Which is better, 4-connectivity or 8-connectivity?

Neither is universally better. The choice depends on whether diagonal contact should join foreground regions in the task.

Does connectivity change the image's pixel values?

The neighborhood rule itself does not. It changes how regions are interpreted or labeled.

Does one connected component equal one object?

Not necessarily. Touching objects can merge, and a single object's visible pixels can be disconnected.

AI Glossary / Computer vision and multimodal AI / Beginner

Pixel connectivity

Definition

Pixel connectivity is the rule used to decide which neighboring pixels can belong to the same connected region. In a two-dimensional square grid, common choices are 4-connectivity, which uses shared edges, and 8-connectivity, which also includes diagonal neighbors.

What it means

A binary image can mark foreground pixels without yet defining how touching pixels form objects. Two foreground pixels meeting only at a corner illustrate the ambiguity. Under 4-connectivity they are separate unless another edge-connected path joins them; under 8-connectivity they connect directly. The choice changes object counts, boundaries and holes even though no pixel values change. Connectivity is therefore part of the analysis definition, not a cosmetic display option. It should match the meaning of contact in the application and be used consistently throughout the processing pipeline.

How it works

Start with pixels satisfying a foreground or label criterion. A connected component contains pixels joined by a chain of allowed neighbor steps. A labeling algorithm assigns the same component identifier to all pixels linked by such a chain. With 4-connectivity, a step can go up, down, left or right. With 8-connectivity, the four diagonals are also allowed.
Related neighborhood choices exist in three-dimensional voxel grids, where face, edge and corner contact can be treated differently. Foreground and background connectivity also interact in digital topology: choosing conventions carefully prevents contradictory interpretations of boundaries. A software parameter named connectivity may encode the neighborhood indirectly, so its documentation should be checked rather than assuming the number literally means four or eight.

Three practical examples

An inspection program counts separate foreground defects in a thresholded image.
A document-analysis tool groups neighboring ink pixels into components before further character analysis.
A segmentation workflow removes small connected regions while preserving larger structures according to a specified neighborhood rule.

Think of it this way

Think of tiles on a floor. One rule lets you move between tiles only across a shared side; another also lets you step across a corner. The reachable groups differ under those rules. The analogy clarifies adjacency, but it does not tell you whether a one-pixel corner touch corresponds to meaningful physical contact in the original scene.

Worked example

Consider a 2 × 2 binary image with foreground pixels at the upper-left and lower-right corners, and background at the other two positions. Under 4-connectivity, there are two foreground components because the foreground pixels share no side. Under 8-connectivity, there is one component because the pixels are diagonal neighbors.
Add a foreground pixel at the upper-right corner. Now all three foreground pixels form one component under either rule: the upper-left connects to upper-right, which connects to lower-right. This demonstrates that connectivity depends on the existence of a chain, not on every pair of pixels directly touching one another.

Illustrative case study

In a fictional print-quality project, a team counts isolated ink specks in scanned samples. A change in library defaults unexpectedly lowers the reported component count. Reviewing small binary examples reveals that corner-touching specks are now grouped together. The team explicitly sets and documents the intended connectivity, then rechecks its counting thresholds. It also recognizes that thresholding and scan resolution influence whether separate marks touch at all. The case illustrates a reproducibility issue without claiming that either neighborhood rule is universally correct.

Common misconceptions and mistakes

Connectivity is not the same as similarity in color or intensity. A region criterion determines which pixels are eligible, while connectivity determines how eligible pixels can join. Diagonally touching pixels are not automatically one object under every convention. A connected component is also not necessarily a semantic object: two objects can touch, and one object can be broken into several regions by occlusion or thresholding. Changing connectivity alone does not correct those segmentation errors.

Business use

For automated counting or shape measurement, include connectivity in the documented measurement protocol. Test corner contacts, narrow bridges, holes and image borders. Review whether image resizing, thresholding or denoising changes the topology before component analysis. If the count drives a business decision, retain a visualization of the labeled regions so an operator can understand why marks were joined or separated. Stable conventions help teams compare results across software versions and datasets.

Check your understanding

Three foreground pixels form a diagonal line with no other foreground pixels. How many components exist under 4-connectivity and 8-connectivity?

Answer: There are three components under 4-connectivity and one under 8-connectivity, because only the latter permits the diagonal steps.

Related terms

Learn next

Learn connected-component labeling next. Draw a few small binary grids containing corner contacts, bridges and holes, and label them under both neighborhood rules before interpreting the corresponding shapes as real objects.

Sources

scikit-image, official measure.label documentation and connectivity examples: https://scikit-image.org/docs/stable/api/skimage.measure.html Primary documentation search excerpts consulted; full source and editorial review remain pending.

Previous: Pivoted document length normalization Next: Pixel-shift