SEGMENTATION OF MERGED CHARACTERS BY NEURAL NETWORKS AND SHORTEST-PATH

A major problem with a neural network-based approach to printed charac ter recognition is the segmentation of merged characters. A hybrid met hod is proposed which combines a neural network-based deferred segment ation scheme with conventional immediate segmentation techniques. In t he deferred segmentation, a neural network is employed to distinguish single characters from composites. To find a proper vertical cut that separates a composite, a shortest-path algorithm seeking minimal-penal ty curved cuts is used. Integrating those components with a multiresol ution neural network OCR and an efficient spelling checker, the result ing system significantly improves its ability to read omnifont documen t text.