基于部件HMM级联的联机手写体汉字识别方法  

An on-line free handwritten Chinese character recognition method based on component cascaded HMMs

作　　者：赵巍[1] 刘家锋[1] 唐降龙[1] 

机构地区：[1]哈尔滨工业大学计算机科学与技术学院,黑龙江哈尔滨150001

出　　处：《哈尔滨工业大学学报》2004年第5期570-573,共4页Journal of Harbin Institute of Technology

基　　金：国家高技术研究发展计划资助项目(2001AA114041).

摘　　要：为了对自由手写汉字进行有效地表征和识别,提出了一种识别自由手写体汉字的级联HMM方法,在部件HMM模型基础上将各模型按照统计概率连接,它扩展了HMM的模式描述方式,允许在级联模型上表征状态的跳跃、转移和驻留等.通过共享手写汉字部件模型来描述级联状态转移概率,可以更可靠地刻画自由手写体的行为特点.采用面向级联的Viterbi算法,无需做部件的分割和标注.通过一定条件下的对比实验训练与识别表明,该方法的第一候选识别率为87 89%,而基于分段HMM识别方法的第一候选识别率为86 17%,降低错误识别率12 4%.This paper presents a cascaded Hidden Markov Model(HMM),which allows state's transition, skip and duration.The cascaded HMM extends the way of HMM pattern description of Handwritten Chinese Character(HCC)and depicts the behavior of handwritten curve more reliably in terms of the statistic probability.Hence the character segmentation and labeling are unnecessary. viterbi algorithm is integated in the cascaded HMM after the whole sample sequence of a HCC is input. Compared with the Segment HMMs approach, the recognition rate of this model for the first candidate is 87.89% and the error rate could be reduced 12.4%.

关 键 词：HMM级联 手写体 汉字识别方法 VITERBI算法 第一候选识别率 

分 类 号：TP391.43[自动化与计算机技术—计算机应用技术]