Incremental learning for linear fusion of handwritten Chinese character classifiers

Describes an incremental learning technique for linear fusion of experts in the recognition of handwritten simplified Chinese characters from paper records. Each expert has been designed using a specific feature extraction method and a classifier paradigm. A tuple-based histogramming approach and discrete hidden Markov models have been used. The recognition accuracy achieved for all 3755 common simplified Chinese characters in GB1 is 88% for uniform coefficients and 97.60% after using the proposed linear fusion method for determining the weighting of these combination coefficients. An error reduction of 80% in achieved. The method recognizes isolated characters only and not words or phrases