基于多尺度注意力和HFS的路面圖像裂縫檢測(cè)方法

基于多尺度注意力和HFS的路面圖像裂縫檢測(cè)方法基于多尺度注意力和HFS的路面圖像裂縫檢測(cè)方法

摘要：路面裂縫是道路結(jié)構(gòu)退化的主要表現(xiàn)之一，目前在路面維護(hù)和安全管理中起著至關(guān)重要的作用。本文提出了一種基于多尺度注意力和HFS（譜聚類的半監(jiān)督方法）的路面圖像裂縫檢測(cè)方法。首先，使用多尺度注意力機(jī)制對(duì)路面圖像進(jìn)行分割，得到裂縫和非裂縫區(qū)域的圖像分割結(jié)果。然后，使用HFS方法將未標(biāo)記數(shù)據(jù)集中的樣本與標(biāo)記數(shù)據(jù)集進(jìn)行聚類，利用聚類結(jié)果對(duì)多尺度注意力機(jī)制進(jìn)行優(yōu)化，從而提高路面圖像裂縫的檢測(cè)精度。在公開數(shù)據(jù)集上的實(shí)驗(yàn)結(jié)果表明，本文提出的方法具有較高的檢測(cè)精度和魯棒性，可用于路面裂縫檢測(cè)領(lǐng)域的實(shí)際應(yīng)用。

關(guān)鍵詞：路面圖像；裂縫檢測(cè)；多尺度注意力；HFS；聚類

簡(jiǎn)介

路面裂縫是道路結(jié)構(gòu)退化的主要表現(xiàn)之一，它們的形成和擴(kuò)展過(guò)程是公路損傷評(píng)估和維護(hù)管理的主要關(guān)注點(diǎn)。因此，在公路養(yǎng)護(hù)和管理方面，路面裂縫檢測(cè)是一項(xiàng)至關(guān)重要的任務(wù)。

路面裂縫檢測(cè)技術(shù)主要基于計(jì)算機(jī)視覺(jué)和模式識(shí)別技術(shù)，其中深度學(xué)習(xí)技術(shù)已經(jīng)成為最先進(jìn)的裂縫檢測(cè)方法之一。但是，當(dāng)前的深度學(xué)習(xí)方法仍存在一些問(wèn)題，例如對(duì)數(shù)據(jù)量和質(zhì)量的依賴性較強(qiáng)、參數(shù)數(shù)量較大等。因此，有必要研究一種新型的裂縫檢測(cè)方法，以解決這些問(wèn)題。

在這篇論文中，我們提出了一種基于多尺度注意力和HFS的路面圖像裂縫檢測(cè)方法。該方法主要包括以下幾個(gè)步驟：首先使用多尺度注意力機(jī)制對(duì)路面圖像進(jìn)行分割，得到裂縫和非裂縫區(qū)域的圖像分割結(jié)果；然后，利用HFS方法將未標(biāo)記數(shù)據(jù)集中的樣本與標(biāo)記數(shù)據(jù)集進(jìn)行聚類，從而獲得更準(zhǔn)確的圖像分割結(jié)果；最后，我們利用聚類結(jié)果對(duì)多尺度注意力機(jī)制進(jìn)行優(yōu)化，以提高圖像分割的精度。

實(shí)驗(yàn)結(jié)果表明，我們提出的方法在公開數(shù)據(jù)集上具有較高的精度和魯棒性，可用于路面裂縫檢測(cè)領(lǐng)域的實(shí)際應(yīng)用。

方法

多尺度注意力機(jī)制

圖像分割是裂縫檢測(cè)的關(guān)鍵步驟。在我們的方法中，我們采用多尺度注意力機(jī)制進(jìn)行圖像分割。

為了獲取多個(gè)分辨率的圖像特征，我們首先使用圖像金字塔對(duì)原始圖像進(jìn)行處理。然后，通過(guò)CNN網(wǎng)絡(luò)對(duì)每個(gè)尺度下的圖像進(jìn)行特征提取。

接下來(lái)，我們采用注意力機(jī)制來(lái)加權(quán)不同尺度的特征圖，以獲得更準(zhǔn)確的圖像分割結(jié)果。特別地，我們使用全局平均池化層將每個(gè)特征圖轉(zhuǎn)化為一個(gè)特征向量，接著使用一組多層感知機(jī)對(duì)特征向量進(jìn)行加權(quán)，然后將加權(quán)后的特征向量加權(quán)求和，得到最終的特征表示。該注意力機(jī)制的數(shù)學(xué)形式如下：

$F_{a}=\sum_{i=1}^{n}W_{i}F_{i}$

其中，$F_{i}$表示第$i$個(gè)尺度下的特征圖，$W_{i}$表示該特征圖的注意力分?jǐn)?shù)。

HFS方法

HFS方法是一種基于譜聚類的半監(jiān)督學(xué)習(xí)方法，它可以利用未標(biāo)記的數(shù)據(jù)來(lái)提高模型的表現(xiàn)。

在我們的方法中，我們利用HFS方法來(lái)獲得更準(zhǔn)確的圖像分割結(jié)果。具體地，我們使用已標(biāo)記數(shù)據(jù)集和未標(biāo)記數(shù)據(jù)集來(lái)訓(xùn)練CNN模型，其中已標(biāo)記數(shù)據(jù)集用于監(jiān)督訓(xùn)練，未標(biāo)記數(shù)據(jù)集用于半監(jiān)督學(xué)習(xí)。

HFS方法的步驟如下：

1.構(gòu)建圖：對(duì)于已標(biāo)記數(shù)據(jù)集和未標(biāo)記數(shù)據(jù)集中的每個(gè)樣本，計(jì)算它們之間的相似度，并將相似度矩陣轉(zhuǎn)化為鄰接矩陣。

2.譜聚類：對(duì)鄰接矩陣進(jìn)行譜聚類，得到聚類結(jié)果。

3.半監(jiān)督學(xué)習(xí)：使用已標(biāo)記數(shù)據(jù)集進(jìn)行監(jiān)督學(xué)習(xí)，并利用聚類結(jié)果對(duì)未標(biāo)記數(shù)據(jù)集進(jìn)行半監(jiān)督學(xué)習(xí)，從而獲得更準(zhǔn)確的圖像分割結(jié)果。

優(yōu)化

最后，我們使用聚類結(jié)果對(duì)多尺度注意力機(jī)制進(jìn)行優(yōu)化。具體來(lái)說(shuō)，我們計(jì)算每個(gè)聚類的權(quán)重，然后將這些權(quán)重用于注意力機(jī)制中，以提高圖像分割的精度。該優(yōu)化的數(shù)學(xué)形式如下：

$F_{a}=\sum_{i=1}^{k}\sum_{j=1}^{n}W_{i,j}F_{i,j}$

其中，$W_{i,j}$表示第$i$個(gè)聚類中第$j$個(gè)尺度的權(quán)重，$F_{i,j}$表示第$i$個(gè)聚類中第$j$個(gè)尺度下的特征圖。

實(shí)驗(yàn)

我們使用了公開數(shù)據(jù)集CRACK500和MERL,并比較我們的方法與其他基于深度學(xué)習(xí)的方法。實(shí)驗(yàn)結(jié)果表明，我們的方法在兩個(gè)數(shù)據(jù)集上均取得了最好的效果，具有較高的檢測(cè)精度和魯棒性。

結(jié)論

本文提出了一種基于多尺度注意力和HFS的路面圖像裂縫檢測(cè)方法。該方法利用多尺度注意力機(jī)制對(duì)路面圖像進(jìn)行分割，然后使用HFS方法對(duì)未標(biāo)記數(shù)據(jù)進(jìn)行半監(jiān)督學(xué)習(xí)，最后使用聚類結(jié)果對(duì)多尺度注意力機(jī)制進(jìn)行優(yōu)化。實(shí)驗(yàn)結(jié)果表明，我們的方法在公開數(shù)據(jù)集上具有較高的精度和魯棒性，可用于路面裂縫檢測(cè)領(lǐng)域的實(shí)際應(yīng)用。Introduction

Crackdetectioninpavementimagesisanimportanttaskinthefieldoftransportationengineering.Accuratedetectionofcrackscanhelppreventaccidentsandprolongthelifespanofroads.Inrecentyears,deeplearningmethodshavebeenappliedtothetaskofcrackdetection,withpromisingresults.However,thesemethodsoftenrelyonalargeamountoflabeleddata,whichcanbeexpensiveandtime-consumingtoobtain.Inthispaper,weproposeasemi-supervisedcrackdetectionmethodusingmultiple-scaleattentionandHFS.

Methodology

Ourproposedmethodconsistsofthreemainsteps:multiple-scaleattention,HFS-basedsemi-supervisedlearning,andoptimizationusingclusteringresults.

Multiple-scaleattention:

Weadoptamulti-scaleattentionmechanismtosegmentthepavementimages.Thismechanismenablesthemodeltofocusondifferentscalesoffeatures,whichiscriticalfordetectingcracksofvarioussizes.Specifically,wefirstadoptaU-Net-likenetworktoobtainfeaturemapsatdifferentscales.Then,weapplytheattentionmechanismtoeachscaleoffeaturemapstoobtainattentionmaps.Finally,wecombinetheattentionmapstogenerateasegmentationmap.

HFS-basedsemi-supervisedlearning:

WeuseHFStoleverageunlabeleddataandenhancetheaccuracyofthecrackdetectionmodel.HFSisagraph-basedsemi-supervisedlearningmethodthatcanpropagatelabelsfromlabeleddatatounlabeleddatabasedonthegraphstructure.Weconstructagraphusingthefeatureembeddingsofthelabeledandunlabeleddata,andapplyHFStoobtainlabelsfortheunlabeleddata.

Optimizationusingclusteringresults:

Wefurtheroptimizethemultiple-scaleattentionmechanismusingclusteringresults.Specifically,wecalculatetheweightsofeachscaleoffeaturemapsaccordingtotheclusteringresults,andusetheseweightsintheattentionmechanismtoimprovetheaccuracyofthesegmentation.

ExperimentalResults:

Weevaluateourproposedmethodontwopublicdatasets,CRACK500andMERL.Ourmethodoutperformsseveralstate-of-the-artmethodsonbothdatasetsintermsofaccuracyandrobustness.

Conclusion

Inthispaper,weproposeasemi-supervisedcrackdetectionmethodusingmultiple-scaleattentionandHFS.Ourmethodleveragestheadvantagesofmultiple-scaleattentionforsegmentationandHFSforsemi-supervisedlearning.Experimentalresultsshowthatourmethodachieveshighaccuracyandrobustnessonpublicdatasets,andcanbepotentiallyappliedtoreal-worldcrackdetectionapplicationsintransportationengineering。Futuredirections

Despitethepromisingresultsachievedbyourproposedmethod,therearestillseveraldirectionsthatfutureresearchcanexploretoimprovetheperformanceandpracticalityofcrackdetectionintransportationengineering.Somepotentialdirectionsinclude:

1.Multi-modaldatafusion:Integratingdifferenttypesofdata,suchasRGBimages,thermalimages,andpointclouds,canprovidecomplementaryinformationandenhancetheaccuracyandrobustnessofcrackdetection.

2.Transferlearning:Thepre-trainedmodelsincomputervisiontaskscanbefine-tunedforcrackdetectiontasks,whichcanimprovetheperformanceofdeeplearningmodelswithlimitedlabeleddata.

3.Real-timedetection:Theproposedmethodneedstosegmenttheinputimageintosmallpatches,whichmakesitdifficulttoapplyinreal-timedetectionscenarios.Futureresearchcanexplorereal-timecrackdetectionmethods,suchasmulti-scaleslidingwindow,toachievereal-timeperformancewithoutsacrificingaccuracy.

4.Deploymentinthefield:Itisessentialtoverifytheperformanceofcrackdetectionmethodsinreal-worldscenarios.Futureresearchcaninvestigatethepracticalityandfeasibilityofdeployingcrackdetectionmethodsonunmannedaerialvehiclesorground-basedinspectionvehicles.

Overall,crackdetectionintransportationengineeringisanessentialtaskthatcancontributetothemaintenanceandsafetyoftransportationinfrastructure.Theproposedsemi-supervisedcrackdetectionmethodusingmultiple-scaleattentionandHFSshowspromisingresultsonpublicdatasetsandprovidesinsightsforfurtherresearchinthefield。Tofurtherextendtheapplicabilityofcrackdetectionmethods,itisworthexploringthedeploymentofsuchmethodsonunmannedaerialvehicles(UAVs)orground-basedinspectionvehicles.Theseautomatedinspectionvehiclescanhelptoreducetheinspectiontimeandcost,improvethedetectionaccuracy,andenhancethesafetyoftheinspectionprocess.

ForUAVs,theyhavetheadvantageofaccessingdifficult-to-reachareas,coveringlargeareasinashorttime,andprovidinghigh-resolutionimagesforcrackdetection.ThecrackdetectionalgorithmscanbeintegratedintotheUAVs'onboardsystems,andtheUAVscanflyoverthetransportationinfrastructureandcaptureimagesofthesurfaces.Theimagescanthenbeprocessedinreal-timeonboardortransmittedtoagroundstationforfurtheranalysis.TheUAVscanalsobeequippedwithsensorssuchasLiDARtoimprovethedetectionaccuracy.

However,therearealsochallengesindeployingcrackdetectionmethodsonUAVs.TheUAVsneedtomaintainastablepositionandaltitudeduringtheinspectionprocesstoensurethequalityoftheimages.TheUAVsalsoneedtonavigatethroughcomplexterrainandavoidobstaclessuchastreesandbuildings.ThewindandweatherconditionscanalsoaffectthestabilityandsafetyoftheUAVs.

Forground-basedinspectionvehicles,theycanmoveonroadsorrailsandperforminspectiontaskscontinuously.Thevehiclescancarryvarioussensorsandcamerastodetectcracksandcaptureimagesofthesurfaces.Theimagescanbeprocessedonboardortransmittedtoacentralstationforfurtheranalysis.Thevehiclescanalsobeequippedwithartificialintelligencealgorithmstodetectcracksautomaticallyandprovidereal-timefeedbacktotheoperators.

SimilartoUAVs,therearealsochallengesindeployingcrackdetectionmethodsonground-basedinspectionvehicles.Thevehiclesneedtoensureastableandsmoothmotionduringtheinspectionprocesstoavoidblurrinessanddistortionintheimages.Thevehiclesalsoneedtoavoidcollisionswithobstaclessuchaspedestriansandothervehicles.Themaintenanceandrepairofthevehiclescanalsobecostlyandtime-consuming.

Insummary,thedeploymentofcrackdetectionmethodsonUAVsorground-basedinspectionvehiclescanbringmanybenefitstothetransportationengineeringfield.However,thechallengesinensuringthestability,safety,andefficiencyoftheinspectionvehiclesneedtobeaddressedtomakethedeploymentfeasibleandpractical.Withfurtherresearchanddevelopment,theautomatedinspectionvehiclescanbecomeanindispensabletoolfortransportationinfrastructuremaintenanceandsafety。Onepotentialchallengeindeployingcrackdetectionsystemsoninspectionvehiclesistheneedforaccurateandreliabledataprocessing.AsUAVsorground-basedvehiclesmovealongtheinfrastructure,theycapturelargeamountsofdata,whichmustbeanalyzedandinterpretedinreal-timetoidentifycracksandotherdefects.Thisrequiressophisticatedalgorithmsandsoftwarethatcanrapidlyprocessdatastreamsandaccuratelydetectanomalies.

Anotherchallengeisensuringthesafetyandreliabilityoftheinspectionprocess.Ascrackdetectionsystemsaredeployedonmovingvehicles,thereisariskofcollisionswithotherobjects,aswellaschallengesinnavigatingthroughcomplexanddynamicenvironments.Additionally,thesystemsmustbeabletooperatereliablyunderawiderangeofweatherandlightingconditions,aswellasinharshenvironmentsthatmaycontaindust,debris,andothercontaminants.

Toovercomethesechallenges,researchersareexploringavarietyofinnovativeapproachestocrackdetection,includingusingAIandmachinelearningalgorithms,advancedsensorsandcameras,andothertechnologiesthatcanenhancetheaccuracyandreliabilityoftheinspectionprocess.Forexample,someresearchersareusingdeeplearningtechniquestotrainneuralnetworkstorecognizethesubtlepatternsofcracksandotherdefectsininfrastructuresurfaces,whileothersareexploringtheuseof3Dimagingtechnologytomapandmodeltheinfrastructureinreal-time.

Inadditiontothesetechnicalchallenges,theremayalsoberegulatoryandethicalconsiderationsthatmustbeaddressedwhendeployingcrackdetectionsystemsonpublicinfrastructure.Forexample,theremaybeconcernsaboutprivacyanddatasecurity,aswellasconcernsabouttheimpactoftheinspectionprocessonlocalcommunitiesandtheenvironment.

Despitethesechallenges,thepotentialbenefitsofautomatedcrackdetectionsystemsareclear.Byenablingmorefrequentandaccurateinspections,thesesystemscanhelpidentifydefectsandstructuralweaknessesintransportationinfrastructurebeforetheybecomemajorsafetyhazards.Moreover,byreducingtheneedformanualinspections,theycanhelpsavetimeandcostswhileimprovingoverallinfrastructuremaintenanceandmanagement.

Inconclusion,crackdetectionsystemsareapromisingareaofresearchinthetransportationengineeringfield,andthedeploymentofsuchsystemsonUAVsorground-basedinspectionvehiclescanbringmanybenefitstoinfrastructuremaintenanceandsafety.Withcontinuedresearchanddevelopment,thesesystemshavethepotentialtobecomeanessentialtoolforensuringtheintegrityoftransportationinfrastructure,whilereducingcosts,improvingefficiency,andenhancingpublicsafety。Inconclusion,crackdetectionsystemsareacriticalcomponentoftransportationinfrastructuremaintenanceandsafety.ThedeploymentofthesesystemsonUAVsandground-basedinspectionvehiclescansignificantlyenhancetheefficiencyandeffectivenessofinfrastructuremaintenanceprogramswhilereducingoverallcosts.

Advancedcrackdetectionsystemscanprovidereal-timemonitoringofinfrastructureconditions,identifyingpotentialissuesbeforetheybecomeseriousproblems.Thiscanultimatelyleadtosafertransportationsystemsandreducedrepaircosts.

However,despitethepotentialbenefitsofthesesystems,thereisstillaneedforcontinuedresearchanddevelopmenttoimprovetheiraccuracyandreliability.Inaddition,regulatoryandlegalframeworksmustbeputinplacetoensurethesafeandresponsibleuseofthesesystems.

Overall,thefutureofcrackdetectionsystemslookspromising,andcontinuedinnovationinthisareahasthepotentialtorevolutionizetransportationinfrastructuremaintenanceandsafety.Astheworldbecomesincreasinglydependentontransportationforsocialandeconomicprogress,investinginadvancedcrackdetectionsystemswillbecomemorecriticalthaneverbefore。Inadditiontothebenefitsdiscussedabove,crackdetectionsystemsalsohavethepotentialtoreducetransportationcostsandimproveoverallefficiency.Byidentifyingissuesearly,repairscanbescheduledbeforetheybecomemoreextensiveandtime-consuming,reducingtheoveralldowntimeoftransportationsystems.This,inturn,canleadtoincreasedproductivityandrevenuegeneration.

However,implementingcrackdetectionsystemsonalargescalecanbeexpensive,andthus,itisessentialtodevelopcost-effectivemodelsthatcanbeimplementedwithoutstrainingbudgets.Somecost-effectivesolutionsincludeusingopticalfibersensors,whichcanbeinstalledduringtheconstructionoftheinfrastructure,andwirelesssensorsthatdonotrequireextensivewiringandcanbeinstalledeasily.

Furthermore,aswithanytechnology,thereisaneedtoensurethattherearenounintendednegativeconsequencesofcrackdetectionsystems.Forinstance,theuseofwirelesssensorsmayraiseprivacyconcerns,asthesensorscangatherdataonthemovementofpeoplewithint