AI Glossary / Computer vision and multimodal AI / Intermediate
Central moment
Definition
A central moment is a summary of a distribution calculated from deviations around its mean. The second central moment is the variance; higher moments describe other aspects of shape. In image analysis, spatial central moments similarly summarize a shape or intensity distribution around its centroid.
What it means
Raw moments use powers of values measured from an origin. Central moments first subtract the mean, so they describe spread and shape relative to the distribution's own center. This removes dependence on a simple shift of location.
The order matters. The first central moment is zero when the mean exists. The second measures spread, while third and fourth moments contribute to measures of asymmetry and tail or peakedness behavior after suitable standardization.
Image moments use spatial coordinates and pixel weights. Their formulas and normalization differ from a one-dimensional statistical average, so the word moment should always be interpreted in the context of the data and convention.
How it works
For a random variable X with mean mu, the kth central moment is E[(X−mu)^k], provided the expectation exists. A direct empirical version averages these powers over observed values. Bias-corrected estimators and standardized moments are separate constructions.
For a two-dimensional image, compute a centroid from the chosen pixel weights, then sum products of centered coordinate powers weighted by intensity or a binary shape mask. These spatial moments are translation-invariant under appropriate coordinate shifts, but not automatically invariant to scale or rotation. Higher orders can be sensitive to outliers, noise and numerical scaling, so choose moments that serve a clear analytical purpose.
Three practical examples
- A statistical report uses the second central moment to describe variability around an average.
- A shape-analysis system uses spatial central moments to summarize a binary object's spread around its centroid.
- A quality-control study checks whether a distribution's asymmetry changes over time using suitably defined moment-based summaries.
Think of it this way
Think of measuring how far weights lie from their balance point, then combining those distances in different ways. Squared distances emphasize spread, while signed odd powers can reflect asymmetry. The analogy helps explain centering, but higher moments require mathematical existence and careful normalization; they are not automatically intuitive descriptions of every dataset.
Worked example
Take the values [1,2,3], whose mean is 2. Their centered deviations are [−1,0,1]. The empirical first central moment is (−1+0+1)/3=0. The second is (1+0+1)/3=2/3. The third is (−1+0+1)/3=0, and the fourth is again (1+0+1)/3=2/3.
Now shift every value upward by 10, giving [11,12,13]. The mean becomes 12, but the centered deviations and all these central moments remain unchanged.
The second central moment here uses division by 3. The usual unbiased sample-variance estimator for an independently sampled population uses division by n−1 and would equal 1. These are related but distinct quantities. A report must state whether it describes the empirical distribution or estimates a population parameter with a particular correction.
Illustrative case study
Fictional case: Spruce Vision compares object shapes using image moments. A developer assumes that central moments make the descriptors invariant to every image transformation. Shifted objects compare consistently, but enlarged and rotated versions do not.
The team reviews translation, scale and rotation separately, chooses appropriate normalized or invariant descriptors where needed and tests them on controlled transformations. It also checks whether intensity weighting or a binary mask best matches the task. This case illustrates the limits of centering, not a claim that one moment descriptor is sufficient for every object-recognition problem.
Common misconceptions and mistakes
Central moments are not all standardized moments. Skewness and kurtosis typically divide by powers of standard deviation, and kurtosis conventions may subtract a constant. The empirical second central moment is not identical to every sample-variance estimator.
Higher moments may not exist for some probability distributions and can be unstable in finite samples. Translation invariance does not imply rotation or scale invariance. In images, a zero total weight makes the usual centroid undefined. Changing segmentation or pixel weighting can materially change spatial moments even when the visible object seems similar.
Business use
Central moments can provide compact summaries for statistical monitoring, shape measurement and feature engineering. Their value is greatest when the chosen moment has a clear interpretation linked to the task.
Document whether the data are scalar values, a probability distribution or an image-weight field. State centering, weighting and normalization conventions. Check sensitivity to outliers, segmentation errors and measurement units. Avoid interpreting a higher-order moment as a complete description of distribution shape: different distributions can share the same few moments. Compare summaries with visual or task-specific evidence before using them to trigger operational decisions.
People also ask
Is the second central moment variance?
For a probability distribution, yes, when it exists.
Does centering remove scale differences?
No. Multiplying values changes central moments according to their order.
Are image central moments the same formula as scalar moments?
They use the same centering idea, but involve spatial coordinates and image weights.
Check your understanding
Every observation is increased by the same constant. What happens to its central moments, assuming the mean is shifted consistently?
Answer: They remain unchanged because each observation's deviation from the mean is unchanged. Raw moments generally do change under that shift.
Related terms
Learn next
Learn raw, central and standardized moments side by side. Calculate them for a small dataset before and after a shift and a rescaling. For images, repeat the exercise with a binary shape to see which transformations each descriptor does and does not ignore.
More in Computer vision and multimodal AI
- CAPTCHA - CAPTCHA means “Completely Automated Public Turing test to tell Computers and Humans Apart.
- Catmull - Rom - Catmull - Rom is a family of interpolating cubic constructions.
- Center-surround - Center-surround describes a receptive-field or filter arrangement in which a central region and its surrounding neighborhood have opposing influences.
- Chroma keying - Chroma keying is a compositing technique that separates a subject from a background using color information, commonly from a green or blue screen.
- Circular convolution - Circular convolution combines two finite sequences while treating their indices as periodic, so values that pass one end wrap around to the other.
- Closed-form matting - Closed-form matting is an image-matting method that estimates foreground opacity by constructing and solving a quadratic optimization problem under local...
Sources
SciPy, scipy.stats.moment: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.moment.html OpenCV, Moments: https://docs.opencv.org/4.x/d8/d23/classcv_1_1Moments.html Official documentation search performed; formulas and examples are original explanations. Full convention-specific and editorial review remain pending.
Previous: Central limit theorem Next: Centralized computing