CMOS ACTIVE PIXEL SENSOR FOR A POLARIZATION-
NSF Summer Undergraduate Fellowship in Sensor Technologies
Gregory J. Barlow (Electrical Engineering) – North Carolina State University
Advisors: Dr. Nader Engheta and Dr. Jan Van der Spiegel
Polarization-sensitive vision is well documented as serving in navigation for many animals, but some types of biological polarization-sensitive vision may enhance object visibility in scattering media. Because neither the human eye nor conventional cameras are polarization-sensitive, artificial polarization vision systems must be designed to exploit the polarization of light; artificial polarization-difference imaging has been shown to be capable of enhancing target detection in scattering media. Previous polarization-sensitive cameras required external processing, were not real-time, and used relatively large amounts of power. A CMOS active pixel sensor is presented for use in a low power, portable, real-time polarization-difference camera. Pixels were designed for integration with a diffractive optical element polarization analyzer. Column readout circuits include fixed pattern noise suppression. In addition, a scaling methodology to enhance system performance and to correct for non-ideal polarization analyzers is presented.
Humans rely heavily on their visual systems to understand the world around them. Human vision is based on brightness and color, which are extremely efficient under normal conditions.  In optically scattering media, however, such as underwater, in fog, or in rain, variations of brightness and color are low, diminishing object contrast and lowering the effectiveness of the visual system.  While polarization-sensitive vision is well documented as serving in many invertebrate navigation systems, Rowe et al. hypothesize that the green sunfish uses polarization vision to enhance target detection underwater.  Artificial polarization-sensitive vision has been shown to enhance the visibility of target objects in scattering media, using a method called polarization-difference imaging (PDI), inspired by the visual system of the green sunfish. 
Polarization-sensitive cameras have been previously demonstrated, but have not been designed to operate at full video rates. [4,5,6,7] In addition, these schemata require the use of significant external computing resources, limiting the portability of the system. Nor have these cameras been designed for low power use. The design of a low power, portable, real-time polarization-sensitive electronic camera-on-a-chip is desirable for use in a variety of applications.
This paper covers a variety of areas related to the design and implementation of a polarization-difference camera. Section 2 gives a brief review of polarization and PDI, background on electronic cameras and diffractive optical elements (DOE), and
information on charged coupled devices and CMOS pixels. Section 3 covers the overall design of the camera. Section 4 addresses the design and layout of the CMOS active pixel sensor (APS) used in this camera, with specific consideration of the design constraints of a PDI camera. Section 5 contains the designs of the readout circuitry for the camera and methods of fixed pattern noise (FPN) reduction. Section 6 details the operation of the camera, with some specifications for control, timing, and drive systems. Section 7 addresses intensity scaling to correct for the non-ideal DOE polarization analyzer. Section 8 gives the simulation results for the designed circuits, while Section 9 contains discussion of results and project conclusions. Future work and recommendations are addressed in Section 10.
Light has three properties detectable by vision systems: intensity, wavelength, and polarization. While the human eye can perceive both intensity and wavelength, it is polarization-blind. For this reason, conventional electronic cameras are not designed to extract polarization information from a scene. 
Light is a transverse electromagnetic wave; its electric and magnetic fields are orthogonal to the direction of propagation. The path traced by the tip of the electric field as the wave propagates defines the polarization of light. Light sources, such as the sun, are usually randomly polarized, while the media that the light encounters tend to alter its polarization.
2.2 Biological Basis
Polarization vision has been extensively studied in many classes of invertebrates. Bees, ants, and other invertebrates use polarization for navigation.  While some vertebrates are capable of extracting polarization information from visible light, the physical mechanism is not as well understood as in invertebrate systems. The sensitivity of the green sunfish to variations in polarization led to the hypothesis that polarization was used to enhance underwater vision.  The potential for vision enhancement makes polarization sensitivity desirable to incorporate into electronic cameras.
2.3 Polarization-difference Imaging
PDI is one method of extracting polarization information from a scene. In this method, images of a scene are captured at two orthogonal linear polarizations. The pixel-by-pixel sum of the two images forms the polarization-sum (PS), and the pixel-by-pixel difference of the two images forms the polarization-difference (PD).  Color is used to map the PD into the visual realm. 
(i,j) and I(i,j), where (i,j) If the two image intensity distributions are symbolized as I12
represents the pixel location and I and Ihave orthogonal linear polarizations, then the 12
PS and PD are as follows:
I(i,j) = I(i,j) + I(i,j) (1a) PS12
I(i,j) = I(i,j) - I(i,j) (1b) PD12
The PS image is equivalent to a conventional image if the linear polarizer is ideal. For non-ideal linear polarizers, corrective scaling must be implemented. It should be noted that the PD image depends on the polarization axes. 
PDI is qualitatively better than conventional imaging for target detection in scattering media; detection enhancement has been demonstrated at observable degrees of polarization of less than 1%.  PDI is inherently capable of common-mode rejection for background light, which further enhances target detection. Because PDI only requires relatively simple computations, it is extremely suitable for use in a polarization sensitive electronic camera.
2.4 Polarization Camera
Previous polarization-sensitive cameras were not real-time, though near-video rates have been achieved. [4,5,6,7] Previous polarization-sensitive cameras have also required intensive external processing. PDI is a suitable method for use in a real-time electronic camera-on-a-chip which is polarization-sensitive.
A polarization-difference camera extends the functionality of a normal digital electronic camera. A conventional electronic camera generally consists of eight stages.  These are (1) optical collection of photons via a lens, (2) discrimination of photons, generally based on wavelength (3) detection of photons via a photodiode or photogate, (4) readout of detectors, (5) timing, control, and drive electronics, (6) signal processing electronics, including FPN, (7) analog to digital conversion, and (8) interface electronics. The order of these stages is not necessarily fixed; some signal processing may occur after analog to digital conversion.
2.5 Diffractive Optical Elements
A DOE is a pattern of microstructures that can transform light in a predetermined manner.  For example, a Fresnel zone lens DOE can be used to focus light. While the Fresnel zone lens requires a variation of the height of the surface, a sub-wavelength binary DOE can exhibit the same behavior.  Figure 1 shows images of both types of lenses. In addition, sub-wavelength binary DOEs are polarization selective, which is advantageous for this project. The DOE designed for this project focuses the incoming light along a line parallel to the grooves of the DOE. This DOE is polarization sensitive.
Figure 1 - (a) Fresnel zone lens and (b) sub-wavelength binary lens
2.6 Image Sensors
The imager technology used in an electronic camera is instrumental in determining the capabilities of the final system. Low noise, large array sizes, high frame rates, and power dissipation are preferred for a polarization-difference camera.
2.6.1 Charge-coupled Device Image Sensors
Charge-coupled device (CCD) technology, currently the most popular sensor technology, is capable of producing high-quality images.  Small, low-resolution CCD cameras are also relatively inexpensive. CCD technology's relative freedom from FPN is one of its most attractive characteristics. However, CCD-based systems often consume several watts of power, can be accessed only a single pixel at a time, and are difficult to integrate with processing circuitry.
2.6.2 CMOS Image Sensors
MOS image sensors were demonstrated in the 1960's, but work fell off with the introduction of the CCD, which displayed much less FPN than MOS sensors.  The need for smaller and less expensive imaging technology has led to a resurgence in the popularity of CMOS image sensors. CMOS imager technology also has the advantage of low power, random and row based pixel access, and easy integration with processing circuitry. There are three main approaches to CMOS pixels: passive pixels, photodiode APSs, and photogate APSs. APSs can be designed for operation in either voltage or current-mode.
2.6.3 CMOS Passive Pixel Sensors
The passive pixel sensor is very simple, consisting of a photodiode and a transfer transistor.  Passive pixel sensors have high quantum efficiency and extremely small pixel size; however, noise levels are quite high, and this pixel type does not scale well.
2.6.4 CMOS Active Pixel Sensors
An active pixel includes at least one active transistor within the pixel cell.  An active amplifier within the pixel helps to improve performance over that of the passive pixel, allowing larger arrays and faster readout speeds. Since the amplifier within a pixel draws power only when the pixel is being read out, power dissipation remains low. Many APSs have been designed, some with very high quality and extremely high frame rates. [13-23] Active pixels are either photodiode or photogate based; readout is either voltage-mode or current-mode.
2.6.5 Photodiode APS
The most common photodiode-type active pixel,
shown in Figure 2, includes three transistors and the
photodiode. A reset transistor resets the photodiode,
and a source follower buffers the voltage on the
photodiode. A row select transistor enables readout to
the column bus. This design has high quantum
efficiency, as the diode is not covered by polysilicon.
 Because the integration and reset nodes are the
same, correlated double sampling (CDS), a FPN
suppression method, is more difficult to implement. Figure 2 – Schematic of CMOS photodiode APS
2.6.6 Photogate APS
Identical to the photodiode-type active pixel in the layout of the reset transistor, source follower, and row select transistor, the photogate-type active pixel uses a photogate followed by a transfer gate in place of the photodiode.  The quantum efficiency of a photogate is less than that of a photodiode, because of the polysilicon covering.  FPN is also higher. The primary advantage of the photogate-type active pixel is its ability to do CDS, which is very effective in suppressing reset noise, 1/f noise, and FPN. 
2.6.7 Voltage and Current-mode APS
A majority of the active pixels that have been investigated are voltage-mode pixels. [13-18] These pixels have demonstrated excellent performance, with high speeds and low noise levels. [15,16] FPN suppression techniques have been developed, and on-chip analog-to-digital conversion has been demonstrated. 
While less work has been done on current-mode active pixels, many have been demonstrated. [19-23] Current-mode active pixels show potential for improvements in readout speed and lower readout noise, though none yet seriously rival voltage-mode pixels. Effective FPN reduction for current-mode active pixels is still being investigated. On-chip analog-to-digital conversion has been presented. [19,22]
3. CAMERA DESIGN
A polarization-difference camera extends the
necessary functions of a normal digital electronic
camera. After the optical collection of photons with
a lens, DOE polarization analyzers are used to
separate the light into polarized components. A pixel
then converts the photons to electrons. For each
pixel of output, it is necessary for the sensor array to
have two input pixels, one for each component of
polarization. After detection, the pixels are read out
on a column bus into CDS circuitry, which
suppresses FPN. The analog signal is then converted
to a digital signal using an on-chip analog-to-digital
converter. Because it is not possible to design an
ideal DOE polarization analyzer, output signal
corrective scaling is necessary. Scaling to enhance Figure 3 - Stages of the system performance is also desirable. For each polarization-difference camera signal pair, the output signal undergoes a combined
scaling process that incorporates both corrective and performance-enhancing scaling. The scaled intensities are then summed and differenced to obtain the polarization-sum and polarization-difference. The polarization-sum is effectively the intensity for that pixel. The polarization-difference is mapped into color. Video interface electronics then display the image. A schematic of the stages of the polarization-difference camera is shown in Figure 3.
This work focused on the design of the APS, the readout circuitry, and the scaling methodology to correct for the non-ideal nature of the polarization analyzer.
4. CMOS APS
The pixel type selected for this design was a voltage-mode photogate APS. Using a voltage-mode pixel simplified FPN suppression circuitry and layout considerations. The use of photogate-type active pixels enabled the use of CDS.
Figure 4 shows a schematic of the pixel
circuit design. This is one of the most
common types of APSs.  The photogate
is separated from the floating diffusion node
by a transfer gate, which allows for CDS.
The transfer gate is biased at a constant
voltage, while the photogate is pulsed. (See
Section 6) The pixel unit also contains a reset
transistor, an in-pixel source follower, and a
132 Figure 4 - Schematic of CMOS photogate
row selection transistor.
While a standard electronic camera has one input pixel for every pixel of output resolution, this polarization camera requires two inputs for every output, since each “half pixel” represents one component of the linear polarization. The sum and difference of the two pixels form the polarization
sum and polarization difference,
which make up the polarization-
difference image. The diffractive
optical element layer above the sensor
array is arranged in stripes;
neighboring pixels in a given row are
overlaid with DOEs oriented
orthogonal to one another, as shown
in Figure 5.
Figure 5 - Schematic for orientation of the DOE overlay of the pixel array 4.2 Layout
The orientation of the DOEs in a striped pattern places certain constraints on the layout of the APS. The DOE focuses incoming light along a line in the center of the DOE, and the pixel should be designed to capture as much light as possible. To maximize the photogate area and to keep the pixel vertically symmetrical for integration with the DOE, the photogate is T-shaped, as seen in Figure 6. A light shield, which also acts as the power rail, covers all areas of the pixel except the photogate. This limits crosstalk between neighboring pixels. The pixel has been designed to overlap with its neighboring pixel to maximize detector area. The integration of the pixels with the DOE is shown in Figure 7. The left DOEs focus light on the vertical section of the corresponding pixels, while the right DOE’s focus light on the horizontal section of the corresponding pixels. When combined into the sensor array, each pixel is 18 microns by 18 microns using an AMI 0.6;m double poly, three metal process. The total photogate area in each pixel is ~130 square microns. All transistors in the pixel have W/L ratios of 6/2.
Figure 6 - CMOS photogate APS layout, using a double poly, three metal CMOS process
Figure 7 - DOE unit overlap of pixels
5. NOISE SUPPRESSION
For an APS, FPN must be suppressed or the performance of the sensor will be limited.  A light shield has been used as described to limit cross talk between pixels, but the primary sources of FPN will be reset voltage variations between pixels and column-to-
column variation resulting from the column readout structure. Two techniques for FPN suppression have been implemented.
5.1 Correlated Double Sampling
CDS is a method for suppressing reset noise. A schematic for a basic readout circuit with CDS is shown in Figure 8. CDS first samples the reset voltage of a pixel and then samples the signal voltage. The difference between the two is the output voltage. CDS suppresses reset noise from variations in reset voltage between pixels as well as threshold variations from the source follower transistor within the pixel.  CDS also reduces low frequency noise.
Figure 8 - Schematic of CDS circuit
5.2 Column Reference Subtraction
To correct for column-to-column variations caused by the column readout structure, a row of dark reference pixels is added to the sensor array, completely covered by the light shield. This row is read out in the same manner as the normal pixels. In the signal processing stage, which occurs after analog to digital conversion, this row is subtracted from each of the other rows, serving as a column reference. For this reason, this row
should be read out first, so that column reference subtraction can occur before the polarization-sum and polarization-difference are formed.
Control circuits for this camera have not been designed, but are relatively straightforward. The power rail, VDD, is carried on the light shield layer that covers the entire array and is set to 5 volts. The transfer gate within the pixel, TX, is set to 2.5 volts, the gate of transistor MLN is biased at 1.5 volts, and the gates of MLP1 and MLP2 are biased at 2.5 volts. (See Figure 7) The first period for operation is the signal integration period, during which the photogate (PG) is biased at 5 volts. The reset transistor, MR, is biased at 2.5 volts as an antiblooming drain. Row select switches are off. After signal integration, the sensor array should be read out row by row. The row to be read out is selected by enabling the row select switch, MS. Then, the reset transistor is pulsed to 5 volts in order to reset the floating diffusion node of the pixel (FD). The reset voltage is then sampled by selecting the sample-and-hold switch MSHR. This stores the reset voltage on capacitor CR. The photogate is then pulsed low to 0 volts, which transfers the signal charge to the floating diffusion node. The signal voltage is sampled by selecting the sample-and-hold switch MSHS. This stores the signal voltage on capacitor CS. Setting the column-select switches MY1 and MY2 low scans out the stored reset and signal voltages. A partial timing sequence is shown in Figure 9.
Figure 9 - Timing for pixel and read out operation
7. INTENSITY SCALING
Because the DOE does not function as an ideal polarization analyzer, corrective intensity scaling is necessary to obtain the true components of linear polarization. The intensity of output light from a DOE is different for transverse electric (TE) waves and transverse magnetic (TM) waves. The ratio of TE to TM for an ideal polarization analyzer is 1 to 0.