Photon Noise, Read Noise, and Reset Noise in CCD Image Sensors
In this article, we’ll take a look at electrical and optical conditions that contribute to reduced image quality in CCD camera systems.
In the previous article, we discussed dark noise in CCD sensors, which results from variation in dark current generated by the sensor's semiconductor material. This is an important noise source in CCD applications, and it has a direct influence on system design because it can be effectively controlled by cooling the sensor.
In this article, we'll talk about two other major contributors to CCD image quality (or lack thereof): photon noise and read noise. We’ll also briefly consider reset noise, which is not a major factor in image quality because it is virtually eliminated by a specialized signal-processing technique.
Before moving on, you may wish to catch up with the rest of this series, which covers the breadth of topics below:
- The basics
- Readout and output signals
- Frame rate
In our examination of dark noise, I pointed out that it is governed by the discrete nature of electric charge and follows the Poisson relationship. We use the Poisson distribution to model phenomena that consist of separate, independent events that exhibit unpredictable exact timing but occur at a consistent average rate. If we count a certain number of events and apply Poisson statistics, the standard error associated with the phenomenon is calculated as the square root of the count.
Photons are discrete “particles” of light, and any array of photosensitive elements is subject to the noise—i.e., the random variation—that characterizes the arrival of photons.
Illumination and illumination-induced generation of electric charge are quantum phenomena governed by the discrete behavior of photons and electrons.
Thus, even if a CCD is illuminated by light that appears to be perfectly uniform, pixel-to-pixel intensity variations caused by photon noise will be observed. When I say “pixel-to-pixel,” that can refer to both spatial and temporal variations: neighboring pixels in a single frame will exhibit tonal differences despite uniform illumination, or a single pixel exposed to steady illumination will exhibit tonal differences from one frame to the next.
These variations are quantified by calculating the Poisson standard error, meaning that photon noise is the square root of the total number of incident photons. Thus, if the scene illuminates a portion of a sensor with light that generates an average of 1000 electrons in each pixel during the integration period, the physical nature of this incident light results in noise of approximately 32 electrons RMS. This random variation in photon arrival is imposed by nature and makes it impossible for any image sensor to have zero noise.
I find photon noise particularly interesting, because in theory it affects the human eye as well. Why do we even think of it as “noise” if it’s an inevitable and omnipresent feature of our visual perception? There’s probably a long, complicated answer to that question, but I suspect that the explanation derives primarily from two important differences between human vision and electronic sensors: our eyes have much higher “resolution,” especially in relation to light-sensitive area, and our visual system includes complex filtering mechanisms.
The term “read noise” (or “readout noise”) is a convenient way of referring to other types of noise—namely, thermal noise and flicker noise—that degrade the CCD signal by way of on-chip and off-chip signal-processing circuitry. We reduce off-chip read noise by incorporating standard low-noise design practices and techniques. On-chip read noise is generated by the CCD’s output amplifier.
I discussed read noise in my article on CCD binning, which is a technique that allows us to trade resolution for noise performance. Binning is the process of combining light-generated charge from neighboring pixels; this reduces the effect of read noise because the signal level of a binned pixel increases while the quantity of read noise stays the same.
This diagram conveys the process of combining the charge packets from four separate pixels into one binned pixel.
As with other types of CCD noise, we can report read noise in electrons. I believe that typical values for read noise are in the range of about 2 to 20 electrons RMS per pixel, with CCD systems for non-specialized applications being closer to 20 electrons RMS.
Reset Noise vs. kTC Noise
We touched on this topic a while back in the article that covers correlated double sampling, but I called it “kTC noise” instead of “reset noise.” The former term refers to the origin of this noise: it is influenced by temperature and capacitance in the CCD’s output circuitry. The latter term refers to the effect, since kTC noise causes pixel-to-pixel variations in the CCD signal’s reset level.
The data level depends on the reset level, so random variations in the reset level would translate to random variations in the light intensity associated with each pixel.
A typical value for reset noise is 50 electrons RMS. This would make a significant contribution to total noise if it weren’t for correlated double sampling, which allows the system’s ADC to measure the difference between the reset voltage and the data voltage of each pixel. This technique reduces reset noise to negligible levels.
Next: Managing Noise in a CCD Imaging System
I hope that you’re enjoying our ongoing discussion of CCD noise. In the next article, we’ll explore how photon noise, read noise, and dark noise interact in the overall operation of a CCD imaging system.