Vectorized image segmentation via trixel agglomeration

We present a broad algorithmic framework for transforming an image comprised of pixels into a vectorized image segmented into polygons that can be subsequently used in image analysis, understanding, and high-level representation. We introduce a new entity—the trixel—that is an adaptive, local, triangular unit of region-contour picture description. Adjacent trixels are merged based on certain elementary heuristic criteria of perceptual organization, modeled as geometric filters on trixel edges. The connected components of a trixel grouping graph, with vertices representing trixels and edges between vertices corresponding to merged adjacent trixels, yield polygons that segment the image.