A 201.4 GOPS 496 mW Real-Time Multi-Object Recognition Proce(5)

Fig.16showsperformancecomparisonsoftheproposedpro-cessorwithpreviousvisionprocessors[2]–[4],[20].Fig.16(a)showspoweref ciencycomparison.TheGOPS/W,whichnormalizestheGOPSperformancewiththepower,isadoptedasaperformanceindexwherethe1operationmeans16-bit xed-pointoperation.Theproposedprocessorachieves290GOPS/W,whichis1.36timeshigherthanthepreviousvisionprocessors.Fig.16(b)showsenergyef ciencycomparisoninobjectrecognition,whichisobtainedbyenergyconsumptionpereachframe.With60frame/secoperationbythepipelinedarchitectureandunder0.5Wpowerconsumptionbytheworkload-awaredynamicpowermanagement,theproposed

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

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

43

Fig.16.(a)GOPS/Wcomparison.(b)Energy/frame

comparison.

Fig.17.Demonstrationsystem.

TABLEICHIPS

UMMARY

TABLEII

POWERBREAK-D

OWN

processorachieves8.2mJenergydissipationperframeforVGAsizedvideoinput,whichis3.2timeslowerthanthebestofthepreviousobjectrecognitionprocessor.

Forthevalidationofthefabricatedchip,ademonstrationsystemforreal-timeobjectrecognitionisdevelopedasshowninFig.17.Itiscomposedoftargetobjects,videocamcorder,evaluationboard,andLCDdisplay.Theevaluationboardiscomposedofthree oors,whichareforhostprocessor,videodecoderandfabricatedrecognitionchip,andperipheralinterfacessuchasLCDdisplay,serial,USB,andEthernet,respectively.Inthedemonstrationsystem,thefabricatedchipisusedasavisionprocessingacceleratorwhilethehostprocessorcontrolsthewholeprogramsequencesandaccessesperipheralmodulestodisplaytheresultsandtointerfacewiththeexternaldevices.Theoverallobjectrecognitionisperformedbythreesteps.First,theinputimageofthetargetobjectsiscapturedfromthevideocamcorderanddecodedtothree-channelRGBpixeldatabythevideodecoder.Then,st,the nalrecognitionresultsaredisplayedwiththekey-pointsattheLCDscreenbythehostprocessor.

VIII.CONCLUSION

Inthiswork,wehaveproposedareal-timemulti-objectrecognitionprocessorwithathree-stagepipelinedarchitec-ture.Thevisualperceptionbasedmulti-objectrecognitionalgorithmhasbeendevelopedtogivemultipleattentionstomultipleobjectsintheinputimage.Forhuman-likemulti-ob-jectperception,aneuralperceptionenginehasbeenproposedwithbiologicallyinspiredneuralnetworksandfuzzylogic

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

44circuits.Inhardwarearchitecture,athree-stagepipelinedar-chitecturehasbeenproposedtomaximizethethroughputofrecognitionprocessing.Thethreeobjectrecognitiontasksareexecutedinthepipelineandtheexecutiontimesofthethreetasksarebalancedforef cientpipeliningbasedonintelligentworkloadestimations.Inaddition,a118.4GB/smulti-castingnetwork-on-chiphasbeenproposedforcommunicationarchi-tecturewithincorporatingoverall21IPblocksoftheprocessor.Finally,workload-awaredynamicpowermanagementwasperformedforlow-powerobjectrecognition.The49mmchipcontains3.7Mgatesand396KBon-chipSRAMina

0.13mCMOSprocess.Withademonstrationsystem,thefabricatedchipachieves60frame/secmulti-objectrecognition

upto10differentobjectsforVGA

(640

480)videoinputwhiledissipating496mWat1.2V.Theobtained8.2mJ/frameenergydissipationis3.2timeslowerthanthestate-of-the-artrecognitionprocessor.

REFERENCES

[1]S.Kyoetal.,“A51.2GOPSscalablevideorecognitionprocessorfor

intelligentcruisecontrolbasedonalineararrayof1284-wayVLIWprocessingelements,”IEEEJ.Solid-StateCircuits,vol.38,no.11,pp.1992–2000,Nov.2003.

[2]A.Abboetal.,“XETAL-II:A107GOPS,600mWmassively-parallel

processorforvideosceneanalysis,”IEEEJ.Solid-StateCircuits,vol.43,no.1,pp.192–201,Jan.2008.

[3]D.Kimetal.,“An81.6GOPSobjectrecognitionprocessorbasedon

NoCandvisualimageprocessingmemory,”inProc.IEEECustomIn-tegratedCircuitsConf.(CICC),Apr.2007,pp.443–446.

[4]K.Kimetal.,“A125GOPS583mWnetwork-on-chipbasedparallel

processorwithbio-inspiredvisualattentionengine,”IEEEJ.Solid-StateCircuits,vol.44,no.1,pp.136–147,Jan.2009.

[5]J.-Y.Kimetal.,“A201.4GOPS496mWreal-timemulti-objectrecog-nitionprocessorwithbio-inspiredneuralperceptionengine,”inIEEEISSCCDig.Tech.Papers,Feb.2009,pp.150–151.

[6]D.G.Lowe,“Distinctiveimagefeaturesfromscale-invariantkey-points,”puterVision,vol.60,no.2,pp.91–110,Jan.2004.

[7]S.Leeetal.,“Thebrainmimickingvisualattentionengine:An802

60digitalcellularneuralnetworkforrapidglobalfeatureextraction,”inIEEESymp.VLSICircuitsDig.,Jun.2008,pp.26–27.

[8]L.Ittietal.,“Amodelofsaliency-basedvisualattentionforrapidscene

analysis,”IEEETrans.PatternAnal.MachineIntell.,vol.20,no.11,pp.1254–1259,Nov.1998.

[9]M.Kimetal.,“A22.8GOPS2.83mWneuro-fuzzyobjectdetection

engineforfastmulti-objectrecognition,”inIEEESymp.VLSICircuitsDig.,Jun.2009,pp.260–261.

[10]S.A.Nene,S.K.Nayar,andH.Murase,ColumbiaObjectImageLi-brary(Coil-100),ColumbiaUniversity,NewYork,TechnicalReportCUCS-006-96,Feb.1996.

[11]S.Agarwaletal.,“Learningtodetectobjectsinimagesviaasparse,

part-basedrepresentation,”IEEETrans.PatternAnal.MachineIntell.,vol.26,no.11,pp.1475–1490,Nov.2004.

[12]J.-Y.Kimetal.,“A66frame/sec38mWnearestneighbormatching

processorwithhierarchicalVQalgorithmforreal-timeobjectrecogni-tion,”inProc.I …… 此处隐藏：5851字，全部文档内容请下载后查看。喜欢就下载吧 ……