Split-and-link algorithms for image segmentation

Abstract Pavlidis has investigated a “split-and-merge” approach to image segmentation based on recursive subdivision into quadrants; an image block is split into quadrants if it is too inhomogeneous, or the quadrants are merged into a single block if the result is sufficiently homogeneous. This approach involves a “pyramid” of successively reduced-resolution versions of the input image, defined by 2-by-2 block averaging. More recently, an approach to segmentation has been studied that makes use of an overlapped “pyramid”, defined by 4-by-4 block averaging with 50%; overlap in each direction, and creates linked subtrees in this pyramid based on similarities between blocks and their subblocks. This paper compares the two approaches and proposes a hybrid “split-and link” approach that combines features of both of them.