一种多模型超图用于手写汉字识别算法  

HANDWRITTEN CHINESE CHARACTER RECOGNITION BASED ON MULTI-MODEL HYPERGRAPH

作　　者：魏炳辉[1] 谢晖慧 邓小鸿[1] Wei Binghui;Xie Huihui;Deng Xiaohong(College of Applied Science,Jiangxi University of Science and Technology,Ganzhou 341000,Jiangxi,China)

机构地区：[1]江西理工大学应用科学学院

出　　处：《计算机应用与软件》2019年第7期192-196,201,共6页Computer Applications and Software

基　　金：国家自然科学基金项目(61762046);江西省教育厅科研项目(GJJ161569);江西省自然科学基金项目(20161BAB212048)

摘　　要：随着银行业提出手填票据自动化处理需求后,对手写汉字的识别技术研究推向新的高潮。由于手写汉字形体复杂多样、训练样本不多,从而导致识别率难以提高。设计一种多模型的超图学习算法来识别手写汉字块,根据训练样本间距离关系构建样本关系阵;以样本的稀疏表示参数为样本间的关系紧密性权重构建另一个样本关系阵;以样本约束法则为基础,以标记样本间的关系权重构建标记样本间的关系阵,融合这几个关系矩阵成为多模型的超图学习框架。通过迭代学习,找出最优的手写汉字块类别归属,在手写汉字块的实验中表现出一定的优势。With the requirement of automatic handwritten bill processing put forward by the banking industry,the handwritten Chinese character recognition technology has reached a new climax. However,due to the complexity and diversity of handwritten Chinese characters and the lack of training samples,it is difficult to improve the recognition rate. For this reason,this paper designed a multi-model hypergraph learning algorithm to recognize handwritten Chinese character blocks. We constructed a sample relationship matrix according to the distance relationship between training samples. Then another sample relationship matrix was constructed based on the sparse representation of samples as the weight of the relationship between samples. Based on the principle of sample constraints,a relationship matrix between marked samples was constructed by the relationship weights between marked samples. These relationship matrices were fused to a multi-model hypergraph learning framework. By iterative learning,the optimal classification of handwritten Chinese character blocks can be found and the proposed method shows certain advantages in the experiment of handwritten Chinese character blocks.

关 键 词：手写汉字识别 多模型超图 成对约束 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]