【作者】 刘鑫；

【导师】 魏宏喜；

【作者基本信息】 内蒙古大学 ， 计算机技术， 2018， 硕士

【摘要】 蒙古文古籍是研究蒙古族历史文化的重要资源,为了加强对蒙古文古籍的保护,提升蒙古文古籍在研究中的利用率,越来越多的蒙古文古籍被转化成数字图像形式。但是蒙古文古籍图像不能直接编辑,无法进行有效分析、统计和检索,因此对蒙古文古籍图像进行识别并将其转化为可编辑的电子文档是一项十分有意义的工作。然而,蒙古文古籍采用木刻雕版印刷,因此蒙古文古籍存在着字体形变过大的特点,并且蒙古文古籍成文年代久远,因此,存在着不同程度的污损、缺失与褪色,由此导致了蒙古文古籍单词图像质量低下,这给蒙古文古籍单词图像的识别带来了极大的不便。现今,对蒙古文古籍识别的研究相对较少,主要基于字元切分的方法,即,将待识别的文字切分成字元,以字元作为识别单位,并在此基础上的多知识策略的方法,其存在如下问题:第一、只能适用于高质量,噪声较少的蒙古文古籍单词图像。第二、字元切分的准确性是决定识别结果的重要因素。本文主要研究基于深度学习的蒙古文古籍整词的识别技术,目的是为了解决以往在蒙古文古籍单词图像识别中存在的字元难以有效切分及对图像噪声敏感的问题,希望通过对蒙古文古籍整词进行识别来进一步提升识别正确率。本文以蒙古文古籍《甘珠尔经》为素材,对蒙古文古籍整词识别技术进行了探索,本文的主要研究内容如下:(1)本文使用的实验素材来源于内蒙古大学图书馆的数字化《甘珠尔经》,从中随机选取100页,经过版面分析、二值化、最终切分为20176幅蒙古文古籍单词图像作为实验样本,并将其分为1336类,并且每个样本都进行了人工的标注,由于存在个别类中样本数量过少的问题,即数据集分布不平衡问题,本文采用SMOTE技术(Synthetic Minority Over-sampling Technique)进行样本的扩充。扩充后样本总量达到267200张,作为本文实验的新数据集。(2)针对蒙古文古籍单词图像质量不佳、单词图像进行字元切分时难度较大的问题。根据深度学习在图像识别领域取得的优异性能,以LeNet-5模型作为实验基础模型,本文提出了改进的基于卷积神经网络(Convolutional Neural Network,CNN)的蒙古文古籍整词识别技术和基于循环神经网络(Recurrent Neural Network,RNN)的蒙古文古籍整词识别技术,分别通过改变图像尺寸、训练次数、数据集分布平衡性及循环神经网络中输入粒度大小等因素来观察对模型识别能力的影响,并对卷积神经网络与循环神经网络的表现性能进行了实验结果的对比。(3)上述的整词识别方法,不能解决集外词识别的问题。在此基础上,结合长短期记忆网络(Long Short-Term Memory,LSTM)的优越性,本文又提出了基于CNN-LSTM的蒙古文古籍整词识别方法,虽然识别正确率达到84.5%,低于CNN和LSTM的识别正确率,但成功的解决了集外词的问题。更多还原

【Abstract】 Historical Mongolian documents are important resources for the study of Mongolian history and culture.In order to strengthen the protection of historical Mongolian documents and improve the utilization rate of historical Mongolian documents in the study,more and more historical Mongolian documents have been transformed into digital image form.However,the images of historical Mongolian documents can not be directly edited,and can not be effectively analyzed,counted and retrieved.Therefore,it is a very meaningful work to recognize the images of historical Mongolian documents and convert them into editable electronic documents.But,the historical Mongolian documents are printed by woods where words were carved by different craftsmen,so the same word of historical Mongolian documents has difference in shape.In addition,the historical Mongolian documents have a long history.Therefore,there are different degrees of stain,loss and fading,which leads to the low quality of the word images of historical Mongolian documents.It brings great inconvenience to the recognition of the word images of historical Mongolian documents.Nowadays,there are relatively few researches on recognition of historical Mongolian documents,and it is mainly about the segmentation-based approach,that is,the words to be recognized are segmented into the corresponding glyphs,which are used as the recognition unit.Another approach is called multi-knowledge strategy method which is based on the segmentation approach.It often has the following problems: First,it can only be applied to high-quality and less-noisy historical Mongolian documents word images.Second,the segmentation accuracy of the character is a key factor in determining the recognition result.This paper mainly studies the holistic recognition technology of historical Mongolian document word images based on deep learning.The purpose is to solve the problem that the word images are difficult to be effectively segmented and sensitive to noise,hoping to further improve the recognition accuracy of historical Mongolian documents word images.The historical Mongolian documents named Kanjur are used as material to explore the word images recognition technology of historical Mongolian documents.The main contents of this paper are as follows:(1)The experimental material used in this paper are from the digitized Kanjur which is stored in the Library of Inner Mongolia University.We randomly selected 100 pages from the Kanjur.After the layout analysis,binarization,we cut them into 20176 word images as experimental samples,and divided them into 1336 categories,and each sample is labeled manually.Due to the problem of too few samples in some classes,the SMOTE technique(Synthetic Minority Over-sampling Technique)is used to expand the samples.After extension,the size of the total samples is up to 267200,which is used to be the new data set of this experiment.(2)In view of the poor quality of historical Mongolian documents word images,and problem of word images cannot to be segmented correctly.According to the excellent performance of deep learning in the field of image recognition,the classical convolution neural network LeNet-5 is used as the basic model of the experiment.In this paper,a holistic recognition approach for words of historical Mongolian documents based on an improved convolution neural network(CNN)and a holistic recognition approach for words of historical Mongolian documents based on the recurrent neural network(RNN)are proposed.And by changing the size of the image,the number of training,the balance of the data set distribution and the size of the input particle in recurrent neural network to observe the factors in affecting the model recognition ability.Also,the performance of the convolutional neural network and the recurrent neural network is compared with the experimental results.(3)To the problem of out-of-vocabulary word recognition.Combined with the superiority of Long Short-Term Memory(LSTM),this paper proposes the holistic recognition technology for words of historical Mongolian documents based on CNN-LSTM.It makes the recognition accuracy reach 84.5%.Although,the recognition accuracy is lower than CNN or RNN,it can successfully solves the problem of out-of-vocabulary words.更多还原