Selective feature-to-feature adhesion for recognition of cursive handprinted characters

A structural-feature-to-structural-feature configuration is naturally constructed using a set of sampled features from a cursive pattern. These features are sampled by maximally fitting bended ellipses in local strokes. This configuration is transformed into an undirected graph to resolve the asymmetric difficulty. The compatibility associated with the graph is further formulated into a devised Hopfield network, where both interfeature and interlink similarities are incorporated into the compatibility. We operate this network to recognize a radical as a whole in a handprinted pattern to accomplish the selective attention task.