Shopping

A Robust Neural Network Based Object Recognition System and Its SIMD Implementation

Description
A Robust Neural Network Based Object Recognition System and Its SIMD Implementation
Categories
Published
of 10
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Share
Transcript
  ARobustNeuralNetworkBasedObject RecognitionSystemanditsSIMDImplementation  AlfredoPetrosino  1  ,GiuseppeSalvi  2  1  INFM-UniversityofSalerno ViaS.Allende,84081Baronissi(Salerno),ITALY alfredo@synapse.irsip.na.cnr.it  2  IstitutoperlaRicercasuiSistemiInformaticiParalleli,IRSIP-CNR ViaP.Castellino111,80131Naples,ITALY  Abstract. Recognitionofobjectsisaparticularlydemandingproblem,if oneconsidersthateachimagemustbeinterpretedinmilliseconds(usually30 or40frames/second).Theproblembecomesmoredicultiftheobjectsare distortedand/orpartiallyoccluded.Inthiscaseasequenceoflocalfeatures aretobeextracted,combinedinaglobalshapedescriptionandclassiedas belongingtopre-denedsetsofknownshapes(referenceshapes).Inthispa- perweproposeamassivelyparallelobjectrecognitionsystem,whichmakes useofthemultipolygonalapproximationschemefortheextractionofro- tationandtranslationinvariantshapefeatures,inconnectionwitharticialneuralnetworksfortheparallelclassicationoftheextractedfeatures.The systemhasbeensuccessfullyappliedforrecognizingaircraftshapesindif- ferentsizes,orientations,withtheadditionofnoisedistortionandocclusion.TimingsontheConnectionMachine200arealsoreported. 1Introduction  Recognitionofshapesisafundamentaltaskincomputervision22].Itsuseproves tobeusefulinmanyapplications,suchasautomatictargetrecognition(ATR),char- acterizationofbiomedicalimagesandidenticationofindustrialpartsforproduct assembly.Objectrecognitionisinitselfstillanopenproblem.Inthelastdecades,successfulattemptshavebeenmadefortherecognitionofisolatedobjectswith completeandwell-denedboundaries.Theusualprocessconsistsintheboundary extractionbyapplyingsomesegmentationtechniquestotheoriginalgreyscaleim- ageandthenapplyingedgedetectorsandthinningalgorithmstothebinaryimage toretaintheboundarywidthassmallaspossible.Oncetheboundaryhasbeenes- tablishedthefeatureextractiontaskbeginswithextractingrelevantfeaturesfrom eachobjectpresentinthescene.Thefeaturesarechosensoastobeinvariantwith respecttotheobjectposition,sizeandorientation.Inadditionmanyapplications,includingautomatictargetrecognition(ATR),characterizationofbiomedicalimages,identicationofindustrialpartsforproduct assembly,etc.requiretherecognitiontotakeplaceinreal-timeandalsoinstrongly noisyandocclusionconditions.Thisproblemisveryinterestingifweconsidera sequenceofimageswitharesolutionof(512    512)pixelsatcolours(8-24bitsfor pixel)whicharetobehandledatarateof30frames/second.Theprocessingand therecognitionofthecontentsofsuchimagescanbecomparedtotheprocessingof 23.6millionofbytes/second.Thus,theuseofmoreandmoresophisticatedparallel architecturesandproceduresofcomputationisnaturalandstronglyneeded,capa- bletodealwithdierentdatastructuresatthedierentlevelsofprocessingand interpretation.However,theuseofparallelarchitecturesgivesrisetonewproblems concerningthechoiceofthemodalityofparallelcomputation(SIMD,MIMD)and themostecientprogramminglanguage.Tooperateasuitablechoice,benchmarks ondierentimagesarenecessary,analizingandreportingnotonlythetotaltime   oftheentireprocessofrecognition,butalsothetimeemployedbyeachprocessing step.Allthiswiththeaimtokeepundercontrolboththequalitativeaccuracyof theprocessandthetimeemployedforrealizingit.Currentliteratureonparallel objectrecongitionessentiallydealwithgraphmatchingapproachesimplemented inparallel1]ontreesearchalgorithms10],geometrichashing3,24]andparallel hypothesisgeneration23].Hereweshowthatacompletesetofcomputervision tasksimplementingltering,featureextractionandstatisticclassicationcanbe ecientelyimplementedasawholesystemonamassivelyparallelSIMDarchitec- ture.Therearemanytechniquesavailabletodescribeanobjectbasedontheirbound- aries,includingamongothersShapeNumbers8,7],Moments12,6],FourierDe- scriptors25,17,9],HoughTransform2],MedialAxisTransform(MAT)4].Among others,thepolygonalapproximation(speciallytheatofhighorder)16]combined withtheCircularAutoRegressive(CAR)modelapproach,formerlyproposedby KashyapandChellappain198114],representsagoodsolutiontohandletheprob- lemsofocclusion,distorsionandnoise,whileremainingalowcomputationalcost methodWeshalluseamodiedversionoftheCARmodelincombinationwith ArticialNeuralNetworks(ANNs)asnonparametricclassiers.Themodications introducedintheCARmodelarebasedontheconsiderationofonesetofpre- dictiveparameters,butrelatedtodierentmodelsoftheshapeunderexam,each correspondingtoadierentpolygonalapproximation.Weshallrefertoitasthe Multi-PolygonalAutoRegressiveModel(MPARM).Inaddition,ArticialNeural Networks(ANNs)areadoptedduetotheirabilitytoadjustwhengivennewin- formation,neuronalmassiveparallelismtypicaloftheSIMDparallelmode,fault tolerancetomissing,confusingand/ornoisydata.Thepaperisorganizedasfollows.Section2brieydecribestheproposedap- proachwhereasthethirdsectionreportsanensableofexperimentalrecognition resultsandtimestakenontheConnectionMachine200. THINNINGFILTERINGGREY LEVELIMAGENEURALCLASSIFIER CAR MODELMULTI - POLYGONALFEATUREVECTOR FEATUREEXTRACTIONDYNAMICCLASSTHRESHOLDINGGRADIENT BASEDEDGE DETECTIONBINARYIMAGE Fig.1. Theoverallschemeoftheobjectrecognitionsystem. 2TheAlgorithm  TheoverallsystemisdepictedinFig.120]andtheoperationscanbesummarized asfollows.Therststageisdedicatedtodetecttheobjectboundary,aftercleaning outthenoisefromtheimagebyapplyingalinearsmoothinglteringandastrongly noiseindependentsegmentationprocedure,basedontheimageentropyoptimization   betweenforegroundandbackground13].Theboundaryisthencodedbyusing elementarybutsalientfeatures.Weadoptandcomparetwostrategiesforthefeature extraction:therstapproachusesthevariationalanglesequence15](seeFig.2 a)),whilethesecondandmorecommonapproachusesthesequenceofeuclidean distancesofpointsalongtheshapeboundaryfromthecentroid5](seeFig.2b)). 1 ABC θθ 2 1 d 3 ABC d 2 d a)b)  Fig.2. Theadoptedshapedescriptionmodels:angleofvariationa)anddistancefrom centroidb)  ThesequenceofshapefeaturesismodelledasaCircularAuto-Regressive(CAR) process14],whichisaparametricequationthatexpresseseachsampleofanordered setofdatasamplesasalinearcombinationofaspeciednumberofprevioussamples plusanerrorterm.Sincethesequenceisassumedtobecircular,thenitisinvariant torotationandtranslation.Theformofthemodelis: y  i  =    0  +  m  X  k  =1    k  y  i  ?  k  8  i  =0  ; 1  ;:::;n  ?  1(1) where  y  i  isthecurrentprimaryfeature; y  i  ?  k  isthefeaturedetected  k  timesbefore thecurrentfeatures;   0  ;  1  ;  2  ;:::;  m  aretheunknownCARcoecients; m  isthe modelorder.Letusindicatewith    =    0  ;  1  ;:::;  m  ]theleastsquareerror(LSE) estimateoftheCARmodel.Toimprovetherepresentativityofthissolutionthe followingmethodisadopted.Ifwex  n  ,thenumberoflinesegmentsintheshape polygonalapproximation,then   p  =  b  (  B=n  )  c  pixelswilllieonthecontourbetween twoendpoints,and(   p  ?  1)polygonalapproximationsoftheshapethroughmodel(2) arepossible,dependingfromthestartingpoint.Thesequencesofprimaryfeatures generatedforeachofthemmaybeslightlydierent.Toovercomethisproblemwe adoptaniterativeschemeconsistingofsolving   p  ?  1systemseachhaving  n  equations and  m  +1unknowns.Itisbasedontheconsiderationthatthe   p  ?  1sequencesare obtainedinsomeorder(clockwiseorcounterclockwise);rstly,theCARvectorfor therstsequenceofprimaryfeaturesiscomputed,thenfortwosequences,three sequencesandsoon.Theprocessisrepeated   p  ?  1times,accordingtothenumberof sequencesorpolygonalapproximations.Inparticular,denotedwith    j  thesolution ofthe  j  -thsystemthusconstructed, j  =1  ;:::;p  ?  1,letusdene   "  j  =    T j    j  ?  1  k    j  k  2  k    j  ?  1  k  2  j  =2  ;:::p  ?  1and  "  1  = 1  k    1  k  2  ameasureofsimilarity.Thenalfeaturevector      issettothesolutionofthe  s  -th system,where  s  =  arg  min  0    j    p  ?  1  f  "  j  g  .Weshallrefertothiswayofproceedingas theMulti{PolygonalAuto{RegressiveModel(MPARM).Forsakeofclearness,we denoteinthefollowing,as  SchemeI  theMPARMappliedtothevariationalangles,whereas  SchemeII  istheMPARMappliedtoeuclideandistancesforthecentroid.ThesequencesofCARparametersarelastlyfedintoanArticialNeuralNet- work(ANN)trainedtoclassifytheplanarobjectswiththeConjugate-Gradient (CG)algorithm21].Theprocessisrepeatedfortheoverallsetofreferenceim- ages.Afterlearning,anunknownobjectpassingthroughallthepreviousstagesis classiedascorrectornotfromtheANNwithfrozenweights.Theoverallsystemdescribedabovehasbeenimplementedonarecongurable SIMDmachine.Theimplementedalgorithmswere:convolutionona3  D  mesh,thresholdingona2  D  mesh,thinningandfeaturedescriptionona2  D  mesh,conju- gategradientforneurallearningona1  D  mesh).Wereportinthenextsectionthe recognitionperformanceandtheparallelexcecutiontimesoftheoveralsystemand referto18,19]foradetaileddescriptionoftheparallelalgorithms. Fig.3. Referenceshapes. 3ExperimentalResults  3.1RecognitionPerformance  Thetaskfacedwastherecognitionofvedierentlyshapedaircrafts(Fig.2).The datasetwasformedbychangingeachaircraftinsize,orientationandposition.Specically,theobjectswere5  0  rotatedandtranslatedinrandompositionswithin theimageboundaries.Thesizeofeachaircraftwasvariedfrom0.25to1.25times thesizeoftheoriginal.Bydoingso,5040shapesweregenerated.Foreachshape therepresentativeCARvectorwasthenextractedbyusingbothfeatureanglesof variationandeuclideandistancesfromthecentroidaselementaryfeatures.Thedatasetwasconsequentlysubdividedin3600shapesforthetrainingandin 1440shapesforthetestoftheneuralclassiers.WecarriedoutvariousexperimentswithCARmodelswithdierentorders  m  .Thevalue  m  =20turnedouttobeoptimalforourpurposes.Table1showsthe classicationperformanceonthetrainingandtestsetsforselectedclassierswith dierentnumberofhiddenunitsandshapedescriptionschemes.  Fig.4. Atypicalrecognitionprocess:(a)theoriginalimagerotated,translatedandscaled withGaussiannoiseadded(    =30);(b)afterapplyingmedianlteringanddynamic thresholding;(c)afterapplyingtheedgedetection;(d)afterapplyingthinning;(e)the shaperecognized. RecognitionResultsonNoisyObjects  Noiseanddistortioneectswerein- troducedbyaddingrandomnoisetotheboundarypointsof360objectshapes.To eachcontourpixelwassassignedaprobability   p  ofretainingitsoriginalcoordinates intheimageplaneandprobability  q  =(1  ?   p  )ofbeingrandomlyassignedtothe coordinatesofoneofitsneighboringpixels.Thedegreeofnoisewasaugmented withincreasedvaluesofthenoiselevel  q  .Wechose  q  equalto1  =  3,i.e.onepixel overthreewasselectedtochangeitsownposition.Thesetofnoisyshapeswas subdividedin270shapesforthetrainingandin90shapesforthetestoftheneural classiers.Table2reportstheclassicationperformanceonthe90testnoisyshapes fortheneuralclassiersutilizedinthesubsection5.1.Asexpected,theclassication performanceobtainedwassubstantiallydegraded.Togetbetterclassicationperformancewere-trainedthepreviouslyneuralclassi- ersonlywiththe270noisyshapesand,afterconvergence,theyweretestedonthe 90testnoisypatterns.TheresultsobtainedarereportedinTable3.Thebestperformancewasachievedwith35hiddenneuronsbyusingaselementary featuresboththeanglesofvariationandtheeuclideandistances. RecognitionResultsonOccludedObjects  Amoderateamountofocclusion isaccommodatedinourexperiments.Theoccludedshapeshavebeenobtainedby cuttingoaportionofthegeneratedshapeswithastraightline.Therestriction onocclusionisthatallthemajorportions,orbranches,shouldremain;therefore,nomajorgeometricpropertyischanged.Thesetofoccludedshapeshasbeensub- dividedin540shapesforthetrainingand180shapesforthetestoftheneural classiers.Table4showstheclassicationperformanceon180occludedshapesfor theneuralclassiersshowninthesubsection5.1.Toimprovetheclassicationperformancefortheneuralclassierswetrainedthe 
Search
Similar documents
View more...
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks