基于卷积神经网络的藏文手写数字和字母识别研究  

Research on Tibetan handwritten numbers and letter recognition based on convolutional neural network

作　　者：周毛克 才让先木 龙从军 闫晓东[2] ZHOU Mao-ke;CAI Rang-xianmu;LONG Cong-jun;YAN Xiao-dong(University of Chinese Academy of Social Sciences,Beijing 102488,China;Minzu University of China,Beijing 100081,China;Institute of Ethnology and Anthropology,Chinese Academy of Social Sciences,Beijing 100081,China)

机构地区：[1]中国社会科学院大学(研究生院),北京102488 [2]中央民族大学,北京100081 [3]中国社会科学院民族学与人类学研究所,北京100081

出　　处：《青海师范大学学报（自然科学版）》2019年第4期34-39,共6页Journal of Qinghai Normal University(Natural Science Edition)

基　　金：国家语委重点项目(ZDI135-17);国家社科基金后续资助项目(18FYY003);中国社科院创新工程项目(2019MZSCX005)

摘　　要：在文字识别领域中,手写体识别比印刷体识别更具挑战性.藏文手写体识别已经成为重要的研究课题之一.本文提出了一种基于卷积神经网络LeNet-5模型的藏文手写数字和字母识别方法.分别采集藏文数字手写体样本和字母手写体样本17768和77636例,并对其进行预处理;然后按8∶2划分成训练集和测试集,并在CNN(LeNet-5)模型上进行训练.经过测试,数字和字母识别准确率分别达到98.81%和97.89%.In the field of text recognition,the recognition of handwriting is more difficult than the recognition of printed matter.Text recognition has become one of the important topics.in Tibetan information processing.This paper proposes a handwritten Tibetan numeral and letter recognition technology based on the convolutional neural network LeNet-5 model.We collected 17768 cases and 76,636 cases of Tibetan numerals and Tibetan letters,and preprocessed the data samples.Then divide the training set and test set by a ratio of 8∶2 and Training on the CNN(LeNet-5)model,the experimental results verify that the accuracy of numbers and letters on the test set reached 98.81%and 97.89%,respectively.

关 键 词：文字识别 卷积神经网络 藏文手写数字字母 LeNet-5模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]