Noise. We all know about it. We all hate it. Digital noise, the one thing about our photography we as photographers collectively love to hate. It degrades image quality, can detract from our vision, shows up at the most unexpected times in the most unexpected places, and can often be difficult to eliminate.
As noise is one of the most ubiquitous “annoyance” factors in modern digital photography, I decided it should be the focus if a few knowledge center articles. For one, it is an often misunderstood factor of image detail and image quality, and two because there are many ways to deal with noise, and deal with it quite effectively, in order to improve the quality of your images. Before I delve into the myriad noise management techniques (which could quite possibly extend this series into a multitude of parts), I thought it was important to properly describe what noise is, where it can come from, and why it might not always be a bad thing.
Noise is a modern digital camera is an unavoidable fact. It is, primarily, the result of the physical nature of light, and therefor in many respects is beyond our control, and even beyond the control of camera and sensor manufacturers. But what is noise, really? Historically, in the film era, noise was called grain, and yet grain was not nearly as universally loathed as noise. Why?
To answer these questions, it might first help to understand the difference between film photography and digital photography. With film, you are using an analog material to sense light and record image detail, where as with digital photography, you are using an electronic image sensor that uses digital electronics to sense and record image detail. One of the key differences between analog and digital processes is randomness. With film, while it still suffers from noise, it is microscopic crystals of silver halide, which maintain non-uniform size and shape, that determine the size and shape of each “grain” of noise in a photograph. In contrast, a digital sensor is manufactured as a neatly ordered perfect grid of pixels of well defined, well known size and shape.
As humans, our minds are designed to recognize patterns. That’s ultimately what our minds are, neural networks tuned for pattern recognition. Though we may not always recognize it on a conscious level, our minds recognize that the noise in a digital image is ordered and well defined, where as the grain of a film photograph is random and largely undefined. True randomness, however, tends to be more pleasing to the mind. We recognize patterns, but we revel in randomness. Which is why so many photographers expend so much effort in an attempt to reduce digital noise, and why so many film photographers explicitly sought out special films with various kinds of visible grain. It’s also the reason why we create so many “rules” about composition…like off-centered subject positioning and uneven triangular distribution and random yet convergent lines things like that. Order, and chaos, used simultaneously despite how much at odds they are with each other.
Philosophy aside, there are two key reasons why noise exists in our photos, deriving from the two key sources of noise. In every digital photo there are two primary forms of noise. The first and most prevalent is photon shot noise, which derives from the random physical nature of light itself, and therefor cannot be directly controlled. The second and less common, but also frequently the more frustrating, is read noise. Read noise is caused by the nature of the electronics in our digital cameras, and in many ways is under our direct control. Neither form of noise can be entirely eliminated, however both can be mitigated and reduced to one degree or another.
Photon shot noise is noise caused by the random nature by which individual photons strike the pixels of an image sensor. As I mentioned before, digital sensors utilize an ordered grid of pixels. When an exposure is made, at a macroscopic scale the shutter opens allowing light to reach the sensor. At a microscopic scale, a “rain” of photons pours through the lens down onto the sensors pixels. The total amount of photons are largely randomly distributed, however there is a certain probability with a given rate of error that any given photon will strike a given pixel. Not every pixel will encounter the same number of photons. Even though two pixels may be direct neighbors, and should both encounter exactly the same number of photons, one is likely to encounter slightly more, and the other slightly less. This difference, this deviation, is what results in the vast majority of noise in digital photographs.
Under less illumination, less light enters the lens and less light reaches the sensor. The relative deviation of photon shot noise vs. how many photons each pixel theoretically “should” encounter increases as light levels drop, and decreases as light levels increase. The average deviation grows slowly as light levels increase, which is why the appearance of noise becomes less when you photograph in good light. From a mathematical standpoint, the amount of photon shot noise in a photo is roughly equal to the square root of the signal strength. Oh, what is signal strength? Well, here is where things get a little more technical, and less philosophical. 😉
Every photograph we make with a digital camera is originally the result of an electronic signal being formed within an image sensor as a result of light striking the sensor’s surface. As photons strike each pixel, an electronic charge builds up. Taken as a whole, each individual charge in each individual pixel observed together in their neat little order produces an image. The maximum charge level of each pixel determines the signal strength of the pixel. The randomness of photons results in slight deviations from what that charge should be. Mathematically, that deviation is the square root of the charge level.
An average full-frame digital image sensor today supports an atomic charge of about 60,000 electrons (or ‘e-‘). An average exposure might result in the average charge of each pixel being 45,000e- (some more, some less, depending on how bright or dark that region of the image is.) The amount of photon shot noise in the image, the intrinsic noise, is:
N = sqrt(S)
N = sqrt(45000)
N = 212
Where N is the noise and S is the signal strength. With a signal strength of 45,000, our noise is 212. Our signal to noise ratio (SNR) then is 45000:212, or about 212:1! What happens if we expose the same scene again, but increase the shutter speed one stop? We now have half as much light reaching the sensor, so our signal strength is 22,500e-, and our noise is 150e-, an SNR of, yup, 150:1. Increase shutter speed by another stop, and our SNR becomes 106:1. As you can see, as we reduce the light levels reaching the sensor, the ratio between our signal strength and noise drops. The lower the ratio, the more apparent noise will be. This becomes more apparent if we take the noise to signal ratio, which becomes a percent. A signal of 45000e- has 0.47% noise, 22500e- has 0.67% noise, 11250e- has 0.94% noise, 5625e- has 1.3% noise, etc. A simple graph demonstrates a little better (note, the primary scale is logarithmic):
Increasing shutter speed by one stop is similar to keeping the shutter the same, and narrowing the aperture by a stop…or reducing light levels by a factor of two and increasing ISO by one stop. Reducing light levels, and increasing ISO by a similar factor, results in more noise. Based on the explanation of noise above, it should become apparent now WHY increasing ISO results in more noise. The ISO setting itself is not actually the cause…the lower amount of light is the cause. Increasing ISO simply makes the fact that there is more noise because there was less light more apparent. We will get into this in more detail in the next part of this series with some visual examples for comparison.
Photon shot noise is the most prevalent form of noise in a photograph, but it is not the only form. The nice thing about photon shot noise is its randomness…it tickles our pleasure at randomness, and as such, when it is at moderately low levels, we often don’t mind it. It can resemble film grain more so than digital noise, so we can be ok with it. A more sinister form of noise lurks within digital sensors, however. READ NOISE, and its derivatives. Read noise is noise introduced by the electronic current flowing through the image sensor and the various other electronic parts of a digital camera, and it can often present in very non-random ways. Lumped together with read noise are a few forms of noise that are not actually electronic in nature…things like salt and pepper noise and hot pixels, which are the result of microscopic defects on the silicon materials the sensor itself is made out of.
The most atrocious form of noise is banding. Banding is a periodic oscillation, regular brightening and darkening in thick or thin bands, both horizontal and vertical, and can occur in stationary and non-stationary forms. Banding results in very recognizable, and unnatural patterns that our minds pick up in a heartbeat. Banding noise usually does not conform with the structure of detail in our images, and as such results in the most loathed interference that noise can create. Thankfully, banding noise, like all forms of read noise, exists in the lower fraction of our image signal…in the darker shades, the shadows. It will usually not exhibit unless we, for whatever reason, must “lift” the shadows during processing to bring out shadow detail.
Banding aside, many forms of read noise are easily manageable. Salt and pepper noise and hot pixels are largely random in their occurrence. Salt and pepper, which are caused by pixels either spiking “hot” (full brightness) or blacking out (full blackness) are often managed automatically by our image editors (like Adobe Lightroom), and if not, they will usually disappear when we scale our images down. Hot pixels, and dead pixels (which can also be a source of pepper noise) are a fixed form of noise, resulting from tiny defects in the silicon of the sensor. Being fixed, most cameras these days offer a feature that will map, memorize, and automatically eliminate hot pixels, and even without such a feature, they can be removed by generating your own hit pixel maps that can be subtracted from your images during processing. Another form of largely random noise is caused by “dark current”, a low level electrical current that always flows through an image sensor. Dark current is largely managed these days by automatic corrective circuitry built right into the sensor. Any remnant not delt with by that circuitry usually shows up as random noise in the deep shadows, and is much like photon shot noise in aesthetics.
For the most part, if you never lift your shadows while processing your images, most forms of read noise will rarely be a problem. In general, read noise is most prevalent at low ISO settings, like ISO 100 and 200, and as ISO is increased, read noise levels drop considerably. At high ISO settings like 800, 1600, and higher, read noise becomes such an infinitesimally small percentage of the image signal that it largely becomes a non-factor, but may leave behind a remnant that is relatively easy to clear up: color noise. Color noise results from the basic design of most image sensors, the bayer color filter array. Unlike film, which was able to detect all primary colors at all locations, bayer sensors can only sense one color at each pixel. Pixels are arranged in pairs of rows, one row of red, green, red, etc. and one row of green, blue, green, etc. Four of these colored pixels are blended to form a single RGB pixel in a JPEG or TIFF image. Because each pixel of the sensor can only sense a single color, a form of color noise will often exhibit as reddish, greenish, and bluish blotches when you load up a RAW image to edit it. For the most part, color noise is automatically managed by tools like Adobe Lightroom, so you will rarely if ever actually have to deal with it.
So there you have it…noise. It’s the result of two processes, one purely natural, one electrical. Photon shot noise, the most prevalent form, derives from the random physical nature of light and the way it interacts with a discrete grid of sensor pixels. Read noise, comprised of banding, hot and dead pixels, “salt and pepper” noise, and dark current noise, is the result of electrical current flowing through the sensor circuitry, and due to microscopic imperfections in the sensor material itself. Both types of noise can be mitigated, however neither can be completely eliminated.