基于深度學(xué)習(xí)的輪對激光光條圖像修復(fù)研究

基于深度學(xué)習(xí)的輪對激光光條圖像修復(fù)研究基于深度學(xué)習(xí)的輪對激光光條圖像修復(fù)研究

摘要：

輪對激光光條圖像的質(zhì)量直接影響到鐵路運輸?shù)陌踩托剩虼藢ζ溥M(jìn)行修復(fù)具有重要意義。本文提出了一種基于深度學(xué)習(xí)的輪對激光光條圖像修復(fù)方法。首先，我們收集了大量原始和瑕疵圖像，用于構(gòu)建訓(xùn)練和測試數(shù)據(jù)集。接著，利用卷積神經(jīng)網(wǎng)絡(luò)進(jìn)行圖像修復(fù)，該網(wǎng)絡(luò)由編碼器、解碼器和反卷積操作組成。在訓(xùn)練階段，我們采用自編碼器和殘差學(xué)習(xí)以增強(qiáng)網(wǎng)絡(luò)的修復(fù)效果。在測試階段，根據(jù)網(wǎng)絡(luò)的輸出進(jìn)行自適應(yīng)像素分類，通過分別對不同的像素分配優(yōu)先級來保證修復(fù)效果優(yōu)良。實驗結(jié)果表明，本文提出的方法可以有效地修復(fù)輪對激光光條圖像，提高了圖像質(zhì)量及細(xì)節(jié)信息的恢復(fù)能力。

關(guān)鍵詞：輪對激光光條圖像，深度學(xué)習(xí)，自編碼器，殘差學(xué)習(xí)，自適應(yīng)像素分類

Abstract:

Thequalityofthewheel-raillaserstripeimagedirectlyaffectsthesafetyandefficiencyofrailwaytransportation.Therefore,itsrepairisofgreatsignificance.Inthispaper,weproposeadeeplearningbasedmethodforrepairingwheel-raillaserstripeimages.Firstly,wecollectedalargenumberoforiginalanddefectiveimagestoconstructtrainingandtestingdatasets.Then,aconvolutionalneuralnetworkisusedforimagerestoration,whichconsistsofanencoder,adecoder,anddeconvolutionoperations.Inthetrainingphase,weusetheautoencoderandresiduallearningtoenhancetherestorationeffectofthenetwork.Inthetestingphase,weadaptivelyclassifypixelsbasedonthenetworkoutput,andassigndifferentprioritiestodifferentpixelstoensuretherestorationeffectisgood.Experimentalresultsshowthattheproposedmethodcaneffectivelyrepairwheel-raillaserstripeimagesandimprovetheabilitytorestoreimagequalityanddetailinformation.

Keywords:wheel-raillaserstripeimage,deeplearning,autoencoder,residuallearning,adaptivepixelclassificationRailwaytransportationplaysasignificantroleinmoderntransportinfrastructure,andthesafetyandreliabilityofrailwaysystemsareessentialfactors.Onekeycomponentofrailwaysystemsisthewheel-railsystem,andthemonitoringofthewheel-railinterfaceisbecomingincreasinglyimportant.Laser-basedopticalmeasurementtechnologyhasbeenwidelyusedtomonitorthegeometryofrailtracks,includingthewheel-railcontactarea.Wheel-raillaserstripeimagingtechnologycanbeusedtoextractthecontactgeometryinformation,andithasbeenappliedinmanyrailwayinspectionscenarios.

However,thewheel-raillaserstripeimagescanbeseriouslydegradedbyvariousfactors,includingenvironmentalchanges,sensornoise,andotherartifacts.Thedegradedimagescanaffecttheaccuracyandreliabilityofrailmonitoringandcanresultinmisleadingresults,whichcouldimpactthesafetyoftherailwaysystem.Therefore,itiscrucialtodevelopeffectivemethodsforrestoringthedegradedwheel-raillaserstripeimages.

Inrecentyears,deeplearningmethodshaveachievedremarkablesuccessinvariousimagerestorationtasks,includingimagedenoising,super-resolution,andimageinpainting.Inthisstudy,weproposeanautoencoder-baseddeeplearningmethodforwheel-raillaserstripeimagerestoration.Inparticular,wedesignaresidualautoencodernetworkthatcaneffectivelycapturethecompleximagefeaturesandrestorethedegradedimagedetails.

Toovercomethelimitationsoftraditionaldeeplearningapproaches,weproposeanadaptivepixelclassificationschemetoprioritizetherestorationofdifferentimagepixels.Theproposedschemecanassignhigherprioritiestoimagepixelswithmoresignificantrestorationpotential,therebyensuringtherestorationqualityandretainingthecrucialinformationintheoriginalimage.

Experimentalresultsshowthattheproposedmethodcaneffectivelyrestorethedegradedwheel-raillaserstripeimagesandimprovetheimagequalityanddetailinformation.Ourapproachoutperformsotherstate-of-the-artimagerestorationmethodsintermsofrestorationaccuracyandcomputationalefficiency.Overall,ourproposedmethodcancontributetothesafeandreliableoperationofrailwaysystemsbyenhancingrailmonitoringaccuracyandreliabilityMoreover,theproposedmethodcanalsohavepotentialapplicationsinotherfields,suchasrobotics,manufacturing,andmedicalimaging,wherelaserstripeprojectioniscommonlyusedfor3Dsurfacemeasurementandinspection.Byrestoringthedegradedlaserstripeimages,ourapproachcanhelpimprovetheaccuracyandreliabilityofsurfacereconstructionanddefectdetection,whicharecriticalforqualitycontrolandproductevaluation.

Inadditiontotheproposedmethod,therearealsosomefutureresearchdirectionsthatcanbeexploredtofurtherimprovetheperformanceoflaserstripeimagerestoration.Forexample,incorporatingmorepriorknowledgeorconstraintsintotheimagerestorationprocess,suchasthegeometricstructureofthelaserstripeorthestatisticalcharacteristicsofthenoise,canhelpenhancetherestorationaccuracyandrobustness.Moreover,multi-viewormulti-frequencylaserstripeprojectioncanbeusedtoobtainmoreinformationaboutthesurfacetextureandshape,whichcanbeexploitedforbetterimagerestorationandfusion.

Overall,theproposedmethodpresentedinthispaperservesasapromisingsolutionforrestoringthedegradedwheel-raillaserstripeimages,whichcansignificantlybenefittherailwayindustrybyimprovingthesafety,efficiency,andreliabilityofrailmonitoringandmaintenance.TheproposedmethodcanalsohavebroaderapplicationsinotherfieldsthatinvolvelaserstripeprojectionandimagerestorationInadditiontotheapplicationsmentionedabove,theproposedmethodcanalsobeappliedtoothertypesoflaserstripeimages,suchasthoseproducedinmanufacturingandindustrialsettings.Forexample,laserstripesensorsarecommonlyusedin3Dscanningandmeasurement,wheretheycaptureobjectsurfaceinformationforinspectionandanalysis.However,thecapturedlaserstripeimagescanbeaffectedbyvariousfactorssuchasnoise,occlusion,andgeometricdistortion,whichcandegradethequalityoftheacquireddata.

Theproposedmethodcanbeadaptedtoaddressthesechallengesandenhancetheaccuracyandreliabilityof3Dmeasurementandinspection.Byeffectivelyremovingnoiseanddistortionfromthelaserstripeimages,theproposedmethodcanhelptoimprovetheprecisionandcompletenessofobjectsurfacereconstruction,whichiscriticalforqualitycontrolanddefectdetectioninmanufacturingandindustrialprocesses.

Moreover,theproposedmethodcanbeintegratedwithotherimageprocessingtechniques,suchasfeaturedetectionandtracking,toenablereal-timeanalysisandfeedbackindynamicenvironments.Forinstance,inroboticsandautomation,laserstripesensorscanbeusedtoguidethemotionandmanipulationofroboticarmsandtools.Theproposedmethodcanhelptoimprovetheaccuracyandrobustnessofthesensingandcontrolsystem,byprovidingreliableandaccuratefeedbackoftheobjectsurfacecharacteristics