Using Binary Pyramids to Create Multiresolution Shape Descriptors

The analysis of a 2D graphical document can be accomplished by using a suitable linear representation, e.g. the skeleton, of the pattern included in the document. Multiresolution representation and description are desirable for pattern recognition applications, as it reduces the complexity of the matching phase. In this paper, multiresolution shape descriptors of 2D graphical documents are obtained by using binary AND-pyramids. A multiscale representation is first obtained by simply extracting the skeleton of the pattern at all resolution levels of the pyramid. The so obtained skeletons are then transformed into multiresolution structures by suitably ranking skeleton subsets, based on their permanence at the various scales. The two different types of hierarchy built in this way both contribute to facilitate recognition. In fact, the skeleton is available at various scales, so one could initially match roughly using only skeletons at lower scales, where only the most significant parts of the pattern are represented. In turn, at each scale, skeleton subsets are furthermore ranked according to their permanence along the pyramid levels, thus reducing the number of prototypes for which a more detailed comparison is necessary.