splitmerge

Splitmerge is a class of image segmentation and clustering algorithms that operate by iteratively splitting regions that are internally heterogeneous and merging neighboring regions that are sufficiently similar. The approach is typically described as a combination of bottom-up over-segmentation and top-down refinement, aiming to produce coherent regions that align with image structure while avoiding excessive fragmentation.

In a typical splitmerge workflow, the process starts with an initial partition of the image into small regions or superpixels (over-segmentation). Each region is evaluated using descriptors such as color statistics, texture features, or edge information. If a region fails a homogeneity test, it is split along its boundaries into smaller regions. Conversely, neighboring regions are examined for similarity using a region adjacency graph or comparable representations; if neighboring regions meet a similarity criterion, they are merged to form larger, more uniform segments. The splitting and merging steps are repeated until a stopping criterion is reached, such as no further splits or merges being possible, or a predefined number of regions is achieved.

Splitmerge methods rely on local statistical tests, model-based assessments, or distance measures between region feature distributions. They can vary in how splitting is performed (e.g., along principal axes, along color or texture boundaries) and how merging is evaluated (e.g., histogram intersections, likelihood ratios, or boundary confidence). The approach is valued for its flexibility in handling irregular region shapes and for accommodating different feature modalities, but it can be sensitive to parameter choices and initial over-segmentation quality.

Applications include general image segmentation, remote sensing, medical imaging, and any domain requiring adaptive region-level analysis without a fixed region count. See also region-based segmentation and graph-based segmentation.