eNEWSLETTER SIGNUP

Q+A CENTER

The majority of motion analysis and industrial imaging applications require synchronous image acquisition of all pixels while often operating under difficult lighting conditions. Through innovative pixel design, CMOS image sensors meet these preconditions, establishing a solid basis for future requirements. Custom design enables sensor solutions optimized for specific applications.

Over the last several years, CMOS sensors have established themselves for fast and high-resolution image acquisition. They can be produced at low cost on standard manufacturing lines, monolithically integrating analog and digital functions on the same chip. By saving external components with their inherent board space overhead and mounting expenditure, such subsystems primarily reduce cost. At the same time they consume less power because drivers for external circuits and components are obsolete.

The advantages and successes of industrial image processing are widely known—leading to ever more novel applications. These novel applications put even stronger demands on image capture systems. The new requirements surpass speed or resolution (number of pixels). In the industrial arena there is a growing demand for synchronism and operation under unfavorable lighting conditions.

Image acquisition systems featuring a global shutter and the widest possible dynamic range are in demand. This demand shifts the center of the design targets of application-specific CMOS image sensors from integration of additional functionality to novel ways of shaping the pixels and the electronic circuitry in close proximity.

Global Shutter

Industrial image processing, when applied to process control, requires the capture of the state of a process at a certain point in time to detect a fault and determine the appropriate corrective measures. This requires cameras with synchronous acquisition of all pixels and holding this information uncorrupted until the entire image information has been read out. This was feasible with interline transfer CCD imagers without many problems. The CCD photodiodes, after exposure, transferred their charges simultaneously to CCD storage elements. The photodiodes were ready for the next exposure immediately afterwards, enabling pipelined global shutter operation. In this scheme, every pixel has two storage elements: the diode for integrating the photon-induced charges, and one within the CCD MOS capacitor chain.

CMOS imagers with the most simple active pixels architecture contain three transistors, the so-called 3T cells (Figure 1a). During operation the photodiode is first reset and charged to Vdd (supply voltage), then exposed to light (which will lower the diode voltage) and finally the resulting signal is taken out via a source follower and delivered to the output line. The only storage element present in a pixel, however, is the photodiode. Therefore, the whole pixel array cannot be read out at the same time (like a storage element) but only line by line, resulting in a rolling shutter operation. This means that fast moving objects will become distorted and a flash exposure might illuminate only a part of the scene.

A global shutter operation requires, as in CCD sensors, a second storage element for the photodiode voltage. In a four-transistor cell (Figure 1b) it is present, but for a different purpose: to reduce the kTC noise (noise related to charging a capacitor, also referred to as reset noise. K=Boltzman's constant, T=absolute temperature, C=capacitance). A charge is transferred inside the pixel from the photodiode via a transfer gate to an FD (floating diffusion) point, acting as a small capacitor. The sensor element is a pinned photodiode reducing the dark current. Thus two storage elements are available: the photodiode and FD capacitor. Under normal operating conditions they are used for noise cancelation by correlated double sampling (CDS) since the reset level signal is held in FD. The question is whether the FD also is suited for storing the image. The answer is yes if the following steps are performed: 

• Reset all FDs
• Transfer all photodiode charge simultaneously to the FDs
• Sample all FDs and read out the image, which unfortunately prohibits CDS
• Capture the next image

The major drawback is that the FD itself resembles a photodiode and as such is sensitive to light. For long readout periods compared to the integration time this leads to unwanted parasitic light sensitivity. This can be reduced to about 0.5 percent by appropriate countermeasures such as light screens, potential barriers within the silicon and microlenses directed at the photodiodes away from the FD.

Being a surface potential barrier, the FD has a large leakage current but, at the same time, the FD capacity is very small. The dark current therefore will corrupt the signal for long readout times. So it's worthwhile to relieve the FD from the long term storage tasks and hand this over to an additional large capacitor C/sample. This leads to the six-transistor cell (Figure 1c), resulting in pipelined global shutter operation with low parasitic light sensitivity. By a double sampling of the voltage on C/sample and the FD reset value the fixed pattern noise (FPN) can be corrected on chip. Finally, the 7T cell also can permit a correlated double sampling (CDS) for noise suppression.

Increasing Dynamic Range

The linear dynamic range of an image sensor is calculated as the ratio of the linear output saturation level and the dark noise level. With this, the available methods for increasing the dynamic range already are suggested: increase the saturation value and lower the noise floor. However, the question arises whether a substantial increase of the linear dynamic range makes sense at all. The human eye, for instance, is able to capture differences in illumination spanning more than 100 dB but, on the other hand, only can differentiate about 64 shades of grey, more or less independent of the absolute light level. With this observation in mind, increasing the dynamic range of CMOS image sensors will result in a sensor with non-linear conversion characteristics of light intensity to an output voltage. Non-linear behavior can be achieved by the following methods: 

• Logarithmic photodiode response
• Emulation of logarithmic characteristics by straight-line approximation
• 4T pixel tricks
• Smart reset pixels
• Combination of different pixel sizes and exposure times
• Logarithmic Characteristics

Logarithmic response is the non-linear correlation between photodiode voltage and photocurrent that exhibits an ascending slope of 60 mV per decade of light intensity. A voltage swing of 360 mV covers six decades or 120 dB. The extremely small currents in low light cause substantial readout delays due to the necessary recharging of the photodiode capacitance. Another serious drawback is the high fixed-pattern noise due to the absence of a dark reference.

straight line sections
A sufficiently accurate approximation to the logarithmic characteristic is achieved by integration of linear segments of the characteristics in standard 3T pixels (Figure 2). The photodiode is partially reset by additional reduced-amplitude pulses at the reset gate. This permits a flexible adaptation of the linear sections' bends and slopes to the sensor's application. However, fitting the 4T pixels with this functionality is more complex since the charges must be transferred via the transfer and reset gates.

 4T Pixel Tricks

A 4-transistor pixel (Figure 1b) has two storage elements, photodiode and FD, both of which can be light-sensitive. While the parasitic light sensitivity of the FD hampers global shutter operation, it can be used for capturing images. Reading out the dark areas of the photodiode is done by correlated double sampling, whereas the light areas are double-sampled by the signal of the FD "photodiode." In this regard, dark noise has no impact. On the other hand, the FD is not a good photodiode because of its very large leakage current.

Smart Reset Pixels

An additional comparator and logic circuitry inside each pixel (Figure 3) resets the photodiode after it overshoots a certain value during exposure. As its output value the pixel delivers the number of reset events and the residual voltage remaining in the photodiode after the last reset. This internal feedback establishes a very wide dynamic range. The additional circuitry consumes area resulting in less efficient or larger pixels for a given CMOS fabrication technology.

Mixed Forms

Figure 4 shows another approach to higher dynamic range: by deploying photodiodes of different sizes in the same image sensor. Large pixels capture the dark areas whereas the small ones are used for the lighter areas. Both pixel classes can share the same readout circuitry so that the additional silicon real estate expenditure remains limited. The sensitivity ratio cannot be altered since it is geometrically defined by the ratio of the two pixel areas.

The approach is different in a time-controlled setup as shown in Figure 5 for a color sensor with modified Bayer pattern. The pixel is composed of a red (R) and a blue (B) sensor area, each having their own average exposure times, plus a green Gs (green short) area with short exposure and another green Gl (green long) with long exposure. No additional control lines are required within the pixel. Such multiplexed operation, by driving the transfer gate, cannot be arbitrarily expanded because the FD capacitance will grow with each added connection.

The design of CMOS imagers with global shutter and high dynamic range is determined by multiple parameters: technological boundary conditions, application demands and, last but not least, by the solution-oriented capabilities of the manufacturer and designers. They must bridge customer demands and technical feasibility, continuously pushing the limits. Custom image sensors will offer better solutions than off-the-shelf products for certain application requirements. This, of course, requires investment that must be justified case by case by a rigid ROI analysis.

Guy Meynants, Ph.D. is co-founder and CEO of CMOSIS (Antwerp, Belgium). He earned his Doctorate in electronics from Catholic University in Leuven, Belgium. He was a founder of FillFactory NV (later acquired by Cypress), where he developed various CMOS sensors for industrial and space applications.