Tree pattern matching for 2D multiresolution objects

Multiresolution representations have been proposed for object recognition since they provide the framework for emulating the focus of attention strategy typical of biological systems. The authors consider the problem of matching two-dimensional objects described by multiresolution representations. Each object is modeled as a tree, in which notes correspond to boundary segments and arcs connect nodes at successive levels of resolution. The children of a given node describe the structural change occurred to a given segment between consecutive resolution levels. A two-dimensional object O can be recognized as an instance of M, possibly rotated and translated, if the corresponding tree T/sub O/ can be mapped into the tree T/sub M/ by cyclically rotating the children of the root. The authors describe two algorithms for matching two-dimensional objects represented by labelled trees. The first algorithm is for exact matching and determines whether one tree can be mapped into the other by cyclically rotating the children of the root. The algorithm uses a top-down strategy and takes time linear in the number of nodes of the trees. The authors then outline an exact tree matching algorithm that finds the best correspondence between nodes at all levels of the two trees according to a given figure of merit and satisfies certain constraints.