Erosion

Problem

Sometimes the mask error goes in the opposite direction:

• bright specks should disappear
• boundaries are too fat
• tiny protrusions should be trimmed away

You want the foreground to contract rather than expand.

Intuition

Erosion asks:

Are all pixels inside the structuring element foreground?

If even one required neighbor is background, the output becomes background.

So erosion behaves like a local min over a binary mask.

The operator

Using a 3 x 3 structuring element, each output pixel becomes 1 only if every pixel in that neighborhood is 1.

In pseudocode:

output[x, y] = 1 if all input neighbors in the window are 1

That causes bright regions to shrink inward.

Worked example

Suppose a 1D mask is:

0 1 1 1 0

With a radius-1 erosion:

0 0 1 0 0

Only the interior survives. Borders contract.

In 2D, that removes thin protrusions and small bright islands quickly.

What erosion is good at

• removing isolated bright noise
• trimming overly thick masks
• stripping away small protrusions
• preparing for opening

But erosion can also destroy thin legitimate structures if you use too large a radius.

Complexity

For an image with $W \times H$ pixels and a structuring element of size $k \times k$:

• direct erosion costs $O(WHk^2)$

As with dilation, small fixed kernels are common in practice.

GPU implementation note

Erosion is the mirror image of dilation in GPU terms:

• each output pixel scans a local window
• the operator reduces the neighborhood with a min-like rule
• radius changes are very easy to expose interactively

So it is a great pair with dilation for teaching morphology.

When to choose it

Choose erosion when:

• the foreground should shrink
• tiny bright defects should disappear
• object boundaries are too fat

Choose something else when:

• gaps need to close rather than widen
• regions are already thin and fragile
• a one-step shrink is too destructive

Key takeaways

• Erosion shrinks foreground regions
• It is the binary-mask counterpart of a local min
• It removes small bright structures aggressively
• It is the opposite primitive from dilation
• Opening uses erosion first to remove noise before regrowing the surviving structure

Practice ideas

• Erode the same mask with multiple radii and note when legitimate thin structures vanish
• Compare erosion and dilation on a mask that contains both holes and bright specks
• Use erosion as the first step of opening, then see why the following dilation does not simply undo everything

Relation to other topics

• Thresholding often creates the mask erosion operates on
• Dilation expands masks, while erosion contracts them
• Opening and closing combine both primitives, with order determining behavior