An integrated optical transient sensor - Circuits and …...of biologically-inspired image-processing circuits that perform such functions as motion-sensing [13], attentional selection

612 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 49, NO. 9, SEPTEMBER 2002

An Integrated Optical Transient SensorJörg Kramer

Abstract—The implementation of a compact continuous-timeoptical transient sensor with commercial CMOS technology ispresented. In its basic version, this sensor consists of a photodiode,five transistors and a capacitor. The proposed circuit producesseveral output signals in parallel. These include a sustained,logarithmically compressed measure of the incoming irradiance,half-wave rectified and thresholded contrast-encoding measuresof positive and negative irradiance transients, and a signal thatshows a combination of the sustained and the bidirectional tran-sient response. The particular implementation reported in thiswork responds to abrupt irradiance changes with contrasts downto less than 1% for positive transients and 25% for negativetransients. Circuit modifications leading to more symmetric con-trast thresholds around 5% are also described. Due to theircompactness these transient sensors are suitable for implemen-tation in monolithic one- or two-dimensional imaging arrays.Such arrays may be used to sense local brightness changesof an image projected onto the circuit plane, which typicallycorrespond to moving contours.

Index Terms—Active pixel, analog VLSI, CMOS image sensor,focal-plane sensor, neuromorphic sensor, temporal processing.

I. INTRODUCTION

T HE HIGH density of visual information in our environ-ment makes real-time image transmission and processing

a major challenge, in spite of the ever increasing speed of avail-able electronic processing circuits. However, image data tendsto be highly redundant and, thus, compressible without infor-mation loss. Furthermore, for a given application, part of theimage information may be irrelevant. In order to alleviate thebandwidth requirements of multiplexed data transmission andprocessing channels, suitable prior coding and reduction of theimage data is highly desirable. This can partly be achieved byperforming local image processing operations concurrently ateach location of visual data acquisition (pixel) in the focal plane.A very basic and useful image compression function is the en-hancement or extraction of fast temporal changes in the image.This is particularly true for highly correlated time-variant im-ages with large static regions. Not surprisingly, temporal changedetection techniques are widely used in video compression al-gorithms, such as those defined by the MPEG standards.

As a substrate for electronic imaging and focal-plane imageprocessing CMOS technology has been gaining new interestin the past years as a consequence of its rapid development, asdriven by the computer industry [1]. This leads to continuingminiaturization and improvement coupled with inexpensive

Manuscript received June 13, 2001; revised November 23, 2002. This paperwas recommended by Associate Editor A. Andreou. This work was supportedin part by the Swiss National Foundation under Research Grant 5002-57809/1.

The author, deceased, was with the Institute of Neuroinformatics, Universityof Zurich and ETH Zurich, 8057 Zurich, Switzerland.

Digital Object Identifier 10.1109/TCSII.2002.807270

fabrication and low power consumption. Due to the essentiallytwo-dimensional structure of CMOS and most other integratedcircuit technologies, image acquisition and processing elementshave to be interlaced in the focal plane and there is a tradeoffbetween image resolution and parallel processing power. Giventhe resulting requirement of high processing densities and theanalog nature of the optical input signal analog approaches forparallel focal-plane image processing compare favorably withdigital solutions for the implementation of various functions.

In this article, we present a compact integrated circuit suitablefor use in a focal-plane image processing pixel. Its main functionis the enhancement of local brightness transients as a means forimage preprocessing and compression. The article is organizedas follows: Section II describes some of the circuit features withregard to imaging applications, Section III presents the configu-ration of the circuit. The circuit characteristics are theoreticallyanalyzed in Section IV and measurement results are shown inSection V. Section VI provides a summary and conclusions.

II. CIRCUIT FEATURES

The presented circuit combines an adaptive photoreceptor[2] with a rectifying and thresholding differentiating element[3], [4]. It computes different functions of the optical inputsignal at different terminals in continuous time. The irradianceimpinging onto the circuit plane is transduced linearly into asustained current and logarithmically into a sustained voltage.Positive orON transients and negative orOFF transients areencoded on separate output nodes. Finally, a voltage signalis provided that superimposesON and OFF transients ontoa logarithmic sustained response. A variation of the circuitincludes a gain stage providing an additional node that amplifieschanges of the sustained voltage. The amplified signal slowlyadapts back to the new sustained value. This modified circuit,therefore, provides transient enhancement at two different timescales. Similar encoding schemes exist in biological early visualprocessing structures that have evolved to handle the sametypes of images that artificial image processing circuits aretypically presented with. SeparatedON and OFF channels arefound in the retina and the lateral geniculate nucleus andsustained and transient responses in the parvocellular andmagnocellular pathways projecting from the retina into thevisual cortex.

The ON and OFF transient nodes are less prone to dc off-sets than nodes with bipolar signal coding, because they re-spond to a static input with a signal at the boundary of theiroutput range (null signal), rather than in its middle. This hasthe additional benefit that their steady-state power consumptionis very small. The transient responses are thresholded, such thatlow-amplitude, high-frequency noise is suppressed, while larger

1057-7130/02$17.00 © 2002 IEEE

KRAMER: INTEGRATED OPTICAL TRANSIENT SENSOR 613

and slower signals are emphasized. For small irradiance tran-sients above the threshold, the output currents at theON andOFF

terminals show a power law dependence on the irradiance rela-tive to its initial value, while larger transients evoke a responsethat approximates the rectified first temporal derivative of thelogarithm of the irradiance. To a first approximation, the tran-sient responses, thus, encode the irradiance transients relativeto irradiance, i.e., the temporal contrast, and are insensitive toabsolute irradiance. At the same time, the difference of the log-arithmic sustained responses of adjacent pixels directly encodesspatial contrast.

For image processing applications, encoding of contrast isgenerally more useful than encoding of brightness. This is dueto the fact that most objects in our environment do not generatelight but merely reflect it in a more or less diffuse manner. Sincethese processes are quite linear, the global image brightnessmainly conveys information about the light source, while theavailable information on object surfaces is found in local imagecontrasts. The illumination of natural environments varies byseveral orders of magnitude and the contrast-sensitivity of linearirradiance sensors changes strongly over such a dynamic range,such that in darker parts of the image no features can be detected.Even if the coding is logarithmic, most of the dynamic range ofan irradiance sensor with a purely sustained response must bedevoted to cover variations of illumination. Variable lens aper-tures alleviate this problem in the case of homogeneous illumi-nation conditions, but such sensors generally fail in the presenceof large irradiance variations within a single image, e.g., due tothe presence of dark shadows in a sunlit environment or due tospecular reflections of metallic surfaces. Enhancement of tran-sient responses and slow local adaptation to a mean brightnesslevel provide a method to make better use of the available dy-namic range.

Most previous implementations of optical transient sensorseither do not provide rectifiedON and OFF output channels[5]–[10], or they are significantly larger than the proposedcircuit [11], [12]. The presented sensor is also more compactand less noise-prone than a previously reported circuit using asimilar temporal differentiation technique [3], [4]. The sensorhas already successfully been used as a front-end for a varietyof biologically-inspired image-processing circuits that performsuch functions as motion-sensing [13], attentional selection[14], and orientation tuning [15]. The basic version of thetransient sensor and some modifications to it are presented inthe following section.

III. CIRCUIT CONFIGURATION

A. Basic Circuit

The proposed circuit in its basic version is shown in Fig. 1.It consists of a photodiode in series with a transistorin source-follower configuration and a negative feedback loopfrom the source to the gate of [2]. The feedback loopconsists of a high-gain inverting amplifier in common-sourceconfiguration ( , ) [2] and a thresholding and rectifyingtemporal differentiator stage ( , , ) [4].

The photocurrent linearly encodes irradiance over severalorders of magnitude. For imaging applications under typical in-

Fig. 1. Circuit diagram of the basic version of the transient sensor.

door or outdoor lighting conditions, conventional lens aperturesand circuit geometries will result in values small enoughto keep in weak inversion (subthreshold domain) and thegate-to-source voltage of , thus, changes logarithmicallywith irradiance. The high-gain feedback keeps the source poten-tial almost constant and ensures that irradiance variationsare logarithmically encoded in the gate voltage. The feed-back loop significantly speeds up the circuit response with re-spect to a simple source-follower configuration, since the smallphotocurrent does not have to charge the large photodiodecapacitance when is clamped [2].

Temporal differentiation is provided by the current onto thecapacitor connected to the gate node of and the sourcenodes of and . For positive irradiance transients,the output signal of the inverting amplifier activatessuch that gets charged by a current proportional to thetemporal derivative of . For negative irradiance transients,

activates to discharge with a current en-coding the temporal derivative of . Note that if the bulks of

and would both be connected to the correspondingpower rails the body-effect would be larger for than for

, because is only two diode voltage drops above thelower power rail. If an n-well process is used to implement thecircuit the body effect of can be avoided by connectingthe well to the source, as shown in Fig. 1. In order to get thesame benefit from a p-well process the types of all transistorshave to be exchanged. A BiCMOS process, providing a baseimplant of the opposite type from the well implant, allows to im-plement one of these two transistors as a bipolar and the other asa well-connected MOSFET, such that none of them has a bodyeffect and they are, thus, both insensitive to the voltage of theoperating point, as set by the irradiance and the amplifier biasvoltage . Irradiance transients that are so small that the dif-ferentiating currents they produce are lower than the leakagecurrents of and are not detected. This implements


Fig. 2. Circuit diagram of a modified transient sensor with a cascode, anadaptive gain stage and leakage current symmetrization.

a thresholding function. Given typical irradiances and a reason-ably small capacitor , and will always stay inweak inversion, even in response to large and sudden irradiancechanges, due to the limited bandwidth of the circuit.

In an ideal circuit, the voltage at the output of the in-verting amplifier has approximately the same steady-state valueas with a small offset, as determined by the different char-acteristics of and . As we will see, however, thereis an additional offset because these two transistors are affectedby leakage currents to the substrate. With respect to its baseline,

logarithmically encodes and , thus, providing anemphasized bipolar transient response on top of the sustainedresponse of .

B. Circuit Variations

The performance of the basic version of the transient sensorcan be improved by incorporating additional circuitry. Threedifferent modifications of the basic circuit are shown in Fig. 2.Each one of these modifications can be implemented indepen-dently of the others and serves a different purpose, as describedbelow.

For sufficiently large amplifier bias currents the bandwidthof the circuit is limited by the fact that has to charge theMiller capacitance between the gate and the drain of. TheMiller effect can be avoided by introducing a cascode toshield the drain of against the large variations of [2].Given that the bandwidth of the photodiode linearly decreaseswith irradiance, the use of a cascode is particularly beneficial ifthe circuit is to be operated under low-irradiance conditions.

The offset of with respect to due to leakage currentsresults in different contrast thresholds for the response of thedifferentiating currents to positive and negative transients. Thisasymmetry can be reduced by providing a constant currenttothe differentiator node that counterbalances the leakage current.

The response of the basic circuit can be enhanced by pro-viding an additional capacitive gain stage in the feedback loop[2]. The voltage variations on the differentiator node are then

Fig. 3. Basic version of the transient sensor with logarithmiccurrent-to-voltage converters at theON andOFF transient outputs.

amplified with respect to the variations of by the capacitivedivider ratio and so are the transient cur-rents and . The resistive element ensures that thevoltage at the differentiator node eventually adapts to a dcvalue close to . The adaptation dynamics are determined bythe characteristics of the resistive element.

A symmetric resistive element can be constructed from afloating well and two diffusions of the opposite doping typefrom the well. It corresponds to two oppositely directed diodesin series. In the ideal case, the current is limited to a low valueby the reverse-biased diode, such that the element exhibits asymmetric saturating sigmoidal current–voltage characteristic.It supports arbitrarily large voltage drops within the limits of thereverse-breakdown voltages of the diodes and provides largeadaptation time constants. However, for small separations ofthe two diffusions the device operates as a lateral p-n-p bipolartransistor with a floating base. The base current, supplied bythe well-substrate leakage current, is then amplified to providean increased adaptation current, which may considerably speedup the adaptation. The total dark leakage current to the sub-strate is small, but photogenerated minority carriers from thesubstrate are collected by the floating well and, thus, providean irradiance-dependent adaptation rate. Furthermore, theparasitic vertical p-n-p bipolars into the substrate are activated.The resulting leakage currents cause an offsetand make the adaptation properties asymmetric. The leakagecurrents have to be supplied by in the circuit and increasethe steady-state value of .


(a) (b)

(c)

Fig. 4. Steady-state characteristics of the basic circuit of Fig. 1, as measured in the configuration of Fig. 3, showing (a) the logarithmic irradiance encoding in thevoltageV at the feedback node, (b) the dependence of the voltageV at the amplifier output node onV , and (c) the irradiance-dependence of the differentnode voltages. The dotted lines indicate the response to an irradiance of 38.2 mW/m, the baseline irradiance for all transient measurements. An irradiance of1 W/m by the 590 nm LED correponds to a flux of about3e6 photons/�m /s, or an illumination of approximately 300 lux by white light.

IV. CIRCUIT ANALYSIS

A. Adapted Steady State

The current through the photodiodecan be computed as

(1)

where denotes the wavelength of the incoming photons,the quantum efficiency of the conversion of photons into

electron–hole pairs contributing to the photocurrent at, thePlanck constant, the speed of light, and the incomingoptical power per unit wavelength at. In the case of an opticalimaging system we find that

(2)

where denotes the aperture of, the optical trans-mission of the imaging system at, the -number of theimaging system, and the irradiance of the system per unitwavelength at .

The transistor is operated in saturation and, for typicalirradiances, in weak inversion. Neglecting its Early effect wethen obtain

(3)

where is the current-scaling parameter and the sub-threshold slope factor of and denotes the thermalvoltage , given by the absolute temperature, the Boltz-mann constant , and the elementary charge. The voltage

is set by the bias current through the inverting amplifier.Assuming a bias voltage that puts and into weakinversion and again neglecting Early effects we obtain

(4)

where and are the current-scaling parameters of and, respectively, and are the corresponding subthreshold

slope factors and is the potential of the positive power rail.In this approximation is independent of . The actual de-pendence is due to the Early effects of and , i.e., to thelimited gain of the inverting amplifier.

Neglecting the leakage currents of and and as-suming weak inversion and saturation for these transistors we


(a) (b)

(a)

Fig. 5. Steady-state characteristics of the circuit of Fig. 2 showing (a) the irradiance response of the voltageV at the differentiator node, (b) the dependenceof the voltageV at the amplifier output node onV with an abrupt transition ofV from a negative to a positive offset with respect toV at the baselineirradiance, as tuned with the leakage currentI , and (c) the irradiance-dependence of the different node voltages.

can compute the steady-stateON andOFF currents of the basiccircuit (Fig. 1) as

(5)

(6)

where and are the current-scaling parameters ofand , respectively, and and are the corre-

sponding subthreshold slope factors. From the steady-state con-dition that we find that

(7)

(8)

As already mentioned in Section III-A, however, this equilib-rium is disturbed by leakage currents in the transient pathways.In the present circuit, such leakage currents mainly affect thediodes to the substrate, i.e., the source and drain diodes of

and the well-to-substrate diode of . The source

diode of does not contribute a leakage current becauseit is shorted and the leakage current of the drain diode of

is small. Furthermore, the parasitic currents at the draindiodes of and do not influence the channel cur-rent, but add to the current drawn from subsequent devices.The source leakage current of and the well leakagecurrent of , however, result in a parasitic currentfrom the differentiator node into the substrate, which has to bebalanced by an increased current through and, therefore,an increased offset of with respect to . Accordingto Kirchhoff’s current law, . Sinceis typically much larger than the current through the idealcircuit, as computed by (7), . Therefore

(9)

If and are implemented in the vicinity of the pho-todiode on the same silicon substrate, photoinduced minoritycarriers also contribute to the leakage and dominate it for largeirradiances, such that it becomes roughly proportional to thephotocurrent, i.e., , where denotes the ratio ofthe electrons collected by the source of and the well of

to those collected by the photodiode. The dependenceof on can be determined from the slope of the


(a) (b)

(c) (d)

Fig. 6. Response of the basic circuit to exponential irradiance transients at (a) the feedback node, (b) the amplifier output node, (c) theON transient node, and(d) theOFF transient node. The different traces in each figure represent the responses to exponentialON andOFF transients with time constants ofj� j = 10, 5.62,3.16, 1.78, 1, 0.562, 0.316, 0.178, and 0.1 s. In (a) and (b) the response toON transients is above the baseline with increasing response to decreasing values of�

and the response toOFF transients is below the baseline with increasing negative response to decreasing values ofj� j. In (c) the solid lines show the increasingresponse of theON transient node toON transients with decreasing� and the dashed lines show the spuriousOFF responses. Conversely, (d) shows the increasingresponse of theOFF transient node toOFF transients with decreasingj� j as solid lines and the spuriousON responses as dashed lines.

versus characteristic. The expected slope for a given rela-tionship between and can be obtained by expressing

in terms of using (3) and then substituting it into (9).For the modified circuit of Fig. 2 the currents on the differen-

tiator node satisfy . If is signifi-cantly larger than , and

(10)

If an additional capacitive gain stage is used in the feedbackloop, as shown in Fig. 2, and the circuit is fully adapted,

and the same analysis applies. The offset betweenandis determined by the leakage currents of the chosen resistive

element, as described in Section III-B. These leakage currentsadd to at the differentiator node.

B. Transient

In the following, we will make a transient analysis of the cir-cuit without considering parasitic capacitances and the effectsof leakage currents in the transistors. How these nonidealitiesaffect the transient operation of the circuit and how the intro-

duction of a cascode and of a current source at the differentiatornode change their effects will be described at the end of thissection. The analysis presumes that the circuit variables havereached an equilibrium state before a transient change in thephotocurrent is applied. The effect ofON andOFFtransients willbe treated separately. First, the basic circuit will be analyzed andafterwards the modification of the obtained results by a capac-itive gain stage will be presented, assuming that no adaptationoccurs in the considered time window.

The absolute value of the gain of the inverting amplifier isdetermined by the Early effects of and , and is given by

(11)

where and are the Early voltages of and , re-spectively.

Differentiating (3) and substitutingyields

(12)


(a) (b)

(c)

Fig. 7. Fitted parameters to the data shown in Fig. 6 as a function of the inverse of the irradiance transient time constant. (a) Temporal derivatives ofthe voltagesignals in response to exponentialON transients. (b) Temporal derivatives of the voltage signals in response to exponentialOFFtransients. (c) Closed-loop responsesof different nodes toON andOFF transients. For display purposes, theOFF response of the amplifier output node is offset by�0.88 V. For an ideal circuit, alltemporal derivatives would be inversely proportional toj� j, while the differential voltages would depend logarithmically onj� j.

Differentiating (5) and (6), respectively, gives

(13)

(14)

If leakage currents in the differentiator stage are neglected,the capacitor current is given by

(15)

It follows from (3) that

(16)

In the closed-loop domain, where the feedback loop is activated,the term can be neglected if the loop gain is much largerthan unity ( ). The difference in the transient cur-rents is then proportional to the temporal derivative of the log-arithm of the photocurrent, i.e., to the relative transient of thephotocurrent.

Considering anON transient which is sufficiently large that, (15) reduces to

(17)

Eliminating the and terms from (12), (13), and (17)we end up with the differential equation

(18)

where

(19)

is the loop gain forON transients. We can solve (18) by inte-grating it twice, which yields

(20)

where

(21)


(a) (b)

(c) (d)

Fig. 8. Response of the modified circuit to exponential irradiance transients at (a) the differentiator node, (b) the amplifier output node, (c) theON transient node,and (d) theOFFtransient node. Refer to the caption of Fig. 6 for details. For largeOFFtransients the circuit starts to adapt within the considered time frame and forthe largestOFFtransient the response ofV additionally saturates near the voltage of the low power supply rail, thus, causing a saturation ofV and terminatingthe response ofV . The sinusoidal modulation of the signals for large transients is due to the discrete approximation of the stimulus to an exponential irradiancetransient.

is the small-signal differentiator time constant forON transientswith given by (7). TheON transient current can be com-puted from (17) and (20) as

(22)

From (5), (20), and (22) we obtain the amplifier output voltage

(23)

As we can see, the transient changes in these circuit variablesonly depend on the relative photocurrent change ,i.e., on the irradiance contrast and not on absolute photocurrentvalues.

For small signals and time delays, where

(24)

we can neglect the effect of the feedback and the circuit operatesin open loop. In this domain, (20), (22), and (23), respectively,reduce to

(25)

(26)

(27)

The transient current, thus, has a power-law dependence and theamplifier output voltage has a logarithmic dependence on therelative photocurrent change in the open-loop domain.

For larger signals and time delays, where

(28)

the feedback loop is closed and (20), (22), and (23), respectively,simplify to

(29)


(a) (b)

(c) (d)

Fig. 9. Irradiance step response of the basic circuit at (a) the feedback node, (b) the amplifier output node, (c) theON transient node, and (d) theOFF transientnode. The different traces in each figure represent the responses toON andOFFstep transients with a rise time of 1 ms corresponding to optical contrasts of 0.625%,4.99%, 9.97%, 24.5%, 46.2%, 76.2%, and 90.5%. In (a) and (b) the response toON transients is above the baseline and the response toOFF transients is belowthe baseline, with increasing distance from the baseline for increasing contrasts. In (c) the solid lines show the increasingON response of theON transient node toincreasing contrasts and the dashed lines show the spuriousOFFresponses. Conversely, (d) shows the increasingOFFresponse of theOFFtransient node to increasingcontrasts as solid lines and the spuriousON responses as dashed lines.

(30)

(31)

An analogous analysis forOFFtransients with the assumptionthat reducing (15) to and using(12) and (14) leads to the differential equation

(32)

where

(33)

is the loop gain forOFF transients. Solving (32) and using (6)yields

(34)

(35)

(36)


where

(37)

is the small-signal differentiator time constant forOFFtransientswith given by (7). For the small-signal and large-signalregimes we can make analogous approximations as for theON

transients.In the presence of an additional capacitive gain stage with a

gain of in the feedback loop, as shownin Fig. 2, the transient characteristics are similar to the onesof the basic circuit, as long as the adaptation through the re-sistive element of the gain stage can be neglected. In order totake into account the gain stage in the above analysis,hasto be substituted for in all equations except for (12). In theidealized circuit described above the gain stage does not affectthe open-loop behavior, but it delays the onset and increases thegain of the closed-loop regime. Considering the fact that thevoltage transient at the differentiator node is amplified by thecapacitive divider ratio with respect to the voltage transient atthe feedback node, according to , the differenceof the closed-loop transient currents, as given by (16), and alsothe closed-loop gain of are amplified by . The loopgains (19) and (33) are reduced toand , respectively. This boosts the timeconstants and of the differentiator node by approx-imately , such that the transition from the open-loop to theclosed-loop regime is delayed and the excursion of fromits steady state at the transition point is increased by a term pro-portional to .

Taking the parasitic capacitances into account, we note thatthe response of the circuit is low-pass filtered by the finite timeconstant of the photodiode node, as determined by the capaci-tance of the photodiode and the Miller capacitance between gateand drain of . The Miller capacitance can be eliminated andthe bandwidth, thus, increased by introducing a cascode tran-sistor , as shown in Fig. 2. Similarly, the response of theamplifier output node is low-pass filtered by parasitic capaci-tances, but the time constant of this filter can be set by the am-plifier bias current via . In order to reduce ringing effects thetwo time constants should be sufficiently different. Typically,the amplifier time constant is set to be smaller than that of thephotodiode node at the largest irradiance the circuit is speci-fied for, such that the circuit is stable over the entire irradiationrange. However, for certain applications, where, for example,light flicker has to be filtered out, it may be advantageous tolimit the bandwidth of the circuit to a low value by low-pass fil-tering the amplifier response.

The parasitic capacitance between the amplifier outputnode and the differentiator node limits the open-loop responseof , because it instantaneously activates the feedback loopwith a transient loop gain of , where isthe capacitive divider ratio . On a short timescale, where the charge on the differentiator node is constant,

changes through capacitive coupling, according to, such that . With

increasing the response of decreases. If, has to be replaced in (27) by . If a capac-

itive gain stage is used, the loop gain is andthe open-loop response is enhanced by, provided that it isstill limited by .

The leakage current to the substrate influences the circuitdynamics. In order to account for the leakage current (15) has tobe modified to . ForON transients,the response is increased and the transition from the open-loopregime into the closed-loop regime happens earlier, so thatis more responsive to small transients at the expense of a largersteady-state value, whileOFFtransients have to overcome a largeopen-loop regime in order to evoke a significant response.For smallOFF transients, is entirely discharged by and

does not change. This increases the threshold for the de-tection ofOFF edges. Balancing with reduces the in-fluence of the leakage at the differentiator node on the circuitcharacteristics. TheON andOFF response thresholds of a fullybalanced circuit are due to the leakage currents at the transientoutput nodes and to the slow response to small transients.

V. EXPERIMENTAL SETUP

Different versions of the transient sensor have been imple-mented with a variety of CMOS technologies. The measure-ments presented here were performed on two circuits, oneinstantiating the basic sensor of Fig. 1 in a 2m n-wellprocess and the other the modified version of Fig. 2 in a1.2 m n-well process, using the resistive element describedin Section III-B and a single p-type MOSFET as a tunablecurrent source onto the differentiator node.

The output currents and were sensed as voltagesand by on-chip logarithmic current-to-voltage

converters implemented with diode-connected MOSFET’sand , as shown in Fig. 3 for the basic version of

the circuit. The transient currents can be computed from themeasured voltages as

(38)

(39)

where and are the current-scaling parameters andand are the subthreshold slope factors of and , re-spectively. These additional transistors neither affect the opera-tion of the circuit nor degrade the measurements significantly.In addition to , , and , was measured in thebasic circuit and in the circuit with the gain stage. Thepower supply voltage was set to 5 V and the bias voltage

to a fixed value for all measurements, such thatandwere operated slightly above threshold. Subthreshold values of

provide a larger amplifier gain, but introduce ringing effectsand lower the dc operating point, thereby limiting the outputrange of for OFF transients.

The measurements presented in Section VI-A–VI-C weredone by illuminating the entire CMOS chip through a dif-fuser with a radiance-modulated light-emitting diode (LED),operating around an emission wavelength of 590 nm, whileshielding the chip from ambient light. A forward voltage wasapplied to the LED with a series resistor, such that the currentand, thus, the radiance of the LED were approximately linear


(a)

(b)

Fig. 10. Fitted parameters to the data shown in Fig. 9 as a function of thelogarithm of the irradiance step size. (a) Closed-loop voltage decay constantsfor different nodes. (b) Decay time constants of the transient node voltages.For an ideal circuit without parasitic capacitances and leakage currents all theseparameters would be independent of irradiance step size.

with the applied voltage in a certain range. A radiance-voltagecalibration was performed using a photometer.

For the steady-state measurements, the irradiance range wasexpanded by inserting neutral density (ND) filters with atten-uation factors of 10, 100, and 1000 between the diffuser andthe chip. For transient measurements the LED was modulatedusing the buffered and low-pass filtered signal from a wave-form generator. The low-pass filtering was necessary to reduceartifacts due to the discretization of the waveform generator’soutput signal. The time constant of the low-pass filter was setto 1 ms. The circuits were allowed to reach a steady state ata default irradiance of 38.2 mW/m(corresponding to an il-luminance of 19.75 lx) before the onset of each transient. Themeasured transient voltages are plotted relative to their adaptedvalue. The voltage measured at theON transient node is invertedin most plots, such that the plotted data is positive for positivetransients. In the dc characteristics each data point representsthe average of 1000 measurements, while for the transient char-acteristics ten voltage traces were averaged for each stimulustransient.

The parameter ranges for the LED modulation experimentswere chosen to generate irradiance patterns similar to those

(a)

(b)

Fig. 11. (a) Peak responses of different nodes of the basic circuit as a functionof the logarithm of the irradiance step size. The response is quite symmetricwith respect toON and OFF transients, but the axis of symmetry is slightlyshifted due to asymmetric leakage currents at the feedback node. While theON transient node responds to contrasts from 0.6%, the response threshold ofthe OFF transient node is at 24.5%. (b) Onset delays to half peak amplitudeof the responses at the same nodes, limited by the 1 ms transition time of theirradiance steps.

found in typical imaging applications, where dynamic scenesconsisting of nonluminous light-reflecting objects under nearlyconstant illumination are imaged onto an array of photosensors.For example, the maximum Michelson (or Rayleigh) contrast

for transitions between two irradiance levels and ,which is defined as

(40)

was 90%, corresponding to the maximum contrast obtainedfrom typical diffusely reflecting surfaces, as given by thereflection coefficients of black and white surface portions. Thechosen default irradiance was in a range that would be typicalfor an image of a scene under room lighting conditions using astandard imaging system. The minimum irradiance rise time of1 ms, as given by the time constant of the used low-pass filter,corresponds to the transition time of a sharp contrast boundarymoving at a velocity of 30 mm/s across the photodiode ofthe basic circuit, which has a width of 30m. Assuming animaging system with a focal length of 8 mm, such a velocity


(a) (b)

(c) (d)

Fig. 12. Irradiance step response of the modified circuit at (a) the differentiator node, (b) the amplifier output node, (c) theON transient node, and (d) theOFF

transient node. Refer to the caption of Fig. 9 for details. Saturation and adaptation effects can be seen for largeOFFsteps. They cause small spuriousON responses.

would be measured for a person walking across the field ofview at a distance of about 40 cm.

For the experiments described in Section VI-D the circuit wasdirectly tested as an imaging device by stimulating it with theimages of moving stripe gratings printed by a laser printer onpaper wrapped around a rotating cylinder. A square-wave and asine-wave grating were used, each with a contrast of 89% anda period of 50 cm. The optical part of the imaging system wasa surveillance camera lens with a focal length of 8 mm and an

-number of 1.2. The gratings were placed at a distance of 20 cmfrom the lens, resulting in an image magnification of 0.04 and agrating period in the image of 2 cm. The experiments were per-formed under standard office lighting conditions, provided byac-driven fluorescent tubes with a flicker frequency of 100 Hz.All the measured voltages are plotted with respect to a base-line level in the middle between the steady-state responses tothe bright and dark stripes. The corresponding irradiance on thechip surface was approximately 7.5 mW/m. Ten voltage traceswere averaged for each stimulus pattern.

VI. EXPERIMENTAL RESULTS

A. Steady-State Response

The voltage at the feedback node of the basic circuitshows the predicted logarithmic behavior over a large irradi-ance range, as shown in Fig. 4(a). Toward low irradiances the

slope flattens out due to the effect of leakage currents. Fig. 4(b)shows the dependence of on over the same irradiancerange. The offset of with respect to is much largerthan predicted by (8). The major part of the offset is, there-fore, due to leakage currents to the substrate. As we saw inSection IV-A, the slope of the versus curve revealswhether these leakage currents are dark currents or due to pho-togenerated charge carriers. For constant leakage currents, (9)predicts a slope of . If we assume that the leakage currentsare proportional to the photocurrent we can substitute (3) into(9) to find the slope . The obtained slopeclearly indicates that dark currents dominate the leakage in theconsidered irradiance range. Fig. 4(c) shows the irradiance de-pendence of , , and . For irradiances that arein the range of the chosen baseline level for the transient mea-surements (dotted line) or smaller, is independent of irradi-ance, which confirms that the leakage currents are dominated bydark currents. Toward larger irradiance values, startsto increase, suggesting a growing influence of photoinducedleakage currents. The value of remains small throughoutthe measured range. In the dark-leakage domain the increase of

is per -fold increase inirradiance.

The dependence of on irradiance for the modified cir-cuit is shown in Fig. 5(a) and the relationship between and

in Fig. 5(b). The bias current was set such that it was


(a)

(b)

Fig. 13. (a) Peak responses of the different nodes of the modified circuit as afunction of the logarithm of the irradiance step size. The asymmetry betweenON

andOFFresponses is decreased with respect to that of the basic circuit, due to thesymmetrization of the leakage currents at the differentiator node. The contrastthresholds are 6.2% forON transients and 2.5% forOFFtransients. The responseof the amplifier output node saturates for largeOFFsteps, because it approachesthe power supply rail. (b) Onset delays to half peak amplitude of the responsesat the same nodes, limited by the 1 ms transition time of the irradiance steps.

approximately equal to at the baseline irradiance. With de-creasing irradiance slightly increases due to the Early effectof the MOSFET implementing the current source, whileslightly decreases due to the smaller concentration of photogen-erated minority carriers and due to the reduced reverse bias onthe leaking diodes. Hence, for irradiances smaller than the de-fault irradiance and theOFF pathway is activatedwhile for irradiances larger than the default and theON pathway is activated. This explains the abrupt transition of

from a negative offset to a positive offset with respect toaround the default irradiance. The steep slope of the upper

part of the curve indicates that photoinduced leakage currentsdominate the characteristics there. The increased light-sensi-tivity of this circuit with respect to the basic circuit can be ex-plained by the additional irradiance-dependent parasitic currentthat has to be supplied to the resistive element, as explainedin Section III-B. With decreasing irradiance the light-inducedleakage currents should become negligible with respect to darkleakage currents and we expect the slope to approach unity forsmall irradiances, according to (10), which is confirmed by a

fit. The values of and are approximately con-stant for irradiances smaller than the baseline value, as can beseen from Fig. 5(c). The measured value of remained al-most constant throughout the entire range. However, it has to benoted that the lowest measured value for represents thelower limit of the on-chip instrumentation circuitry rather thanthe true value. For irradiances larger than the baseline value wenotice a significant increase of with irradiance, con-firming the strong contribution of photoinduced components tothe leakage currents. The photoinduced leakage current at thelargest measured irradiance is orders of magnitude larger thanthat of the basic circuit.

B. Response to Exponential Irradiance Changes

According to (1) and (2) irradiance is linearly converted intophotocurrent. An exponential irradiance transient, thus, resultsin a photocurrent of the form

(41)

where is the Heaviside function and where the timeconstant of the exponential is positive forON transients andnegative forOFF transients. For such a stimulus and neglectingparasitic effects, (26), (27), and (38) predict linear changeswith respect to time for and in response toON

transients in the open-loop regime. The same behavior ispredicted for and in response toOFF transients.The temporal derivatives of these voltages are given by

, ,and . In theclosed-loop domain, the transients of and arealso linear, as can be seen from (29) and (31) forON tran-sients. They are ,

, and. For the basic

circuit and and the ON transientslope of converges to and the otherclosed-loop slopes to , while a gain stage amplifiesall these slopes by . Note that all transients are inverselyproportional to . In the closed-loop regime, theON andOFF currents are constant over time in response toON andOFF transients, respectively. The corresponding voltages are

, according to (30)and (38), and .Extrapolation of the closed-loop response to the time of the startof the transient ( ) yieldsfor ON transients and for OFF

transients for and , respectively. Notethat these equations relate to the ideal circuit and that in realcircuits especially shows a large offset with respect tothe ideal value. However, assuming constant leakage currents,these relationships still hold differentially between differentvalues.

The measured response of the instrumented nodes of the basiccircuit to an exponentially changing irradiance with differenttime constants, according to (41), is shown in Fig. 6. Since theON pathway is partially activated by leakage currents in steady


(a) (b)

(c)

Fig. 14. Responses of different nodes of the basic circuit to the images of 89% contrast gratings. (a) Voltage traces for a square-wave grating moving at a speedof 15 mm/s. (b) Voltage traces for a sinusoidal grating moving at a speed of 15 mm/s. (c) Peak responses as a function of image speed for the square-wave grating.The dependence of peak amplitude on speed is logarithmic, as expected for closed-loop behavior.

state, a negative irradiance transient first suppresses this leakage[dashed lines in Fig. 6(c)] before it activates theOFF pathway.This leads to a significant delay of the response of theOFFtran-sient node [Fig. 6(d)] and no response at all for the slowest of theused stimuli with 10 s. Accordingly, the open-loop regimeof the amplifier output signal is much larger for negative tran-sients than for positive transients, as can be seen from Fig. 6(b).

The temporal derivatives of the measured voltage signals, asobtained from linear fits to the different curves, are shown as afunction of in Fig. 7. The open-loop slopes of the transientnode voltages are much smaller than predicted by theory, due tothe parasitic capacitances. As we noted in Section IV-B, the am-plifier gain extracted from the open-loop measurement ofis limited by parasitic capacitances and yields an open-loopgain of rather than . The voltage offsets in theclosed-loop regime as a function of are shown in Fig. 7(c).

The responses of the different nodes of the modified circuit tothe same exponential irradiance transients are shown in Fig. 8.We note the enhanced response of with respect to . Thisamplification and the cascode increase the open-loop responseof , which is now limited by the amplifier gain, rather thanby the parasitic coupling. The response of theON transient nodeis also increased. The larger open-loop response and the more

symmetric leakage currents result in a faster activation of thefeedback loop forOFFtransients and, therefore, shorter time de-lays of the signal. The amplitude is underestimateddue to the already mentioned fact that the reference voltage for

is elevated by the limitations of the instrumentation cir-cuitry. For largeOFF transients the response becomes nonlinearon a longer time scale due to adaptation of toward . Forthe largestOFF transient hits the lower power supply railand , therefore, suddenly starts to decay.

C. Response to Abrupt Irradiance Changes

The ideal response of the photocurrent to an irradiance steptransient from to can be described by

(42)

where is the photocurrent before the transient andis thephotocurrent after the transient. For the actual measurements thetransition times between and were limited by those ofthe irradiance changes, rather than by the finite response timeof .

The open-loop behavior forON transients, described by(25)–(27) and (38), predicts that the steps in and are


(a) (b)

(c)

Fig. 15. Responses of different nodes of the modified circuit to the images of 89% contrast gratings. (a) Voltage traces for a square-wave grating moving at aspeed of 15 mm/s. (b) Voltage traces for a sinusoidal grating moving at a speed of 15 mm/s. (c) Peak responses as a function of image speed for the square-wavegrating.

linearly dependent on the logarithm of the relative irradiancestep , increasing by and ,respectively, per -fold increase in . For OFF tran-sients, decreases by and increases by

per -fold decrease in . In theclosed-loop regime, the transient voltages converge logarith-mically with time to the steady-state values. ForON steps thevoltage changes for an-fold increase in time arefor , if , and for , according to(29)–(31) and (38); forOFFsteps they are for , if

, and for . The decay time constantsand are the times it would take the circuit to reach its

new steady state after a step transient, under the assumptionsmade for the circuit analysis. A circuit with an additional gainstage ideally shows the same response characteristics at thetransient nodes, with rescaled values of and , as longas the adaptation can be neglected.

The measured responses of the instrumented nodes of thebasic circuit toON andOFF step transients with different con-trasts are shown in Fig. 9. The open-loop behavior significantlydeparts from the ideal one, due to the finite rise time of the ir-radiance change and the effect of parasitic capacitances. Thecapacitive coupling between the amplifier output node and thefeedback node pulls toward its new steady state. This effect

can be seen in Fig. 9(a), where it induces some ringing for largetransient steps, which can also be observed at the other nodes.For the chosen value of the ringing effects were small enoughas not to induce significant spuriousOFF responses at theterminal and spuriousON responses at the terminal. Theparasitic capacitances and the finite rise time of the irradiancetransients result in peak responses that are significantly smallerthan those predicted by the theory for the ideal circuit.

In the closed-loop regime, the logarithmic decay ofand after the initial ringing effects can clearly be seen inthe traces of Fig. 9(c) and (d). The logarithmic parts of thetraces in Fig. 9(b) are less pronounced, in particular forOFFtran-sients, where the decay characteristic becomes linear when thecircuit reaches the open-loop regime, because it is then governedby the constant leakage current that discharges the differentiatornode to steady state.

The fitted values of the voltage decay constants as a functionof the relative irradiance step change are shown in Fig. 10(a).The decay time constants and are shown in Fig. 10(b)as a function of the relative irradiance step. They are the valuesof the time-axis intersections of the fits to and .

The peak responses of the , , and nodes asa function of the irradiance step size are plotted in Fig. 11(a).The response curve of is quite symmetric and so is the


curve with respect to the curve, but the symmetryaxis is displaced from the origin due to the asymmetry in theleakage currents. For small irradiance steps the peak values of

are almost linear in , as expected in open loop,while for larger steps the feedback loop is activated before theopen-loop peaks are reached. The response threshold offor ON transients is at 0.6% contrast and the threshold offor OFF transients is at 24.5% contrast. The onset delays of thedifferent signals with respect to the step transient as a functionof step size are plotted in Fig. 11(b). The minimum delays forlargeON steps are limited by the finite irradiance rise time.

The measured responses of the modified circuit are shown inFig. 12. The circuit exhibits more symmetric time delays forON

andOFFtransients, less ringing, and a larger gain than the basiccircuit. Saturation and adaptation effects can be observed in the

and traces [Fig. 12(a) and (b)] after high-contrastOFFsteps. The adaptation effects cause small spurious responsesin the ON pathway [Fig. 12(c)]. In order to suppress spuriousresponses it is, therefore, important that the adaptive gain stageis designed to work at significantly longer time scales than thedifferentiator stage.

The peak responses at the different nodes of the modified cir-cuit are shown in Fig. 13(a) and the onset delays of the signals inFig. 13(b). The response thresholds are now at contrasts of 6.2%for ON transients and 2.5% forOFFtransients. The response de-lays toON transients are slightly larger than for the basic circuit,while the delays of the signal toOFFtransients are reducedto about 30% of those of the basic circuit. Part of this improve-ment is due to the symmetrization of the leakage currents andthe other part to the activation of the cascode, which reduces theresponse delays by about 25% at the considered irradiance level.Increasing makes the circuit more sensitive toOFFtransientsbut less sensitive toON transients, i.e., the relativeON andOFF

thresholds and delays can be tuned with. The slope offor small steps is much larger than for the basic circuit and lim-ited by rather than by , as observed before.

D. Moving Contrast Grating

Fig. 14 shows the time courses of the different voltagesignals of the basic circuit in response to the image of thesquare-wave and of the sine-wave grating at an image speed of15 mm/s and the peak amplitudes of these signals in responseto the square-wave stimulus as a function of image speed. Thetransient slope of the irradiance is expected to be inverselyproportional to the transition time across the photodiode,i.e., proportional to the stimulus velocity. Hence, the peakamplitudes of the different signals should be linear with speedif the peaks occur in the open-loop regime and logarithmic withspeed if they occur in the closed-loop regime. The data shownin Fig. 14(c) suggests that the closed-loop approximation holdsover the entire speed range for the used stimulus. The responseof the modified circuit to the same stimulation patterns is shownin Fig. 15. The reduced peak response of to ON transientsat high speeds is due to incomplete adaptation.

VII. CONCLUSION

We have presented two different versions of an integratedsensor that shows contrast-encoding sustained and transient re-sponses to optical stimulation. The sensor is very compact and,thus, suitable for integration in medium-density focal-plane par-allel image processors. The circuit has been characterized withdifferent stimulation patterns. At the chosen irradiance level, themore sensitive version of the tested circuit implementations re-sponds to sharp positive and negative irradiance transients downto contrasts around 5%.

ACKNOWLEDGMENT

A version of the presented circuit was originally conceived ina discussion with R. Sarpeshkar. The author would like to thankA. Stocker for helpful discussions.

REFERENCES

[1] E. R. Fossum, “CMOS image sensors: Electronic camera-on-a-chip,”IEEE Trans. Electron Devices, vol. 44, pp. 1689–1698, Oct. 1997.

[2] T. Delbrück and C. A. Mead, “Analog VLSI adaptive logarithmic wide-dynamic-range photoreceptor,” inProc. IEEE Int. Symp. Circuits andSystems, London, U.K., May 1994, pp. 339–342.

[3] J. Kramer, R. Sarpeshkar, and C. Koch, “An analog VLSI velocitysensor,” inIEEE Int. Symp. Circuits and Systems, Seattle, WA, May1995, pp. 413–416.

[4] , “Pulse-based analog VLSI velocity sensors,”IEEE Trans. CircuitsSyst. II, vol. 44, pp. 86–101, Feb. 1997.

[5] M. A. Sivilotti, M. A. Mahowald, and C. A. Mead, “Real-time visualcomputations using analog CMOS processing arrays,” inProc. StanfordConf. on Advanced Research in VLSI, Stanford, CA, 1987, pp. 295–312.

[6] T. Delbrück and C. A. Mead, “An electronic photoreceptor sensitiveto small changes in intensity,” inAdvances in Neural Information Pro-cessing Systems, D. S. Touretzky, Ed. San Mateo, CA: Morgan Kauf-mann, 1989, vol. 1, pp. 720–727.

[7] M. A. Mahowald, “Silicon retina with adaptive photoreceptors,” inProc.Int. Soc. Optical Engineers—Visual Information Processing: From Neu-rons to Chips, vol. 1473, Apr. 1991, pp. 52–58.

[8] J. Mann, “Implementing early visual processing in analog VLSI: Lightadaptation,” inProc. Int. Soc. Optical Engineers—Visual InformationProcessing: From Neurons to Chips, Apr. 1991, vol. 1473, pp. 128–136.

[9] T. Delbrück and C. A. Mead, “Time-derivative adaptive silicon pho-toreceptor array,” inProc. Int. Soc. Optical Engineers—Infrared Sen-sors: Detectors, Electronics, and Signal Processing, 1991, vol. 1541,pp. 92–99.

[10] C. P. Chong, C. A. T. Salama, and K. C. Smith, “Image motion detectionusing analog VLSI,”IEEE J. Solid-State Circuits, vol. 27, pp. 93–96,Jan. 1992.

[11] R. Etienne-Cummings, S. Fernando, N. Takahashi, V. Shtonov, and J.Van der Spiegel, “A new temporal domain optical flow measurementtechnique for focal plane VLSI implementation,” inProc. Computer Ar-chitecture for Machine Perception, M. Bayoumi, L. Davis, and K. Vala-vanis, Eds., New Orleans, LA, 1993, pp. 241–250.

[12] O. Landolt, A. Mitros, and C. Koch, “Visual sensor with resolution en-hancement by mechanical vibrations,” inProc. 19th Conf. Advanced Re-search in VLSI, Salt Lake City, UT, Mar. 2001, pp. 249–264.

[13] G. Indiveri, A. M. Whatley, and J. Kramer, “Reconfigurable neuromor-phic VLSI multi-chip system applied to visual motion computation,” inProc. 7th Int. Conf. Microelectronics for Neural, Fuzzy, and Bio-InspiredSystems, Granada, Spain, Apr. 1999, pp. 37–44.

[14] G. Indiveri, R. Mürer, and J. Kramer, “Active vision using an analogVLSI model of selective attention,”IEEE Trans. Circuits Syst. II, vol.48, pp. 492–500, May 2001.

[15] S.-C. Liu, J. Kramer, G. Indiveri, T. Delbrück, T. Burg, and R. Douglas,“Orientation-selective aVLSI spiking neurons,”Neural Networks, vol.14, pp. 629–643, 2001, Special issue.


Jörg Kramer was born in Zurich, Switzerland.He received the degree in physics from the SwissFederal Institute of Technology, Zurich (ETHZ), in1987, with a diploma project on planar waveguides,the M.S. degree in applied optics from ImperialCollege of Science and Technology, London,U.K., in 1988, where he worked on photon-limitedimaging, and the Ph.D. degree in physics fromETHZ, in 1993, for a project carried out at the PaulScherrer Institute Zurich (PSIZ).

At PSIZ, he developed a real-time video 3-Dimaging system and designed optoelectronic devices for metrology applicationsusing CMOS technology. From 1993 to 1996, he was a Postdoctoral Fellowin Christof Koch’s group at the California Institute of Technology, Pasadena,where he designed analog VLSI velocity sensors and velocity-based sensingsystems for image segmentation and time to contact measurements. From 1996to 2002, he was a Staff Scientist at the Institute of Neuroinformatics, Zurich,where he worked on the design of early visual processing sensors and theirimplementation on mobile platforms for orientation and navigation purposesand also taught courses in theory and design of analog VLSI circuits.

Jörg was an avid and experienced solo hiker. In July 2002, he died of heatexhaustion in Death Valley after climbing from Badwater to Telescope Peak.He was 39. His eager collaboration, gentle humor, and quiet leadership will begreatly missed by his friends around the world.

Documents

An integrated optical transient sensor - Circuits and …...of biologically-inspired image-processing circuits that perform such functions as motion-sensing [13], attentional selection