An algorithm is described that is capable of decomposing homogeneous image regions into a complete set of convex parts ordered by size. The process is information-preserving and computationally efficient, and does not require prior edge smoothing. Shape decomposition is useful for parsing undersegmented image regions into which parts correspond to separate semantic entities. The algorithm is efficient because it operates on a list-oriented form of shape representation. Results of decomposing several different shapes are presented and discussed.>

The authors describe a connectionist network which is not a neural net in which knowledge is distributed globally via the connections of functionally simple nodes. The technique combines constraint propagation with analogical or eidetic representations. It represents an alternative to the atomistic signals-to-symbols paradigm in which a one-to-one correspondence is assumed to exist between symbols and the concepts they represent.>