To some extent, the popular demand for increasingly higher ISO performance is an expression of photographers’ desire for higher speed. Here, the word alludes to the same concept referred to by lens speed, being the shutter. When digital cameras became widely available, the flexibility of variable ISO values offered digital photographers a unique glimpse into the world of possibilities for low-light handheld photography that would otherwise require using a tripod or flash in the preceding years. With every new release, image sensor manufacturers continue to push their technological advancements towards higher amplification while delivering less electronic noise than the outgoing model. Despite this, it’s helpful to understand what electronic noise is, how it manifests in your pictures, and what you can do to alleviate it, both during and after exposure.
Signal, noise, and the signal-to-noise ratio
Signal. In photography, light is the signal recorded by the image sensor to form a photograph. Specifically, light is the portion of the electromagnetic spectrum that’s visible to human eyes. Increasing the intensity of light by manipulating exposure produces a stronger signal that carries more information.
Noise. Every analogue system designed to receive or transmit signals will exhibit some degree of noise. Noise is undesirable because it interferes with and obscures portions of the signal. It manifests in several varieties, all of which are undesirable. There are generally three broad causes of noise in digital photography. Shot noise is random noise caused by the uneven emission of photons from their source during any sufficiently short period. Dark noise, like shot noise, is produced by the inherent variations in thermally generated electrons within the image sensor during a sufficiently short period. Lastly, there’s read noise, which is created by intrinsic imperfections in the design of all electronic components that carry, interact with, and amplify the signal before its digitization.
Read noise may be reduced with better engineering and greater precision in manufacturing the electronic components. In scientific applications, such as astronomy, dark noise is reduced by actively cooling image sensors to cryogenic temperatures. Since this isn’t feasible in practical photography, manufacturers work to reduce thermal generation by designing their image sensors to be more energy efficient, which reduces heat generation, and by implementing passive cooling, such as internal heat dissipators. Unfortunately, shot noise is caused by an inherent property of light and cannot be reduced by camera design. The purpose of this semi-technical description is meant to demonstrate that despite the seemingly unstoppable trend of low-light high-ISO performance pursued by modern camera manufacturers, there looms are very real and very hard barrier to this steady progress.
Signal-to-noise ratio. The signal-to-noise ratio, or SNR, is the amount of signal divided by the amount of noise. At levels above 1:1, there’s more signal than noise in the system. A higher SNR produces results with less visible noise. The only way to increase the SNR is by increasing exposure.
Types of image noise
Image sensors exhibit three different yet common types of image noise: random, fixed-pattern, and banding.
Random noise. This type of noise produces statistically random variations in the brightness and colour of an image both above and below the signal’s level. Although the pattern of random noise varies between photos shot with identical exposure settings, the magnitude of the noise will remain the same. Random noise is present in all digital images and is especially prevalent in high ISO and short duration exposures. It’s primarily caused by shot noise and read noise.
Fixed-pattern noise. Fixed-pattern noise is often associated with the phenomenon of “hot pixels,” which are pixels that exhibit greater intensity than both the signal and the surrounding random noise. While exceptions occur, most fixed-pattern noise manifests during long exposures and hot temperatures. Hot pixels and other less pronounced types of fixed-pattern noise are generally found in the same position across multiple images when the shooting conditions are similar.
Banding noise. Banding is a type of semi-fixed-pattern noise that manifests as faint vertical or horizontal stripes. This type of noise is highly dependent on the specific image sensor employed in each camera. Some image sensors—and thus all camera models fitted with them—are highly prone to banding noise generated during the data transfer process. Banding is most evident in the shadows of images shot with high ISO values, or in underexposed low ISO images that have been excessively pushed.
Noise distribution across an image
Shadows, mid-tones, and highlights. When examining a photo that contains a variety of tones, you’ll notice that the apparent distribution and magnitude of noise isn’t entirely uniform. The brighter parts of an image will exhibit less visible noise than the darker parts. This uneven distribution of noise is caused by localized differences in the signal-to-noise ratio. Since the brighter regions of an image receive more light, they record a higher signal and thus have a higher SNR than the darker areas, which receive less light and record a comparatively weaker signal.
Characteristics of random image noise
The noise from any colour-capable digital camera can be separated into two distinct elements, luminance noise and chrominance noise.
Luminance noise presents as fluctuations in the tonal intensity of the noise pattern and is likened to film-grain by many photographers. Chrominance noise presents as fluctuations in the hues of the noise pattern. Chrominance noise is unique to digital photography as there’s nothing analogous in film, and is commonly considered undesirable. The chief cause of chrominance noise is the difference in light transmission between the three colours of the colour filter array found on image sensors.
In-camera. Every camera offers some form of built-in noise reduction algorithms. Most bridge, enthusiast, and professional cameras will allow users to disable noise reduction, or to specify its strength. The algorithms are often broad in scope and apply evenly to the entire image area. Experiment with varying degrees of noise reduction at different ISO settings to see how it affects image quality, specifically, the balance between reducing image noise and preserving fine details and textures. Lastly, in-camera noise reduction only affects the JPEG/TIFF output files; images recorded in the RAW format aren’t changed beyond the embedded preview.
Reducing luminance and chrominance noise. When editing images on a personal computer, some software applications allow users to separate luminance and chrominance noise reduction into discrete processes. Luminance noise is generally more challenging to reduce, and the process tends to smooth out the detail in high-frequency patterns, such as those of fabrics, fur, and hair. Small to medium amounts of chrominance noise is generally straightforward to remedy on a personal computer. Although colour noise reduction does not have a significant adverse impact on textures, it will decrease the fine spatial resolution of colours.
Dark frame subtraction, or long-exposure noise reduction, reduces fixed-pattern noise caused by long exposures. Immediately following a long-exposure, the camera takes a second and equally long “exposure” with the shutter closed. The process records an image of the prevailing dark and read noise patterns and automatically subtracts it from the previous photograph. Dark frame subtraction must be implemented manually when shooting photos in RAW format.
Sharpening and noise. In general, sharpening an image will also sharpen the image noise, thereby increasing its apparent magnitude. This is especially problematic when in-camera noise reduction is disabled and sharpening remains active.