【作者】 杨帆；

【导师】 飞龙；

【作者基本信息】 内蒙古大学 ， 计算机技术， 2021， 硕士

【摘要】 手写识别是智能化人机交互的重要课题之一,根据识别的方法可分为在线手写识别和离线手写识别。传统蒙古文作为我国蒙古族的语言文字,是我国少数民族的文化瑰宝。蒙古文的文字识别研究开展于二十一世纪初。如今,蒙古文离线识别的研发方法趋于成熟。在市场上已经出现用于传统蒙古文识别的OCR软件,可以进行蒙古文古籍、印刷体文档识别等任务。但是,对于蒙古文在线手写识别,由于手写文字写法随意、有效数据收集困难、蒙古文手写字符难以分割等原因,研究工作还存在着诸多不足。研究蒙古文在线手写识别可以加强科技信息技术在民族地区的普及和应用,有助于蒙古文智能信息化技术的发展与应用,对于传统蒙古文的传承和保护具有重要意义。本文围绕蒙古文在线手写识别任务展开深入研究,主要工作如下:1.构建了蒙古文手写数据语料库,并改进预处理流程中的重采样处理方法,提出了一套适用于蒙古文在线手写坐标序列的预处理流程。实验结果表明,该方法可以较好的优化手写样本的数据表示能力,并有效的提高蒙古文手写识别的准确率。2.本文提出了结合卷积网络、自注意力模型和注意力机制的序列到序列蒙古文在线手写识别模型,实现了模型自主提取笔段和笔画特征并进行无分割字符级识别。同时,提出了基于词典引导的集束搜索算法进行解码实现蒙古文整词识别。实验对比结果表明,本文提出的识别模型提高了蒙古文在线手写单词的识别率,Top10预测结果在测试集上正确率提高到了89.77%。3.搭建了蒙古文在线手写识别云服务系统。系统采用浏览器/服务器架构(Browser/Server,B/S)设计,利用Python第三方库Tornado开发了高并发量网络接口服务,并嵌入到蒙古文智能整词输入法等软件中得到了广泛的应用。更多还原

【Abstract】 Handwriting recognition is one of the significant topics of intelligent human-computer interaction,which can be divided into online handwriting recognition and offline handwriting recognition according to the method of recognition.Traditional Mongolian,as the language and script of Mongolian people in China,is a cultural treasure of minority people in China.The research of Mongolian handwriting recognition was conducted at the beginning of the 21 st century.Nowadays,the researches of offline Mongolian recognition are becoming mature.OCR software for traditional Mongolian recognition has appeared in the market for tasks such as recognition of Mongolian ancient books and printed documents.However,for Mongolian online handwriting recognition,there are still many shortcomings in the researches due to the arbitrary handwriting writing style,the difficulty of collecting effective data,and the difficulty of segmenting Mongolian handwritten characters.The study of Mongolian online handwriting recognition can strengthen the popularization and application of scientific and technological information technology in ethnic areas,help the development and application of Mongolian intelligent information technology,and is of great significance to the inheritance and protection of traditional Mongolian.This thesis presents an in-depth study of the task of online handwriting recognition in Mongolian,and the main work is as follows:1.A corpus of Mongolian handwriting data is constructed,and the resampling processing method in the preprocessing process is improved.A set of preprocessing processes applicable to Mongolian online handwriting coordinate sequences is proposed.The experimental results show that the method can optimize the data representation ability of handwritten samples and effectively improve the accuracy rate of Mongolian handwriting recognition.2.In this thesis,a sequence-to-sequence Mongolian online handwriting recognition model combining convolutional network,self-attentive model and attention mechanism is proposed to extract stroke segment and stroke features autonomously and perform segmentation-free character-level recognition.Meanwhile,a lexicon-guided beam search algorithm based on decoding is proposed to achieve Mongolian word recognition.The experimental comparison results show that the recognition model proposed in this thesis significantly improves the recognition rate of Mongolian online handwritten words,and the Top10 prediction results improve to89.77% on the test set.3.A Mongolian online handwriting recognition cloud service system was built.The system was designed with browser/server architecture,and a high concurrency web interface service was developed using Tornado,and embedded into Mongolian intelligent whole-word input method and other software to be widely used.更多还原