A new technique is proposed for the recognition of hand-printedLatin characters using artificial neural networks together withconventional techniques. One advantage of this technique is that itcombines rule-based (structural) and classification tests. Anotheradvantage is that it it more efficient for large and complex sets. Thetechnique can be divided into five major steps: (1) digitization of theimage; (2) thinning of the binary image using a parallel thinningalgorithm; tracing of the tree; (3) tracing of the skeleton of the imageand construction of a binary tree; (4) extraction of features from thestructural information; and (5) classification of the segmenteddescriptions as particular characters by a feedforward neural networktrained by backpropagation
展开▼