基于深度学习的印刷体文档字符识别的研究
2020-12-23徐长英赖伟财陈英
徐长英 赖伟财 陈英



摘 要: 针对传统基于模板匹配光学识别效果存在不理想的状态,提出改进的深度学习模型的印刷体文档字符识别算法。首先,生成包括一级字库、部分二级字库、英文大小写字母和标点符号的图片数据集,其数量大约为500万张;然后,在Lenet?5网络模型的基础上进行改进和重新构造,提出一种增强型的深度学习模型Lenet?5Pro,该模型可提高印刷体文档的识别率;最后,对比实验结果表明,该模型可以更加有效地提高印刷体字符识别的准确率,其字符识别准确率达到98%以上。
关键词: 印刷体字符识别; 深度学习; 图片数据集; Lenet?5Pro; 字符增强; 仿真分析
中图分类号: TN911.73?34; TP391 文獻标识码: A 文章编号: 1004?373X(2020)23?0072?04
Abstract: In view of the unsatisfactory effect of the traditional optical recognition based on template matching, a printed document character recognition algorithm based on improved deep learning model is proposed. A picture dataset including the first?level font library, part of the secondary font library, English upper and lower case letters, and punctuation marks is generated, in which about 5 million pieces of pictures are collected. An enhanced deep learning model Lenet?5Pro, by which the recognition rate of printed documents can be improved, is proposed based on the improvement and reconstruction of Lenet?5 network model. The comparative experimental results show that the accuracy of character recognition can be improved by the proposed model, and its character recognition accuracy is over 98%.
Keywords: printed document character recognition; deep learning; image dataset; Lenet?5Pro; character enhancement; simulation analysis
0 引 言
印刷体文档字符识别是光学字符识别(Optical Character Recognition,OCR)技术的重要组成部分,印刷体文档的识别基本用途是把图片输入计算机,计算机输出识别字符,实现人与计算机信息的交互。文献[1]利用卷积神经网络构建了印刷体汉字识别模型,使用集成了传统OCR识别技术的MODI(Microsoft Office Document Imaging)进行身份证汉字识别。文献[2]提出了一种基于深度信念网络融合模型对手写汉字识别的方法,简单的汉字使用基于SVM(Support Vector Machine)的二次判别函数分类器识别,使用深度信念网络模型处理较为复杂的汉字图像。文献[3]使用基于神经网络反馈的方法对所提取的文本行基于像素点进行判断而进行二值化,结合垂直投影方法对字符切分,提高OCR识别率。文献[4]提出了一种无分割的端到端神经模型,用于离线光学字符识……
