"?2017Computer-AidedCivilandInfrastructureEngineering","note":"_eprint:/doi/pdf/10.1111/mice.12334","page":"731-747","source":"WileyOnlineLibrary","title":"AutonomousStructuralVisualInspectionUsingRegion-BasedDeepLearningforDetectingMultipleDamageTypes","volume":"33","author":[{"family":"Cha","given":"Young-Jin"},{"family":"Choi","given":"Wooram"},{"family":"Suh","given":"Gahyun"},{"family":"Mahmoudkhani","given":"Sadegh"},{"family":"Büyük?ztürk","given":"Oral"}],"issued":{"date-parts":[["2018"]]}}}],"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[2]開發(fā)了一種基于FasterR-CNN的結(jié)構(gòu)損傷檢測方法，用于檢測混凝土裂縫、鋼筋腐蝕、螺栓腐蝕和鋼筋分層5種表面損傷。Jia等ADDINZOTERO_ITEMCSL_CITATION{"citationID":"uaTKkA4C","properties":{"formattedCitation":"\\super[3]\\nosupersub{}","plainCitation":"[3]","noteIndex":0},"citationItems":[{"id":734,"uris":["/users/local/Mc1qNPH1/items/ECAUMCZK"],"itemData":{"id":734,"type":"paper-conference","abstract":"AdrainagepipedefectdetectionmodelbasedonimprovedYOLOv5isproposedfortheproblemofleakageanderrordetectioninthemanualinspectionofdrainagepipes.First,theproposedCSSPPFmoduleisaddedtotheYOLOv5networktoenhancemulti-scalefeaturelearning,enhancefeaturemapdetailinformationextraction,andimprovethedetectionprecisionofsimilardefectobjects.Secondly,theattentionmechanismCBAMisintroducedtosuppresstheinterferenceofirrelevanttubewallbackgroundinformationandimprovethedetectionaccuracyofthemodel.Finally,theYOLOv5backbonenetworkstructureismodifiedtointroduceasmallobjectdetectionlayertoobtainalargerfeaturemapandimprovetheprecisionofthenetworkfordetectingsmallobjects..Theexperimentalresultsshowthattheprecision,recall,andmAPoftheimprovedmodelareimprovedby0.5%,2%,and2.5%,respectively,comparedwiththeoriginalYOLOv5networkmodel.Theexperimentalresultsshowthattheprecision,recall,andmAPoftheimprovedmodelareimprovedby0.5%,2%,and2.5%,respectively,comparedwiththeoriginalYOLOv5networkmodel.ComparedwiththemainstreamdetectionalgorithmsYOLOv4,SSD,andYOLOv7,themAPoftheimprovedmodelimprovedby26.6%,12.4%,and7.5%,respectively.Andithasstrongerfeatureextractionabilityandhigherrecognitionratefordrainagepipesmallobjectdefectdetection,whichmeetsthedemandforefficientandintelligentdrainagepipedefectdetectionandrecognition.","container-title":"20238thInternationalConferenceonIntelligentComputingandSignalProcessing(ICSP)","DOI":"10.1109/ICSP58490.2023.10248934","event-title":"20238thInternationalConferenceonIntelligentComputingandSignalProcessing(ICSP)","page":"1950-1955","source":"IEEEXplore","title":"DetectionModelofDrainagePipeDefectBasedonImprovedYOLOv5","URL":"/document/10248934","author":[{"family":"Jia","given":"Pengtao"},{"family":"Guo","given":"Tong"},{"family":"Guo","given":"Fengjing"},{"family":"Wang","given":"Bin"},{"family":"Xiong","given":"Qi"}],"accessed":{"date-parts":[["2024",3,22]]},"issued":{"date-parts":[["2023",4]]}},"label":"page"}],"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[3]提出了一種基于改進(jìn)YOLOv5的排水管道缺陷檢測方法，提出跨級(jí)金字塔池化模塊（CSSPPF）來替代原有YOLOv5模型中的SPPF模塊，并引入小目標(biāo)檢測層和注意力機(jī)制CBAM，提高模型的整體檢測性能，實(shí)現(xiàn)對(duì)13種排水管缺陷的準(zhǔn)確高效檢測。Du等ADDINZOTERO_ITEMCSL_CITATION{"citationID":"CmQo8VNl","properties":{"formattedCitation":"\\super[4]\\nosupersub{}","plainCitation":"[4]","noteIndex":0},"citationItems":[{"id":751,"uris":["/users/local/Mc1qNPH1/items/JYWW6SNT"],"itemData":{"id":751,"type":"article-journal","abstract":"Toensurethesafeoperationofhighwaytrafficlines,giventheimperfectfeatureextractionofexistingroadpitdefectdetectionmodelsandthepracticabilityofdetectionequipment,thispaperproposesalightweighttargetdetectionalgorithmwithenhancedfeatureextractionbasedontheYOLO(YouOnlyLookOnce)algorithm.TheBIFPN(BidirectionalFeaturePyramidNetwork)networkstructureisusedformulti-scalefeaturefusiontoenhancethefeatureextractionability,andVarifocalLossisusedtooptimizethesampleimbalanceproblem,whichimprovestheaccuracyofroaddefecttargetdetection.IntheevaluationtestofthemodelintheconstructedPCD1(PavementCheckDataset)dataset,themAP@.5(meanAveragePrecisionwhenIoU=0.5)oftheBV-YOLOv5S(BiFPNVarifocalLoss-YOLOv5S)modelincreasedby4.1%,3%,and0.9%,respectively,comparedwiththeYOLOv3-tiny,YOLOv5S,andB-YOLOv5S(BiFPN-YOLOv5S;BV-YOLOv5SdoesnotusetheImprovedFocalLossfunction)models.Throughtheanalysisandcomparisonofexperimentalresults,itisprovedthattheproposedBV-YOLOv5Snetworkmodelperformsbetterandismorereliableinthedetectionofpavementdefectsandcanmeettheneedsofroadsafetydetectionprojectswithhighreal-timeandflexibilityrequirements.","container-title":"Sensors","DOI":"10.3390/s22093537","ISSN":"1424-8220","issue":"9","language":"en","license":"/licenses/by/3.0/","note":"number:9\npublisher:MultidisciplinaryDigitalPublishingInstitute","page":"3537","source":"","title":"ImprovementofLightweightConvolutionalNeuralNetworkModelBasedonYOLOAlgorithmandItsResearchinPavementDefectDetection","volume":"22","author":[{"family":"Du","given":"Fu-Jun"},{"family":"Jiao","given":"Shuang-Jian"}],"issued":{"date-parts":[["2022",1]]}}}],"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[4]提出了一種基于YOLOv5S的增強(qiáng)特征提取的輕量級(jí)目標(biāo)檢測算法BV-YOLOv5S。采用BIFPN（BidirectionalFeaturePyramidNetwork）網(wǎng)絡(luò)結(jié)構(gòu)進(jìn)行多尺度特征融合，增強(qiáng)特征提取能力，采用變焦損耗優(yōu)化樣本不平衡問題，提高道路缺陷目標(biāo)檢測精度。同樣，基于卷積神經(jīng)網(wǎng)絡(luò)(CNN)的目標(biāo)檢測算法在鐵路領(lǐng)域的基礎(chǔ)設(shè)施檢測和維護(hù)中發(fā)揮著重要作用。Liu等ADDINZOTERO_ITEMCSL_CITATION{"citationID":"zfEavSzZ","properties":{"formattedCitation":"\\super[5]\\nosupersub{}","plainCitation":"[5]","noteIndex":0},"citationItems":[{"id":745,"uris":["/users/local/Mc1qNPH1/items/G39IJZ2Y"],"itemData":{"id":745,"type":"article-journal","abstract":"Theapplicationofcomputervisiontechnologyindefectdetectionofindustrialproductsisapopularresearchdirectioninrecentyears.Thisarticlepresentsthepyramidfeatureconvolutionalneuralnetwork(CNN)fordefectdetectionofrailsurfaces.First,multi-scalefeaturemapsareextractedbasedonthecharacteristicsofdefectsandbackgroundsbythepyramidfeatureextractionmodule(PFEM).Thenthefeaturemapsareinputtoalightweightnetworkconsistingofasmallnumberofparameters.Thenetworkistrainedwithonly40%dataofthedatasetusingbinarycross-entropylossfunctionandtheintersectionofunion(IOU)lossfunction.Intheexperiment,theperformanceoftheproposedmethodisevaluatedusingtherailsurfacedefectdataset(RSDD)datasetbycomparingitwithothermethods.Theexperimentalresultsshowthatthesegmentationperformanceandreal-timeperformanceoftheproposedmethodarebetterthanthoseofothermethods.","container-title":"IEEETransactionsonInstrumentationandMeasurement","DOI":"10.1109/TIM.2022.3165287","ISSN":"1557-9662","note":"event-title:IEEETransactionsonInstrumentationandMeasurement","page":"1-10","source":"IEEEXplore","title":"ARailSurfaceDefectDetectionMethodBasedonPyramidFeatureandLightweightConvolutionalNeuralNetwork","volume":"71","author":[{"family":"Liu","given":"Yu"},{"family":"Xiao","given":"Huaxi"},{"family":"Xu","given":"Jiaming"},{"family":"Zhao","given":"Jingyi"}],"issued":{"date-parts":[["2022"]]}}}],"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[5]提出了基于金字塔特征的輕量級(jí)CNN（PFCNN）來檢測鋼軌表面缺陷，利用金字塔特征提取模塊（PFEM）基于缺陷特征和圖像背景提取多尺度特征圖，輸入到輕量級(jí)CNN中，具有較好的分割性能和實(shí)時(shí)性。Zhang等ADDINZOTERO_ITEMCSL_CITATION{"citationID":"BqvX6eYF","properties":{"formattedCitation":"\\super[6]\\nosupersub{}","plainCitation":"[6]","noteIndex":0},"citationItems":[{"id":748,"uris":["/users/local/Mc1qNPH1/items/TTZFEXPS"],"itemData":{"id":748,"type":"article-journal","abstract":"Defectsonrailsurfaces,whichhavebecomecriticalproblems,needtobedetectedandremovedasquicklyaspossibletoensurethefast,safe,andstableoperationoftrains.Atpresent,althoughmanysolutionshavebeenproposedtoaddresstheseproblems,thecomprehensiveness,rapidity,andaccuracyofdefectdetectionremainunsatisfactory.Thisstudyaimstoresolvetheseexistingproblemsandaccordinglyproposesamulti-modelrailsurfacedefectdetectionsystembasedonconvolutionalneuralnetworks(MRSDI-CNN)fromthestandpointofstudyingthesquatontherailsurface.TheconvolutionalneuralnetworksutilizedincludetheimprovedSingleShotMultiBoxDetector(SSD)andYouOnlyLookOnceversion3(YOLOv3)—twotypesofone-stagenetworks.Weexpoundedandanalyzedtheperformanceoftheconvolutionalneuralnetworksaswellastheirapplicabilitytorailsurfacedefectdetection.Weusedadiverserangeofraildefectsizestoimprovethedetectionperformanceofthetwodeeplearningnetworks,followingwhichtheycouldidentifythreetypesofsquatsinparallelwithimprovedaccuracyandwithoutreductionofthedetectionspeed.Theexperimentalresultsconfirmtheeffectivenessandsuperiorityoftheproposedmethodoverthoseofpreviousstudies.","container-title":"IEEETransactionsonIntelligentTransportationSystems","DOI":"10.1109/TITS.2021.3101053","ISSN":"1558-0016","issue":"8","note":"event-title:IEEETransactionsonIntelligentTransportationSystems","page":"11162-11177","source":"IEEEXplore","title":"MRSDI-CNN:Multi-ModelRailSurfaceDefectInspectionSystemBasedonConvolutionalNeuralNetworks","title-short":"MRSDI-CNN","volume":"23","author":[{"family":"Zhang","given":"Hui"},{"family":"Song","given":"Yanan"},{"family":"Chen","given":"Yurong"},{"family":"Zhong","given":"Hang"},{"family":"Liu","given":"Li"},{"family":"Wang","given":"Yaonan"},{"family":"Akilan","given":"Thangarajah"},{"family":"Wu","given":"Q.M.Jonathan"}],"issued":{"date-parts":[["2022",8]]}}}],"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[6]提出了一種基于改進(jìn)的SSD[43]和YOLOv3深度學(xué)習(xí)網(wǎng)絡(luò)的快速、高精度和實(shí)時(shí)的多網(wǎng)絡(luò)組合鋼軌表面缺陷檢測系統(tǒng)。通過兩個(gè)子網(wǎng)的并行檢測，兩種單階段算法的相互作用，降低了錯(cuò)誤檢測率，提高了檢測精度。Xu等ADDINZOTERO_ITEMCSL_CITATION{"citationID":"AkHJROKd","properties":{"formattedCitation":"\\super[7]\\nosupersub{}","plainCitation":"[7]","noteIndex":0},"citationItems":[{"id":706,"uris":["/users/local/Mc1qNPH1/items/GEUT76EQ"],"itemData":{"id":706,"type":"article-journal","abstract":"Railwaysubgradedefectistheseriousthreattotrainsafety.Vehicle-borneGPRmethodhasbecomethemainrailwaysubgradedetectiontechnologywithitsadvantagesofrapidnessandnondestructiveness.However,duetothelargeamountofdetectiondataandthevarietyindefectshapeandsize,defectrecognitionisachallengingtask.Inthiswork,themethodbasedondeeplearningisproposedtorecognizedefectsfromthegroundpenetratingradar(GPR)profileofsubgradedetectiondata.BasedontheFasterR-CNNframework,theimprovementstrategiesoffeaturecascade,adversarialspatialdropoutnetwork(ASDN),Soft-NMS,anddataaugmentationhavebeenintegratedtoimproverecognitionaccuracy,accordingtothecharacteristicsofsubgradedefects.TheexperimentalresultsindicatesthatcomparedwithtraditionalSVM+HOGmethodandthebaselineFasterR-CNN,theimprovedmodelcanachievebetterperformance.Themodelrobustnessisdemonstratedbyafurthercomparisonexperimentofvariousdefecttypes.Inaddition,theimprovementstomodelperformanceofeachimprovementstrategyareverifiedbyanablationexperimentofimprovementstrategies.Thispapertriestoexplorethenewthinkingfortheapplicationofdeeplearningmethodinthefieldofrailwaysubgradedefectrecognition.","container-title":"ScientificProgramming","DOI":"10.1155/2018/4832972","ISSN":"1058-9244","language":"en","note":"publisher:Hindawi","page":"e4832972","source":"","title":"RailwaySubgradeDefectAutomaticRecognitionMethodBasedonImprovedFasterR-CNN","volume":"2018","author":[{"family":"Xu","given":"Xinjun"},{"family":"Lei","given":"Yang"},{"family":"Yang","given":"Feng"}],"issued":{"date-parts":[["2018",6,14]]}}}],"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[7]提出一種集成特征級(jí)聯(lián)、對(duì)抗性空間丟棄網(wǎng)絡(luò)（ASDN）、Soft-NMS和數(shù)據(jù)增強(qiáng)的改進(jìn)FasterR-CNN的鐵路路基缺陷自動(dòng)識(shí)別方法，提高識(shí)別準(zhǔn)確率，改進(jìn)模型能夠獲得更好的性能和魯棒性。Bai等ADDINZOTERO_ITEMCSL_CITATION{"citationID":"JYzvaCDE","properties":{"formattedCitation":"\\super[8]\\nosupersub{}","plainCitation":"[8]","noteIndex":0},"citationItems":[{"id":737,"uris":["/users/local/Mc1qNPH1/items/Y9ZDKMMN"],"itemData":{"id":737,"type":"article-journal","abstract":"Accuratefastenerpositioningandstatedetectionformtheprerequisiteforensuringthesafeoperationofrailtrack.Thedemandsforintelligent,fastandaccuratedetectioncannotbesatisfiedbytraditionalmethodsusingimageprocessingandfastenerclassification.Inviewofthis,atwo-stageclassificationmodelbasedonthemodifiedFasterRegion-basedConvolutionNeuralNetwork(FasterR-CNN)andtheSupportVectorDataDescription(SVDD)algorithmsisproposedinthepaperforfastenerdetection.Firstly,thedatasetofdetectionimagesisbuiltwiththeimagesbeinglabeled,andtheclassificationanddetectionmodelbasedonFasterR-CNNisconstructedaccordingtothecharacteristicsofpracticalfastenerimages.Theanchorboxoptimizationfunctionisestablishedbylabeleddatasettooptimizetheboxofregionproposalnetworkinthemodel,toenhancethedetectionrateandaccuracyofdetection.Then,accordingtothedetectionresultbyFasterR-CNN,theSVDDalgorithmisappliedforthesecondstageclassificationofdeviatedfasteners,whichavoidsinaccurateclassificationcausedbydifferentdeviatedanglesoffasteners.Throughtheverificationandanalysisofpracticaldetectioncase,itisverifiedthattheproposedmethodcanimprovetheefficiencyandprecisionoffastenerdetectionwithhigherdetectionratesandaccuracyincomparisonwithotherbaselinedetectionmethods,makingitsuitableforfastandaccuratedetectionoffastenerstates.","container-title":"Measurement","DOI":"10.1016/j.measurement.2021.109742","ISSN":"0263-2241","journalAbbreviation":"Measurement","page":"109742","source":"ScienceDirect","title":"AnoptimizedrailwayfastenerdetectionmethodbasedonmodifiedFasterR-CNN","volume":"182","author":[{"family":"Bai","given":"Tangbo"},{"family":"Yang","given":"Jianwei"},{"family":"Xu","given":"Guiyang"},{"family":"Yao","given":"Dechen"}],"issued":{"date-parts":[["2021",9,1]]}}}],"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[8]提出一種基于FasterR-CNN模型的優(yōu)化RPN網(wǎng)絡(luò)進(jìn)行緊固件檢測，建立了基于FasterR-CNN和SVDD算法的兩階段分類模型，用于優(yōu)化鐵路緊固件缺陷檢測。Chen等ADDINZOTERO_ITEMCSL_CITATION{"citationID":"GV4J50Z9","properties":{"formattedCitation":"\\super[9]\\nosupersub{}","plainCitation":"[9]","noteIndex":0},"citationItems":[{"id":701,"uris":["/users/local/Mc1qNPH1/items/GKJSDYAX"],"itemData":{"id":701,"type":"article-journal","abstract":"Theexcitationandvibrationtriggeredbythelong-termoperationofrailwayvehiclesinevitablyresultindefectivestatesofcatenarysupportdevices.Withthemassiveconstructionofhigh-speedelectrifiedrailways,automaticdefectdetectionofdiverseandplentifulfastenersonthecatenarysupportdeviceisofgreatsignificanceforoperationsafetyandcostreduction.Nowadays,thecatenarysupportdevicesareperiodicallycapturedbythecamerasmountedontheinspectionvehiclesduringthenight,buttheinspectionstillmostlyreliesonhumanvisualinterpretation.Toreducethehumaninvolvement,thispaperproposesanovelvision-basedmethodthatappliesthedeepconvolutionalneuralnetworks(DCNNs)inthedefectdetectionofthefasteners.OursystemcascadesthreeDCNN-baseddetectionstagesinacoarse-to-finemanner,includingtwodetectorstosequentiallylocalizethecantileverjointsandtheirfastenersandaclassifiertodiagnosethefasteners'defects.ExtensiveexperimentsandcomparisonsofthedefectdetectionofcatenarysupportdevicesalongtheWuhan-Guangzhouhigh-speedrailwaylineindicatethatthesystemcanachieveahighdetectionratewithgoodadaptationandrobustnessincomplexenvironments.","container-title":"IEEETransactionsonInstrumentationandMeasurement","DOI":"10.1109/TIM.2017.2775345","ISSN":"1557-9662","issue":"2","note":"event-title:IEEETransactionsonInstrumentationandMeasurement","page":"257-269","source":"IEEEXplore","title":"AutomaticDefectDetectionofFastenersontheCatenarySupportDeviceUsingDeepConvolutionalNeuralNetwork","volume":"67","author":[{"family":"Chen","given":"Junwen"},{"family":"Liu","given":"Zhigang"},{"family":"Wang","given":"Hongrui"},{"family":"Nú?ez","given":"Alfredo"},{"family":"Han","given":"Zhiwei"}],"issued":{"date-parts":[["2018",2]]}}}],"schema":"/citation-style-language/schema/raw/master/csl-citation.json"}[9]結(jié)合了SSD、YOLO等網(wǎng)絡(luò)方法構(gòu)建了一個(gè)從粗到細(xì)的級(jí)聯(lián)檢測網(wǎng)絡(luò)對(duì)高鐵線路緊固件進(jìn)行缺陷檢測，在復(fù)雜環(huán)境下能夠?qū)崿F(xiàn)較高的檢測率，具有良好的適應(yīng)性和魯棒性。Cui等ADDINZOTERO_ITEMCSL_CITATION{"citationID":"RmvrEVMZ","properties":{"formattedCitation":"\\super[10]\\nosupersub{}","plainCitation":"[10]","noteIndex":0},"citationItems":[{"id":729,"uris":["/users/local/Mc1qNPH1/items/ZHUMEC49"],"itemData":{"id":729,"type":"article-journal","abstract":"Noisebarriersplayacriticalroleinreducingnoiseandpreventingforeignobjectfrominvadingrailway.Noisebarrierstructuraldefectssuchasrustedcolumn,deterioratedmortarlayerandotherdamagesmakeitsstructureunstable,therebythreateningseriouslyrailwayoperationsafety.Unfortunately,existingnoisebarrierinspectionmethodsstillrelyheavilyonmanualinspection,whicharelow-efficiency,subjectiveanddifficulttodetecttheexternalstructureofnoisebarriers.Tosolvetheseproblems,thisstudyproposesanautomaticinspectionmannerfornoisebarrierusingUAVimages,anddevelopsafullyconvolutionalnetwork(FCN)-basednoisebarrierdefectdetectionapproachnamedskipconnectionYOLOdetectionnetwork(SCYNet),whichfocusesonthreeaspects:networkstructure,lossfunctionanddataaugmentation.First,askip-connectedfeaturestructureSimi-BiFPNisincorporatedintothenetworktofullyfusethefeaturesextractedfromvariousscalelayerswithoutaddingmuchcomputationaloverhead.Second,aNoiseIoUlossforboundingboxregressionisdesignedtoimproveexistingIoU-basedlossesandgetbetterperformanceonsmalldataset.Thirdly,amixedsampledataaugmentationmethodnamedAutoFMixisproposedtoeliminatetheover-fittingissuecausedbyexcessivesimilaritybetweensamples,andfurtherimprovethedetectionaccuracy.Finally,experimentsconductedontheUAVrailwaynoisebarrierdatasetshowthattheproposedSCYNetmodelachieves92.2mAPand78.7FPS,respectively,whichoutperformothermodelsintermsofaccuracyandprocessingspeed.Thefast-processingspeedandhighdetectionaccuracycanquicklyturnUAVimagesintousefulinformationtoassistrailwaymaintenance,therebyimprovingthesafetyoftrainoperation.","container-title":"IEEETransactionsonIntelligentTransportationSystems","DOI":"10.1109/TITS.2023.3292934","ISSN":"1558-0016","issue":"11","note":"event-title:IEEETransactionsonIntelligentTransportationSystems","page":"12180-12195","source":"IEEEXplore","title":"SkipConnectionYOLOArchitectureforNoiseBarrierDefectDetectionUsingUAV-BasedImagesinHigh-SpeedRailway","volume":"24","author":[{"family":"Cui","given":"Jing"},{"family":"Qin","given":"Yong"},{"family":"Wu","given":"Yunpeng"},{"family":"Shao",