Must seeds be selected by a person?

No. Seeds can be selected manually, generated from another algorithm or derived from known geometry.

Does it always use intensity?

No. Color, texture, surface orientation and other features can define acceptance.

Why can a region leak across a boundary?

The boundary may not create a sufficient change under the chosen criterion, especially with noise or gradual transitions.

AI Glossary / Computer vision and multimodal AI / Intermediate

Region growing

Definition

Region growing is a segmentation approach that starts from one or more seed locations and expands regions by adding neighboring elements that satisfy a chosen criterion. The criterion may use intensity, color, texture or another property relevant to the application.

What it means

Rather than classify every pixel independently, region growing builds connected groups outward from selected starting points. A seed inside a uniform-looking object can help identify the surrounding area with similar properties. The result depends on the seeds, the neighborhood rule and the acceptance criterion. Similarity can be measured against the original seed, a neighboring pixel or statistics of the evolving region, and those choices are not equivalent. Region growing is a general family of methods rather than one fixed algorithm.

How it works

Initialize a region with a seed and place candidate neighbors on a processing list. Test candidates against the defined criterion; accepted pixels join the region and can introduce new candidates. Continue until no eligible neighbors remain or another stopping condition is met.
With several seeds, regions may compete or later be merged. Connectivity determines which steps are allowed, such as shared-edge or diagonal adjacency. If the criterion changes as region statistics update, processing order can influence some implementations. Noise, weak boundaries and gradual intensity changes can cause leakage into unwanted areas, while an overly strict rule can leave gaps or split a valid region.

Three practical examples

An image editor selects a connected area of similar color from a clicked seed.
An inspection workflow grows a region around a known part interior before measuring its visible area.
A point-cloud segmentation method expands groups of neighboring points with compatible surface properties.

Think of it this way

Think of spreading paint from a starting tile, allowing it to cross only into neighboring tiles that meet a rule. The starting point and rule determine where the spread stops. The analogy is limited because a numerical region can use changing statistics or several competing seeds, and similarity does not necessarily correspond to a real object boundary.

Worked example

Consider a one-dimensional row [10,11,12,20,21], with a seed at the first value 10. Use direct neighbors and accept values within 2 of the seed value. The region grows through 11 and 12, then stops at 20 because |20 − 10| = 10. The selected region contains the first three positions.
Now imagine a longer ramp [10,12,14,16]. A rule comparing each candidate only with its immediately preceding accepted neighbor could accept the entire ramp with tolerance 2. A fixed-seed rule would stop after 12. This distinction shows why the reference used for similarity must be stated, not just the tolerance number.

Illustrative case study

In a fictional packaging-inspection project, a team seeds the center of a printed label and grows a region using color similarity. It works when the label contrasts with the package, but leaks through a low-contrast edge on one design. The team reviews the seed placement, adds an appropriate boundary cue and evaluates multiple package colors. It does not treat the first successful example as evidence that the same tolerance works everywhere. The case illustrates how segmentation assumptions interact with actual design variation.

Common misconceptions and mistakes

Region growing is not guaranteed to find a semantic object. A connected region of similar pixels may include parts of several objects, while one object may contain very different colors. A threshold alone does not fully specify the method; the comparison reference and connectivity matter. Seed choice can strongly affect the result. Another mistake is assuming that a region that stops growing must be correct: it may have stopped at noise, an internal texture change or an overly restrictive criterion.

Business use

For practical use, define the region property that matters and choose seeds that support it. Test low contrast, noise, internal texture and touching objects. Review both leakage and under-segmentation, and retain overlays that show selected pixels for human inspection where useful. If a measurement depends on the mask, evaluate its sensitivity to seed and tolerance changes. A simple, interpretable growing rule can be valuable when its limitations match the workflow and failures are detectable.

Check your understanding

Two methods use tolerance 2 but one compares with the seed and another with the latest accepted neighbor. Must they select the same pixels?

Answer: No. A gradual ramp can remain locally similar while moving far from the seed value, producing different regions.

Related terms

Segmentation
Pixel connectivity
Boundary leakage
seed point
flood fill
region merging
similarity criterion
connected component
region statistics

Learn next

Learn flood fill and connectivity next. Grow regions on a small image using both fixed-seed and region-dependent criteria, then compare where each leaks, stops too soon or changes with seed placement.

Sources

scikit-image, official flood and morphology documentation: https://scikit-image.org/docs/stable/api/skimage.morphology Primary documentation search excerpts consulted; full source and editorial review remain pending.

Previous: Region connection calculus Next: Region of interest