Why does Poisson noise depend on brightness?

In a counting model, the expected signal count determines the variance, so changing the mean changes the noise distribution.

Can Poisson noise be negative?

A sampled count is nonnegative, but its residual relative to the expected count can be negative.

Is a Gaussian approximation always suitable?

No. It may work at sufficiently large counts for some purposes, but low-count behavior and signal-dependent variance require care.

AI Glossary / Computer vision and multimodal AI / Intermediate

Poisson noise

Definition

Poisson noise is random variation modeled using a Poisson distribution, commonly associated with counts of independently occurring events over a fixed interval or region. In imaging, it often models fluctuations in detected photon or photoelectron counts, making the noise depend on the signal level.

What it means

A Poisson model uses one nonnegative parameter, λ, representing the expected count. Both the mean and variance equal λ. This differs from a simple additive noise model with the same variance at every brightness level. Brighter regions can have greater absolute counting variation but smaller relative variation. The model describes uncertainty in repeated measurements under its assumptions; it does not imply that every camera artifact is Poisson-distributed. Electronic read noise, fixed-pattern effects and image-processing artifacts may need separate treatment.

How it works

For a nonnegative integer count k, the Poisson probability is exp(−λ) λ^k/k!. A simulation can draw a count for each pixel from a mean determined by the expected signal. To simulate image formation meaningfully, first define how image values correspond to expected counts; arbitrarily treating normalized display values as counts changes the noise strength.
The observed count has variance λ, while the noise residual is the observed count minus its expectation. At high counts, a Gaussian approximation may be useful, but its variance still depends on the count level. Real cameras often require a combined Poisson - Gaussian model, plus calibration of gain and offsets, rather than an unqualified Poisson assumption.

Three practical examples

A denoising researcher tests a method on synthetic count data with several expected signal levels.
A camera model combines photon-count variation with a separate electronic read-noise contribution.
A low-light imaging workflow evaluates how longer exposure changes relative counting uncertainty.

Think of it this way

Think of counting arrivals at a service desk during equal intervals when arrivals are independent and the average rate is steady. The count varies even though the underlying average has not changed. The analogy is useful for the distribution, but real customers can arrive in groups and real light can have non-Poisson statistics, so the assumptions must fit the setting.

Worked example

Suppose the expected count is λ = 4. The standard deviation is √4 = 2, and the probability of observing zero events is exp(−4), approximately 0.0183 or 1.83%. If the expected count rises to 100, the standard deviation becomes 10.
The relative standard deviation is 2/4 = 50% in the first case and 10/100 = 10% in the second. Absolute variation increases, but the measurement becomes more precise relative to its mean. These are distribution properties, not predictions that every observed count will lie within one standard deviation. A simulated count must also remain a nonnegative integer before any gain or normalization step.

Illustrative case study

In a fictional inspection-model experiment, a team adds noise directly to images normalized between 0 and 1. The resulting samples become much more erratic than the intended camera scenario. The team revises the simulation to map brightness to an explicit expected photoelectron range, samples counts, adds separately modeled read noise and converts back to the desired image representation. It then compares the assumptions with calibration data. The case illustrates why a named distribution alone is insufficient to create realistic test images.

Common misconceptions and mistakes

Poisson noise is not generated by simply adding a fixed amount of Gaussian noise to every pixel. Mean and variance equality applies to the underlying counts in appropriate units, not automatically to offset or scaled digital values. The Poisson equation used in image blending is unrelated to this counting model despite sharing a name. Another mistake is interpreting a denoised result as a perfect recovery: reducing visible variation can also suppress weak structures or introduce model-dependent estimates.

Business use

When noise simulation informs a product decision, document the expected-count scale and any conversion to digital values. Include a range of brightness levels, because a method that works well on bright regions may fail in dark ones. Compare simulated statistics with real captures where possible. Separate appearance-based evaluation from downstream measurement accuracy, and preserve clean references or calibration evidence when available. A realistic noise model makes evaluation more informative but does not replace testing on the intended deployment data.

Check your understanding

A sensor model doubles the expected count from 25 to 50. Does the Poisson standard deviation double?

Answer: No. It changes from 5 to √50, about 7.07, a factor of √2 rather than two.

Related terms

Poisson distribution
Count data
Read noise
Variance
Denoising
photon shot noise
signal-dependent noise
Gaussian noise
sensor gain

Learn next

Learn the difference between a noisy observation and its residual next. Simulate counts at several means, compare empirical mean and variance, and then examine how gain and offsets change those statistics in digital units.

Variance

Sources

NumPy, official Poisson random-sampling documentation: https://numpy.org/doc/stable/reference/random/generated/numpy.random.Generator.poisson.html EMVA, camera noise and characterization resources: https://www.emva.org/standards-technology/emva-1288/ Primary documentation and standard search excerpts consulted; full source and editorial review remain pending.

Previous: Poisson (gradient-domain) blending Next: Poisson regression