文献翻译——基于2t-dram的紧凑的数字像素传感器（DSP）

第1页外文文献资料ACompactDigitalPixelSensor(DSP)Using2T-DRAM1.Introduction.Complementarymetal-oxide-semiconductor(CMOS)imagesensorsarenowpartofoureverydayslifecoveringawidespectrumofapplicationsfromcell-phonecameras,webcams,digitalcameras,videogamestosecurityandautomotiveapplications1.ThesuccessfuldeploymentofCMOSimagesensorsoverthechargecoupleddevices(CCD)technologyismainlyduetoreducedcostandpowerconsumption,aswellashigherintegrationandon-chipprocessingcapabilities,whicharecriticalformobileapplications.AmongCMOSimagesensors,twomainstreamarchitecturesaretobedistinguisheddependingonwheretheanalog-to-digitalconversion(ADC)isachieved.Firstisactivepixelsensors(APS),whichperformtheADCoutsidethepixelarrayusingaper-arrayorper-columnreadoutscheme.Secondisdigitalpixelsensors(DPS),whichintegratetheADCatthepixellevel.DPS,whilegivingflexibilityofuse,performmassiveparallelanalogtodigitalconversionatthepixellevelandenablethepromisingpixel-levelimageprocessing,whichisveryattractiveforrealtimeapplicationssuchasautomotiveandsurveillancesystems.Inaddition,duetoaveryearlyanalog-to-digitalconversion,DPScanofferimprovednoisefigureanddynamicrange2.PreviousDPSimplementationsarebasedoneitherpulsefrequencymodulation(PFM)3orpulsewidthmodulation(PWM)4schemetodigitizethevoltageofthephotodiodessensingnodewithavoltagecomparator.ThePWMschemeusesthecomparatoroutputsignaltowriteaglobalcountervalueinthepixel-levelmemory,whilethePFMschemeusesthecomparatoroutputtoenableanincrementofthepixel-levelcounter.BothPFMandPWMschemesaretypicallypowerhungry,duetothehighswitchingactivityatthepixellevel.Inaddition,amajordrawbackofbothschemesistherequirementofarea-第2页hungrypixellevelmemoryparticularlyforhighresolutionimaging.Asaconsequence,theareaandpoweroverheadduetothepixellevelmemoryaresubstantialandcanaccountforupto50%ofthetotaloverhead.Thisstronglyimpactsthepixelsfillfactorandthusitslightsensitivityaswellasthepowerandpixelarea.Severalattemptsweremadetoreducethememoryneedsatpixellevel.Amongtheothersolutions,theuseofanaddresseventrepresentation(AER)scheme5completelyremovesthememoryfromthepixelatthecostofincreasedcomplexityandtheintroductionoftimingerrorsduetocollisionwhichoccurswhenseveralpixelsattempttoaccessthedatabussimultaneously.Anotherreportedsolutionconsistsofcompressingthedataevenbeforestorageasproposedin6.Thislastsolutiondrasticallyreducesthepixelareabutatthecostofreducedsignal-to-noiseratio(SNR)duetotheuseoflossycompressionscheme.In7,theoutputofthecomparatorissampledandstoredusingasinglebitregistercellperpixel.Allregistercellsareconnectedinseriestoformascanchainandthewholearrayisscannedaftereachsamplingoperation.Whileonlyonebitpixel-levelmemoryisrequired,thisarchitectureincreasespowerconsumptionasthenumberofsampleswhichisastrongfunctionoftheresolution.Inthiswork,aDPSarchitectureusing2TDRAMasthestorageelementandamulti-resetintegrationmethodologyareproposedinordertoreduceboththememoryneedsandthememorysizeatthepixel-levelandreducethepoweroverhead.Thepaperisorganizedasfollows:Section2introducesthePWMtimedomainDPSanditsconversiontimeanalysis.Section3isdevotedtothemulti-resetintegrationmethodologyandadiscussionabouthetrade-offsinvolvedintermsofspeed,areaandmemorysize.Section4describesthemulti-resetintegrationDPSarchitecture,thepixelcircuitryandthe2TDRAMimplementationandsensingscheme.Section5providesananalysisofpowerconsumptionaswellaspowerreductiontechniques.Section6presentstheprototypeimplementationandexperimentalmeasurements.Section7concludesthiswork.第3页2.Time-DomainPWMDPS2.1ConventionalArchitectureTheconventionalpixelarchitectureofaDPSusingthePWMtechniqueisshowninFigure1.ThepixelcomprisesaphotodiodePd,aresettransistorAR,avoltagecomparator,amemoryunitandafeedbackcircuittoperformtheauto-resetofthephotodiode.Outsidethepixelarray,atimingcontrolunitandaglobalcounterarerequiredtoperformtheconversionofthelightintensityintoadigitalcode.Aread-outcircuitisalsoneededtoreadthecontentsfromthememoryandtooutputthedatatotheprocessingunit.Theintegrationphasestartsbydisablingtheprechargetothesupplyvoltageofthephotodiode,i.e.whentheresetsignaltransitsfrom0toV-d-d.Meanwhile,aglobalcounterlocatedoutsidethepixelarrayissimultaneouslyenabled.ThevoltageV-dofthephotodiodenodeloadedwithacapacitanceCpdstartsdecreasingproportionallytolightintensityduetothephoto-generatedcurrentIdthroughthephotodiodePd.WhenthevoltageV-dreachesareferencevoltageV-ref,theoutputofthecomparatorswitcheshighandthecountervalueiswrittentothepixel-levelmemory.Theresetofthephotodiodeisperformedautomaticallyafterthephotodiodevoltagehas第4页reachedthereferencelevelV-ref.ThewrittencodeinthememoryisadigitizedvalueofthetimerequiredforV-dtoreachV-ref,whichisafunctionofthelightintensity.Usingfirstorderapproximation,theintegrationtimeisexpressedas:,FromEquation(1),onemustnoticethattheintegrationtimeisinverselyproportionaltothephoto-generatedcurrentId.Toperformthequantizationoftheintegrationtime,theglobalcounterisusedtoprovidethequantizationboundariesforthetimetodigitalconversion.Thenon-linearitybetweenthephotocurrentandtheintegrationtimecanbecompensatedbyadaptingthefrequencyoftheglobalcounterasshowninFigure2.Ononehand,Figure2(a)representsauniformtimedomainquantizationleadingtoanon-linearresponsefromthesensorasthephotocurrentboundaryquantizationstepsarenon-linear.Ontheotherhand,thenon-uniformtimedomainquantizationdepictedinFigure2(b)enablestolinearizetheconversionofphotocurrentintodigitalcodeasthephotocurrentboundaryquantizationstepsarenowlinear.2.2ConversionTimeAnalysis第5页PWMcodingschemeencodestheilluminationlevelinformationofeachpixelusingasinglepulse.ThispulsewidthrepresentstimeTinttodischargeaphotodiodefromV-d-dtoV-ref.Inordertoconvertthepulseintodigitalcode,atimecodegeneratedbytheglobalcounteriswrittenintotheembeddedmemoryoncethepulseisdetected.TherearetwoquantizationapproachestoconvertthePWMpulsesignalintodigitalcode.Thefirstapproach,referredtoasuniformtimedomainquantization(UQ),whichprovidessamplingtimes(orboundaryquantizationlevels)fromT-mintoT-maxuniformly.Thesecondapproachreferredtoasnon-uniformtimedomainquantization(NUQ),resolvesthesamplingtimeinordertoformauniformlyquantizedphotocurrents.Anon-lineartime-domainquantizeristhereforerequiredinordertocompensatefortime-to-photocurrentnon-linearity.2.2.1UniformTime-DomainQuantizationUniformtimedomainquantizationschemedividesthetimewithintheboundariesT-minandT-maxequallyinto2nquantizationlevels,wherenrepresentsthenumberofbitsorresolution.Therefore,thetimewidthofeachquantizationlevelTcanbeexpressedas:Whiletherelationshipbetweenthequantizationlevel-UQ(n,t)andthetimewidthofthePWMpulsetisillustratedinthefollowingequation:UsingEquation(3)theconversionbetweenthequantizationlevel-UQ(n,Id)andthedischargephotocurrentIdofthephotodiodecanbeexpressedasfollowing:AssumingthatIdmaxIdmin,Equation(4)canbeapproximatedas:第6页Equation(5)suggeststhat,thequantizationlevelisinverselyproportionaltothedischargecurrentundertheUQscheme,asIdmin/Idmaxisaconstant.Indeed,italsosuggeststhattheUQschemeissensitivetothedischargingcurrentclosetoI-d-min,showninFigure3(b).ThissuggeststhattheUQschemefocusesonquantizingthelowilluminationlevels,whilethehighilluminationlevelsarenotproperlycovered.2.2.2Non-UniformTime-DomainQuantizationNon-uniformtimedomainquantizationschemeresolvesthequantizationlevelsinordertoformlinearlydistributedphotocurrentsIwithintheboundariesI-d-minandI-d-max.ThephotocurrentIdcanbeconvertedtoitsdedicatedquantizationlevel-NUQ(n,Id)asfollows:第7页FromEquation(7),wecannotethattherelationshipbetweenthetimeandthequantizationlevel-NUQ(n,t)canbeexpressedas:IncontrasttoUQscheme,Equation(8)suggeststhat,theNUQquantizationlevelsareinverselyproportionaltothesamplingtime.Figure3(b)showsthattheNUQprovidesanevenlydistributedphotocurrentsamplingboundaries.3.TheProposedMulti-ResetIntegration(MRI)SchemeInordertoreducesiliconareaofthepixelandtoimprovethefillfactor,amulti-resetintegration(MRI)schemeisproposedinthisworktoreducethememoryneedsatthepixellevel.ThissectionpresentstheconceptoftheMRIanddiscussesthetrade-offsintermsofdelayoverheaddependingonthesizeofthepixelmemory.3.1MRIConceptTheproposedintegrationschemetakesadvantageofthesequentialwayhowtheilluminationlevelisdigitized.TheMRIschemecanbeinterpretedas第8页performingtheintegrationprocessseveraltimesinordertoresolveeachbitoftheilluminationcodesequentially,fromhemostsignificantbit(MSB)totheleastsignificantbit(LSB),asshowninFigure4.Duringeachnewintegrationphase,onlyasub-setofthenbitsarestoredatthepixelreducingthememoryrequirements.Then,betweentwointegrationperiods,thecontentofthepixelmemoryisscannedoutofthearrayallowingfortheremainingbitsofthecodetobestoredduringthesuccessiveintegrationphases.Therequirednumberofbitsforthememoryisthereforereducedbyafactorproportionaltothenumberofiterations.Forexample,consideringaresolutionof8bitsandasinglebitmemory,8successiveresetswillberequired.Inordertoreducethedelayoverheadcausedbythesuccessiveintegrationperiods,itispossibletodefinetimingboundariesforwhichthevalueoftheconcernedbitofthecodeisresolvedallowingtooptimizethecorrespondingintegrationdurationforeachbit.Assumingthephotocurrentisquantizedinto2nvalues,withnastheresolution.ThequantizationlevelsNQTcoveredbyeachbitcanbeexpressedas:第9页wherebit(0)representstheLSBandbit(n1)representstheMSB.ThequantizationlevelscoverageoftheMSBisonlyhalfofthetotalquantizationlevels.Therefore,theoptimizedintegrationtimeforthebitsclosertotheMSBismuchshorterthanthosetowardstheLSB.Indeed,derivedfromEquation(1),theoptimizeddurationofthepartialintegrationphaseTintrequiredtoobtainthebitnumberiofthecodeisgivenby:whereTmax=a/Iministhemaximumintegrationtimefornbitsresolution.UndertheImaxImincondition,Equation(10)illustratesthatintegrationtimedecreasesexponentiallywheniincreases.Assuming1-bitpixelmemory,gettingtheNbitssuccessivelywillgiveatotalintegrationtimeTtotalthatcorrespondstothesumofthedurationsofeachpartialintegrationphaseTint.Therefore,theactualintegrationtimeoftheproposedschemeisexpressedby:Fromthisequation,thetotaltimerequiredtorealizetheintegrationwiththemulti-resetintegrationschemeismuchshorterthannT-max.Itcanalsobenotedthatthetimerequiredtoscanthevaluesoutisnotaccountedforinthisequation.However,thistimecanbereducedsignificantlyusingparallelandhigh-speedreadoutandremainsnegligiblecomparedtothetotalintegrationtime.Inaddition,theread-outphasecanbeinterleavedwiththeintegrationphase.Figure5takes2-bitresolutionasexampleandassumesImax/Imin=100.UsingEquation(10),TintMSB=1/26Tmax,TintLSB=T-max.ItispossibletoresolvetheMSBata1/26fractionoftheintegrationperiodfollowedbythesecondbitortheLSBafteroneintegrationperiod.Originally,undertheMRIscheme,2Tmaxarerequiredto第10页obtaintheMSBandtheLSBvalues.Inthisscheme,usingtheoptimumtimingboundaries,thetotalintegrationtimeisreducedby48%.Theseoptimizedtimingboundariesareusedtoreseteachpartialintegrationthereforegreatlyreducethetimingoverheadforhighresolutionimagers.Inaddition,overintegrationperiodsareavoidedleadingtoreducedpowerconsumption.3.2Trade-offAnalysisDependingonthememorysizeintegratedatthepixellevel,differentarea-delaytrade-offscanbeachieved.Indeedatrade-offcanbemadebetweentherequirednumberofbitsforthepixel-levelmemoryandtherequirednumberofresetoftheintegrationphases.MATLABsimulationswereperformedtoextractthetrendsinpixelareaversusthesizeofthememoryembeddedatthepixel-levelandtheresultsareshowninFigure6.Inthesimulation,asquarepixelwith30%fill-factorisassumed.TheareaisestimatedforCMOS0.35mtechnology.Arealimitedbythemetalwiresandthetotalsizeofthetransistorsarealsoconsidered.Areaofpixel-levelembeddedmemorieswith6TSRAMand2TDRAMarebothconsidered.TheFigureillustratesthattheareaofthepixelwithembedded6TSRAMislineartothenumberofembeddedbits,sinceitismainlyconstrainedbythetransistorsarea.Whiletheareaofthepixelwithembedded2TDRAMismainlyconstrainedbymetalwires.Thesizeofapixelusing4bit2TDRAMisonly72%ofthepixelusing4bit6TSRAM.第11页MATLABsimulationswerealsoperformedtoextractthetrendsinthedelayoverheadversusthesizeofthememoryembeddedatthepixel-level.Figure7showstheinterpolatedcurveofthedelayoverheadintermsofintegrationtimeversusthememorysizenforan8-bitresolutionbyassumingImax/Imin=250.Fromtheseresults,itappearsthatusinga4-bitpixel-levelmemoryrequiresonlyoneresetandleadstolessthan10%overheadonthetotalintegrationtime.Consideringboththeareaandtheintegrationtimeoverheads,usingthe4bits2TDRAMoptimallymaintainsasmallpixelareaandkeepsarelativelyhighoperationspeed.第12页中文翻译稿1介绍互补金属氧化物半导体（CMOS）图像传感器，是我们现在日常生活的一部分涵盖面广，从手机摄像头，网络摄像头，数码相机，视频游戏的安全性和汽车应用1。在电荷耦合器件（CCD）的技术之上成功部署的CMOS图像传感器，主要是由于降低了成本和功率消耗，以及更高的集成度和片上的处理能力，并且是移动应用程序的关键。在CMOS图像传感器之间，两种主流架构是为了区别在模拟-数字转换器（ADC）的实现。第一种是有源像素传感器（APS），它执行使用每个阵列或每列读出方案的像素阵列外部的ADC。第二种是数字像素传感器（DPS），并且在像素级中集成的模数转换器。DPS，同时给予的灵活性使用时，在像素级执行并行模拟到数字的大规模转换，并能启用有前途像素级的图像处理，这是一种对于实时应用，如汽车和监视系统是非常有吸引力的。此外，基于非常早期的模拟-数字转换，DPS可以提供改进的噪声系数和动态范围2。过去的DSP实现是基于两种脉冲频率调制（PFM）3或脉冲宽度调制（PWM）4方案对用电压比较器中的光电二极管的传感节点的电压进行数字化。该PWM结构采用比较器输出信号，以在像素级存储器写一个全局计数器的值，而PFM方案使用比较器的输出，以使像素级计数器的增量。由于在像素级的高开关活动中，PFM和PWM方案这两种通常耗电。此外，两种方案的一大缺点是特别对于高分辨率成像区饥渴像素级别内存的需求。因此，基于像素级存储器面积和功耗是相当大的，能占到总开销的高达70。这很大的影响了像素的填充素从而影响了其光灵敏度，以及电源和像素面积。做多次尝试是为了减少在像素级别的存储需求。在其他解决方案中，地址事件完全表示（AER）的方案的使用在增加的复杂性的价值上和当几个像素试图同时访问该数据总线因碰撞发生时间误差的采用完全消除了内存。另一报告的解决方案包括甚至在所建议的存储6压缩数据。最后的解决方案除了使用了了有损压缩的成本方案降低的信号-噪声比（SNR）的价值意外，大大减少了所述像素区域。在存储器7中，比较器的输出是使用每一个位寄存器单元像素进行采样和存储。所有寄存器单元被串联连接，以形成一个扫描链和之后的整个阵列进行扫描每个采样操作。虽然只有一个位像素级的内存是必需的，这种架构作为该决议的一个强大的功能案例的数目增加了耗电量。在这项工作中，一个使用2TDRAM作为存储元件和一个多复位积分方法的DPS架构是为了减少存储器需求和第13页存储容量都在像素级别和减少建议电力架空。本文结构如下：第2节介绍了PWM时域DPS和它的转换时间分析。第3节是专门用于多复位积分方法和有关的讨论权衡参与速度，面积和内存大小方面。第4节介绍了多重组DPS架构，像素电路和2TDRAM的实施和检测方案。第5节提供电力消耗的分析以及降低功耗的技术。第6节介绍了原型实现和实验测量。第7节总结这项工作。2时域的PWM的DPS2.1传统的体系结构使用PWM技术的一个DPS的传统的像素结构如示于图1。该像素包括一个光电二极管PD，复位晶体管AR，一个电压比较器，一个存储单元和一个反馈电路来执行光电二极管的自动复位。外像素阵列，定时控制单位和一个全局计数器被需求来执行光强度转换成一个数字代码。一个读出电路还需要读存储器中的从所述的内容并且将数据输出处理单元。图1传统的像素和相应的时序图的示意图。当复位信号状态从0到Vdd，该整合阶段开始是通过通过禁用预充电到光电二极管的电源电压。与此同时，位于像素阵列外部的全局计数器同时启用。Vdof装有一个电容CPD光电二极管节点的电压开始按比例递减，基于光生电流Id，通过光电二极管钯来来第14页点亮光的强度。当光的当电压Vd达到基准电压Vr，器的输出为高和计数器的值被写入到像素级的存储器。光电二极管的电压达到参考电平Vref之后，进行光电二极管的复位自动。在内存中的代码编写是所需Vdto达到Vref的时间的一个数字化的价值，这是由于光的功能强度。采用一阶近似，积分时间被表示为：从公式（1）中，我们必须注意到积分时间是反比于光生电流Id。要执行的积分时间的量化，全球计数器用于提供量化边界的时间-数字转换。非线性的光电流和积分时间之间，通过适应全局的频率如图2计数器可以被补偿。一方面，图2（a）表示来自传感器的光电流边界的量化步长导致非线性响应一个均匀的时域量化有非线性的。另一方面，在图2所示的非均匀的时域量化（二），以线性光电流转换为数字代码作为光电流边界量化步骤现在是线性的。图2均匀的和非均匀的时域量子化4。2.2转换时间分析PWM编码方案使用单脉冲的每个像素的照明级信息进行编码。该脉冲宽度表示时间色调从VDD到VREF到光电二极管放电。为了使脉冲转换成数字代码，由全局计数器生成的时间码一旦检测到脉冲就被写入到嵌入式存储器。有两个量化方法的PWM脉冲信号转换成数字代码。第一种方法，被称为均匀时域量化（UQ），它提供从TMIN采样时间（或边界量化电平）到最高温度均匀。第二种方法被称为非均匀时域量化（NUQ），解决了采样时间，以便形成均匀量化的光电流。一个非线性时域量化器是必需的，以便为了补偿时第15页间到光电流的非线性。2.2.1统一的时域量化统一的时域量化方案划分的时间界限T最小和T最大内平均为2n个量化电平，其中n表示位或分辨率的数目。因此，每个量化电平的时间宽度T能够被表示为：当量化级-UQ（N，T）之间的关系以及PWM脉冲t为时间宽度的公式如下：使用公式（3）的量化级-UQ（3）之间的转换（N，ID）和光电二极管的光电流放电Id可表示如下：假设Id最大值大于Id最小时，方程（4）可以近似为：等式（5）表明，假如Id最小值/Id最大值是常数，量化电平是成反比下UQ计划放电电流.确实。确实，它也表明，UQ方案是放电电流接近的Id最小值，如图3（b）所示。这表明该UQ计划集中在量化的低照度水平，而高照明度没有被覆盖。2.2.2非均匀时域量化非均匀时域量化方案解决量化电平为了形成线性分布的光电流I在Idmin和Idmax范围内。光电流Id可以转化为它的专用的量化电平NUQ（N，ID），如下所示：图3.（一）3比特UQ和NUQ在时间上;（二）3比特UQ和NUQ在以下方面放电电流。第16页从式（7）中，我们可以注意到，时间和量化电平之间的关系NUQ（N，T）可以表示为:相反，UQ计划，方程（8）表明该NUQ量化电平成反比正比于采样时间。图3（b）所示该NUQ提供均