Q14Digital Image Processing

Region-based segmentation partitions an image directly by grouping pixels into coherent regions based on predefined similarity criteria such as intensity, color, or texture, in contrast to edge-based methods that look for discontinuities.

Region Growing

Region growing starts from one or more seed points, chosen either manually or by an automatic criterion. Neighboring pixels of the seed are examined and appended to the growing region if they satisfy a similarity predicate, for example if the intensity difference from the region's mean is below a threshold $T$. The process repeats until no more pixels can be added. For example, seeding a bright tumor region in an X-ray with pixels within $\pm 10$ intensity levels grows outward until it meets the surrounding darker tissue boundary.

Region Splitting and Merging

This approach starts with the entire image as one region and subdivides it into quadrants using a quadtree structure. If a region is not homogeneous (fails a uniformity predicate $P$, e.g., maximum-minimum intensity difference exceeds a threshold), it is split into four sub-quadrants; this repeats recursively. After splitting, adjacent regions (even from different branches of the quadtree) that together satisfy the predicate $P$ are merged. This combination avoids the over-segmentation of pure splitting and the sensitivity to seed choice of pure growing, producing a compact set of homogeneous regions that adapt to the image content.

For example, an image containing a single dark object on a light background will initially fail the homogeneity test at the root and be split into four quadrants; a quadrant lying entirely within the background or entirely within the object satisfies $P$ and is not split further, while a quadrant straddling the object boundary continues to subdivide until each block is uniform. The merging pass then recombines the numerous small background blocks into one large region, since they satisfy $P$ jointly, while preserving the finer subdivision near the true edge. The final segmentation is therefore adaptive: coarse in flat areas and fine near boundaries, unlike a fixed grid partition.