A natural stroke-based structural approach to loosely-constrained handwritten Chinese character recognition

Proposes a natural stroke-based recognition approach to deal with the structural deformation problem in off-line, loosely-constrained handwritten Chinese character recognition. A layered, modular neural network architecture is employed to address problems like shifting, distortion and scaling. Knowledge initialization of the network is speeded up by direct mapping of the character structure knowledge (which is expressed as rules) onto the network. Natural strokes extracted by a fuzzy stroke extractor are input to the recognizer. The proposed rule-mapped network model closely resembles the hierarchical nature of the Chinese character set. 120 categories of handwriting samples are tested, and our recognizer seems to deal with deformations among the samples satisfactorily