Skip to main content
BVDART Logo
BVDART
← All guidesFilters & transforms

Blur, denoising and sharpening

Neighborhood filters smooth or emphasize local variation. Averaging, rank selection and detail amplification solve different image problems.

Separate a signal into blur and signed detail

Original I
Original I
Low frequencies B
Low frequencies B
Sharpened I + aD
Sharpened I + aD
Cut through the image
Original and blurred row
IBI + aD
Signed detail residual D = I − B
Positive and negative detail around zero

Positive lobes brighten one side; negative lobes darken the other. Large amounts create halos.

D = I − Gσ(I)   ·   I′ = I + 1.0DThe plotted values keep overshoot. Only the rendered images clip to 0–255. At a = 0, the sharpened image equals the original.

Broaden the blur, then increase the detail added back to reveal overshoot and halos. Use any control to take over.

What to try
  1. Change sigma and move the inspected row through the image.
  2. Increase the detail amount, then set it to zero to restore the original.
Open in the Lab ↗

What to look for: Unsharp masking adds I − blur(I) back to I. Positive and negative lobes create contrast and, at high strength, halos; plots retain overshoot that the displayed images clip.

How it works

Box blur averages a uniform neighborhood. Gaussian blur weights nearby samples more strongly and is controlled by sigma. Median filtering chooses a middle value and can remove isolated outliers without averaging every boundary. The Lab’s box and Gaussian implementations filter premultiplied color to avoid transparent-edge fringes.

Unsharp masking subtracts a softened copy to estimate detail, then adds some of that detail back. Its threshold limits weak differences. High-pass filtering shows the difference around a neutral gray. These operations can emphasize noise as well as meaningful structure.

Use it in the Lab

Choose median for isolated specks, Gaussian for smooth attenuation, and unsharp masking for controlled detail emphasis. Preview at actual output scale and use selection bounds for local treatments.

Gaussian blur ↗Box blur ↗Median ↗Unsharp mask ↗High pass ↗

Limits & distinctions

Radius and sigma are different measures. The same numeric setting in two filter families does not imply the same blur strength.

Sources & references

Guide updated