1108 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 58, NO. 4, APRIL 2011
Linear–Logarithmic CMOS Pixel With
Tunable Dynamic Range
Monica Vatteroni, Pietro Valdastri, Member, IEEE, Alvise Sartori, Member, IEEE,
Arianna Menciassi, Member, IEEE, and Paolo Dario, Fellow, IEEE
Abstract—A CMOS pixel with linear–logarithmic response and
programmable dynamic range (DR), based on a tunable transition
point, has purposely been designed for endoscopic applications.
A theoretical model of the pixel was developed and validated.
A chip with a 100 × 100 pixel array and a 12-b digital output
was fabricated in a 0.35-μm technology and was fully tested, thus
demonstrating state-of-the-art performance in terms of DR and
noise. Intraframe DR proved to be extendable to more than 110 dB
through a logarithmic compression of the signal in the light ir-
radiation power density (LIPD) range. The measured temporal
noise (pixel noise) was less than 0.22% over the full range. The
architecture presented limited fixed pattern noise (FPN) due to
the scheme adopted, which allowed its correction over the full
signal range: FPN was 0.83% (1.37%) in the linear (logarithmic)
region. Although the test chip was designed mainly for endoscopic
applications, the technology may also be applied to other fields,
e.g., robotics, security and industrial automation, whenever high
DR is a crucial feature.
Index Terms—CMOS imager, endoscopy, logarithmic response,
pixel.
I. INTRODUCTION
S INCE 1879 [1], vision systems have widely been used inbiomedical applications, mainly for endoscopic inspection
and visualization enhancement in surgery. Despite this condi-
tion, the first fiber endoscope was developed more than 70 years
later when Hopkins and Kapani published the use of a gastro-
scope based on coherent glass fibers in 1954 [2]. One additional
milestone was reached in this field with the introduction of
television, which allowed binocular vision from a convenient
distance by several observers. In 1988, a digital vision system
that was placed on the distal part of the instrument eliminated
the need for optical fibers that run through the entire shaft [1].
Nowadays, as a consequence of the wide diffusion of minimally
invasive diagnostic and surgical techniques, endoscopic vision
systems have dramatically strengthened their role in the surgical
field [3], [4].
Manuscript received June 21, 2010; accepted December 26, 2010. Date of
publication February 14, 2011; date of current version March 23, 2011. This
work was supported in part by the European Commission through the frame-
work of VECTOR FP6 European project EU/IST-2006-033970 (www.vector-
project.com). The review of this paper was arranged by Editor J. R. Tower.
M. Vatteroni, P. Valdastri, A. Menciassi, and P. Dario are with the
BioRobotics Institute, Scuola Superiore Sant’Anna, 56025 Pisa, Italy
(e-mail: m.vatteroni@sssup.it; p.valdastri@sssup.it; a.menciassi@sssup.it;
p.dario@sssup.it).
A. Sartori is with NeuriCam SpA, 38100 Trento, Italy (e-mail: sartori@
neuricam.com).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TED.2011.2106787
Minimally invasive surgery allows surgeons to perform pro-
cedures through small incisions, thereby offering several ben-
efits to the patient, e.g., reduced trauma and faster recovery,
compared to traditional open techniques. Since its introduction,
surgical endoscopy has played a major role in improving the
safety, precision, and reliability of medical interventions. The
most desirable features of an endoscopic vision system are
high image resolution, low noise, bright illumination, and low
working temperature. Saturated regions must particularly be
prevented, because they limit and, sometimes, hamper image
comprehension.
The disruptive progress that was achieved by the CMOS im-
ager industry in the last few years, mainly driven by consumer
electronics, has allowed all these requirements to simultane-
ously be met [5]. This case is particularly true for saturation,
which can be prevented by using high-dynamic-range (HDR)
image sensors [6].
HDR image sensors allow a wide range of light irradiation
power density (LIPD) to be mapped in the same picture. Tech-
niques for obtaining HDR have widely been investigated and
can be classified as interframe and intraframe techniques. In the
former case, images are combined together to obtain a single
HDR image [7]–[10], whereas in the latter case, the whole light
dynamic is mapped in the same image.
With regard to the interframe approach, the multi-integration
technique is currently the most commonly used approach at
the industrial level, because it is based on a simple concept. In
particular, this approach is easily implementable with a CMOS
sensor.
Among existing intraframe techniques, the simplest ap-
proach is the logarithmic technique [11], [12]. The array is con-
tinuously read out and the photodetector output is compressed
with a dynamic metal–oxide–semiconductor field-effect tran-
sistor (MOSFET) load. This case results in a logarithmic
relationship between the LIPD input and the voltage output
covering more than six decades of LIPD. The main advantages
of this technique are its simple architecture and HDR. On
the other hand, poor response at low light intensity and high
residual fixed pattern noise (FPN) are the main drawbacks.
One alternative and more recent technique for obtaining
intraframe HDR is the time-to-saturation approach [13]. In this
case, the output is obtained by combining a standard linear out-
put and a signal related to the time required by the photodiode
to saturate. The DR can be extended to more than 120 dB,
depending on the function used to map the saturation time into
the output voltage. A detailed reference list with a comparative
theoretical study of main HDR architectures is reported in [14].
0018-9383/$26.00 © 2011 IEEE
VATTERONI et al.: LINEAR–LOGARITHMIC CMOS PIXEL WITH TUNABLE DYNAMIC RANGE 1109
Focusing on endoscopic applications, the main requirements
for an HDR vision system are chip size and image quality. Con-
sequently, the cited techniques could hardly be implemented
in endoscopy. In fact, the multi-integration technique is based
on the acquisition of the same image, with different integration
times, to obtain an HDR image from their superposition. This
case is not always guaranteed during endoscopic inspections
because of possible movements (e.g., peristalsis, respiration,
and tool motion) during image acquisition. With regard to
intraframe techniques, the time-to-saturation pixel guarantees
good image quality, at the cost of a relevant pixel pitch. There-
fore, high resolution cannot be achieved, unless a large silicon
area is used. The logarithmic pixel is an intraframe approach
that can be implemented in a small pitch; however, the response
at low LIPD is poor. This condition is a crucial drawback in the
event of a light- and temperature-controlled environment, e.g.,
endoscopy.
Based on these considerations, a novel image sensor with
linear–logarithmic response was developed by merging the
simple design and the HDR of the logarithmic technique in
case of high LIPD with a linear response at low intensity
light. This case is not the first time that a linear–logarithmic
pixel is presented in the literature [15]. The main innovative
feature of the proposed design consists of the integration of
a hard reset structure that gives rise to an FPN subtraction in
hardware—completely when the response is in the linear region
and partially when the behavior is logarithmic. Moreover, the
power responsivity curve can be adjusted through a tunable
analog reference to dynamically achieve the desired tradeoff
between HDR and high pixel resolution [16], [17].
To preliminarily evaluate the performance of this new pixel
architecture, one model of the pixel response was developed
and is presented in Section II. The new technique was im-
plemented in a 100 × 100 pixel image sensor, fabricated
in a standard 0.35-μm, 3.3-V CMOS technology. Details are
given in Section III. Section IV gives concrete information on
performance by illustrating the results of the electro-optical
tests that were carried out and by comparing them to the
proposed model. This section also describes the performance
of endoscopic image acquisition through ex vivo experiments.
II. PIXEL ARCHITECTURE AND MODEL
The linear–logarithmic pixel that was integrated in the pro-
posed chip is shown in Fig. 1. The core is the same as that
of a linear pixel. The photodiode (D1), a first MOSFET that
works as a reset switch (M1), and a second MOSFET (M5), in
source–follower configuration, shield the photosensitive node
from the load of the readout chain. A global shutter (SH), which
consists of a transistor (M4) that was driven by an external
signal, completes the basic block. Note that, during preliminary
testing, the sensitivity of the pixel was relevantly affected by the
switching of the M4 transistor. Therefore, it was excluded from
the signal chain by leaving it always switched on and will not
be considered in the following discussion. HDR functionality
is guaranteed by a series of transistors (M2 and M3), with the
gate connected to the drain in a diode configuration. Placing a
second transistor in series to the first one doubles the gain in
Fig. 1. Schematic of the linear–logarithmic pixel.
the logarithmic region, thus enhancing the signal-to-noise ratio
(SNR) [18].
This part of the circuitry works as a voltage-light dependent
active load, and is responsible for the logarithmic compression.
Such a load is active only when the voltage VPH at the photo-
sensitive node (PH) exceeds a threshold which can be tuned by
the operator through an external reference voltage VLOG.
The operation of the pixel starts with a reset status. During
this phase, the reset transistor M1 is switched on, and the
photosensitive node (PH) is pulled to the reference voltage
VRES, which can be adjusted by the user. The following in-
tegration phase begins with the release of PH by switching
M1 off. PH is now isolated. and its voltage (VPH) starts to
decrease as a consequence of the charge carriers drained by the
photogenerated current. The higher the photosensitive current
is, the faster the discharge will be. If the combination of light
intensity and integration time is not enough to allow the pixel
to reach the threshold at which the active load starts to work
(i.e., VLOG − 2∗VTH, where VTH is the transistor threshold
voltage, then the signal VPH at the photosensitive node will be
proportional to the photogenerated current (iPH), increased by
the dark current (idark), and to the integration time (tINT) as
follows [6]:
VPH =
tINT(iPH + idark)
CIN
VPH > VLOG − 2∗VTH (1)
where CIN is the capacitance at the photosensitive node, which
is made up of the intrinsic photodiode capacitance and the
parasitic one.
If VPH reaches the threshold that was tuned by VLOG, then
the active load starts to work by draining part of the photogen-
erated current. The result is a logarithmic relation between iPH
and VPH.
This behavior is obtained due to the subthreshold working
condition of the M2 and M3 load transistors, guaranteed by
the shortcut between the gate and the drain (VGS = VDS).
Indeed, for VDS < VTH, the transistor should be switched off.
However, from an analog standpoint, the transistor is in weak
inversion, and a small current of minority carriers (iD) flows in
the MOSFET channel.
1110 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 58, NO. 4, APRIL 2011
The equation for iDS is [19]
iDS = Dn
kT√
2qNAΨS
εSi
(
W
L
)(
n2i
NA
)
e
qΨS
kT
(
1− e− qVDSkT
)
(2)
where ni and NA are, respectively, the intrinsic and the donor
doping concentration, Ψs is the surface potential, Si is the
silicon dielectric constant, W and L are the MOSFET width
and length, and Dn is the electron diffusivity. At room tem-
perature, VDS > 3kT/q ∼ 75 mV, and the last term of this
equation is around 1. Furthermore, VDS = VGS = nΨS , and all
the constant parts of the formula can be collected in one single
term as
i0 = Dn
kT√
2qNAΨS
εSi
(
W
L
)(
n2i
NA
)
(3)
where Dn = (kT/q)μn, and μn is the electron mobility.
Therefore, (2) can be simplified as
iDS = i0e
VDS
kT
q . (4)
This formula can be inverted to find the source voltage as
logarithmically dependent on the iDS current. We have
VS = VD − (kT/q)∗ ln(iDS/i0) (5)
with VS = VPH, VD = VLOG, and iDS = iPH. Equation (5)
represents the logarithmic relation between the photogenerated
current and the voltage at the photosensitive node for a single
transistor load and for VPH ≥ VLOG − VTH. In the configura-
tion presented in Fig. 1, the load transistors are M1 and M2, and
the transfer function in the logarithmic region is
VPH =VLOG − 2∗(kT/q)∗ ln(iDS/i0) VPH≤VLOG − 2∗VTH.
(6)
To relate these results to the physical implementation
presented in this paper, the electrical design rules from
ON-Semiconductor (formerly AMI) 0.35 μm [20] may be
considered. This standard CMOS technology was chosen for
the proposed sensor, because it may partially be customized,
thus enhancing the overall sensor performance. In particular,
a nonstandard shield to silicide implant and to complementary
p-well diffusion was added at the array level [21].
To define a model that can practically be used, a series of
approximations must be assumed. In particular, i0, which is
temperature dependent as T 2 [11], is considered constant over
the whole temperature range. This assumption is acceptable for
endoscopic applications, where the temperature is supposed to
be constant. As a result, based on the data from the silicon
foundry, the i0 current is in the order of 0.01 fA, and the
logarithmic response may be calculated as a function of the
photogenerated current.
The model was developed by using Microsoft Excel 2007,
and the results are shown in Fig. 2. As further reported,
the model was validated on the developed prototype. Once
Fig. 2. Simulated power responsivity for the linear–logarithmic pixel at
different VLOG and TINT values. ADCOUT is the ADC digital output word
as defined in Section III.
experimentally verified, the model will allow the theoretical
extrapolation of several parameters that would otherwise be
difficult to measure, e.g., the capacitance at the photosensitive
node and the actual dark current.
As shown in Fig. 2, the response in the logarithmic region
maintains the same trend, regardless of the integration time. In
particular, VPH is not affected by CIN [see (5)]. This condition
leads us to conclude that the DR is not related to the full-
well capacitance, as in linear pixels, but is potentially limited
only by the voltage range of the readout channel and the
silicon secondary effects. The CIN effect is only relevant for
the linear region, where a small CIN value would increase the
slope of the power responsivity, as highlighted in (1), thus
resulting in an improved discrimination between different levels
of input signal. To better fit the linear region, a dependency
of the photodiode capacitance from the applied voltage was
considered in CIN [6]. This approach results in a better fit in
the low-light region.
III. IMPLEMENTATION OF THE PIXEL IN A TEST CHIP
The linear–logarithmic technique presented in this paper was
implemented in a monolithic active-pixel gray-level camera-
on-a-chip sensor. The photo-sensitive matrix consists of a
100 × 100 pixel array based on an n-well/p-substrate photo-
diode, with a pixel pitch of 9.4 μm × 9.4 μm and a fill factor
(FF) of 30% [6]. The sensor integrates a pixel array, a number
of column data sampling (CDS) blocks, a data double sampling
(DDS) block at the array level [10] with double sampling and
preamplification capability, a 12-b analog-to-digital converter
(ADC), and several other support blocks (see Fig. 3).
When a row-decoder addresses a row, every pixel of the
selected line simultaneously transfers its output value to a CDS,
which stores the information. Then, the pixels of the selected
row are reset, and their reset values are transferred to the
associated bit-line CDS. The CDS subtracts this value from the
stored one and makes the result available at the output. This
subtraction between two noncorrelated samples mainly allows
pixel FPN reduction. Depending on the response, either linear
VATTERONI et al.: LINEAR–LOGARITHMIC CMOS PIXEL WITH TUNABLE DYNAMIC RANGE 1111
Fig. 3. Simplified block diagram of the camera.
or logarithmic, this correction is performed, respectively, on
the whole signal chain or just on a part of it. In particular,
when the pixel is in the reset condition, the level of VPH is
such that the active load is not working. This condition means
that M2 and M3 are not involved in the signal path during
the acquisition of the reset value; thus, their FPN contribution
cannot be subtracted.
The sampling operation of the CDS block can be summarized
as follows:
VSigCDSOut = [(VResPix − VSigPix) + VbCDS]×GCDS (7)
where VSigCDSOut is the output of the CDS, VResPix and
VSigPix are the reset and signal output of the pixel, respectively,
VbCDS is an external reference voltage, and GCDS is the
CDS gain.
Each CDS block is then reset to generate a reference value
for the following DDS block. The reset output of the CDS is
VResCDSOut = VbCDS ×GCDS. (8)
The following operation is the sequential selection of each
CDS block by a column decoder. After the selection, the output
value of each CDS is transferred to the DDS twice: one time
for the signal and another time for the reset value. The DDS
is a fully differential switched capacitor block that performs a
subtraction between the signal and the reset value from each
CDS, the addition of a threshold voltage (VREF), and a mul-
tiplication by a gain factor (GAIN). Such a gain can be set by
the user, ranging from 1 to 4, by changing the input capacitance
of the switched capacitor circuit. The resulting DDS output is
differential as described in the following relation:
VOutDDS± = VCM ±GAIN
× (VREF − (VResCDSOut − VSigCDSOut)) (9)
where VOutDDS± are the DDS outputs, and VCM is the com-
mon mode signal.
Fig. 4. Complete (a) imager and (b) pixel layout.
The DDS differential output signals are sent to a 12-b
pipelined ADC [22] and digitized into 4096 levels. Therefore,
each pixel is described by a 12-b word as follows:
ADCOUT =
V +OutDDS − V −OutDDS
VREFP − VREFN ∗ 2
11 (10)
where VREFP and VREFN are the ADC reference voltages, and
ADCOUT is the ADC digital output word.
IV. EXPERIMENTAL RESULTS
The image sensor presented in this paper was fabricated in
the standard 0.35-μm mixed-signal CMOS technology from
ON-Semiconductors. The chip layout is shown in Fig. 4(a).
The die size is 6.36 mm× 3.68 mm, including test structures,
more than five 100 × 100 pixel arrays, not described in this
paper. Only the first 100 × 100 array based on a standard
layout design [see Fig. 4(b)] was considered and characterized.
Electro-optical tests were carried out on a dedicated electro-
optical bench to characterize the imager performance. For
this purpose, a custom system was developed (see Fig. 5)
to interface the sensor with a personal computer (PC). The
system consists of two boards. The first board (referred to as
the “daughter board” in Fig. 5) is a custom board, particularly
designed to interface the vision sensor with the second board
(referred to as “mother board” in Fig. 5) through a couple of
50-pin stripes. The second board features a 100-Kgates field-
programmable gate array (FPGA; Xilinx, Spartan). The code
1112 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 58, NO. 4, APRIL 2011
Fig. 5. Customized test boards.
that was loaded on the FPGA enables imager control and
interfacing of the imager with a PC-based user console.
With regard to electro-optical characterization, the overall
integration time was set to 30 and 100 ms, and the VLOG
reference ranged from 1 V to 3.3 V to assess the power
responsivity modulation. The reference signals of the sensor
were set to have an offset value of 500 LSBs, thus preventing
downward saturation of the readout channel. This condition
guarantees accurate sensitivity estimation, which measures the
capability of the sensor to detect low LIPD. Note that sensitivity
can be identified by the LIPD value, where SNR = 1 for the
given configuration of the sensor parameters [6].
As shown in Fig. 6(a), a DR of 110 dB can be achieved over
an effective number of bits of 11.5 by setting TINT = 30 ms
and VLOG = 2.3 V. For this specific case, 11.2 b resolves
the linear region of the sensor, thus covering 2.25 decades of
irradiance (45 dB). The rest of the dynamics, i.e., 65 dB, which
is equivalent to more than three decades, is logarithmically
mapped over 9.2 b, with a significant decrease in resolution.
To test power responsivity modulation, VLOG was set at
three different values, i.e., 1.8, 2.0, and 2.3 V. As represented
in Fig. 6(a), an increase in VLOG results in a shift of the
transition point from the linear to the logarithmic region in
the power responsivity curve. Furthermore, when changing the
VLOG value from 1.8 V to 2.3 V, the resolution loss in the
linear region is negligible if compared to the overall gain of
one decade in the DR.
Note that the 110-dB power response limit is due to the
measurement setup, which cannot provide the sensor with a
LIPD over 103 W/m2. This limit can theoretically be extended
over 120 dB with a proper sensor setting, as detailed later on
in this paper. Furthermore, pixel sensitivity can be improved
by increasing the integration time, similar to a standard linear
active pixel sensor (APS). This feature was both modeled (see
Fig. 2) and experimentally assessed by setting the integration
time up to 100 ms, thus obtaining an improvement in sensitivity
by a factor of 30.
Noise performance of the sensor was extracted from ex-
perimental results as a percentage of the signal range. The
average FPN value was measured at different LIPD levels as
the standard deviation of pixel values over an image acquired
without a focusing lens on top of the chip. To exclude pixel
Fig. 6. (a) Power responsivity and (b) SNR at different VLOG and TINT
values over the whole LIPD range.
noise (PN), the standard deviation was calculated over an image
that results from the average of 100 acquisitions at constant
environmental and setting conditions. The resulting value for
FPN is 0.83% in the linear region and 1.37% in the logarithmic
region [see Fig. 7(a)]. PN was measured over the whole LIPD
range as the standard deviation over 100 pixel acquisitions
taken at constant environmental and setting conditions. To
avoid FPN contribution, each value was taken as the average
pixel over a whole image. The resulting value for PN is less
than 0.22% of the whole signal range [see Fig. 7(b)].
The measured SNR, defined as the signal over the PN (in
decibels), is plotted in Fig. 6(b) for the entire dynamic range
(DR). Note that variations in VLOG do not affect the SNR
response. Furthermore, there is no appreciable decrease in the
SNR curve for strong incident light, thus potentially allowing
us to achieve a maximum DR larger than 120 dB.
Examples of ex vivo biological images of a porcine stomach,
acquired with the described chip, are shown in Fig. 8. The same
image was acquired with different VLOG levels to show the
increasing number of details that can be obtained by tuning
VLOG. To quantify this variation, the histogram of each image
is included. The additional information content that can be
recovered by increasing VLOG is highlighted with a circle. Note
that no processing was applied to the images shown in Fig. 8.
Residual FPN in the high-light power region is 1.37% of the
full signal range, i.e., four times lower than with a standard
VATTERONI et al.: LINEAR–LOGARITHMIC CMOS PIXEL WITH TUNABLE DYNAMIC RANGE 1113
Fig. 7. (a) FPN and (b) PN at different VLOG and TINT values over the
whole LIPD range.
logarithmic pixel [23]. The main performance of the image
sensor is summarized in Table I.
The power responsivity results were compared with the
model presented in Section II. Parameters that cannot di-
rectly be measured on chip, e.g., the photodiode capacitance,
were tuned around values that were calculated considering the
0.35-μm ON-Semiconductor process parameters (e.g., CIN =
16 fF and iD = 52 fA). Fig. 9 shows a good fit between the
model and the measurements, both in the logarithmic and the
linear regions.
This fit was quantified as percentage error for each LIPD step
and is plotted in Fig. 10. The average percentage error over the
LIPD range, excluding the saturated regions, is around 0.8%
for all the considered sensor configurations, with a peak in the
transition region.
Once model output reliability has been demonstrated, the DR
that may be achieved by the sensor can further be discussed.
As reported in Fig. 11, by setting VLOG = 2.3 V, the model
predicts a LIPD saturation of 4 × 106, thus achieving more
than 120 dB in the DR. This result is further supported by the
aforementioned SNR stability at high LIPD.
V. CONCLUSION
A novel linear–logarithmic pixel for intraframe HDR image
sensor, featuring a 12-b digital output, has been presented.
This pixel configuration was integrated in a 100 × 100 pixel
Fig. 8. Comparison between ex vivo images of a porcine stomach with differ-
ent dynamic ranges, from low (VLOG = 2.0 V) to high (VLOG = 2.70 V)
threshold values.
1114 IEEE TRANSACTIONS ON ELECTRON DEVICES, VOL. 58, NO. 4, APRIL 2011
TABLE I
MAIN CHARACTERISTICS AND PERFORMANCE OF
THE IMAGE SENSOR DEVELOPED
Fig. 9. Modeled and experimental power responsivity of the linear–
logarithmic pixel at different VLOG and TINT values over the whole LIPD
range.
array within a monochrome imager. A DR over 110 dB was
experimentally measured, whereas an extension over 120 dB,
not measurable due to measurement setup limitations, is ex-
pected on account of the modeling results and SNR stability at
high LIPD. Measurement results show good PN (0.22% rms)
and FPN (0.83–1.37%) performance, both in the linear and
the logarithmic regions. Furthermore, the DR and, accordingly,
the resolution of the power responsivity curve can simply be
adjusted by tuning the point of transition between the linear
and the logarithmic domains through the setting of an analog
reference signal (i.e., VLOG).
A behavioral model of the linear–logarithmic pixel was
presented and compared with the experimental results. The
Fig. 10. Percentage error over the LIPD range in the fit of the measured results
by the model for different VLOG and TINT.
Fig. 11. Percentage error over the LIPD range in the fit of the measured results
by the model for different VLOG and TINT.
average percentage error in the fit between experimental and
theoretical data was 0.8%. Given such consistency, the model
can be used to estimate the DR with different settings of the
reference signals or to extrapolate specific device parameters
from experimental data (e.g., the photodiode capacitance and
the dark current), which would otherwise be impossible to
experimentally quantify.
The sensor was tested by acquiring images from an ex
vivo porcine stomach. Due to the sensor’s ability to be self
referential by changing VLOG and, accordingly, the DR, the
efficiency of an HDR response in comparison with a standard
DR one was proven. The main result consists of additional
details in the HDR image if compared to the standard DR
one. In endoscopic imaging, HDR is useful in case of high
reflective regions, mainly in correspondence of wet mucosa. In
the case of the presented pixel, these regions are mapped in the
logarithmically extended dynamic, thus avoiding saturation of
the scene and providing better detail to the user.
To complete the characterization of the pixel, a spectral
responsivity measurement will be performed as the next step
to gather information about the spectral performance and the
quantum efficiency.
VATTERONI et al.: LINEAR–LOGARITHMIC CMOS PIXEL WITH TUNABLE DYNAMIC RANGE 1115
The presented HDR pixel technology will be implemented
in a second prototype with extended resolution. Filters will be
added to obtain a color image stream, which is of paramount
importance in the targeted biomedical application.
ACKNOWLEDGMENT
The authors would like to thank Carmela Cavallotti, Daniele
Covi, and Luca Clementel for their valuable support and activ-
ity in the optical characterization phase.
REFERENCES
[1] G. Berci and M. Paz-Partlow, “Electronic imaging in endoscopy,” Surg.
Endosc., vol. 2, no. 4, pp. 227–233, Dec. 1988.
[2] H. Hopkins and N. S. Kapany, “A flexible fiberscope using static
scanning,” Nature, vol. 173, no. 4392, pp. 39–41, Jan. 1954.
[3] AltaSens—Medical Imaging. [Online]. Available: http://www.altasens.
com
[4] OmniVision—Medical Imaging. [Online]. Available: http://www.ovt.
com/applications/application.php?id=5
[5] M. Bigas, E. Cabruja, J. Forest, and J. Salvi, “Review of CMOS image
sensors,” Microelectron. J., vol. 37, no. 5, pp. 433–451, May 2006.
[6] A. Theuwissen, “Digital imaging: Image capturing, image sensors, tech-
nologies and applications,” CEI-Europe2004.
[7] D. Yang, A. El Gamal, B. Fowel, and H. Tian, “A 640× 512 CMOS image
sensor with ultrawide dynamic range floating-point pixel-level ADC,”
IEEE J. Solid-State Circuits, vol. 34, no. 12, pp. 1821–1834, Dec. 1999.
[8] T. Lulé, B. Schneider, and M. Bohm, “Design and fabrication of a high-
dynamic-range image sensor in TFA technology,” IEEE J. Solid-State
Circuits, vol. 34, no. 5, pp. 704–711, May 1999.
[9] B. J. Hosticka, W. Brockherde, A. Bussmann, T. Heimann, R. Jeremias,
A. Kemma, C. Nitta, and O. Schrey, “CMOS imaging for automotive
applications,” IEEE Trans. Electron Devices, vol. 50, no. 1, pp. 173–183,
Jan. 2003.
[10] J. B. Chun, H. Jung, and C. M. Kyung, “Dynamic-range widening in a
CMOS image sensor through exposure control over a dual-photodiode
pixel,” IEEE Trans. Electron Devices, vol. 56, no. 12, pp. 3000–3008,
Dec. 2009.
[11] M. Loose, K. Meier, and J. Schemmel, “A self-calibrating single-chip
CMOS camera with logarithmic response,” IEEE J. Solid-State Circuits,
vol. 36, no. 4, pp. 586–596, Apr. 2001.
[12] S. Kavadias, B. Dierickx, D. Scheffer, A. Alaerts, D. Uwaerts, and
J. Boagaerts, “A logarithmic-response CMOS image sensor with on-chip
calibration,” IEEE J. Solid-State Circuits, vol. 35, no. 8, pp. 1146–1152,
Aug. 2000.
[13] D. Stoppa, M. Vatteroni, A. Baschirotto, A. Sartori, and A. Simoni, “A
120-dB dynamic range CMOS image sensor with programmable power
responsivity,” IEEE J. Solid-State Circuits, vol. 42, no. 7, pp. 1555–1563,
Jul. 2007.
[14] S. Kavusi and A. El Gamal, “Quantitative study of high-dynamic-range
image sensor architectures,” Proc. SPIE, vol. 5301, pp. 264–275, 2004.
[15] J. Hynecek, E. C. Fox, and D. R. Dykaar, “Sensor pixel with linear and
logarithmic response,” U.S. Patent 6 323 479, Nov. 27, 2001.
[16] A. Sartori and M. Vatteroni, “Photo-sensitive element for electro-optical
sensors operating under various illumination conditions,” U.S. Patent
7 119 320, Oct. 10, 2006.
[17] A. Sartori and M. Vatteroni, “Photo-sensitive element for electro-optical
sensors,” European Patent, EP1533995, Feb. 13, 2008.
[18] G. Tecchiolli and A. Sartori, “Photo-sensitive element for electro-optical
sensors,” U.S. Patent 6 891 144, May 10, 2005.
[19] M. Loose, “A self-calibrating CMOS image sensor with logarithmic re-
sponse,” Ph.D. dissertation, Institut fur Hochenergiephysik, Universitat
Heidelberg, Heidelberg, Germany, 1999.
[20] Design Rule Manual C035M-A, ON-Semiconductor, Phoenix, AZ, 2002.
ID DS 13337, Rev. 03.
[21] H. S. Wong, “Technology and device scaling considerations for CMOS
imagers,” IEEE Trans. Electron Devices, vol. 43, no. 12, pp. 2131–2142,
Dec. 1996.
[22] S. Hamami, L. Fleshel, and O. Yadid-Pecht, “CMOS image sensor em-
ploying 3.3-V 12-bit 6.3-Ms/s pipelined ADC,” Sens. Actuators A: Phys.,
vol. 135, no. 1, pp. 119–125, Mar. 2007.
[23] M. Vatteroni, D. Covi, D. Stoppa, B. Crespi, and A. Sartori, “High-
dynamic-range CMOS image sensors in biomedical applications,” in
Proc. IEEE EMBS Conf., Lyon, France, 2007, pp. 2819–2822.
Monica Vatteroni received the M.S. degree in electrical engineering from the
University of Pisa (I), Pisa, Italy, in 2001 and the Ph.D. degree in physics from
the University of Trento (I), Trento, Italy, in 2008.
From 2002 to 2008, she was with NeuriCam, Trento, as a Pixel Engineer
and an Analog Designer, where she was responsible for CMOS image sensor
development in 2005. She is currently a Postdoctoral Fellow with the Scuola
Superiore Sant’Anna, Pisa, where she is in charge of the research and develop-
ment of image sensors and vision systems for biomedical applications. She is
the author or a coauthor of several conference proceedings and journal publi-
cations and is the holder of three patents. Her research interests include CMOS
image sensors, low-noise analog electronics, high-dynamic-range pixels, and
endoscopic vision systems.
Pietro Valdastri (M’05) received the Ph.D. degree (with honors) in electronic
engineering from the University of Pisa, Pisa, Italy, in February 2002 and the
Ph.D. degree in bioengineering from the Scuola Superiore Sant’Anna, Pisa, in
2006. His Ph.D. dissertation was multiaxial force sensing in minimally invasive
robotic surgery.
He is currently an Assistant Professor with the CRIM Lab, Scuola Superiore
Sant’Anna. His research interests include implantable robotic systems and
active capsular endoscopy. He is working on several European projects for the
development of minimally invasive and wireless biomedical devices.
Alvise Sartori (M’91) received the M.A. degree in physics from the University
of Oxford, Oxford, U.K., in 1978 and the Ph.D. degree in geophysics from
Imperial College, London, in 1983.
He then joined the Central Research Laboratory of Olivetti, where he carried
out research on the modeling of fluido-dynamic systems and the design of
digital CMOS integrated circuits. In 1990, he was with the research institute
Center for Information Technology (IRST), Bruno Kessler Foundation, Trento,
Italy, where he was in charge of the VLSI Design Laboratory. Since 1998, he
has been the President and the Chief Executive Officer (CEO) of NeuriCam
SpA, Trento, a company that he cofounded in 1998, which is active in the
fabless production of chips and systems for computer vision.
Arianna Menciassi (M’00) received the degree (with honors) in physics from
the University of Pisa, Pisa, Italy, in 1995 and the Ph.D. degree in bioengi-
neering, with a research program on the micromanipulation of mechanical and
biological micro objects, from the Scuola Superiore Sant’Anna, Pisa, in 1999.
Her Ph.D. dissertation was microfabricated grippers for the micromanipulation
of biological and mechanical objects.
She is currently a Professor of biomedical robotics with the Scuola Superi-
ore Sant’Anna. Her main research interests include biomedical microrobotics
and nanorobotics, microfabrication technologies, micromechatronics, and mi-
crosystem technologies. She is working on several European and international
projects for the development of microrobotic and nanorobotic systems for
medical applications.
Paolo Dario (F’02) received the Dr.Eng. degree in mechanical engineering
from the University of Pisa, Pisa, Italy, in 1977.
He established and teaches the mechatronics course with the School of
Engineering, University of Pisa. He has been a Visiting Professor with the
Ecole Polytechnique Federale de Lausanne (EPFL), Lausanne, Switzerland,
and Waseda University, Tokyo, Japan. He is currently a Professor of biomedical
robotics with the Scuola Superiore Sant’Anna, Pisa. He is also the Director of
the CRIM Lab, Scuola Superiore Sant’Anna, where he supervises a team of
about 70 researchers and Ph.D. students. His main research interests include
medical robotics, mechatronics, and microengineering, in particular sensors and
actuators for the aforementioned applications. He is the Coordinator of several
national and European projects, the Editor of two books on robotics, and the
author of more than 200 journal papers.
Prof. Dario is a Member of the Board of the International Foundation of
Robotics Research. He is an Associate Editor for the IEEE TRANSACTIONS
ON ROBOTICS AND AUTOMATION, a Member of the Steering Committee of
the JOURNAL OF MICROELECTROMECHANICAL SYSTEMS, and a Guest Ed-
itor of the SPECIAL ISSUE ON MEDICAL ROBOTICS, IEEE TRANSACTIONS
ON ROBOTICS AND AUTOMATION. He is the President of the IEEE Robotics
and Automation Society and a Cochair of the Technical Committee on Medical
Robotics of the IEEE Robotics and Automation Society.