Photocurrent processing

All the information a photosensor can extract from the light distribu-tion in a scene is contained in the spatial and temporal moduladistribu-tion of the photocurrent in the individual pixels. For this reason, it is of much interest to process the pixels’ photocurrents accordingly, in order to ob-tain the relevant modulation parameters in the most efficient manner [7]. Traditionally, only the integrated photocurrent could be extracted;

today a large variety of photocurrent preprocessing is available, mak-ing it possible to optimize the photosensor acquisition parameters to a given problem. In the following, a few examples of such photocurrent preprocessing are presented.

5.3.1 Photocharge integration in photodiodes CCDs

The simplest type of photocurrent processing is the integration of the photocurrent during a certain time, the exposure time. In this way an integrated charge is obtained that is proportional to the number of photons incident on the pixel’s sensitive area during the exposure time.

This functionality is very easy to implement by employing the capaci-tance of the device used for generating the electric field for photocharge separation. Figure5.6illustrates this principle for the two most impor-tant photosensitive structures, thephotodiode(PD) and the metal-oxide-semiconductor (MOS) capacitor as used in the charge-coupled device

5.3 Photocurrent processing 121 a

p-type silicon substrate space-charge region

n+

conductor

b

p-type silicon substrate space charge

p+ ^channelstop

oxide conductor

Figure 5.6:Cross sections through the two major types of electrical field gener-ating and charge storing devices in semiconductors:aphotodiode, consisting of a reverse-biased p-n junction;bMOS capacitance, consisting of a (transparent) electrode on the semiconductor material, separated by a dielectric insulation.

(CCD) image sensors. Both devices are easily fabricated with standard semiconductor processes.

A photodiode consists of a combination of two different conductiv-ity types of semiconductor, as illustrated in Fig.5.6a. In the junction be-tween the two types of semiconductor, an electric field in the so-called space-charge region exists, as required for the separation of photogen-erated charge carriers. At the same time, this space-charge region has a certain capacitance, varying with the inverse of the space-charge region width. Photodiodes are typically operated by biasing (“resetting”) them to a certain potential and exposing them to light. Photocharge pairs en-tering the space-charge region are separated in the PD’s electric field, a photocurrent is produced, and the photocharge is accumulated on the PD’s capacitance, lowering the voltage across it. After the exposure time, the residual voltage is measured, and the voltage difference com-pared with the reset voltage level is a measure of the amount of light incident on the pixel during the exposure time.

The MOS-capacitance illustrated in Fig.5.6b consists of a thin layer of oxide on top of a piece of semiconductor. The oxide is covered with a conductive material, often a metal or highly doped polycrystalline silicon (polysilicon). As in the case of the PD, the MOS structure is biased to a suitable voltage, leading to a space-charge region of a certain extent in the semiconductor. Again, photocharge is separated in the electric field and it is integrated on the MOS capacitance, collected at the interface between semiconductor and oxide.

A typical value for the PD and MOS area capacitance is 0.1 fF/µm². Assuming a maximum voltage swing of a few volts, this implies a stor-age capacity of a few thousand photoelectrons perµm². Once this stor-age capacity is exceeded, additional photocharge in the corresponding

M M M

M C M

V V V

program

reset

reset

Figure 5.7:Schematic diagram of the offset pixel with current source transistor Mcur, reset transistorMr, row-select transistorMsel, and sense transistorMs. The value of the offset current is stored on the switched offset memory capacitor CMwith the programming switchMp[16].

pixel starts to spill over to neighboring pixels. This effect is called blooming, and well-designed image sensors provide special collecting (“antiblooming”) structures for a reduction of this effect [15].

5.3.2 Programmable offset subtraction

Several machine vision and optical metrology problems suffer from small spatial contrast [7]. In such cases in which the spatial signal modulation is small compared to the background light level, one would profit from anoffset subtractionmechanism in each pixel. This can be realized, even programmable in each pixel, with the offset subtraction mechanism proposed by Vietze and Seitz [16]. Each pixel contains a photodiode in series with a programmable current source, as illustrated in Fig.5.7. This current source is easily realized with a MOSFET, whose gate voltage can be preset to a certain voltage level with a second MOS-FET, and by using a capacitance for the storage of this gate voltage. The MOSFET is operated in the so-called weak-inversion regime, where the drain current depends exponentially on the gate voltage; the current typically doubles with each increase of gate voltage by about 30mV. In this way, the offset current can be varied easily between 1 fA up to sev-eral tens ofµA [17]. The same integration mechanism as presented in Section5.3.2is employed for the collection of signal photocharge, rep-resenting the difference between total photocharge minus offset pho-tocharge. Using this method, a dynamic range exceeding 150dB can be reached, and several interesting applications can be realized very eas-ily. An example of this is a simplechange detector, implemented as a two-stage process. In a first stage, the offset current in each pixel is pro-grammed such that the net result is zero; the offset currents effectively cancel the local photocurrents. In a second stage, the image is simply observed for nonzero pixels, indicating that there was a change in the

5.3 Photocurrent processing 123

a b

Figure 5.8: Application example of the offset pixel—motion detector realized with a 26×28 pixel CMOS image sensor [17]:asensor image of a simple scene (black letters “PSI” on white paper) after adjusting the pixels’ individual offset current to a medium gray level;bsensor image after moving the scene slightly downwards and to the right. Pixels with changed values appear either black or white.

present scene compared to the original “reference” scene: a change in the scene has occurred!

The realization of such a change detector is illustrated with an ex-perimental offset pixel image sensor with 28×26 pixels, fabricated with standard CMOS technology [17]. In Fig. 5.8a the result of offset can-cellation for a stationary scene containing the letters PSI is shown: a uniform gray picture. Once the object is moved (the letters are shifted downwards), the resulting pixels appear as bright where the dark object was, or as dark where the bright background was, see Fig.5.8b.

5.3.3 Programmable gain pixels

Another local operation desirable in an image sensor is the individual multiplication of the photocurrent with a programmable factor. This can be achieved with a modification of a simple electronic circuit called current-mirror, consisting of two transistors. In the standard configu-ration, the gate terminals of the two transistors are connected. In the modification proposed in Vietze [17], a voltage difference between the two gates is applied, as illustrated in Fig.5.9. This voltage difference is either fixed (e.g., by semiconductor process parameters), or it can be im-plemented as individually programmable potential differences across a storage capacitor. The photocurrent produced by a photodiode in the first branch of the modified current mirror results in current in the sec-ond branch that is given by the photocurrent times a factor. By using a similar physical mechanism as in the offset pixel, the gain pixel shows a current doubling (or halving) for each increase (decrease) of the voltage difference by about 30mV. In this way, current multiplication (division) by several orders of magnitude can easily be obtained. As before, the

V_diff

VDD V_reset V_DD

reset

select out

Figure 5.9:Schematic diagram of the gain pixel, consisting of a modified cur-rent mirror [17], with which a photocurcur-rent multiplication with a factor ranging between 10⁻⁴up to more than 10⁴can be realized.

multiplied photocurrent is integrated on a storage capacitor and read out using conventional circuitry.

An application of this is a high-sensitivity image sensor as described in Reference [17], in which each pixel has a fixed gain of about 8500.

In this way, a sensitivity (see Section5.5.1for the definition) of 43 mV per photoelectron has been obtained, and an input-referred rms charge noise of better than 0.1 electrons at room temperature. As will be discussed in Section 5.5, this impressive performance must come at a price. In this case it is the reduced bandwidth of the pixel, reflected in the low-pass filter characteristics at low photocurrents with response times of several milliseconds.

5.3.4 Avalanche photocurrent multiplication

The multiplication mechanism described in the foregoing is based strict-ly on the use of electronic circuitry to achieve gain. In semiconductors there is a physical mechanism that can be exploited to multiply charge carriers before they are detected. This effect is called avalanche multi-plication, and it is used in so-calledavalanche photodiodes(APDs) [18].

If the electric field is increased to a few times 10⁵V/cm, charge carriers are multiplied with a strongly field-dependent factor. Depending on the specific doping conditions in the semiconductor, the necessary electric fields correspond to breakdown voltages of between a few volts and a few hundred volts. The strong dependency of the multiplication fac-tor on voltage is illustrated with a model calculation for a breakdown voltage of 40V, shown in Fig.5.10[19].

The APDs are commercially available and, because of high achiev-able gains, they are even suitachiev-able for single-photon light detection [20].

Due to the unusual voltages, the complex voltage stabilization/homoge-nization circuits, and the nontrivial readout electronics in each pixel, most APDs are only of the single-pixel type. The development of APD

5.3 Photocurrent processing 125

0 10 20 30 40

1 10 100 1000

1 10 100 1000

Voltage [V]

Avalanche multiplication gain

Breakdown voltage = 40 V Exponential factor = 4

Figure 5.10:Empirical relationship between applied voltage and obtained cur-rent gain in an avalanche photodiode, for which a breakdown voltage ofVB=40 V and an exponent ofn=4 have been assumed.

line and image sensor arrays has only just started. Nevertheless, the fabrication of reliable APD image sensors with CMOS processes is an active topic of research, and promising results have already been ob-tained (see, for example, Mathewson [21].

5.3.5 Nonlinear photocurrent to signal voltage conversion

Image processing algorithms are often motivated by solutions found in biological vision systems. The same is true for different types of photodetection strategies, especially for the realization of image sen-sors offering a similarly large dynamic range already inherent in animal vision. The fact that the human eye shows a nonlinear, close to logarith-mic sensitivity has been exploited, for example, in the artificial retina described in Mahowald [22].

The realization of CMOS pixels offering a logarithmic sensitivity is particularly easy to achieve: One can use the logarithmic relationship between gate voltage and drain current in a MOSFET operated in weak inversion, already described in Section 5.3.2. The resulting pixel ar-chitecture, shown in Fig.5.11 and exploited in CVA1 [Chapter 8], is particularly easy to implement in a CMOS process because a pixel con-sists of just a photodiode and three MOS transistors [23]. A typical photoresponse of about 40mV per decade of optical input power is obtained with such logarithmic pixels, and their useful dynamic range exceeds 120dB. Practical examples of scenes requiring such a high

dy-M M M V

select

sense

sel log

Figure 5.11: Schematic diagram of a pixel with logarithmic response, consist-ing of just one photodiode and three MOSFETs. Implemented with a standard CMOS process, such a pixel shows an output voltage increase of about 40 mV per decade of incident light power.

a b

c d

Figure 5.12: Pictures taken with a small-area logarithmic image sensor with 64×64 pixels: aElectric light bulb where the glowing filament and the back-ground are visible simultaneously;bBack-illuminated scene of a portrait in front of a window;cParking garage application with its notoriously high dynamic range (headlights compared to dark corners) and lowaverage light levels;d Welding application in which the object and the welding arc can be observed at the same time without blooming.

namic range are illustrated in Fig.5.12, with the actual measurements taken with a logarithmic image sensor exhibiting 64×64 pixels. In the image of a light bulb, the glowing filament as well as the background are clearly visible at the same time. Back-illuminated scenes, such as a por-trait in front of a window, are dreaded by photographers, but they are