Integrated Split/Merge Image Segmentation

Abstract

The KNIFE segmentation algorithm interleaves splitting and merging of regions during mono-chrome or multiband image partitioning. KNIFE splits regions along object boundaries, thus avoiding rectangular quadtree artifacts and establishing a context for good statistical decisions. Its iterative subregion extraction is based on multiband cluster analysis, with histogram-based threshold analysis used as a heuristic shortcut in simple cases. Splitting and merging decisions are based on sloped (rather than constant) surface fits, with successively more powerful thresholds and techniques employed until each region is split or found homogeneous. The user specifies only a desired level of segmentation, which is converted to procedural form by the KNIFE control process. The KNIFE package also offers a region-growing algorithm based on recursive splitting of neighboring regions. Examples of the two techniques are given for the domains of aerial cartography and reconnaissance, target cuing, and navigational vision.