Sensor design part 2


In the second of his two-part series on the processes behind sensor

IN THE first part of this article (AP 14 July), I looked at the essential operating principles of a CMOS image sensor. Armed with that knowledge, I now will look at the design trade-offs that

 quantum efficiency

sensor designers face and how these affect the cameras we buy.

The first matter to be examined is the metrics that sensor designers work with to decide whether a sensor is  good  or not.

Quantum efficiency (QE): Despite the grand name, this is a very simple concept. If a sensor is subjected to, say, 100 photons of light, how many will be counted? Being  counted  requires that the photon strikes a silicon atom and releases a photoelectron, and that the photoelectron so released finds its way to the gate of the read transistor.

The QE is thus the percentage of incident photons that do this. The QE is important because it is a measure of how much light available is being wasted by the sensor, and since the main component of image noise, the photon-shot noise, depends directly on the number of photons counted (the signal-to-noise ratio in an area is actually the square root of the number of photons counted), this affects directly the low-light performance of a camera. Current cameras have a QE of 40-60%, which represents something like a doubling of efficiency over the past ten years.

Read noise: Nothing s perfect, and even if all the incident photons were to make their way to the read transistor it is not guaranteed that the number will be counted with perfect accuracy. In practice there is an error, which varies randomly between read operations (and therefore from pixel to pixel). The error is constant (at a given ISO setting), which means it is usually of small significance when there are a lot of photons (the highlights), but is of more importance when there a few (the shadows).

Pixel response non-uniformity (PRNU):

In a perfect world, all pixels and their associated readout circuitry would be perfectly identical and would respond to light in precisely the same way. Not only is this not the case, but the difference in pixel response can be influenced by which row or column the pixel is in, resulting in distracting  tartan  patterns in the noise. Sometimes, in sensors with a multi-channel readout, differences in the channels can cause regular repeating patterns of noise.

Dark current noise: As well as photons, thermal energy can release electrons from

the structure of the silicon. Once released, these electrons are indistinguishable from the ones due to light. At normal exposures, there are not enough thermal electrons to significantly affect the image, but for longer exposures this may not be the case. The use of live view and video modes will tend to result in the sensor becoming warmer, and increase the problems of thermal noise.


There are several factors that affect quantum efficiency. We will deal first with those that affect the silicon part of the sensor. The photodiode has to be deep enough so that the incident light does not pass straight through it. Red light penetrates up to 6 microns into the surface of silicon.

If the photodiode is not at least that deep, some of the incident light will be wasted and quantum efficiency reduced. Next, it must be ensured that the released electrons are counted. The electrons are shepherded


towards the read transistor by an electrical field, which is maintained by implanting  doping  elements into the silicon. The profile of this field impacts its effectiveness, and must be carefully controlled. Although it is often said that an advantage of CMOS image sensors is they can be produced on standard memory and processor fabrication lines, these two factors alone ensure sensor processes are these days very specialised.

If the silicon is as efficient as possible, then the overall efficiency depends on the number of photons that actually reach it. Thus there are two layers over the silicon-the colour-fitter array and microlens layer. The function of the colour-filter array is to stop two thirds of the light (the primary colours other than the one that the filter is intended to record). The filters can be made to be a little less discriminating, and the quantum efficiency will increase, but at the cost of a little blindness’, which manifests itself as increased chroma noise.

The role of the microlens is to ensure that all the light heading in the direction of a pixel from the exit pupil of the lens strikes the sensitive part of the pixel. Essentially, the microlens must form an image of the exit pupil that fits completely on the photodiode. This can be achieved either by making the photodiode large, or by making the focal length of the microlens short so that it acts as a  wideangle . The limit to this is the layer of wiring over the sensor chip. To ensure good quantum efficiency, this must be as shallow as possible. Alternatively, light guides can be built in to convey the image from a short focal-length microlens to the surface of the silicon, or  back-side illumination’can be adopted, in which the chip is thinned to expose the rear surface of the photodiodes, and the microlens and colour filter array are situated on that side, opposite the wiring layer. Another way to shorten the focal length is by adding an  inner lens  to produce a two-element arrangement. All these processes add significantly to the cost of the sensor.


The measure of read noise that is important is how it looks relative to the light reaching the sensor. It is therefore normally measured in  electrons , its effect in comparison with the photoelectrons making up the image.

For a given amount of electronic noise in the system, its visible effect is determined by the  conversion gain -the number of photoelectrons that represents a volt of sensor output. The higher the gain, the fewer electrons correspond to a given voltage and the less read noise is visible. Thus, the conversion gain needs to be made as large as possible to minimise read noise. Unfortunately, the total amount of light the sensor can measure is generally determined by the maximum possible output voltage, so a high conversion gain results in a sensor that cannot accept many photons-in other words, it has a high  base ISO’.


To an extent, both PRNU and thermal noise can be controlled by processing after capture. If the pattern of the noise is known (since it is not strictly random noise), then

it can be subtracted from the captured image. In the case of PRNU, this is done by building the sensor with a black border that is shielded from the light. It is assumed that the pixels in the border will share the same characteristics as those in the same row or column, so by measuring these black pixels, the PRNU can be estimated and subtracted away. This is usually done before the raw file is written (raw files aren t always so raw). In the case of thermal noise, a  black frame  with the same exposure time as the image can be taken and subtracted from it.


If it were possible, every sensor would be designed to be perfect. Since this is not possible, the design must in all cases be a compromise between conflicting requirements. Once a fabrication process has been optimised with deep and carefully profiled implants, then the designer has limited options. Quantum efficiency can be raised by improving or changing the colour filtration or by improving the microlenses. The latter might involve moving to a new process with thinner wiring (maybe using copper instead of aluminium), by adding inner lenses or by designing a cell configuration with less wiring-as, for instance, is done in the Panasonic sensors. However, this might have the downside of adding to the electronic noise generated.

So far as read noise is concerned, one part of the equation is to minimise the electronic noise. This can be done by using careful design and superior components, or by placing as much circuitry as possible on the sensor so that signal runs are short. However, this also has a downside, because the on-chip circuitry increases the heat dissipated, thus increasing thermal noise.

The second part of controlling read noise is to raise the conversion gain, which is done by minimising the capacitance of the pixel. As stated above, the trade-off is that the amount of light that the pixels can collect is reduced. The way to maintain the overall light-collecting ability of the sensor is then to reduce pixel size, resulting in higher pixel counts-with the knock-on effect of slower read-out times.

The options offered to a sensor designer are relatively limited, and progress depends on the availability of smaller geometries, leading to higher efficiencies with smaller pixels. That is the essential technological driver behind the  megapixel race’.

Comments are closed.