基于FPGA的人脸识别项目

1FacialExpressionRecognitionBasedonfieldProgrammableGateArrayJzau-ShengLin,Shao-HanLiou,Wu-ChihHsieh,Yu-YiLiao,HongChaoWang*,andQingHuaLan*DepartmentofComputerScienceandInformationEngineeringNationalChin-YiUniversityofTechnology*EmbeddedLaboratory,HuaxiaVocationCollege,Xiamen,China2Outline5.Experimentalresults4.HardwareImplementation3.PreprocessingforFFE2.DWT-CMAC1.Introduction6.Conclusion31.IntroductionTheCMAC(JamesAlbus)non-learningmappinglocalgeneralizationlowcomputationcomplexityrapidconvergencespeedfastpracticalhardwarerealizationTheproposedhardwareimplementationofarecognitiontechniquewithFPGAThe2DDWTovertheentirefaceimageasafacialexpressionfeatureextractorTheCMACwithclusteringmemoryasaclassifier.42.DWT-CMAC(1/3)Differenceimage=givenexpressionimage–neutralimageHaarwavelettransformspacedomainfrequencydomainTheimageisdividedintofournon-overlappingmulti-resolutionsub-bandsCMACwithCMAnS-bitinputvectorisencodedtoquantizeintocclusterswhichincludingbbitsTheoutputlayerisk-layerparallelmemoriestomemorizekfacialexpressions52.DWT-CMAC(2/3)TheproposedCMACTheweightcanbeadjustedwiththesteepingdescentruleasfollows:)()()1(kwkwkwyEkw)(cidyyE12)(21)()()1(yykwkwd62.DWT-CMAC(3/3)ArchitectureofthefacialexpressionrecognitionofCMAC73.PreprocessingforFFE(1/2)83.PreprocessingforFFE(2/2)94.HardwareImplementationBlockRAMSRAMVGADWTIPCMACIPRS232104.1DWTIP(1/2)ControllermoduleMemoryofbaseaddress=Row×128＋ColumnDWTModuleOperationmoduleHaarwaveletTransformationDWT_DONE=HIGH114.1DWTIP(2/2)BlockdiagramoftheDWTmodule124.2CMACIP(1/4)ThetrainingphaseTheinputdatalow-frequencyband(32×32bytes)ofHaarDWTtransformationdivideanun-overlap2×2blockasatrainingsample(tosimplifytheinputcodingcircuitoftheCMAC)SRAMTostorethecodingaddresseswithasizeof16×16×12=3072bits.134.2CMACIP(2/4)CMACTrainingcircuit144.2CMACIP(3/4)TherecognitionphaseTheunknowninputdatamemoryTheWeightSummationModuletofindtheweightsTosumthemastotalweightstoexpresstheintensitiesfordifferentfacialexpressions.TousetheMaxmoduletofindthemaximumfacial-expressionintensity.154.2CMACIP(4/4)BlockdiagramofCMACrecognitioncircuit165.Experimentalresults(1/3)ThefacialexpresspatternsJAFFEdatabase120subjectsfortraining60subjectsfortestingTheimageinterfaceFPGASoCplatformwithVirtexIV-ML40XTheparameterofCMACnetworkThelearningratewasset0.05Thesizesofclusterwereset2,4,and6bits175.Experimentalresults(2/3)RecognitionratefortheproposedCMACIPwith6-bitclustersHSPADNRate%H81000180%S08000280%P00700370%A02160160%D12007070%N0000010100%Average76.6%185.Experimentalresults(3/3)HSPADNRate%H1000000100%S09100090%P01800180%A01072070%D00019090%N0000010100%Average88.3%RecognitionratefortheproposedCMACIPwith4-bitclusters196.ConclusionTheproposed2DDWTIPtransformdifferenceimagesfromspatialdomaintofrequencydomaintoyield1024lower-frequencycoefficientsinLL-banddataastheinputsignalsthatwerequantizedasabinarymannerandfedintotheCMACIP.TheCMACIPusedapropersizeof4-bitclusterandalearningrate0.05hassuccessfullyobtainedabetterrecognitionrateabout88.3%.20