Recognition of Handwritten Characters by Voronoi Representations Recognition of Handwritten Characters by Voronoi Representations

We present a new skeletonization algorithm well suited for the problem of handprinted character recognition. Our approach employs a novel algorithm for computing the Voronoi diagram of a polygon with holes. We show that Voronoi skeletons can serve as eecient shape descriptors because they preserve connectivity and Euclidean metrics. Compared to traditional skeletonization techniques, we suggest that shape representations based on Voronoi skeletons may increase the reliability of production quality character recognition systems. A feasibility study is described in which more than 10,000 handprinted characters were recognized with an error rate of 2.34% by a neural network trained using Voronoi skeletons of character shapes from a class of 52 distinct alphanumeric patterns and graphical symbols. These results show that feature vectors extracted from Voronoi skeletons provide for high reliability in handprinted character recognition at a reduced cost of representation.