主要从事的研究方向是无人机、无人系统的环境感知，即通过计算机视觉（SLAM，SfM）构建三维场景，然后通过机器学习来实现高层语义信息的提取。研究的整体架构如下图所示：
MymainresearchdirectionisenvironmentperceptionforUAV,whichisbuildingthree-dimensionalscenesthroughcomputervision(SLAM,SfM),andthenextractinghigh-levelsemanticinformationthroughmachinelearning,soastoprovidedecisionsupportforunmannedsystems.Theoverallstructureofthestudyisshowninthepreviousfigure.
KeLi,ChangqingZou,ShuhuiBu,YunLiang,JianZhang,MinglunGong,PatternRecognition,2018.
Thispaperpresentsamulti-modalfeaturefusionbasedframeworktoimprovethegeographicimageannotation.Toachieveeffectiverepresentationsofgeographicimages,themethodleveragesalow-to-highlearningflowforboththedeepandshallowmodalityfeatures.Itfirstextractslow-levelfeaturesforeachinputimagepixel,suchasshallowmodalityfeatures(SIFT,Color,andLBP)anddeepmodalityfeatures(CNNs).Itthenconstructsmid-levelfeaturesforeachsuperpixelfromlow-levelfeatures.Finallyitharvestshigh-levelfeaturesfrommid-levelfeaturesbyusingdeepbeliefnetworks(DBN).ItusesarestrictedBoltzmannmachine(RBM)tominedeepcorrelationsbetweenhigh-levelfeaturesfrombothshallowanddeepmodalitiestoachieveafinalrepresentationforgeographicimages.Comprehensiveexperimentsshowthatthisfeaturefusionbasedmethodachievesmuchbetterperformancescomparedtotraditionalmethods.
ShuhuiBu,YongZhao,GangWan,andZhenbaoLiu,IEEE/RSJInternationalConferenceonIntelligentRobotsandSystems,2016.
ShuhuiBu,YongZhao,GangWan,KeLi,GongCheng,ZhenbaoLiu,MultimediaToolsApplication,2016.
ShuhuiBu,PengchengHan,ZhenbaoLiu,JunweiHan,PatternRecognition,2016.
Effectivefeaturesandgraphicalmodelaretwokeypointsforrealizinghighperformancesceneparsing.Recently,ConvolutionalNeuralNetworks(CNNs)haveshowngreatabilityoflearningfeaturesandattainedremarkableperformance.However,mostresearchesuseCNNsandgraphicalmodelseparately,anddonotexploitfulladvantagesofbothmethods.Inordertoachievebetterperformance,thisworkaimstodesignanovelneuralnetworkarchitecturecalledInferenceEmbeddedDeepNetworks(IEDNs),whichincorporatesanoveldesignedinferencelayerbasedongraphicalmodel.ThroughtheIEDNs,thenetworkcanlearnhybridfeatures,theadvantagesofwhicharethattheynotonlyprovideapowerfulrepresentationcapturinghierarchicalinformation,butalsoencapsulatespatialrelationshipinformationamongadjacentobjects.Weapplytheproposednetworkstoscenelabeling,andseveralexperimentsareconductedonSIFTFlowandPASCALVOCDataset.TheresultsdemonstratethattheproposedIEDNscanachievebetterperformance.
QinLi,KeLi,XiongYou,ShuhuiBu,ZhenbaoLiu,Neurocomputing,vol.199,pp.114-127,2016.
Effectivefeaturesandsimilaritymeasuresaretwokeypointstoachievegoodperformanceinplacerecognition.Inthispaperweproposeanimagesimilaritymeasurementmethodbasedondeeplearningandsimilaritymatrixanalyzing,whichcanbeusedforplacerecognitionandinfrastructure-freenavigation.Inordertoobtainhighrepresentativefeature,ConvolutionalNeuralNetworks(CNNs)areadoptedtoextracthierarchicalinformationofobjectsintheimage.Inthemethod,theimageisdividedintopatches,thenthesimilaritymatrixisconstructedaccordingtothepatchsimilarities.Theoverallimagesimilarityisdeterminedbyaproposedadaptiveweightingschemebasedonanalyzingthedatadifferenceinthesimilaritymatrix.Experimentalresultsshowthattheproposedmethodismorerobustthantheexistingmethods,anditcaneffectivelydistinguishthedifferentplaceimageswithsimilar-lookingandthesameplaceimageswithlocalchanges.Furthermore,theproposedmethodhasthecapabilitytoeffectivelysolvetheloopclosuredetectioninSimultaneousLocationsandMapping(SLAM).
ShuhuiBu,ZhenbaoLiu,JunweiHan,JunWu,RongrongJi,IEEETransactionsonMultimedia,vol.16,no.8,pp.2151-2167,2014.
Inthispaper,weproposeamulti-level3Dshapefeatureextractionframeworkbyusingdeeplearning.Tothebestofourknowledge,itisthefirsttimetoapplydeeplearninginto3Dshapeanalysis.Experimentson3Dshaperecognitionandretrievaldemonstratethesuperiorperformanceoftheproposedmethodincomparisontothestate-of-the-artmethods.WeimplementadeeplearningtoolboxwithGPUwhichcanboostthecomputationperformancegreatly.Thesourcecodeispubliclyavailableandeasytouse.
ShuhuiBu,ShaoguangCheng,ZhenbaoLiu,JunweiHan,IEEEMultimedia,vol.21,no.4,pp.38-46,2014.
Thispaperpresentsanovel3Dfeaturelearningframeworkwhichcancombinedifferentmodalitydataeffectivelytopromotethediscriminabilityofuni-modalfeature.TwoindependentDeepBeliefNetworks(DBNs)areemployedtolearnhigh-levelfeaturesfromlow-levelfeatures,andaRestrictedBoltzmannmachine(RBM)istrainedforminingthedeepcorrelationsbetweenthedifferentmodalities.Accordingtoourknowledge,wearethefirsttointroducethemulti-modalfeaturefusioninto3Dshapeanalysis.WeimplementadeeplearningtoolboxwithGPUwhichcanboostthecomputationperformancegreatly.Thesourcecodeispubliclyavailableandeasytouse.
ShuhuiBu,PengchengHan,ZhenbaoLiu,KeLi,JunweiHan,TheVisualComputer,vol.30,no.5,pp.867-876,2014.
Inthispaper,wepresentashift-invariantringfeaturefor3Dshape,whichcanencodemultiplelow-leveldescriptorsandprovidehigh-discriminativerepresentationoflocalregionfor3Dshape.First,severaliso-geodesicringsarecreatedatequalintervals,andthenlow-leveldescriptorsonthesamplingringsareusedtorepresentthepropertyofafeaturepoint.Inordertoboostthedescriptivecapabilityofrawdescriptors,weformulatetheunsupervisedbasislearn-ingintoanL1-penalizedoptimizationproblem,whichusesconvolutionoperationtoaddresstherotationambiguityofdescriptorsresultingfromdifferentstartingpointsinrings.Inthefollowingextractionprocedureofhigh-levelfeature,weusethelearnedbasestocalculatethesparsecoefficientsbysolvingtheoptimizationproblem.Furthermore,tomakethecoefficientsirrelevantwiththesequentialorderinring,weuseFouriertransformtoachievecircular-shiftinvariantringfeature.Experimentson3Dshapecorrespondenceandretrievaldemonstratethesatisfactoryperformanceofthepro-posedintrinsicfeature.
ZhenbaoLiu,SicongTang,WeiweiXu,ShuhuiBu,JunweiHan,KunZhou,ComputerGraphicsForum,vol.33,no.7,2014.
Sinceindoorscenesarefrequentlychangedindailylife,suchasre-layoutoffurniture,the3Dreconstructionsforthemshouldbeflexibleandeasytoupdate.Wepresentanautomatic3DsceneupdatealgorithmtoindoorscenesbycapturingscenevariationwithRGBDcameras.Weassumeaninitialscenehasbeenreconstructedinadvanceinmanualorothersemi-automaticwaybeforethechange,andautomaticallyupdatethereconstructionaccordingtothenewlycapturedRGBDimagesoftherealsceneupdate.Itstartswithanautomaticsegmentationprocesswithoutmanualinteraction,whichbenefitsfromaccuratelabelingtrainingfromtheinitial3Dscene.Afterthesegmentation,objectscapturedbyRGBDcameraareextractedtoformalocalupdatedscene.Weformulateanoptimizationproblemtocomparetotheinitialscenetolocatemovedobjects.Themovedobjectsarethenintegratedwithstaticobjectsintheinitialscenetogenerateanew3Dscene.Wedemonstratetheefficiencyandrobustnessofourapproachbyupdatingthe3Dsceneofseveralreal-worldscenes.
ZhizhongHan,ZhenbaoLiu,JunweiHan,ShuhuiBu,TheVisualComputer,vol.30,no.7,2014.
Inthispaper,weproposeanewstyletransfermethodforautomatic3Dshapecreationbasedonnewconceptsofstyleandcontentof3Dshapes.Ourunsupervisedstyletransfermethodcouldplausiblycreatenovelshapesnotonlybyrecombiningexistentstylesandcontentsinasetbutalsobycombiningnew-comingstylesorcontentswiththeexistentonesconveniently.Thisfeatureprovidesabetterwaytoincreasethediversityofcreatedshapes.Wetestourmethodinseveralsetsofman-made3Dshapesandobtainplausiblecreatedshapesbasedonthereasonablyseparatedstylesandcontents.
D.Pickup,X.Sun,P.L.Rosin,R.R.Martin,Z.Cheng,Z.Lian,M.Aono,A.BenHamza,A.Bronstein,M.Bronstein,S.Bu,U.Castellani,S.Cheng,V.Garro,A.Giachetti,A.Godil,J.Han,H.Johan,L.Lai,B.Li,C.Li,H.Li,R.Litman,X.Liu,Z.Liu,Y.Lu,A.Tatsuma,J.Ye,EurographicsWorkshopon3DObjectRetrieval,2014.
SHREC’14isanewbenchmarkingdatasetfortestingnon-rigid3Dshaperetrievalalgorithms,onethatismuchmorechallengingthanexistingdatasets.Thedatasetfeaturesexclusivelyhumanmodels,inavarietyofbodyshapesandposes.3Dmodelsofhumansarecommonlyusedwithincomputergraphicsandvision,andsotheabilitytodistinguishbetweenbodyshapesisanimportantshaperetrievalproblem.Ourdeeplearningmethodistestonthisdataset,andachievegoodresults.
JunweiHan,PeichengZhou,DingwenZhang,GongCheng,LeiGuo,ZhenbaoLiu,ShuhuiBu,JunWu,ISPRSJournalofPhotogrammetryandRemoteSensing,vol.89,no.3,pp.37-48,2014.
Automaticdetectionofgeospatialtargetsinclutteredscenesisaprofoundchallengeinthefieldofaerialandsatelliteimageanalysis.Inthispaper,weproposeanovelpracticalframeworkenablingefficientandsimultaneousdetectionofmulti-classgeospatialtargetsinremotesensingimages(RSI)bytheintegra-tionofvisualsaliencymodelingandthediscriminativelearningofsparsecoding.ComprehensiveevaluationsonasatelliteRSIdata-baseandcomparisonswithanumberofstate-of-the-artapproachesdemonstratetheeffectivenessandefficiencyoftheproposedwork.
ChengliangLi,ZhongshengWang,ShuhuiBu,HongkaiJiang,andZhenbaoLiu,JournalofMechanicalengineeringScience,vol.228,no.6,pp.1048-1062,2014.
Thispaperpresentsanovelmethodformachineryconditionprognosis,namedleastsquaressupportvectorregressionstrongtrackingparticlefilterwhichisbasedonleastsquaressupportvectorregressioncombingwithstrongtrackingparticlefilter.Therearetwomaincontributionsinourwork:first,theregressionfunctionofleastsquaressupportvectorregressionisextended,whichconstructsabridgefortheapplicationofcombiningdata-drivenmethodwitharecursivefilterbasedonextendKalmanfilter;second,anextendKalmanfilter-basedparticlefilterisstudiedbyintroducingastrongtrackingfilterintoaparticlefilter.Theexperimentresultsdemonstratethattheproposedmethodisbetterthanclassicalconditionpredictorsinmachineryconditionprognosis.
ShuhuiBu,ZhenbaoLiu,TsuyoshiShiina,KengoKondo,MakotoYamakawa,KazuhikoFukutani,YasuhiroSommeda,andYasufumiAsao,IEEETransactionsonBiomedicalEngineering,vol.59,no.5,pp.1354-1363,2012.
Thispaperintroducesareconstructionmethodforreducingamplificationofnoiseandartifactsinlowfluenceregions.Inthismethod,fluencecompensationisintegratedintomodel-basedreconstruction,andtheabsorptiondistributionisiterativelyupdated.Ateachiteration,wecalculatetheresidualbetweendetectedPAsignalsandthesignalscomputedbyaforwardmodelusingtheinitialpressure,whichistheproductofestimatedvoxelvalueandlightfluence.Byminimizingtheresidual,thereconstructedvaluesconvergetothetrueabsorptiondistribution.Inaddition,wedevelopedamatrixcompressionmethodforreducingmemoryrequirementsandacceleratingreconstructionspeed.Theresultsofsimulationandphantomexperimentsindicatethattheproposedmethodprovidesabettercontrast-to-noiseratio(CNR)inlow-fluenceregions.WeexpectthatthecapabilityofincreasingimagingdepthwillbroadentheclinicalapplicationsofPAT.
BunichiroShibazaki,ShuhuiBu,TakanoriMatsuzawa,andHitoshiHirose,JournalofGeophysicalResearch,vol.115,B00A19,Apr.2010.
Wedevelopedamodelofshort‐termslowslipevents(SSEs)onthe3‐DsubductioninterfacebeneathShikoku,southwestJapan,consideringarate‐andstate‐dependentfrictionlawwithasmallcutoffvelocityfortheevolutioneffect.WeassumeloweffectivenormalstressandsmallcriticaldisplacementattheSSEzone.Onthebasisofthehypocentraldistributionoflow‐frequencytremors,wesetthreeSSEgenerationsegments:alargesegmentbeneathwesternShikokuandtwosmallersegmentsbeneathcentralandeasternShikoku.Usingthismodel,wereproduceeventsbeneathwesternShikokuwithlongerlengthsinthealong‐strikedirectionandwithlongerrecurrencetimescomparedwitheventsbeneathcentralandeasternShikoku.Thenumericalresultsareconsistentwithobservationsinthattheeventsatlongersegmentshavelongerrecurrenceintervals.