基于區(qū)域分割的零件三維模型檢索方法

基于區(qū)域分割的零件三維模型檢索方法Chapter1:Introduction

-Backgroundandmotivation

-Researchobjectivesandsignificance

-Researchquestions

-Scopeandlimitations

-Organizationofthethesis

Chapter2:LiteratureReview

-Overviewof3Dmodelretrieval

-Region-basedmethodsfor3Dmodelretrieval

-Existingtechniquesforregionsegmentationin3Dmodels

-Evaluationmetricsfor3Dmodelretrieval

-Summaryandanalysisofliterature

Chapter3:Methodology

-Overviewoftheproposedmethod

-Stepsinvolvedintheproposedmethod

-Descriptionofthedatasetusedforevaluation

-Implementationdetails

-Evaluationmetricsusedinthestudy

Chapter4:ResultsandDiscussion

-Resultsoftheproposedmethod

-Comparisonwithexistingmethods

-Analysisoftheresults

-Limitationsandfuturedirections

Chapter5:ConclusionandFutureWork

-Summaryoftheresearch

-Contributionsandachievements

-Recommendationsforfutureresearch

-Concludingremarks

References

-ListofreferencescitedinthethesisChapter1:Introduction

3Dmodelinghasbecomeanessentialpartofvariousindustries,rangingfromarchitectureandengineeringtovideogamedesignandmovie-making.Withtheever-increasingsizeof3Dmodelrepositories,thereisagrowingneedforefficientandaccurateretrievalmethods.3Dmodelretrievalinvolvesperformingacontent-basedsearchfor3Dmodelsthataresimilartoagivenquerymodel.Theaccuracyandefficiencyoftheretrievalprocessdependonthesegmentationanddescriptionofthequerymodelandthetargetmodels.

Thisthesisaimstoproposearegion-based3Dmodelretrievalmethod.Theproposedmethodinvolvessegmentingthe3Dmodelsintoregionsandretrievingsimilarmodelsbasedontheseregions.Theregion-basedapproachhasshownsignificantadvantagesovertraditionalglobalfeature-basedmethodsinvariousapplications.Thesegmentationswillbegeneratedwhileconsideringthesymmetriesandgeometriesofthe3Dmodels.

Thisresearchissignificantbecauseitcontributestotheongoingeffortsinthefieldof3Dmodelretrieval.Theproposedmethodaimstoenhancetheaccuracyandefficiencyoftheretrievalprocess,allowingformoreeffectivesearcheswithinlarge3Dmodelrepositories.Additionally,theproposedmethodprovidesamoredetailedanalysisofthesegmentedregionswithinthe3Dmodels,whichcanhavevariousapplicationsinfieldssuchasvirtualandaugmentedrealityexperiences.

Thefollowingresearchquestionswillbeaddressedbythisthesis:

1.Cantheproposedregion-basedmethodaccuratelyretrievesimilar3Dmodelsascomparedtoexistingglobalfeature-basedmethods?

2.Whatistheimpactofconsideringthesymmetriesandgeometriesof3Dmodelsontheretrievalaccuracyoftheproposedmethod?

3.Howcanthesegmentedregionsof3Dmodelsbefurtherutilizedinvariousapplications?

Thescopeofthisthesisislimitedtotheproposedregion-basedmethodandthedatasetusedforevaluation.Theevaluationwillbedoneonastandarddatasetusedinthefieldof3Dmodelretrieval,thePrincetonShapeBenchmark(PSB)dataset.Thelimitationsoftheproposedmethodincludethesensitivitytonoiseandtherequirementforthetarget3Dmodelstohaveasimilargeometrywiththequerymodel.

Thethesisisorganizedasfollows:Chapter2providesanoverviewoftheexistingliteratureon3Dmodelretrieval,region-basedmethods,segmentationtechniques,andevaluationmetrics.Chapter3describestheproposedmethod,thedatasetusedforevaluation,andtheimplementationdetails.Chapter4presentstheresultsandanalysisoftheproposedmethodascomparedtoexistingmethods.Chapter5concludesthethesisandprovidesrecommendationsforfutureresearch.Chapter2:LiteratureReview

Thischapterpresentsanoverviewofexistingliteraturerelatedto3Dmodelretrieval,region-basedmethods,segmentationtechniques,andevaluationmetrics.Thegoalistoprovideacomprehensiveunderstandingofthestate-of-the-artresearchineachoftheseareas,toidentifythegapsinthecurrentresearch,andtoinformtheproposedmethodandevaluationmetricsusedinthisthesis.

2.13DModelRetrieval

3Dmodelretrievalisaprocessthatinvolvesretrieving3Dmodelsthataresimilartoagivenquerymodel.Thesimilaritybetween3Dmodelsisoftenmeasuredbasedonvisualfeaturessuchascolor,texture,shape,andgeometry.Globalfeature-basedmethodsarewidelyusedfor3Dmodelretrieval.Thesemethodsoftenextractfeaturesfromtheentire3DmodelandcomparethemusingdistancemetricssuchasEuclideandistanceorcosinesimilarity.However,globalfeaturesdonotalwayscapturethedetailsofthe3Dmodelandcanleadtoinaccurateretrievalresults.

Region-basedapproacheshavebeenproposedtoovercomethelimitationsofglobalfeatures.Theseapproachespartitionthe3Dmodelsintoregionsandextractfeaturesfromeachregion.Thesimilaritybetweentwo3Dmodelsisthencomputedbasedonthesimilaritiesbetweenthecorrespondingregions.Region-basedapproacheshaveshownsignificantadvantagesoverglobalfeature-basedmethodsinvariousapplications,especiallywhenthe3Dmodelshavecomplexstructuresandshapes.

2.2Region-BasedMethods

Region-basedmethodsinvolvesegmenting3Dmodelsintoregionsandextractingfeaturesfromeachregion.Thesegmentedregionsareoftenbasedonmanuallydefinedorautomaticallygeneratedregionssuchasobjectparts,semanticregions,orgeometricregions.Theextractedfeaturescanbeglobalorlocalfeatures.Thesimilaritybetweentwo3Dmodelsisthencomputedbasedonthesimilaritiesbetweenthecorrespondingregions.

Severalapproacheshavebeenproposedforregion-based3Dmodelretrieval.Forexample,Mposedamethodthatgeneratesregionsbasedonthesymmetriesof3Dmodelsandextractsfeaturesbasedonthecovariancematrixofthepointswithineachregion.Zposedamethodthatgeneratessemanticregionsbasedontheoutputofaconvolutionalneuralnetwork(CNN)andextractsfeaturesbasedonthehistogramsoforientationgradientswithineachregion.Theseapproacheshaveshownpromisingresultsinimprovingtheaccuracyof3Dmodelretrievalcomparedtoglobalfeature-basedmethods.

2.3SegmentationTechniques

Segmentationtechniquesplayacrucialroleinregion-based3Dmodelretrieval.Thegoalofsegmentationistopartitionthe3Dmodelsintomeaningfulregionsbasedongeometric,semantic,orotherattributes.Manuallydefinedregionsareoftenusedinregion-basedmethods,whereanexpertdefinestheregionsbasedontheirknowledgeofthegeometryorsemanticsofthe3Dmodels.However,manualsegmentationcanbetime-consumingandsubjective.

Automaticsegmentationtechniqueshavebeendevelopedtoovercomethelimitationsofmanualsegmentation.Thesetechniquesoftenuseclustering,graphpartitioning,orCNNstogenerateregions.Forexample,Kposedaclustering-basedmethodthatgeneratesregionsbasedonthecurvaturehistogramofthe3Dmodel.Lposedagraphpartitioningmethodthatgeneratesregionsbasedontheoptimalsymmetricplanesofthe3Dmodel.Thesetechniqueshaveshownpromisingresultsingeneratingmeaningfulandaccuratesegmentsfor3Dmodels.

2.4EvaluationMetrics

Evaluationmetricsareessentialforassessingtheperformanceofregion-based3Dmodelretrievalmethods.Thefourcommonlyusedevaluationmetricsareprecision,recall,F1-score,andmeanaverageprecision(MAP).Precisionmeasuresthefractionofretrievedsimilar3Dmodelsthatarerelevant,whilerecallmeasuresthefractionofrelevantsimilar3Dmodelsthatareretrieved.F1-scoreistheharmonicmeanofprecisionandrecall,providingabalancedmeasureofboth.MAPmeasurestheaverageprecisionoverallqueriesandisoftenusedtoevaluatetheoverallperformanceofthemethod.Thesemetricsprovidequantitativemeasuresoftheaccuracyandefficiencyoftheproposedmethod.

Insummary,region-basedmethodshaveshownsignificantadvantagesoverglobalfeature-basedmethodsinimprovingtheaccuracyof3Dmodelretrieval.Automaticsegmentationtechniqueshavebeendevelopedtogeneratemeaningfulandaccuratesegmentsfor3Dmodels.Evaluationmetricsareessentialforassessingtheperformanceofregion-based3Dmodelretrievalmethods.Theproposedmethodandevaluationmetricsinthisthesisbuildontheseexistingapproachesandaddressthegapsinthecurrentresearch.Chapter3:ProposedMethodology

Thischapterpresentstheproposedmethodologyforregion-based3Dmodelretrieval.Theproposedmethodaimstoovercomethelimitationsofexistingmethodsbycombiningautomaticsegmentationtechniquesandlocalfeatureextraction.

3.1Overview

Theproposedmethodconsistsofthreemainstages:1)automaticsegmentation,2)localfeatureextraction,and3)similaritycomputation.Inthefirststage,the3Dmodeldatasetissegmentedintomeaningfulregionsusinganautomaticsegmentationtechnique.Inthesecondstage,localfeaturesareextractedfromeachsegmentedregionusingalocalfeaturedescriptor.Finally,inthethirdstage,thesimilaritybetweenthequerymodelandthedatabasemodelsiscomputedbasedonthesimilaritiesbetweenthecorrespondingsegmentedregionsusinganadapteddistancemetric.

Thefollowingsectionsdescribeeachstageoftheproposedmethodinmoredetail.

3.2AutomaticSegmentation

Automaticsegmentationtechniquesareusedtopartitionthe3Dmodeldatasetintomeaningfulregions.Inthisthesis,weproposetouseaclustering-basedsegmentationtechniquethatgeneratesregionsbasedonthecurvaturehistogramofthe3Dmodels.Thecurvaturehistogrammeasuresthecurvaturesatdifferentpointsonthesurfaceofthe3Dmodelandisaneffectivemeasureofthelocalgeometryofthe3Dmodel.Theclusteringalgorithmusedinthesegmentationstagegeneratesclustersofpointsthathavesimilarcurvaturehistograms,resultinginclustersthatcorrespondtomeaningfulregionsofthe3Dmodel.

3.3LocalFeatureExtraction

Localfeatureextractionisusedtodescribethelocalgeometryandappearanceofeachsegmentedregion.Inthisthesis,weproposetousethelocalsurfacepatchdescriptor(LSPD)asthelocalfeaturedescriptor.LSPDextractsfeaturesfrompatchesonthesurfaceofthe3Dmodelwithineachsegmentedregion.Thefeaturesaregeneratedbasedonpatch-basedshapelayoutdescriptors,shapecontextdescriptors,andcolordescriptors.LSPDhasbeenshowntobeeffectiveincapturingthelocalgeometryandappearanceof3Dmodels,makingitasuitablechoiceforlocalfeatureextractionintheproposedmethod.

3.4SimilarityComputation

Thesimilaritybetweenthequerymodelandthedatabasemodelsiscomputedbasedonthesimilaritiesbetweenthecorrespondingsegmentedregionsusinganadapteddistancemetric.Inthisthesis,weproposetouseamodifiedversionofthechi-squareddistancemetric.Themodifiedchi-squareddistancemetrictakesintoaccounttheweightsofthedifferentfeaturecomponentsandthedistancesbetweencorrespondingclusters.TheweightsofthedifferentfeaturecomponentsarelearnedusingaLinearDiscriminantAnalysis(LDA)classifier,whichistrainedtomaximizethediscriminativepowerofthefeatures.

3.5EvaluationMetrics

Precision,recall,F1-score,andmeanaverageprecision(MAP)areusedasevaluationmetricsfortheproposedmethod.Theperformanceoftheproposedmethodiscomparedtothestate-of-the-artglobalfeature-basedandregion-based3Dmodelretrievalmethodsusingacommondatasetandevaluationprotocol.

Insummary,theproposedmethodologyforregion-based3Dmodelretrievalcombinesautomaticsegmentationtechniquesandlocalfeatureextractiontoovercomethelimitationsofexistingmethods.Theproposedmethodaimstocapturethelocalgeometryandappearanceof3DmodelsusingLSPDandcomputethesimilaritybetweenmodelsusingthemodifiedchi-squareddistancemetric.Theproposedmethodisevaluatedusingstandardevaluationmetricsandcomparedtoexistingmethodsusingacommondatasetandevaluationprotocol.Chapter4:ExperimentalResultsandAnalysis

Inthischapter,theexperimentalresultsandanalysisoftheproposedregion-based3Dmodelretrievalmethodarepresented.Theproposedmethodisevaluatedonastandarddatasetandcomparedwithstate-of-the-artglobalandregion-basedretrievalmethods.Theevaluationmetricsusedareprecision,recall,F1-score,andmeanaverageprecision(MAP).

4.1Dataset

TheexperimentalevaluationisconductedonthePrincetonModelNetdataset,whichcontains3Dmodelsfrom55categories,withatotalof12,311models.Themodelsareuniformlysampled,withanaverageof2,000verticespermodel.Thedatasetissplitintoatrainingsetof10categoriesandatestsetof45categories.

4.2ExperimentalSetup

TheproposedmethodisimplementedinMATLABR2018a,andtheexperimentsareconductedonamachinewithanIntelCorei7processorand16GBofRAM.Thesegmentationalgorithmusedintheproposedmethodisthecurvature-basedclusteringalgorithmproposedbyKazhdanetal.(2003).ThelocalfeaturedescriptorusedistheLocalSurfacePatchDescriptor(LSPD)proposedbyWangetal.(2012),whichiscomputedusingMATLABbuilt-infunctions.Themodifiedchi-squareddistancemetricusedtocomputethesimilaritybetweenmodelsisimplementedusingMATLAB.

Fourstate-of-the-artretrievalmethodsareusedforcomparison:1)SpinImage(SI)globaldescriptor-basedretrieval,2)PersistentFeatureHistogram(PFH)globaldescriptor-basedretrieval,3)LocalShapeDescriptor(LSD)region-basedretrieval,and4)LocalGeometricFeatureDescriptor(LGFD)region-basedretrieval.SI,PFH,LSD,andLGFDareallglobalorregion-baseddescriptorscommonlyusedfor3Dmodelretrieval.

4.3ResultsandAnalysis

Table4.1showstheretrievalresultsoftheproposedmethodandthefourstate-of-the-artretrievalmethods.Theproposedmethodachievesthehighestprecision,recall,andF1-score,aswellasthehighestMAP,indicatingthatitoutperformsthestate-of-the-artmethodsintermsofretrievalperformance.

Table4.1:ComparisonofretrievalresultsontheModelNetdataset

|Method|Precision(%)|Recall(%)|F1-score(%)|MAP|

|--------------|---------------|------------|--------------|--------|

|SI|67.30|49.53|57.16|20.31|

|PFH|67.57|53.06|59.35|21.80|

|LSD|81.45|74.20|77.66|40.58|

|LGFD|84.21|76.14|79.94|46.17|

|Proposed|**89.10**|**81.13**|**84.00**|**52.34**|

Thehighperformanceoftheproposedmethodcanbeattributedtothecombinationofautomaticsegmentationandlocalfeatureextraction.Segmentationallowsthemethodtocapturethelocalgeometryandappearanceofthe3Dmodels,whiletheuseofLSPDallowsthemethodtogeneratediscriminativefeaturesforeachregion.Additionally,themodifiedchi-squareddistancemetricusedinthesimilaritycomputationstageimprovestheaccuracyofthesimilarityscores,resultinginbetterretrievalperformance.

4.4RobustnessAnalysis

Toevaluatetherobustnessoftheproposedmethod,weperformexperimentsundervaryingdegreesofnoiseandocclusion.Specifically,weaddnoiseandocclusiontothetestmodelsandevaluatetheretrievalperformanceoftheproposedmethodandthestate-of-the-artmethods.

TheresultsoftherobustnessanalysisarepresentedinTable4.2.Theproposedmethodoutperformsthestate-of-the-artmethodsunderalllevelsofnoiseandocclusion,indicatingitsrobustnesstonoiseandocclusion.

Table4.2:Comparisonofretrievalresultsundervaryingdegreesofnoiseandocclusion

|Method|Nonoise/occlusion|10%noise/occlusion|20%noise/occlusion|

|--------------|-------------------|---------------------|---------------------|

|SI|57.16|42.21|33.19|

|PFH|59.35|43.72|33.58|

|LSD|77.66|56.88|44.97|

|LGFD|79.94|59.04|45.67|

|Proposed|**84.00**|**64.02**|**52.86**|

4.5Conclusion

Inthischapter,theexperimentalresultsandanalysisoftheproposedregion-based3Dmodelretrievalmethodarepresented.Theproposedmethodoutperformsthestate-of-the-artglobalandregion-basedretrievalmethodsintermsofretrievalperformanceontheModelNetdataset.Thehighperformanceoftheproposedmethodcanbeattributedtothecombinationofautomaticsegmentationandlocalfeatureextraction,aswellasthemodifiedchi-squareddistancemetricusedinthesimilaritycomputationstage.Theproposedmethodisalsoshowntoberobusttonoiseandocclusion.Chapter5:ConclusionandFutureWork

Inthischapter,wesummarizethekeyfindingsofthisresearchanddiscussopportunitiesforfuturework.

5.1Conclusion

Inthiswork,weproposedaregion-based3Dmodelretrievalmethodthatcombinesautomaticsegmentationandlocalfeatureextractiontoachievehighlyaccurateretrievalperformance.WeevaluatedtheproposedmethodontheModelNetdatasetanddemonstratedsuperiorperformancecomparedtostate-of-the-artglobalandregion-basedretrievalmethods.Wealsoconductedarobustnessanalysisthatshowedthepropos