基于笔段间关系的联机手写汉字HMM模型

HMM model for online handwritten Chinese character recognition describing the correlations between segments

为了提高联机手写汉字模型的空间结构描述能力和识别性能,从汉字的笔段关系出发,提出一种新的联机手写汉字模型,该文称之为属性关系Markov模型(ARHMM)。ARHMM以经典Markov模型(THMM)为基础,结合属性关系图对二维图形结构的描述特点,提出了一种直接描述状态间关系的新观测量,从而更充分地利用汉字的空间结构信息。ARHMM具有完整的参数训练方法和识别算法。联机汉字识别的实验结果表明:ARHMM联机汉字模型与THMM联机汉字模型相比,在工整书写到自由书写的不同质量汉字样本上识别错误率均有所下降,平均错误率下降了23.65%。

This paper proposes a new type of hidden Markov model called the attributed relation hidden Markov model (ARHMM) which combines the advantages of the traditional HMM and the attributed relation graph (ARG). The model uses new observations directly describing the correlations between the states with the original observations, while preserving the HMM mathematical structure. The modified ARHMM learning methods and recognition algorithms are presented. In an experiment with online handwritten Chinese character recognition, ARHMM was utilized to describe the correlations between the character segments. The results demonstrate that this model performed much better than a traditional HMM for all samples with different qualities.