AI Glossary / Computer vision and multimodal AI / Advanced
Robust combination (sigma-clipping)
Definition
Sigma-clipping is an iterative rejection procedure that identifies values sufficiently far from an estimated center relative to an estimated spread. In image combination, it can exclude unusual samples across an aligned image stack before the remaining samples are combined. The threshold, center estimator, spread estimator and stopping rule define the particular procedure.
What it means
A simple average can be pulled by an isolated extreme value. Sigma-clipping attempts to reduce that influence by identifying inconsistent observations. However, an unusual value is not automatically an error: it could represent a real transient event, motion or a different underlying condition. The procedure depends on an appropriate comparison set and noise model. Robust combination therefore involves more than selecting a clipping multiplier. Alignment, calibration and preservation of rejection information matter when interpreting the combined output.
How it works
Estimate a center and spread, reject values outside the chosen lower and upper limits, then optionally recompute and repeat. Center estimates can use a mean or median, while spread estimates can use standard deviation or a robust alternative. Asymmetric clipping can apply different limits on each side. In an image stack, the comparison commonly occurs among observations corresponding to the same scene location. Small sample sets, biased initial estimates and zero estimated spread need deliberate handling. The final statistic may be a mean or another combination of the retained values.
Three practical examples
- An imaging team combines aligned exposures while rejecting isolated extreme sensor events and retaining a map of rejected samples.
- A researcher compares clipping based on standard deviation with clipping based on a robust spread estimate in a contaminated dataset.
- A quality-control exercise checks whether unusual observations are genuine signals before deciding that an automated rejection rule is appropriate.
Think of it this way
Imagine several people measuring the same table. Most readings cluster closely, while one is much larger. A clipping rule asks whether that reading lies outside a stated tolerance relative to the group. It does not explain why the reading differs. Perhaps someone used the wrong units, or perhaps they measured a different object. The analogy separates detecting inconsistency from establishing an error and shows why the comparison set must represent the same intended quantity.
Worked example
In an original teaching example, five values are 9,10,10,11 and 30. Suppose a separately justified center is 10 and spread is 1, with rejection outside center plus or minus 3 spreads. The retained interval is [7,13], so 30 is rejected and the retained mean is (9+10+10+11)/4=10. The unfiltered mean is 14. This example stipulates the center and spread to make the rule transparent. Estimating standard deviation from these five values instead could inflate the spread and change rejection, so the result should not be described as the inevitable outcome of every three-sigma implementation.
Illustrative case study
A fictional night-imaging team combines repeated views of a static test scene. Clipping suppresses some isolated bright artifacts, but it also removes a real brief light event present in only one frame. The team preserves the original sequence and rejection mask, then chooses a combination policy suited to whether its goal is a clean static view or documentation of transient events. This scenario illustrates a task-dependent decision. It does not claim that rejected values are always defective or that a clean composite accurately records every event occurring during capture.
Common misconceptions and mistakes
Sigma-clipping does not prove that an excluded sample is wrong. Another mistake is expecting a conventional standard deviation to remain unaffected by extreme contamination. A small dataset can make center and spread estimates unstable, and repeated clipping can change the retained distribution. Misaligned images can make legitimate edge values look inconsistent. Threshold multipliers also do not imply universal false-rejection probabilities when the distribution is not the assumed one. Finally, replacing rejected samples and masking them are different operations that should be documented explicitly.
Business use
Robust combination can improve repeated-measurement workflows, but teams should define what counts as unwanted contamination and what must be preserved. Record the clipping rule, estimators, iteration limit and rejected-sample counts. Validate behavior on known artifacts and legitimate unusual events. If outputs support investigation or measurement, keep source observations available rather than treating the composite as a complete record. Compare results with and without clipping to reveal its influence. The goal is a justified combination for the intended use, not the removal of every value that looks inconvenient.
People also ask
Must sigma-clipping use the mean as its center?
No. Implementations can support alternatives such as the median.
Is every rejected point a measurement error?
No. Rejection identifies inconsistency under a rule, not its cause.
Why repeat the procedure?
Removing candidates changes center and spread estimates, which can reveal additional candidates under the updated rule.
Check your understanding
Which value was rejected in the stipulated example?
Answer: Thirty.
Why did the retained mean become 10?
Answer: The remaining four values summed to 40.
Why might another implementation keep 30?
Answer: A different center, spread estimate, threshold or iteration rule can produce a different rejection decision.
Related terms
- Robust statistics
- Outlier rejection
- median absolute deviation
- image stacking
- trimmed mean
- uncertainty estimation
Learn next
Apply a stated clipping rule to a small dataset, recording the estimates at every iteration. Compare standard deviation with a robust spread estimate. Then include a legitimate transient observation and decide whether rejection matches the purpose of the analysis.
More in Computer vision and multimodal AI
- Remote photoplethysmography - Remote photoplethysmography, often abbreviated rPPG, estimates a blood-volume-related pulse signal from subtle changes in light recorded by a camera...
- Ridge detection - Ridge detection identifies elongated structures that are locally high or low relative to neighboring values across their width.
- Roberts cross gradient operator - The Roberts cross gradient operator estimates local image-intensity change using two diagonal differences across a 2×2 neighborhood.
- Robust regression - Robust regression is a family of methods for estimating relationships while limiting sensitivity to specified departures from modeling assumptions, often...
- Rolling-shutter correction - Rolling-shutter correction compensates for geometric distortions caused when different parts of an image are exposed at different times while the camera or...
- Saliency - In visual perception and computer vision, saliency describes how strongly a location or feature stands out within a scene under a stated model of attention.
Sources
Astropy, sigma_clip reference: https://docs.astropy.org/en/stable/api/astropy.stats.sigma_clip.html The five-value calculation, table-measurement analogy and night-imaging case are original educational illustrations. The worked example stipulates its center and spread and does not imply that those values arise from default settings on the same sample.
Previous: Robust collaborative filtering Next: Robust control