Is a median filter always better than a mean filter?

No. Their behavior differs, and suitability depends on the noise and structures being preserved.

Can a nonlinear filter be fast?

Yes. Nonlinearity describes a mathematical property, not a required computational cost. Implementation and window size affect speed.

Is a fixed Gaussian convolution nonlinear because its weights form a curve?

No. Fixed convolution is linear in its input under compatible linear boundary handling. The shape of the weight curve does not change that fact.

AI Glossary / Computer vision and multimodal AI / Intermediate

Nonlinear filter

Definition

A nonlinear filter is a signal- or image-processing operation that does not obey the superposition property of a linear filter. Its response to a weighted combination of inputs cannot, in general, be obtained by applying the same weighted combination to the separately filtered inputs.

What it means

For a linear operation F, the relationship F(aX + bY) = aF(X) + bF(Y) holds for allowed inputs X and Y and scalar weights a and b. A nonlinear filter violates that relationship for at least one valid choice.

This is a mathematical distinction, not a judgment about complexity or quality. Selecting the median value of three numbers is simple to describe but nonlinear. A large fixed convolution can involve many calculations while remaining linear.

In images, nonlinear filters can make decisions based on ordering, thresholds or the content being processed. They are useful when a fixed weighted average is poorly matched to a task. They do not automatically preserve every edge, remove every outlier or produce a better result than a linear alternative.

How it works

A median filter gathers values in a neighborhood, sorts them and uses a middle value. Because the selected position depends on the input ordering, this cannot generally be represented by one fixed set of averaging weights applied to every possible input.

Other nonlinear image operations use minima, maxima or input-dependent weights. These procedures behave differently, so choosing a “nonlinear filter” does not specify a complete algorithm. The particular rule, neighborhood, parameters and boundary treatment must be identified.

A filter can also be linear with respect to one input while depending nonlinearly on another quantity. When assessing such a claim, state what is held fixed. Rounding or clipping in an implementation can introduce additional departures from the ideal mathematical operation.

Practical selection begins with the unwanted variation and the useful structures that must survive, followed by evaluation on representative data.

Three practical examples

A median filter suppresses isolated extreme pixels in a scanned image, with checks that small printed marks are not removed as well.
An image-processing workflow uses a maximum or minimum operation to alter bright or dark structures according to a chosen neighborhood.
A sensor analyst compares a median-based treatment with an averaging filter when occasional spikes are more troublesome than ordinary small fluctuations.

Think of it this way

Think of a committee that always averages its members' numbers versus one that selects the middle-ranked number. The first follows fixed averaging weights. The second changes which member's number becomes the answer as the ranking changes.

The analogy explains one source of nonlinearity, not all nonlinear filters. Many operate through smooth formulas rather than selecting one member, and an image neighborhood is not a group making a judgment.

Worked example

Showing why the median is nonlinear

Take two three-value neighborhoods: X = [0, 0, 10] and Y = [0, 10, 0]. The median of each is zero. Adding the separate median results therefore gives zero.

Now add the inputs position by position. X + Y = [0, 10, 10], whose median is 10. Filtering the sum gives 10, while summing the filtered inputs gives zero. This single counterexample establishes that the median operation does not satisfy additivity and therefore is not linear.

A separate practical example shows both usefulness and risk. The neighborhood [20, 200, 21] has median 21, so an isolated large value is replaced by something close to its neighbors. That may remove an unwanted impulse. However, if 200 represents a genuine tiny bright feature, the identical calculation removes useful information.

The filter has no built-in knowledge of which interpretation is correct. Decide using the acquisition process, expected feature size and the downstream task. These are illustrative calculations.

Illustrative case study

Illustrative case study - fictional

A fictional archive digitizes old index cards. Some scans contain isolated bright specks, and a small median filter makes many card backgrounds look cleaner. During review, a staff member notices that punctuation in a faint reference number has also changed.

The team compares the unprocessed scan with the candidate result and tests a smaller processing region. It decides that identifying punctuation reliably is more important than making every background look uniform. Original scans are retained alongside any enhanced viewing copies.

The example shows why a filter should be assessed against information preservation. It does not establish a universally correct window size.

Common misconceptions and mistakes

“Nonlinear” does not mean neural, learned or intelligent. A deterministic median operation is sufficient.
It does not mean the output simply follows a curved tone graph. Neighborhood operations can be nonlinear without being pointwise tone mappings.
Preserving some edges in a test image is not a promise to preserve thin lines, corners or isolated real features.
The order of processing can matter. Applying two nonlinear operations in different orders need not give the same result.
A software function's name is insufficient evidence of mathematical properties. Examine the operation and what inputs are fixed.

Business use

Businesses encounter nonlinear filters in document preparation, image enhancement and inspection pipelines. They can be valuable when a known disturbance calls for something other than ordinary averaging.

Specify an acceptance test that reflects the business purpose. For document work, compare recognition accuracy and small symbols; for inspection, compare false alarms and missed defects. Record the filter, settings and boundary handling so that a later team can reproduce the result. Visual appeal alone is insufficient when content drives decisions.

Check your understanding

If F(X + Y) differs from F(X) + F(Y) for one valid pair of inputs, can F be a linear filter?

Answer: No. A linear filter must satisfy additivity for all valid input pairs in its defined domain.

Related terms

Median filter
Linear filter
Convolution
Noise reduction
image morphology
bilateral filter

Learn next

Calculate the mean and median for several three-value neighborhoods containing a spike, a gradual ramp and a thin feature. Then perform the additivity test above. Keep mathematical classification separate from deciding which output is useful.

Sources

Sources: MathWorks, Median Filtering: https://www.mathworks.com/help/signal/ug/median-filtering.html ; MathWorks, Two-dimensional median filtering: https://www.mathworks.com/help/images/ref/medfilt2.html ; MathWorks, Image Filtering overview: https://www.mathworks.com/help/images/linear-filtering.html . Relevant documentation was reviewed. The counterexample and decision scenarios are original illustrations.

Previous: Nonlinear constraints Next: Nonlinear modeling