When it comes to digital cameras, noise is a primary topic of concern. Whenever a new camera comes out, one of the first questions that usually gets asked is "What are the noise characteristics of the camera?" If an image has too much noise, the photographer who took the image usually wants to decrease the effects of the noise. A person doesn't have to be around digital photography very long to figure out that noise is, generally, not a good thing (although there are exceptions such as when a photographer adds noise to reduce a banding problem).
Since noise is such an important topic, this series looks at the details of noise and several techniques that can be used to reduce the effects of noise.
In order to understand noise, one must first understand the concept of the signal. A signal is simply a measure of some property of an object. For example, when a driver travels down the highway, the speedometer tells the driver how fast she is going. The speedometer is providing a measure of speed. Thus, the speedometer is providing the driver with a signal of how fast the car is moving. If everything is working properly, this signal is an accurate representation of the speed of the car. So, when the car is traveling 65 MPH, the speedometer will provide the driver with a signal that the car is going 65 MPH. In short, the driver can depend on the speedometer to provide accurate signals, which the driver can utilize to make decisions (e.g., I need to slow down).
Now, for those of you that are old enough, think back to the days before digital speedometers. Remember the old needle on the dashboard that told you how fast you were going. One of the problems with the needle type speedometer was that the needle could vibrate a bit if the car was driven on a very rough road. The car might be going 65 MPH, but the needle might bounce between 62 and 68 MPH. This bouncing around is the noise in the speedometer. If you think about the noise in the speedometer reading, two things become apparent. First, the noise is nothing more than error in the system. In other words, if the car is going 65 MPH, but the speedometer briefly shows 67 MPH, there is an error of 2 MPH. Second, the bouncing around of the speedometer needle is random. Basically, this is the essence of noise. Noise is simply random error in a measurement system.
In a camera sensor, there is an analogous situation. The sensor is made of many pixels. Each pixel is essentially a miniature light meter. Thus, each pixel produces a signal that indicates how much light reached the pixel during an exposure. In a perfect world, the signal from the pixel would indicate exactly how much light reached the pixel. Unfortunately, this is not a perfect world as each pixel has not only a signal but also noise. That noise causes random error in the information that each pixel reports. Due to the noise, sometimes a pixel reports more light than it actually received. At other times, a pixel reports less light than it received. In either case, this causes two categories of noise.
Noise appears as a grain or as numerous tiny specks. When the specks vary in tone only, the noise is referred to as luminance noise. Figure 1 shows a 100% crop of an image that has primarily luminance noise.
Digital photographers tend to talk a lot about noise. The problem is that noise, by itself, has little meaning. Just because one pixel has more noise than another does not mean that the pixel with the larger amount of noise will produce a lower quality image. This is demonstrated in Figure 3. In this figure, pixel B has a bit more noise than pixel A. However, pixel B also has a lot more signal. Consequently, the signal to noise ratio (SNR) of pixel B is better than that of pixel A. As a result, pixel B will produce a better quality image than pixel A -- even though it has more noise.
In short, it is not the noise that is important; rather, it is the SNR. A poor (low) SNR will result in noise being visible in an image. A good (high) SNR will produce a higher quality image.
Luminance and color noise are caused by four primary mechanisms. This results in four types of noise: photon noise, dark current noise, readout noise, and pattern noise.
At a simple level, we can say that pixels measure light by counting photons (small packets of light). However, even with a stable light source (i.e., a source that emits a constant amount of light), the number of photons that arrive at any point in a given period of time will fluctuate. Thus, if we have two pixels that are right next to each other and they are illuminated by the exact same light source, the two pixels are likely to receive a different number of photons. This creates what is called photon noise. It is also referred to as Poisson noise. This is because the number of photons that arrive at any one point, over a period of time, follows a statistical, Poisson distribution.
One thing that needs to be understood, with respect to photon noise, is that it has nothing to do with the camera or sensor. It is an inherent property of light. This is not a noise source that the digital camera engineers can eliminate. The engineers may be able to refine how a digital camera deals with photon noise, but the noise will always be there.
One of the properties of photon noise is that the amount of the noise increases when the signal increases. However, the photon noise increases at a slower rate than the signal. Thus, as the signal increases, the SNR of photon noise improves. What this means is that photon noise will be a bigger problem in the shadows and with underexposed images. This is one of the reasons that many photographers maximize the exposure when using digital cameras (for more information on maximizing the exposure, see, Digital Exposure).
A few paragraphs above, it was mentioned that pixels measure light by counting photons. This is a simplified view. What actually happens is that the photons reach the sensor and excite electrons (the photons do this by transferring energy to the electrons). These excited electrons are freed from the molecules to which they are attached. When a voltage is applied, these free electrons create a current and flow into a capacitor. This creates a charge on the capacitor. The charge is then measured to create a voltage measurement. This voltage measurement is processed by the camera to determine how much light reached the pixel during exposure.
The problem is that electrons can also be freed by heat in the sensor. This is because heat transfers energy to the electrons. This excites the electrons and frees some of them from the molecules. The electrons that are freed by heat combine with the electrons that are freed by the photons. This creates a type of noise known as dark current noise. This type of noise exists even when the sensor is not receiving any light due to the fact that dark current noise depends on heat not light.
Dark current noise increases as the temperature of the sensor increases. This is why some sensors, that are used for very exacting applications (e.g., military or scientific), are cooled to very cold temperatures (e.g., with liquid nitrogen). This reduces the dark current noise and improves the SNR of the sensor.
After the freed electrons flow into a capacitor and the voltage of the capacitor is measured, the voltage is amplified before any further processing is performed. However, the amplifiers that perform this function are not perfect. Any particular amplifier will vary the amount it amplifies the voltage from one exposure to the next. Sometimes, an amplifier will amplify a little bit more; at other times, the same amplifier will amplify a little bit less. This variation in amplification, within the same amplifier, creates what is called amplifier or readout noise.
In a CMOS sensor, each pixel has its own amplifier. Unfortunately, there is variation between the amplifiers. As a result, the amount the voltage is amplified varies from one pixel to another. This is called fixed pattern noise.
CCD sensors have only one amplifier for the entire sensor. Thus, the voltages from all of the pixels are run through the same amplifier. Thus, fixed pattern noise is not a problem for CCD sensors.
The temperature of the sensor affects the noise level. Higher sensor temperatures create higher noise levels. This is primarily due to dark current noise. In most cases, this isn't a problem as normal use of a digital camera doesn't cause high levels of heat. However, it can become an issue if one, for instance, leaves a camera in a glove compartment of a hot car in the middle of summer.
Another factor that affects the noise level is the ISO. The way that a digital camera increases the ISO is to apply a greater amount of amplification to the voltages that come from the pixels' capacitors. Thus, low ISO settings apply a small amount of amplification while high ISO settings apply a greater amount of amplification. Since high ISO settings apply a greater amount of amplification, less light is needed to get a shot. The key here is that the SNR improves as a pixel gets more light (largely due to photon noise). Conversely, the SNR degrades as the amount of light collected by a pixel decreases. Consequently, high ISO shots (that require less light) have poorer SNRs than low ISO shots (that get more light). In addition, the large amounts of amplification that are applied to high ISO shots amplify the noise as well as the signal. The result is that high ISO shots have more noise than low ISO shots. This can be seen in Figures 4 and 5. Figure 4 shows a crop of an image shot at ISO 80. Figure 5 shows a crop of the same object shot at ISO 1600. Clearly, the higher ISO shot has much more noise.
There are two areas where noise tends to be the most noticeable: areas of little detail and shadows.
Noise tends to stand out in areas of little detail, such as bland skies, simply because there is no detail to hide the noise. Little detail means little tonal and color variation. Thus, the tonal or color variation caused by noise is more conspicuous.
Noise is also problematic in the shadows. Shadows are areas of little light. Consequently, the pixels in the shadows collect only small amounts of light. Since pixels that have received small amounts of light have poor SNRs, the shadows are likely to have higher levels of noise than the other areas in an image.