IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 47, NO.

1, JANUARY 2000 55

Performance Analysis of a Color CMOS Photogate

Image Sensor
Andrew J. Blanksby, Member, IEEE, and Marc J. Loinaz, Member, IEEE

Abstract—The performance of a color CMOS photogate image these performance limitations is essential to guide the develop-
sensor is reported. It is shown that by using two levels of correlated- ment of the next generation of CMOS APS sensors. In this paper
double sampling it is possible to effectively cancel all fixed-pattern the performance of a color CMOS photogate active pixel sensor
noise due to read-out circuit mismatch. Instead the fixed-pattern
noise performance of the sensor is limited by dark current nonuni- is experimentally evaluated and compared with that of CMOS
formity at low signal levels, and conversion gain nonuniformity at photodiode APS imagers and CCD sensors. CMOS APS im-
high signal levels. It is further shown that the imaging performance ager performance limitations are identified and directions for
of the sensor is comparable to low-end CCD sensors but inferior to research needed to address these issues are presented.
that reported for high-end CCD sensors due to low quantum effi-
ciency, high dark current, and pixel cross-talk. As such the perfor-
mance of CMOS sensors is limited at the device level rather than II. SENSOR ARCHITECTURE AND OPERATION
at the architectural level. If the imaging performance issues can be
addressed at the fabrication process level without increasing cost A. Sensor and Column Circuit
or degrading transistor performance, CMOS has the potential to
seriously challenge CCD as the solid-state imaging technology of A 352 × 288 photogate sensor was fabricated in a Lucent
choice due to low power dissipation and compatibility with camera Technologies nonsilicided 0.8 m CMOS process with pixel di-
system integration. mensions of 16.0 m × 16.0 m [7]. The architecture of the
Index Terms—Active pixel sensor, CMOS image sensor, color. sensor is shown in Fig. 1 and its operation can be explained as
follows. During photocurrent integration the polysilicon photo-
gate is held at and photo-generated carriers are col-
I. INTRODUCTION lected beneath the gate oxide. The transfer device is bi-

C MOS active pixel sensor (APS) technology is emerging

as an alternative solid-state imaging technology to charge
coupled device (CCD) technology that has significant advan-
ased with at 0.7 V, isolating the collected charge under the
photogate from the floating diffusion node . During sensor
read-out each row is addressed in turn with decoding logic (not
tages in terms of low power dissipation, scaling to high reso- shown in Fig. 1) driving the appropriate row line to .
lution formats, and compatibility with camera system integra- The node of the pixels in the read-out row are reset to a
tion [1]–[4]. Traditionally the performance of passive CMOS voltage approximately one threshold voltage drop below
imagers has been limited by fixed-pattern noise (FPN) due to by pulsing high. The reset operation introduces uncer-
mismatch in the read-out circuits, and corruption of the signal tainty into the reset level due to the thermal noise and
by thermal noise during pixel read-out due to the large column threshold voltage of device . The reset level of the node
line capacitance. However, active pixel architectures are com- is buffered by the source follower formed by – and sam-
patible with correlated-double sampling (CDS) techniques that pled on the gate capacitance by pulsing to . Elec-
have proven effective in reducing FPN due to circuit mismatch trons collected under the photogate are then transferred to the
[5]. Furthermore, active pixel sensors can be designed to have node via device by holding at ground for an in-
extremely low read-out noise [6]. This has raised the possi- terval . The signal electrons displace the voltage on
bility of manufacturing high performance image sensors in stan- which is sampled on the gate capacitance by pulsing .
dard CMOS, an outcome that would substantially lower the The sampling of the pixel reset and signal levels by the column
technology barrier and allow many new players to enter the circuits is carried out in parallel for the pixels in the selected
solid-state imaging market. However, despite the intense in- row during the line blanking interval. To generate the image
terest in CMOS active pixel sensors in recent years, a detailed signal each column circuit is then addressed in turn by decoding
performance comparison between CMOS APS technology and logic (not shown in Fig. 1) driving to ground. The column
CCD technology has not been reported. Furthermore, the per- source followers – buffer the sampled reset and signal
formance limitations of the present generation of CMOS APS levels onto common buses and , respectively. To per-
imagers, namely the photogate and photodiode active pixel sen- form CDS the first differential amplifier subtracts the reset and
sors, have not been rigorously established. The identification of signal levels removing thermal noise and offset mismatch due to
and . This signal value is stored in a sample-and-hold
Manuscript received February 23, 1999; revised June 1, 1999. The review of stage by pulsing . During the second level of CDS, referred
this paper was arranged by Editor J. Hynecek. to as crowbar (CB) [5], the signal is driven high and the
The authors are with DSP & VLSI Systems Research Department, Bell offset difference between column followers is stored in a second
Laboratories, Lucent Technologies, Holmdel, NJ 07733 USA (e-mail:
[email protected]). sample-and-hold stage by pulsing . A final differential am-
Publisher Item Identifier S 0018-9383(00)00148-9. plifier delivers the image signal to a 12-bit analog-to-digital con-
0018–9383/00$10.00 © 2000 IEEE
56 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 47, NO. 1, JANUARY 2000

Fig. 1. CMOS photogate sensor architecture.

verter (not shown in Fig. 1) with column offsets removed. The

digitized pixel values are acquired by a PC with a framegrabber
for analysis and display purposes. The layout of the photogate
pixel is given in Fig. 2. Vertical buses in metal 1 are used for
and while horizontal buses in metal 2 are used for
the , , and signals. The voltage is supplied
by a horizontal bus in polysilicon. The large shaded area is the
polysilicon photogate which realizes a drawn pixel fill factor of
35%.

B. Color Filter Array

A color sensor was produced from the monochrome imager
by the deposition of a polyamide color filter array (CFA). The
CFA pattern used was the Bayer checkerboard pattern shown Fig. 2. CMOS photogate pixel layout.
below [8]:

The checkerboard pattern provides superior horizontal resolu-

tion compared to stripe CFA patterns at the expense of diagonal
resolution where the acuity of the eye is poor [9]. A CFA based
on primary colors was used rather than one based on comple-
mentary colors so as to simplify the subsequent stages of digital
color processing. The transmission characteristics of the color
filters are shown in Fig. 3. A color compensating filter was em-
ployed to suppress infra-red. Sensors were also produced with
microlenses to improve the effective pixel fill-factor.

C. Color Processing Fig. 3. Color filter transmission characteristics.

To generate 24-bit color images the digitized pixel values
were spatially interpolated over a 3 × 3 pixel neighborhood to gamma correction where the result was truncated to 8-bits per
using linear low-pass filter routines to generate three color com- color channel as shown in Fig. 4.
ponents for each pixel. A 3 × 3 matrix and gamma correction While the color processing was performed in software on the
were then applied to improve the color rendition and correct for PC, and the differential amplifiers and A/D-conversion were re-
the nonlinearity of the display cathode ray tube (CRT), respec- alized at the board level, this implementation was preferred for
tively [10]. The color processing was performed in software on experimental convenience rather than practical necessity. Un-
a PC using 16-bit integer arithmetic to preserve accuracy prior like CCD technology, CMOS active pixel sensor technology
BLANKSBY AND LOINAZ: COLOR CMOS PHOTOGATE IMAGE SENSOR 57

Fig. 4. Color processing architecture.

Fig. 5. Mean dark signal, pixel fixed-pattern noise (FPN), and dark current Fig. 6. Optical sensitivity of the monochrome sensor and color sensor with
shot noise as a function of exposure time at 25 C. microlenses as a function of faceplate illuminance for 30 ms exposure. Data
for the monochrome sensor is represented by diamonds while the red, green,
and blue pixels of the color sensor with microlenses are represented by circles,
is fully compatible with camera system integration providing squares, and triangles, respectively.
power and miniaturization advantages [3], [4].

III. SENSOR PERFORMANCE

To measure the performance of the photogate sensor statis-
tical definitions were developed for the mean signal level of
the sensor denoted µ, the RMS temporal noise , the RMS
pixel fixed-pattern noise , and the RMS column fixed-pat-
tern noise . At each experimental data point in Figs. 5– 7, 100
frames were acquired and the mean and standard deviation of
each pixel value, and ,were computed. The mean signal µ
level was calculated as the average of the pixel means ’s, and
the RMS temporal noise was computed as the square root of
the mean of pixel variances ’s. To quantify the sensor FPN a
mean image was formed from the ’s. Then the RMS pixel
fixed-pattern noise was measured as the square root of the av-
erage variance down the columns of , while the RMS column
fixed-pattern noise was quantified as the square root of the
average variance of the pixels along the rows of . By knowing
the gains on the board, the system signal and noise components
Fig. 7. Sensor pixel and column FPN as a function of mean signal level with
could be referred back to the sensor output. Furthermore, cali- and without the application of correlated-double sampling (CDS) and crowbar
bration procedures were used to deduce the contribution of noise (CB) for 30 ms exposure. The FPN floor imposed by dark current nonuniformity
sources associated with correlated-double sampling at the board is also shown.
level, and the digital acquisition system.
time was found with a slope of 82 mV/s at 25 C. This
A. Dark Current corresponds to a dark current density of 180 pA/cm ,
The sensor dark current parameters were measured by normalized with respect to the pixel area and using a con-
varying the integration time of the sensor while it remained version gain of 28 V/electron (see Section B). This result
unexposed to light and are plotted in Fig. 5. A linear re- is comparable to values reported for CMOS photogate and
lationship between mean dark signal and integration photodiode APS imagers in similar technology, but almost a
58 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 47, NO. 1, JANUARY 2000

factor of two times larger than values reported for CCD sen- undersized ( and ). To drive the off-chip fixed-pat-
sors without the use of charge pumping or surface pinning, tern noise correction circuits at 30 frames/s it is necessary to
and more than an order of magnitude larger than values re- supply large bias currents resulting in a loss in gain. The con-
ported for CCD sensors using such advanced dark current version gain referred to the pixel floating diffusion node
management techniques [11]–[13]. is 70 V/electron corresponding to a capacitance for the pixel
The dark current nonuniformity of the photogate sensor floating diffusion node of fF. is composed of
was determined by measuring the pixel FPN as a function the parasitic capacitances of the shared drain/source region of
of integration time as shown in Fig. 5. The pixel FPN devices and , together with the gate-drain overlap ca-
can be assumed to consist of independent components due pacitance of . As the bandwidth of the pixel need only satisfy
to dark current nonuniformity , and residual device the read-out line rate all of these devices can be of minimum size
mismatch not suppressed by CDS such as pixel to achieve a large value of conversion gain and to maximize the
gain mismatch according to: pixel fill-factor. In comparison, the bandwidth of a CCD output
stage must accommodate the full pixel rate and so the corre-
(1) sponding source follower device(s) cannot be of minimum size
and the value of conversion gain achievable in a CCD output
remains constant and linearly increases as stage is limited unless feedback is employed [21]. It is also dif-
a function of integration time. At zero integration time it can ficult to achieve high conversion gain in a CMOS photodiode
be assumed that is negligible and extrapolation of the imager but for a completely different reason. In a CMOS photo-
measured data on a linear scale gave V diode sensor the photodiode used for collecting photon gener-
RMS. This is small compared to for practical values ated carriers is also the charge-to-voltage conversion node. As
of integration time. For example, at 30 ms exposure is the area of the photodiode is maximized for photo-carrier col-
found to be 650 µV RMS and rearrangement of (1) gives lection the value of is much larger than that of the corre-
V RMS at 25 C. This demonstrates that CDS sponding photogate sensor yielding a smaller value of conver-
effectively suppresses pixel FPN caused by device mismatch sion gain. As CMOS technology scales for a photodiode
to an insignificant level relative to that due to dark current sensor will increase for a given pixel size due to higher doping
nonuniformity. However, at 30 ms exposure corresponds levels and an increase in pixel fill-factor. In contrast for a
to 23 electrons RMS, about an order of magnitude greater photogate sensor will become smaller as the feature dimensions
than values reported for CCD sensors that employ surface shrink giving a corresponding increase in conversion gain .
pinning [13]. Furthermore, recently published results demon-
strate that with each successive generation of CMOS tech- C. Optical Sensitivity
nology, the dark current performance of the corresponding To determine the optical sensitivity of the photogate sensor a
photogate or photodiode CMOS imager is substantially de- 3200 K tungsten-halogen lamp was used to provide a uniform
graded [14]–[16]. This suggests that to achieve FPN perfor- source of illumination and sensor exposure was controlled by
mance that is competitive with CCD sensors both now and in the use of neutral density filters. A color compensating filter
the future will require intervention at the fabrication process was employed to suppress infrared. The optical sensitivity of
level to improve cleanliness and/or realize pixel designs that the monochrome and color photogate sensors as a function of
support surface pinning [17], [18]. faceplate illuminance for 30 ms exposure are graphed in Fig. 6.
Straight lines have been fitted to the data points in a least squares
B. Conversion Gain sense over the linear portion of the response. The optical sen-
Exploiting the square root relationship between dark current sitivities of the monochrome and color photogate sensors are
shot noise and the mean dark signal (see Fig. 5) the conversion computed by dividing the slope of the mean signal curves by
gain of the sensor was found to be nominally 28 V/electron the exposure time and were found to be 4.25 V/lux s for the
at the chip output. This is significantly higher than values re- monochrome sensor, and 1.30 V/lux s, 1.12 V/lux s, and 0.66
ported for CMOS photodiode1 sensors and CCD output stages V/lux s for the red, green, and blue pixels of the color sensor,
which are typically of the order of 10 V/electron [12], [13]. respectively. The addition of microlenses increased the optical
High conversion gain is advantageous from a signal-to-noise sensitivity of the red, green, and blue pixels by 33, 27, and 15%,
ratio (SNR) perspective as it minimizes the contributions to the respectively.
sensor SNR from the thermal noise of the read-out circuits and To enable a fair comparison between the optical sensitivity
subsequent stages of signal processing (CDS, A/D conversion of the photogate sensor to values reported for other CMOS and
etc.) [20]. Circuit simulations give the combined gain of the CCD image sensors it is necessary to normalize results with
sensor pixel and column source followers to be 0.4 under respect to pixel area and sensor conversion gain. To this end an
the same bias conditions used to obtain the experimental data. optical sensitivity figure of merit was defined:
This relatively low value of gain is a consequence of the column
source follower current sources in the prototype sensor being (2)

1We refer here to the simple three-transistor photodiode pixel [19]. While
four-transistor photodiode pixels yield higher conversion gain, they suffer from where
significant image lag [15]. mean optical sensitivity in V/lux s;
BLANKSBY AND LOINAZ: COLOR CMOS PHOTOGATE IMAGE SENSOR 59

conversion gain in V/electron; The charge capacity is determined by , the doping pro-
pixel area in square microns. files, and the physical dimensions of the CCD stage, and is typ-
ically 60–70K electrons. Techniques such as surface pinning or
Using (2) for the monochrome photogate sensor is 590
charge pumping can be used to reduce such that its con-
electrons/lux s/µm2 while a value of 2400 electrons/lux s/µm2
tribution to (4) is negligible. Under these conditions the noise
can be computed from data reported for a frame-interline
floor is set by the read noise whose contribution to the
transfer (FIT) CCD under similar illumination conditions,
dynamic range is determined by . Hence maximizing the con-
and a value of 3400 electrons/lux s/µm2 can be calculated
version gain increases the sensor dynamic range until the
from data reported for a CMOS photodiode sensor under
output stage saturates.
unknown illumination conditions [12], [22]. The substantially
Unlike a CCD sensor the saturation level of a CMOS photo-
lower optical sensitivity of the photogate sensor results in
gate or photodiode sensor is determined by read-out circuit con-
poor SNR performance when compared to the photodiode and
siderations. The available signal swing at the pixel floating dif-
CCD sensors. As the temporal noise of a solid-state sensor
fusion node is limited by the threshold voltage drops across
is typically dominated by photon shot noise (see Fig. 6),
the reset device and the source follower device . The dy-
the SNR is proportional to the square root of the number of
namic range of a photogate sensor is approximately given by
electrons collected by each pixel [20]. Therefore the factor of
four difference in optical sensitivity between the photogate
sensor and that of the FIT-CCD sensor translates into a factor
of two (or 6 dB) difference in SNR performance. The optical V
(5)
sensitivity of the photogate sensor is fundamentally limited by
low quantum efficiency as will be discussed shortly.

D. Dynamic Range where

The temporal noise floor of the photogate sensor is V threshold voltage of an n-device;
composed of dark current shot noise , and the total read combined gain of the pixel and column source fol-
noise of the sensor that includes contributions from the lowers;
pixel, column, and CDS amplifier circuits (see Fig. 6). As these conversion gain of the pixel floating diffusion node.
components are independent they add in quadrature: In contrast to (4) the read noise components of the CMOS
sensor are effectively increased by the low combined gain
(3) of the source followers ( ) and the dark current
shot noise is amplified by the high in-pixel conversion gain
( V/electron).
Using calibration procedures the values for and
The dynamic range of a photodiode sensor is also given by (5)
were determined to be 260 µV RMS and 445 µV RMS,
but an additional term must be added in quadrature
respectively, for 30 ms exposure at 25 C. The temporal noise
to the denominator to account for the thermal noise introduced
floor of the sensor is therefore 525 µV RMS. The sensor
when the pixel is reset. For a photodiode sensor this uncertainty
saturates at a signal level of 1.38 V or 49 K electrons for
(expressed in electrons RMS) cannot be cancelled by
V and is limited by the read-out circuits rather than
CDS without the use of a frame memory. While and
the charge capacity of the potential well beneath the pixel
are typically the same for corresponding photogate and photo-
photogate. This gives a dynamic range for the sensor of 68 dB.
diode sensors, and are both smaller for the photo-
While this is similar to values reported for other CMOS and
diode sensor resulting in a larger dynamic range [15]. However,
CCD sensors [13], [15], the factors that determine the dynamic
as CMOS technology scales the dynamic range of photodiode
range of CMOS and CCD sensors are quite different. The
and photogate sensors will be limited by the available signal
dynamic range of a CCD sensor can be written in the form:
swing due to reduced voltages. While there is scope at the
circuit level to improve gain in the signal path [3], this will be
offset by the increase in dark current with technology scaling
(4) [14]–[16]. The effect of higher dark current will be exacerbated
in the photogate sensor by increased in-pixel conversion gain
. It is therefore likely that to maintain dynamic range per-
formance that is competitive with CCD sensors intervention at
where the process level will be required to reduce dark current, and
capacity of a CCD stage in electrons; in the case of the photogate limit at a reasonable value
through fabrication or layout techniques.
total RMS read noise voltage of the sensor output
stage; E. Charge Transfer Noise
conversion gain of the output stage, and
During the experiment used to obtain the sensitivity data of
dark current shot noise expressed in RMS elec- Fig. 6 a previously unreported temporal noise phenomenon as-
trons. sociated with charge transfer within the photogate pixel was ob-
60 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 47, NO. 1, JANUARY 2000

served. This component referred to here as charge transfer noise reported value of gain mismatch of 1% for a CMOS photogate
was found to dominate the temporal noise per- APS imager [11]. The cause of pixel gain mismatch may be at-
formance of the sensor at low signal levels with a magnitude tributed to photoresponse nonuniformity and/or conversion gain
of 1–2 mV RMS [23]. However, it was found that the value nonuniformity due to tolerances in the manufacturing process.
of was substantially reduced by increasing the However, as the dimensions of the pixel floating diffusion node
period of time that the pixel photogate node is held are much smaller than that of the photogate it is reasonable to
low during sensor read-out, and for ns the in- conclude that the dominant cause of pixel gain mismatch is con-
clusion of in (3) could be neglected. It is pro- version gain nonuniformity. An RMS variation of 0.36% in con-
posed that associated with image lag, a transfer loss noise sim- version gain corresponds to an RMS variation of fF in
ilar to that found in CCD shift registers is present [24]. The the pixel floating diffusion node capacitance .
structure of the single-poly photogate pixel is such that it is With the application of both CDS and CB the column FPN
subject to lag or incomplete charge transfer. This means that it and pixel FPN cannot be separated by measurement and are
takes a number of read-out cycles before the number of signal measured at the same level as the pixel FPN with one level of
electrons transferred to is equal to the number of CDS. This demonstrates that CB does not alter the pixel FPN
photon generated electrons collected by the level while effectively suppressing column FPN to the same
photogate each cycle. As a consequence the number of elec- level as the pixel FPN, i.e. 0.65 mV RMS (0.29% )
trons held under the photogate is actually greater than at low to moderate signal levels, and increasing linearly at 3.63
in the steady state. These “extra” electrons mV RMS/V for moderate to high signal levels.
that are not transferred introduce shot noise to . How- The results of this FPN investigation are significant because
ever, increasing allows a greater proportion of carriers col- they demonstrate that by applying two levels of CDS it is pos-
lected under the pixel photogate to diffuse to yielding a re- sible to suppress all FPN due to transistor threshold voltage mis-
duction in image lag and charge transfer noise. In CCD sensors match in the pixel and column circuits. Under these conditions
lag is greatly reduced at the device level using surface pinning the FPN performance becomes limited by dark current nonuni-
techniques [17]. formity at low signal levels, and conversion gain nonuniformity
at high signal levels. As it is possible to employ two levels of
F. Fixed-Pattern Noise CDS with a CMOS photodiode APS imager using the read-out
The fixed-pattern noise (FPN) performance of the photogate architecture of Fig. 1, the FPN performance of photodiode APS
sensor was determined by computing the RMS pixel fixed-pat- imagers is also limited by dark current nonuniformity at low
tern noise and RMS column fixed-pattern noise as a signal levels, and conversion gain nonuniformity at high signal
function of illuminance during the optical sensitivity measure- levels.
ment. The effectiveness of correlated-double sampling (CDS)
as a technique for managing FPN was assessed by measuring G. Quantum Efficiency
the pixel and column FPN without CDS, then using one level The absolute quantum efficiencies of the monochrome and
of CDS, and finally two levels of CDS, also known as crowbar color photogate sensors were measured using a monochromator
(CB) [5]. The combined FPN data for 30 ms exposure at 25 C and are shown in Fig. 8. A color compensating filter was used
is shown in Fig. 7 plotted as a function of mean signal level. when determining the quantum efficiency of the color sensor.
Without the use of CDS the sensor FPN is dominated by pixel The quantum efficiency of the monochrome sensor is similar
offset mismatch due to threshold voltage variations in the pixel to that reported for other CMOS photogate sensors [11], but
reset device , and source follower devices . Under these about a factor of three or four lower than that found for CMOS
conditions the pixel and column FPN is 15.4 mV RMS or 6.7% photodiode sensors and interline-transfer CCD sensors [19].
. When compared to a frame-transfer CCD the peak quantum
With the application of the first level of CDS, pixel offset efficiency of the photogate sensor is approximately 70% lower,
mismatch due to threshold voltage variations is cancelled. In- and the quantum efficiency at short wavelengths is almost five
stead column FPN due to offset mismatch between the column times lower [13]. The low quantum efficiency of the photogate
source followers formed by devices – becomes the sensor, particularly at short wavelengths, is largely due to
dominant source of FPN at 4.1 mV RMS or 1.8% . absorption in the polysilicon photogate and the lower fill-factor
At low to moderate signal levels the pixel FPN is suppressed to achieved by the photogate pixel. By way of comparison the
the noise floor imposed by dark current nonuniformity at 0.65 polysilicon of a standard submicron CMOS process is 3–10
mV RMS or 0.29% . As stated in Section A, while times thicker than that used in a frame-transfer CCD. The
this result is similar to values reported for CMOS APS sensors quantum efficiency of the photogate architecture will improve
in comparable technology, it is more than an order of magni- as the feature dimensions of CMOS technology continue to
tude larger than values reported for CCD sensors using surface shrink and a corresponding increase in pixel fill-factor can
pinning [13]. At moderate to high signal levels with one level be realized. However, to substantially enhance the quantum
of CDS the pixel FPN increases as function of the mean signal efficiency of the photogate sensor would require decreasing
due to gain mismatch between pixels. If these data points are the thickness of the polysilicon photogate as a modification
replotted on a linear scale they form a straight line with a slope to the CMOS fabrication process. Furthermore, it is common
of 3.63 mV RMS/V corresponding to an RMS pixel gain mis- practice in sub-micron CMOS technology to form silicide on
match of 0.36% [23]. This compares favorably with a previously the polysilicon gate and/or source/drain regions of devices.
BLANKSBY AND LOINAZ: COLOR CMOS PHOTOGATE IMAGE SENSOR 61

controlling the doping concentration of the substrate to limit the

diffusion length of minority carriers, and optimizing the pixel
structure such as providing channel stops between pixels to ab-
sorb free carriers [20]. However, neither of these options is avail-
able in standard CMOS technology for image sensor design. As
a consequence the cross-talk of the photogate sensor is signifi-
cantly larger than that reported for CCD sensors [27], [28].

IV. SENSOR COLORIMETRIC PERFORMANCE

A. Colorimetric Accuracy
To improve the color rendition of solid-state imagers it is nec-
essary to perform white balance and color correction [10]. White
balance is the operation whereby separate gains , , and
are applied to the red, green, and blue pixel values respectively
so that a reference white in the scene gives equal response in
the red, green, and blue pixels. Color correction is then required
Fig. 8. Quantum efficiency of the monochrome sensor and color photogate
sensor with microlenses. The data has not been divided by the pixel fill-factor. to transform from the native color space of the sensor to the
color space of the display system (e.g., the RGB primaries de-
fined by the NTSC standard). A 3 × 3 color correction matrix
and offsets were applied to the white
Silicide improves the electrical properties of transistors but is balanced and interpolated color components for each pixel of
almost opaque to light in the visible spectrum [16]. While it the photogate sensor to convert them into
was possible to block silicide formation in the manufacture output values according to:
of the photogate sensor reported here, it will be necessary for
the manufacture of CMOS image sensors in future sub-micron
technology that provisions exist to block the formation of
silicide on the photo-sensing devices. However, as this option (6)
is normally provided in analog CMOS technology to support
polysilicon resistors this does not represent a major “nonstan-
dard” requirement. The value of the coefficients and offsets
were determined by numerical
H. Pixel Cross-Talk optimization using the 24 color samples of the Macbeth
ColorChecker chart under D6500K illumination. The samples
Pixel cross-talk is the phenomenon whereby photons incident of the Macbeth chart were measured using a colorimeter and
on one pixel generate carriers that are collected by a different a gradient descent algorithm was used to find the that
pixel [20]. The depth that a photon will penetrate a silicon sub- minimizes the error between the samples and the output of the
strate before generating an electron-hole pair is strongly wave- sensor with matrix applied. The optimization was performed
length dependent, the longer the wavelength the deeper the pen- in the CIE color space to provide a more equal measure
etration depth [25]. This means that many of the carriers pro- of perceived color error across the gamut of colors represented
duced by long wavelength photons are generated below the pixel by the Macbeth chart [29]. The RMS color difference
structure and diffuse to a nearby pixel depletion region to be was used as the objective function as defined by:
collected. As there is no predefined direction for the diffusion
path it is possible for carriers to be collected by pixels other
than those under which they were generated. Pixel cross-talk
degrades image sharpness and results in color mixing that re-
duces colorimetric accuracy [26]. By inspection of Fig. 8 it is
possible to identify significant cross-talk from the red pixels into (7)
the green and blue pixels, in addition to cross-talk from the green where is the measured tristimulus value for sample
pixels into the blue pixels. It was also found that green pixels in of the Macbeth ColorChecker in the CIE color space,
the same rows as the red pixels were subject to greater cross-talk and is the tristimulus value obtained for sample
than those in the same rows as the blue pixels. This is consistent from the sensor after application of (6) and transformation into
with the physical layout of the photogate pixel shown in Fig. 2 the CIE color space. The optimal matrix found by
in which the diffusion length required for cross-talk in the hori- this procedure was
zontal direction is much shorter than that required in the vertical
direction. The corresponding quantum efficiencies in Fig. 8 have
been labeled “green 1” and “green 2,” respectively. Cross-talk (8)
in CCD sensors is managed at the fabrication process level by
62 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 47, NO. 1, JANUARY 2000

(a) (b)

Fig. 9. (a) The ideal NTSC color analysis functions for a D6500K white point and (b) the color analysis functions of the color photogate sensor after color
correction.

with offsets region of the spectrum. This allows a substantial percentage of

normalized to 255 (8-bits). The RMS color difference error the blue response of the sensor to be contributed by cross-talk
over the 24 color samples of the Macbeth chart for rather than carriers generated by short wavelength photons. Fur-
this matrix is 7.7 units RMS. A similar study performed with thermore, the low quantum efficiency at short wavelengths shifts
a CCD-based camera using the same color samples and error the peak of the blue response. In Fig. 3 the peak of the blue
metric reported a value for of 4.8 units RMS [30]. As color filter characteristic is at 450 nm, as required by the ideal
a color difference of 1 unit in the CIE color space blue NTSC color analysis function. However, the shape of the
corresponds to a “just noticeable difference,” the colorimetric quantum efficiency of the monochrome photogate sensor shown
accuracy of the CMOS photogate sensor is slightly poorer than in Fig. 8 shifts this peak to 510 nm. The color correction ma-
that of the CCD sensor. This is due to the CCD sensor having trix is only able to move this peak back to 480 nm with a cor-
superior quantum efficiency in the blue portion of the spectrum responding loss of colorimetric accuracy. The combined effect
and minimal color cross-talk. of cross-talk and low quantum efficiency is that the color sepa-
ration achieved by the photogate sensor is poorer than that re-
B. Color Analysis Functions ported for CCD sensors [27], [28].
Fundamentally the colorimetric accuracy of any solid-state
image sensor is limited by the shape of its color analysis func- C. Signal-to-Noise Degradation
tions. Ideally the shape of the sensor color analysis functions While the application of the color correction matrix
after color correction should be the same as the ideal NTSC provides a large improvement in the color rendition of the
color analysis functions shown in Fig. 9(a) [10]. The color anal- photogate sensor, it has an unwelcome impact on the noise
ysis functions of the photogate sensor are found by normalizing performance. As the temporal and fixed-pattern noise in the
the quantum efficiencies of the red, green, and blue pixels to components are independent, their noise
give white balance for a D6500K illuminant, and then applying variances add and are weighted by the coefficients of to give
the color correction matrix given in (8). The resulting color the noise in the color corrected components [9].
analysis functions are shown in Fig. 9(b) and it is apparent that As many of the off-diagonal coefficients of are comparable
substantial differences exist when compared to the ideal func- to the on-diagonal terms, the SNR of the sensor after color
tions. The most important of these concerns the blue color anal- correction is significantly degraded. For example, at 1 lux
ysis function with its peak at 480 nm rather than 440 nm and faceplate illumination and 30 ms exposure the temporal SNR
its excessive bandwidth and negative excursion. The nonideal of the red, green, and blue channels after interpolation are 33,
shape of the color analysis functions is due to substantial color 32, and 29 dB, respectively. With the application of the color
cross-talk, particularly for the blue pixels which have signifi- correction matrix the corresponding SNR’s become 29, 22, and
cant response for wavelengths 500–750 nm (see Fig. 8). As a 19 dB. Using an f/2.8 lens visible noise is observed in images
consequence the color correction matrix has large negative co- acquired with 30 ms exposure at a room illumination of 150
efficients , and to try to cancel cross-talk from the lux. Such SNR performance limitations can only be overcome
red and green pixels. The effect of cross-talk is exacerbated by with an increase in sensor quantum efficiency, particularly at
the poor quantum efficiency of the photogate sensor in the blue short wavelengths, and a reduction of pixel cross-talk.
BLANKSBY AND LOINAZ: COLOR CMOS PHOTOGATE IMAGE SENSOR 63

TABLE I sonable cost, then CMOS imaging technology will replace CCD
SUMMARY OF PHOTOGATE SENSOR CHARACTERISTICS for a wide range of applications due to its low power dissipation
and compatibility with camera system integration.

ACKNOWLEDGMENT
The authors would like to thank S. Mendis for the sensor de-
sign, D. Inglis for managing the color filter array deposition, D.
Gibbon and R. Kollarits for sharing their knowledge of color, I.
Fujimori for the quantum efficiency data, and B. Ackland and
A. Bouzerdoum for encouraging this work.

REFERENCES
[1] E. Fossum, “Active pixel image sensors—Are CCD’s dinosaurs?,” Proc.
SPIE, vol. 1900, pp. 2–14, Feb. 1993.
[2] B. Ackland and A. Dickinson, “Camera on a chip,” in ISSCC Dig., Feb.
1996, pp. 22–25.
[3] M. Loinaz, K. Singh, A. Blanksby, D. Inglis, K. Azadet, and B. Ackland,
“A 200-mW 3.3 V CMOS color camera IC producing 352 × 288 24 b
video at 30 frames/s,” in ISSCC Dig., Feb. 1998, pp. 168–169.
[4] S. Smith, J. Hurwitz, M. Torrie, D. Baxter, A. Holmes, M. Panaghiston,
R. Henderson, A. Murray, S. Anderson, and P. Denyer, “A single-chip
D. Subjective Comparison 306 × 244-pixel CMOS NTSC video camera,” in ISSCC Dig., Feb. 1998,
pp. 170–171.
In this paper various performance parameters of the photo- [5] S. Mendis, S. Kemeny, R. Gee, B. Pain, Q. Kim, and E. Fossum,
gate sensor have been found and compared with reported values “Progress in CMOS active pixel sensors,” Proc. SPIE, vol. 2172, pp.
for other CMOS and CCD sensors. However, it should be noted 19–29, 1994.
[6] M. Loinaz, K. Singh, A. Blanksby, D. Inglis, K. Azadet, and B. Ack-
that by virtue of their publication most of these sensors are or land, “A 200-mW 3.3 V CMOS color camera IC producing 352 × 288
were state-of-the-art. To gain greater perspective on the per- 24 b video at 30 frames/s,” IEEE J. Solid-State Circuits, vol. 33, pp.
formance of the photogate sensor it was subjectively compared 2092–2103, Dec. 1998.
[7] A. Blanksby, M. Loinaz, D. Inglis, and B. Ackland, “Noise performance
with a low-end color CCD-based multi-media camera under flu- of a color CMOS photogate image sensor,” in IEDM Tech. Dig., Dec.
orescent room lighting. While both sensor performed poorly at 1997, pp. 205–208.
low-light levels it was found that under normal light levels the [8] B. E. Bayer, “Color Imaging Array,” U.S. Patent 3 971 065, July 1976.
[9] K. A. Parulski, “Color filters and processing alternatives for one-chip
photogate sensor provided a cleaner image with superior color cameras,” IEEE Trans. Electron Devices, vol. ED-32, pp. 1381–1389,
rendition. This suggests that the performance of the photogate Aug. 1985.
sensor is comparable or better than that of low-end CCD sen- [10] W. N. Sproson, Color Science in Television and Display Sys-
tems. Bristol: Adam Hilger Ltd., 1983.
sors. [11] R. Nixon, S. Kemeny, B. Pain, C. Staller, and E. Fossum, “A 256 × 256
CMOS active pixel sensor camera-on-a-chip,” in ISSCC Dig., Feb. 1996,
pp. 178–179.
V. CONCLUSION [12] J. Hurwitz, P. Denyer, D. Baxter, and G. Townsend, “An 800 K-pixel
The performance of the CMOS photogate sensor is summa- color CMOS sensor for consumer still cameras,” Proc. SPIE, vol. 3019,
pp. 115–124, 1997.
rized in Table I. The results of this study demonstrate that the [13] J. Bosiers, E. Roks, H. Peek, A. Kleinmann, and A. van der Sijde, “An
imaging performance of CMOS active pixel sensors employing S-VHS compatible 1/3″ color FT-CCD imager with low dark current by
two levels of correlated-double sampling is primarily limited at surface pinning,” IEEE Trans. Electron Devices, vol. 42, pp. 1449–1460,
Aug. 1995.
the device level and not at the architectural or circuit level. While [14] B. Mansoorian, G. Yang, R. Panicacci, C. Wrigley, C. Staller, B.
the image quality delivered by CMOS sensors is comparable to Pain, and E. Fossum, “Megapixel CMOS APS with analog and digital
that of low-end CCD sensors, at present imaging performance readout,” in Proc. IEEE Workshop Charge-Coupled Devices and
Advanced Image Sensors, Bruges, Belgium, June 5–7, 1997.
is inferior to that reported for high-end CCD sensors due to ex- [15] S. Mendis, A. Budrys, J. Lin, and K. Cham, “Active pixel image sensors
cessive dark current nonuniformity, conversion gain nonunifor- in 0.35 µm CMOS technology,” Proc. IEEE Workshop Charge-Coupled
mity, image lag, pixel cross-talk, and low quantum efficiency. Devices and Advanced Image Sensors, June 5–7, 1997.
[16] H. Wong, R. Chang, E. Crabbe, and P. Agnello, “CMOS active pixel
Furthermore, as CMOS technology scales the imaging perfor- image sensors fabricated using a 1.8-V, 0.25 µm CMOS technology,”
mance gap between CMOS and CCD sensors will grow due to IEEE Trans. Electron Devices, vol. 45, pp. 889–894, Apr. 1998.
increased dark current and conversion gain nonuniformity, and [17] N. Teranishi, A. Kohono, Y. Ishihara, E. Oda, and K. Arai, “No image
lag photodiode structure in the interline CCD image sensor,” in IEDM
reduced supply voltages. To support image sensor design it will Tech. Dig., Dec. 1982, pp. 324–327.
therefore be necessary to introduce modifications to standard [18] R. Guidash, T. Lee, P. Lee, D. Sackett, C. Drowley, M. Swenson, L.
CMOS fabrication processing to address these issues. In partic- Arbaugh, R. Hollstein, F. Sharpiro, and S. Domer, “A 0.6 µm CMOS
pinned photodiode color imager technology,” in IEDM Tech. Dig., Dec.
ular it can be expected that techniques employed in CCD sensors 1997, pp. 927–929.
such as surface pinning and the use of channel stops will be de- [19] R. Nixon, S. Kemeny, C. Staller, and E. Fossum, “128 × 128 CMOS
veloped as additional CMOS process modules [18]. If combined photodiode-type active pixel sensor with on-chip timing, control, and
signal chain electronics,” Proc. SPIE, vol. 2415, pp. 1–8, 1995.
with improved process cleanliness these techniques are able to [20] A. Theuwissen, Solid-State Imaging with Charge-Coupled De-
close the gap between CMOS and CCD imaging quality at a rea- vices. Dordrecht, The Netherlands: Kluwer, 1995.
64 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 47, NO. 1, JANUARY 2000

[21] H. Akimoto, H. Ando, H. Nakagawa, Y. Nakahara, M. Hikiba, and H. Andrew J. Blanksby (M’98) was born in Adelaide,
Ohta, “A 1/3-in 410 000-pixel CCD image sensor with feedback field- South Australia, on February 6, 1972. He received the
plate amplifier,” IEEE J. Solid-State Circuits, vol. 26, pp. 1907–1914, B.S. degree in electrical and electronic engineering
Dec. 1991. (with First Class Honors) from the University of Ade-
[22] M. Morimoto, K. Orihara, N. Mutoh, A. Toyoda, M. Ohbo, Y. laide in 1993, and submitted his Ph.D. thesis in June
Kawakami, T. Nakano, K. Chiba, S. Kawai, K. Hatano, K. Arai, M. 1998.
Nishimura, Y. Nakshiba, A. Kohno, I. Akiyama, N. Teranishi, and Y. Since July 1998, he has been Member of Technical
Hokari, “A 1-inch 2-M pixel HDTV CCD image sensor with tungsten Staff in the DSP & VLSI Systems Research De-
photo-shield and H-CCD shunt wiring,” IEEE Trans. Electron Devices, partment at Bell Laboratories, Lucent Technologies,
vol. 42, pp. 50–57, Jan. 1995. Holmdel, NJ. His interests include VLSI design,
[23] A. Blanksby, “Color cameras in standard CMOS,” Ph.D. dissertation, CMOS image sensors, channel equalization, trellis
Univ. Adelaide, South Australia, July 1998. coding, and surfing.
[24] J. Carnes and W. Kosonocky, “Noise sources in charge-coupled
devices,” RCA Rev., vol. 33, pp. 327–343, Jun. 1972.
[25] J. Lavine, E. Trabka, B. Burkey, T. Tredwell, E. Nelson, and C. Anag-
nostopoulosm, “Steady-state photocarrier collection in silicon imaging Marc J. Loinaz (M’95) was born in Manila, The
devices,” IEEE Trans. Electron Devices, vol. ED-30, pp. 1123–1134, Philippines, on August 20, 1967. He received the
Sept. 1983. B.S. degree in electrical engineering from the Uni-
[26] K. Engelhardt and P. Seitz, “Optimum color filters for CCD digital cam- versity of Pennsylvania, Philadelphia, in 1988 and
eras,” Appl. Opt., vol. 32, no. 16, pp. 3015–3023, June 1993. the M.S. and Ph.D. degrees in electrical engineering
[27] Y. Naito, A. Kobayashi, T. Ishigami, S. Nakagawa, Y. Shimohida, A. from Stanford University, Stanford, CA, in 1990 and
Izumi, H. Endo, H. Mizoguchi, Y. Hirotani, and S. Horiuchi, “A 1/3-inch 1995, respectively.
360 k pixel progressive scan CCD image sensor,” IEEE Trans. Consum. Since 1995, he has been Member of Technical
Electron., vol. 41, pp. 442–447, Aug. 1995. Staff in the DSP & VLSI Systems Research De-
[28] O. Nishima, N. Kato, T. Higuchi, Y. Wataya, M. Oohashi, E. partment at Bell Laboratories, Lucent Technologies,
Machishima, J. Kuroiwa, and K. Kusano, “A 1/4-inch 380 k-pixel Holmdel, NJ. His research interests are in the area
IT-CCD image sensor,” IEEE Trans. Consum. Electron., vol. 41, pp. of high-performance analog and digital circuit design with recent emphasis on
430–435, Aug. 1995. CMOS image sensors, sensor interfaces, high-speed clock and data recovery,
[29] G. Wyszecki and W. Stiles, Color Science: Concepts and Methods, and mixed-signal integration issues.
Quantitative Data and Formulae, 2nd ed. New York: Wiley, 1982. Dr. Loinaz was a recipient of the E. Stuart Eichert Memorial Prize from the
[30] R. Kollarits and D. Gibbon, “Improving the color fidelity of cameras for University of Pennsylvania in 1987, and is a member of Tau Beta Pi and Eta
advanced television systems,” Bell Labs, Tech. Memo., May 1, 1990. Kappa Nu.