A 201.4 GOPS 496 mW Real-Time Multi-Object Recognition Proce(3)

KIMetal.:A201.4GOPS496mWREAL-TIMEMULTI-OBJECTRECOGNITIONPROCESSORWITHBIO-INSPIREDNEURALPERCEPTIONENGINE

37

Fig.7.Detailedvisualperceptionalgorithm.

staticfeaturessuchasintensity,color,andorientationandgen-eratethesaliencymapthatcombinestheextractedfeaturemapsthroughrepeatednormalizations.TheODEisproposedtoper-formthe nalROIclassi cationforeachobjectusingthegener-atedsaliencymap.TheRISCcontrollertakesaroleincontrol-lingthethreededicatedenginesandperformingsoftwareori-entedoperationsbetweenthededicatedoperationsoftheen-gines.A24KBmemoryisusedforstoringoriginalimagesanddatacommunicationamongthethreeenginesbysharinginter-mediateprocessingdata.Afterthe nalROIclassi cation,theNPEtransfersinformationoftheobtainedROIgrid-tilestotheSTMthroughaFIFOqueue.

Fig.7showsthedetailedvisualperceptionalgorithmoper-atedbytheNPE,whichbroadlyconsistofsaliencymapgen-erationandROIclassi cation.ThesaliencymapgenerationismainlybasedonItti’ssaliencybasedvisualattention[8]andac-celeratedbytheVAE.First,theRGBchannelsofVGAsized

60pixelsandaninten-inputimagearedown-sizedto

80

sityfeaturemapandtwocolorfeaturemapsaregeneratedbyper-pixel lteringoperations.Fourorientationfeaturemaps,forthedirectionof0,45,90,and135,aregeneratedfromthein-tensityfeaturemapwiththeGabor ltering.Aftergeneratingmulti-scaleGaussianpyramidimagesforeachof7maps,eachimageistransformedbyacenter-surroundmechanismtoen-hancethepartsoftheimagethatdifferfromtheirsurround-ings.Finally,thesaliencymapisgeneratedbyrepeatedcom-binationofnormalizedfeaturemaps.Themotionvectormap,generatedbytheME,isalsocombinedinthisstep.Amongtheseprocesses,computationallyintensiveimage lteringop-erationssuchasGabor,Gaussian,andcenter-surround lteringareacceleratedbythehardwareacceleratorVAE.Thenormal-izationprocesses,whichincludeirregularoperationsandcanbeperformedindifferentways,areperformedbysoftwarebytheRISCcontroller.Aftersaliencymapgeneration,ROIclassi ca-tionisperformedbytheODE.First,the10mostsalientpointsareselectedastheseedpointsoutofthesaliencymap.Then,fromthemostsalientseedpoint,theROIofanobjectgrowsfromneighborpixelsoftheseedthroughrepeatedhomogeneityclassi cations.Fortheclassi cationofeachpixel,anintensity,saliency,andlocationareusedforhomogeneityevaluation.Thesimilaritiesbetweentheseedandtargetpixelaremeasuredforabovethreemetrics,andbasedonthesummatedresult,the nalclassi cationthatthetargetpixelisdeterminedtobejoinedtotheROIornotisdetermined.Incasethattheotherseedpointsareincludedbythegrownregion,theyareinhibitedfromtheseedpointsinthenextROIclassi cation.Afterrepeatingclas-si cationprocessesfor10seedpoints,theROIofeachobjectinpixelunitisquantizedtothesmallsizedgrid-tileunit.

InthedesignoftheVAEandODE,biologicallyinspiredcel-lularneuralnetworksandneuro-fuzzyclassi erareemployedforfastfeatureextractionandrobustclassi cation,respectively.IntheVAE,2-Dcellularneuralnetworksareusedtorapidlyex-tractvariousfeaturesfromtheinputimageusingitsregionalandcollectiveprocessing[7].Fig.8showsoverallblockdi-agram,circuits,andmeasuredwaveformsoftheODE.Item-ploysGaussianfuzzymembershipandsingle-layerneuralnet-

A 201.4 GOPS real-time multi-object recognitionprocessor is presented with a three-stage pipelined architecture.Visual perception based multi-object recognition algorithm isapplied to give multiple attentions to multiple objects in the inputimage. For human-like multi-object perception, a neural perceptionengine is proposed with biologically inspired neural networksand fuzzy logic circ

38IEEEJOURNALOFSOLID-STATECIRCUITS,VOL.45,NO.1,JANUARY

2010

Fig.8.Blockdiagram,circuits,andmeasuredwaveformsofobjectdetectionengine.

workforsimilaritymeasureanddecisionmaking,respectively.Incircuitdesign,theODEexploitsanalog-basedmixed-modecircuitstoreduceareaandpoweroverheadofGaussianfunc-tioncircuitsandneuralsynapticmultipliers.Exceptthedigitallyimplementedlearningpart,dataprocessingpartsoftheODEareimplementedbyanalogcircuits.Toexploittheanalogdataprocessing,8-bitintensity,saliency,andlocationvaluesofthetargetandseedpixelareconvertedtoanalogsignalsbyDACs.Afterthat,threeGaussianfunctioncircuitsmeasurethesimi-laritiesbetweenthetwopixelsforthreemetrics.AGaussianfunctioncircuitisrealizedbythecombinationofMOSdiffer-entialpairandminimumfollowercircuitincurrentmodecon- guration.Thedifferentialpaircircuitoutputsthesymmetricdifferentialsignals,eachofwhichhasexponentialnon-linearitycharacteristics.AndtheminimumfollowercircuitgeneratestheGaussian-likeoutputbyfollowingtheminimumbetweenthesymmetricdifferentialsignals.A2-DGaussianfunctioncircuitcanbeimplementedbytwoconsecutiveGaussianfunctioncir-cuitsbyconnectingtheoutputofaGaussianfunctioncircuittothebiascurrentinputtailofthenextGaussianfunctioncircuit.Finally,current-modeneuralsynapticcircuitsmergethethreemeasuredsimilaritieswithmultiplyingtheirweightvalues,andcomparatorcircuitmakethe naldecisionthroughthresholding.WithaHebbianlearning[14],theweightvaluesoftheneuralsynapticcircuits,whichplayaroleinclassi cationcriteria,areupdatedeverycycle.Asaresult,theODEcompletestheROIdetectionfor1objectwithin

7sat200MHzoperatingfrequency.Anditsanalog-basedmixed-modeimplementationreducestheareaandpowerconsumptionby59%and44%,respectively,comparedwiththoseofdigitalimplementation.

Fig.8alsoshowsthemeasurementwaveformsofmixed-modeODE.TheyincludeDACoutputsignal,Gaussianfunctioncir-cuitoutputsignal,and nalclassi cationsignal.Asshownintheenlargedwaveforms,theGaussianoutputsignalvarieswiththesimilarityoftwoanaloginputsignals,andthe nalclassi -cationsignalismadebasedonit.B.SIMDProcessorUnit

TheSPUisdesignedtoaccelerateparallelimageprocessingtasksofthedescriptorgenerationstage.AsshowninFig.9,theSPUconsistofaSPUcontroller,eightSIMDcontrolleddual-issuedverylonginstructionword(VLIW)PEs,128-bit-widedatamemory,and2-DDMA.TheeightPEsperformpixelparallelimageprocessingoperationsuchasGaussian ltering,localmaximumsearch,andhistogramoperation.TheSPUcon-trollercontrolstheoverallprogr …… 此处隐藏：7277字，全部文档内容请下载后查看。喜欢就下载吧 ……