一种重叠书写中文手写句子实时识别方法.pdfVIP

二〇一二年第二届全国文字与计算学术研讨会 一种重叠书写中文手写句子实时识别方法 1 2 3 3 吕彦飞 ，黄琳琳 ，王大寒 ，刘成林 1 北京航空航天大学自动化科学与电气工程学院， 北京 100191 2 北京交通大学电子信息工程学院，北京 100044 3 中国科学院自动化研究所模式识别国家重点实验室，北京 100190 摘 要：本文提出了一种针对重叠书写的中文手写句子实时识别方法。采用切分和识别集成的字符串识别框 架，首先用支持向量机判断新输入的笔划是否属于新的笔划块，并对笔划块进行合并得到候选字符，进而 用单字分类器对候选字符进行识别得到候选类别，形成一个候选切分-识别网格。每当提笔时间大于设定的 阈值时，系统假设字符串输入完毕，并对候选切分-识别网格进行最优路径搜索得到句子识别结果。本文方 法充分地利用了笔划块的几何信息，针对重叠书写的特点，设计了有效的几何模型加入到路径评价函数中， 并结合单字分类器和语言模型，提高了字符切分和识别正确率。在生成的联机中文重叠书写字符串样本上 进行测试，得到了91.55%的字符正确率，表明了本文方法的有效性。 关键词：重叠书写；实时识别；过切分；几何模型 An Approach for Real-Time Recognition of Online Overlaid Handwritten Sentences 1 2 3 3 LV Yanfei , HUANG Linlin , WANG Dahan , LIU Chenglin 1School of Automation Science and Electrical Engineering, Beihang University, Beijing 100191 2School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing 100044 3National Laboratory of Pattern Recognition, Institute of Automation of Chinese Academy of Science, Beijing 100190 Abstract:This paper proposes an approach for real-time recognition of online overlaid handwritten Chinese sentences. The main feature is a dynamically maintained segmentation-recognition candidate lattice combined with specifically designed geometric contexts for overlaid characters. Whenever a new stroke is input, it is judged by a support vector machine (SVM) whether it is a new character component or not, and the component sequence is updated accordingly. From the component sequence, candidate characters are generated by over-segmentation and are recognized by a character classifier to update the candidate recognition-segmentation lattice. When the pen lift time exceeds a threshold, the string