基于HMM的联机汉字识别系统及其改进的训练方法

本文描述了一个基于HMM模型的联机汉字识别系统的设计思想与实现方法。系统以联机汉字的笔段序列作为观察序列,采用带有多跨越的模型结构消除自由书写汉字笔段序列的冗余与丢失问题。HMM模型的训练是本系统设计的一个重要问题,针对复杂HMM模型参数训练容易收敛于局部最小的情况,本文结合联机汉字识别的特点,提出了一种利用“引导模型”进行训练的改进方法,避免了训练过程收敛于局部最小点的发生。经过大量样本的训练,本系统对规范书写汉字和自由书写汉字均取得了比较令人满意的结果。

A HMM Based On-line Chinese Character Recognition System and Improved Training Algorithm

This paper describes the design and implementation of an on line Chinese Character recognition system, which is based on Hidden Markov Models The strokes of on line Chinese character are regarded as the input observation sequence, and a multi cross left right model structure is employed in order to eliminate the influence caused by redundancy or loosing of strokes The training of HMM models is also an important problem for this system, in order to avoid the training process falls into local minimum, an improved training approach is proposed After sufficient training, this system gains an satisfying result for both ordinary writing characters and free style writing characters