摘要
针对非限制脱机手写体汉字识别率不高的问题,该文提出了一种级联MQDF(改进的二次分类函数)分类器,以提高识别正确率。它是一种基于串行结构的分类器集成算法,并在度量层次对各级分类器结果进行融合。广义置信度被用作评价识别结果好坏的度量。该算法分级构建多个Gauss模型,实现对样本分布精细的描述,达到提高识别率的目的。该文利用最大似然框架论述了该集成算法的工作原理。用该算法在HCL2000及THOCR-HCD手写体汉字数据库上进行试验,识别错误率分别下降了10.75%、9.82%和25.31%,证明了算法的有效性。
The accuracy of unconstrained handwritten character recognition is improved by a cascade MQDF classifier. The cascade structure-based ensemble method combines recognition results at the measurement level. The generalized recognition confidence is used as the measurement of recognition result. The algorithm uses multilayer Gaussian models to elaborately descript the sample distributions to improve the recognition rate. The paper then uses the maximum likelihood model to interpret the ensemble algorithm mechanism. The algorithm was applied to the HCL2000 and THOCR-HCD handwritten Chinese character recognition databases and achieved 10.75%, 9.82% and 25.31% error reductions.
出处
《清华大学学报(自然科学版)》
EI
CAS
CSCD
北大核心
2008年第10期1609-1612,共4页
Journal of Tsinghua University(Science and Technology)
基金
国家自然科学基金资助项目(60472002)
关键词
字符识别
分类器集成
脱机手写体汉字识别
广义置信度
改进的二次分类函数
character recognition
classifier ensemble
offline handwritten character recognition
generalized recognition confidence
modified quadratic discriminant function (MQDF)