A Minimax Classification Approach to HMM-Based Online Handwritten Chinese Character Recognition Robust Against Affine Distortions

We present a minimax classification approach to online recognition of isolated handwritten Chinese characters, which is robust against global affine distortions of an input handwriting sample. According to the nature of features used in our continuous-density hidden Markov model (CDHMM) based handwriting recognizer, four types of affine transformations with different degrees of freedom are proposed to serve as a possible distortion model. The corresponding formulations are derived and presented for minimax classification rule. The effectiveness of the proposed approach is demonstrated by an experimental study on the Nakayosi and Kuchibue Japanese character databases.