3.4. Median Blur

3.4.1. Overview

While the Gaussian blur filter calculates the mean of the neighboring pixels, the Median blur filter calculates the median:

Figure 17.14. Calculating Median

Calculating Median

A 3x3 neighborhood. Values in ascending order. Median surrounded in red.


This calculation does not create a new value, and an unrepresentative pixel in the neighborhood will not affect the result. So, the filter preserves edges and rounds corners. It is used to reduce noise, especially salt and pepper noise, and delete scratches on photographs.

3.4.2. Activate the filter

You can find this filter in the image menu under FiltersBlurMedian Blur…

3.4.3. Options

Figure 17.15. Median filter parameters

“Median” filter parameters

Presets, Preview, Split view

These options are common to GEGL-based filters. Please refer to Section 2, “Common Features”.

Neighborhood

The shape of the neighborhood. Three options: Square, Circle(default), Diamond. Differences are subtle and not predictable: experiment, on-canvas editing makes this easy.

Figure 17.16. Median Neighborhood

“Median” Neighborhood

Radius=3..........Left: Square

Middle: Circle..........Right: Diamond


Radius

The radius of the neighborhood. Increasing radius increases blur. Contrary to the Gaussian filter, edges are not blurred. Corners are rounded and convex surfaces are eroded.

Figure 17.17. Median blur vs Gaussian blur

“Median” blur vs “Gaussian” blur

Left: Origin

Middle: Median

Right: Gaussian


Increasing radius too much can create unwanted effects:

Left: Origin

Right: Radius=100

Percentile

By default, the Median Blur filter finds the median value at the neighborhood of each pixel. In spite of its name, the filter can actually find *any* arbitrary percentile, not just the median (i.e., the 50th percentile). The "Percentile" parameter controls the percentile used for the color components. Lower values shift the image toward darker tones and higher values toward brighter ones.

Left: Origin

Middle: Percentile=0. Image is darker. Stems are enlarged but sharp.

Right: Percentile=100. Image is brighter and much blurred.

Alpha Percentile

To see the effect of this parameter, you need an image with transparency. If the alpha channel is opaque everywhere, the result will also be fully opaque, regardless of the percentile.

Lower values for the Alpha percentile parameter shift the image toward more transparency, and higher values shift the image toward more opacity, where a value of 50% is balanced. Roughly speaking, values less than 50% make the opaque regions of the image smaller, while values greater than 50% make the opaque regions of the image larger.

Left: Origin. The image has an alpha channel. A circle is transparent.

Middle: Percentile=0%. Transparent circle is enlarged.

Right: Alpha percentile=100%. Transparent circle is reduced.

Abyss policy

Abyss policy (border management) is treated with Abyss policy.

High precision

This option avoids clipping and quantization but is slower.

3.4.4. Using Median Blur

Reducing salt and pepper noise

Left: origin (from Wikipedia)

Middle: radius = 1 applied twice

Right: radius = 1 applied three times

Reducing scratches

Left: origin

Middle: radius = 2

Right: radius = 1 applied twice. The image is less blurred.