【作者】 玛依热·依布拉音；

【导师】 艾斯卡尔·艾木都拉；

【作者基本信息】 武汉大学 ， 通信与信息系统， 2013， 博士

【摘要】 联机手写维吾尔文字识别是模式识别领域的难题之一,也是一门综合性技术。近年来,随着智能手机、平板电脑、电子白板等移动终端设备的普及,联机手写输入做为一种自然、方便的输入方法,已经得到了高度重视,并在日常生活中得到了广泛应用。但是,新疆少数民族地区广泛使用的官方语言——维吾尔语的文字识别研究相对滞后,对联机维吾尔文手写识别技术的研究也非常少。维吾尔文字手写输入系统的研究不仅对其他民族文字手写输入系统的研究也有一定的参考价值,而且促进少数民族地区的信息技术发展。维吾尔语文字是一种拼音文字,其书写方式与汉文和西文有很大不同。维吾尔文是从右到左写,每个词中所有字母连着写,且每一个字母在一个词的词首、词中和词尾所取的字形不一样。这些特点给维吾尔语文字的识别带来很大的困难。本文在深入调研国内外联机手写识别技术的研究动态,并通过分析维吾尔文字母与单词自身的结构和书写特点,在联机手写维吾尔文字母识别、维吾尔文单词切分、维吾尔文单词识别等三方面做了一些有益的探索。主要工作包括以下几个方面：一、首次分别建立了手写维吾尔文字母样本库和手写单词样本库。为了支持维吾尔文手写识别方面的研究,收集整理了联机手写数据库。维吾尔文字母样本库包括32个维吾尔文字母的四种不同书写类型(128类),由612个人书写,总共包含78336个样本。手写单词样本库包含常用的1460的单词,由400个人书写,总共有584000个维吾尔文单词。该数据库可用于手写字符识别、文档检索和笔迹鉴别等多方面的研究。二、提出了一种联机手写维吾尔文字母识别方案。选择在手写汉字识别技术中所提出来的归一化、特征提取、以及常用的分类方法,从中找出了最佳的技术选择。在实验对比中,本文采用了线性归一化(LN)、矩归一化(MN)、二分矩归一化(BMN)、中心边界对准(CBA)、改进的中心边界对准(MCBA)以及几种相应的伪二维归一化方法等八种不同的归一化预处理方法,基于坐标归一化的特征提取(NCFE)方法,以及MQDF(改进的二次分类函数)、DLQDF(判别学习型二次判别函数)、LVQ(学习矢量量化)、SVM(支持向量机)四种分类器。同时,再考虑字符在文档中的空间几何特征,进一步提高了识别性能。在128个维吾尔文字母类别、12,800个测试样本的实验中,正确识别率最高达到了89.08%,为进一步研究面向维吾尔文字母特性的识别技术奠定了重要基础。三、设计了一个基于DTW的联机手写维吾尔文字母识别系统框架。对预处理、特征提取、聚类分析和匹配识别等模块进行了较系统地理论和实验研究。在预处理中,采用了维吾尔文字母的线性归一化和基于点密度的一化的方法；特征提取中,采用了结合结构特征和统计特征的方法并提取了均匀采样特征、方向链表特征、网格方向特征、投影特征等;聚类分析中中采用基于最小生成树(MST)的动态聚类算法,分类器采用最近邻分类方法。测试结果表明,独立式、前连式、后连式、双连式字母的识别率分别达到74.67%,70.43%,63.33%,72.02%。出现在5个候选字母中的识别率达到94.34%,94.19%,93.15%,95.86%；四、提出了基于多分类器集成维吾尔文字母识别方法。多分类器组合能够在一定程度上弥补单个分类器的缺陷,因此它在模式识别中得到了广泛的应用。本文利用5种不同的特征提取方法构造了5个独立的分类识别器,采用了等权投票和不等权投票等两种策略将5种分类识别器进行了有效组合。其中，单分类器采用了基于动态时间弯折匹配距离的最近邻分类方法。实验结果表明,本文提出的集成策略的识别率明显高于单分类器的识别率,而且为特征的综合集成提供了多种有效途。五、通过对手写维吾尔文字中的字母连接特点深入研究,提出了一种有效的基于动态规划的联机手写单词分别方案。首先,对单词进行过切分。去掉单词中的附件部分后,通过分析主要笔划书写轨迹的形状,找出潜在的过分割点并合并被切分成的基本块与对应它的附加部分,得到基本字母片段序列;然后,再利用单字母分类信息和基于切分块的几何信息进行融合,采用动态规划算法来进行评价分析,从而寻找出最优的分割路径。对于联机单词样本进行的实验证明,该文所提出的算法对于维吾尔文单词的分割有很好的效果。六、提出了基于多步分割的维吾尔文单词切分方法。本文采用了先把单词分割成连体段,然后把连体段切分成字母的两级切分方案,并针对连体段分割问题和字母切分问题分别提出了连体段分割算法和字母切分算法。七、提出了基于词典驱动的、集成切分与识别的联机手写维吾尔文单词识别框架和方法。系统中把单词识别问题转化为一个词典中的词条与手写单词图像匹配的优化问题。维吾尔文单词识别的主要难点是字母在被识别之前不能准确地切分。解决方法是把字母分割与识别集成起来,通过组合搜索得到最优的切分和识别结果。第一步,利用过切分算法将单词进行切分,对相邻的基本片段进行组合形成切分候选网格。第二步,采用词典驱动的方法,将字母识别信息、几何信息和词典信息一起加入到单词识别系统的路径匹配过程。本文采用用置信度转换的方法,将分类器的输出转换成概率的形式,使参数调整更为方便。动态规划算法应用于单词识别过程中的最优路径匹配过程,得到最优识别结果。词典中的词条分别为100、500、1000、10000的情况下,单词识别实验结果分别为84%、78%、68%和47%。更多还原

【Abstract】 On-line Handwritten Character Recognition is one of challenging important topic in the field of pattern recognition, but also a comprehensive technology. In recent years, with the increasing use of mobile devices such as cell phone, tablet computers and digital pen, on-line handwriting input as a natural, convenient has been attached great importance to and has been widely used in daily life. Uyghur language is an official language and very popular in the Xinjiang province of China. However, the research about the technology for Uyghur handwritten recognition is lagging much behind and little work has been done in this area. The research of recognition techniques for online handwritten Uyghur characters are not only has great reference values for starting the research of other ethnic group’s handwritten scripture recognition, but also has a far-reaching meaning about developing the information technology and national culture of specific ethnic group.Uyghur words are formed by concatenation of the characters, which has a very special written structure different from Chinese and English characters. It is written from right to left; every letter may have different shapes in different positions. All these characteristics bring many difficulties to recognition. In this paper, through in-depth study the research trend of on-line handwriting recognition technology in domestic and abroad, based on analysis of the unique shape and writing styles of Uyghur characters, proposes an approach for online handwritten Uyghur character and word recognition. The major contributions of this dissertation are as follows:1. A handwritten Uyghur character database and a handwritten word database are established for the first time respectively. In order to support the research of Uyghur handwritten recognition, we collected the samples of online Uyghur handwriting. The datasets of Uyghur characters contain78,336samples of128classes (including four different types of32characters set), are handwritten by612volunteers includes students and teachers. The datasets of Uyghur words contain584,000samples of commonly used1460words, individually are handwritten by400volunteers also includes students and teachers. The database can be used for typical research tasks of handwritten document analysis such as handwritten recognition, handwritten document retrieval and writer identification etc.2. An online handwritten Uyghur characters recognition framework have been presented. We evaluate various techniques of normalization, feature extraction and classification that have been successfully applied in handwritten Chinese character recognition. Specifically, we use eight normalization techniques such as liner normalization (LN), moment normalization (MN), bi-moment normalization (BMN), Centroid-boudary alignment (CBA) and several corresponding pseudo2D normalization methods. We use the normalization cooperated feature extraction (NCFE) method with different settings. For classification, we use four classifiers, namely, the modified quadratic discriminant function (MQDF), the discriminative learning quadratic discriminant function (DLQDF), the learning vector quantization (LVQ) classifier, and the support vector classifier with RBF kernel (SVC-rbf). Furthermore, the geometric features which characterizing the spatial context in handwritten documents are extracted for enhance the recognition performance. In experiments on38,400test samples of128classes, the proposed approach achieved an accuracy of89.08%.3. we designed a framework for online handwritten Uyghur characters recognition system based on DTW and carried out a more systematic theoretical and experimental research on its’module, such as pre-processing, feature extraction, Cluster analysis and classifier. In the pre-processing, in order to obtain the structure information of characters, according to handwritten Uyghur character’s feature, we use linear normalization and nonlinear normalization based on dot density method. Taking into account the more similar characters in the Uyghur language, use the feature extraction method of combined with the structural features and statistical features, such as uniform sampling feature, direction feature, grid direction density feature, two directional projection feature. Cluster analysis use the dynamic clustering algorithm based on the minimum spanning tree (MST), and classifier use the nearest neighbor matching classification. The experimental testing has been carried out and the results show that over-all recognition rate for four different character shapes is respectively74.67%,70.42%,63.33%,72.02%; the recognition rate for the handwritten characters which are recognized as in one of the two candidate characters are respectively86.85%,86.09%,80.43%,88.41%, and one of the five candidate characters are94.34%,94.19%,93.15%,95.86%respectively.4. And an online handwritten Uyghur characters recognition method based on the integration of multiple classifiers have been presented. Combination of multiple classifiers, a certain extent, compensate for defects of a single classifier, so it has been widely applied in pattern recognition. In our research, we applies five different feature extraction methods to construct five separate classifier and using voting strategy of ranging from rights to effective combined five kinds of classifier. Each classifier use the nearest neighbor classification method based on dynamic time bending matching distance. Experimental results show that the recognition rate based on integration strategy is significantly higher than the recognition rate of separate classifier, and it also provide a variety of effective ways for the comprehensive integration of features.5. To be enabled to separate the many connected characters in cursive Uyghur handwriting, we present a novel character segmentation method using dynamic programming. Firstly, after removing delayed strokes from the handwritten words, potential breakpoints are detected from concavities and ligatures by temporal and shape analysis of the stroke trajectory. Reconstruct delayed strokes and obtained a sequence of primitive segments. Then, by merging the neighboring blocks, create candidate segmentation paths. Then paths were evaluated by the character recognition and geometric information, and a dynamic programming method is applied to find the best segmentation point for each character. Our preliminary experiments on an online Uyghur word dataset demonstrate that the proposed method can achieve good performance in segmenting cursive handwritten Uyghur characters.6. On the issues of characters segmentation, we adopted the two level segmentation scheme in which the word segmented into conjoined section firstly, and then the conjoined section cut into characters in the next steps. We put forward conjoined section segmentation algorithm and characters segmentation algorithm for conjoined section segmentation problem includes characters segmentation problem.7. The online handwritten Uyghur word recognition approaches based on a lexicon-driven, integrated segmentation and recognition have been presented. Word recognition problem is transformed into matching optimization problems between the dictionary entry and the handwritten word image. There are many connected characters in cursive Uyghur writing, which makes the segmentation and recognition of Uyghur words very difficult. The solution is using of integrating the segmentation and recognition method to obtain the optimal segmentation and recognition results came from combined search. The first step, using of the over-segmentation algorithm to word separation, formed the segmentation candidate grids by combining adjacent fragments. In the second step, using lexicon-driven approach, combined with character recognition information, geometric information and dictionary information into path matching procedure in the word recognition system. Our preliminary experiments on an online Uyghur word dataset demonstrate that the proposed method can give high recall rate of segmentation point detection. Then using the confidence transformation method convert the similarity scores into probabilities, such that the tuning of weighting parameters becomes easier. Dynamic matching between characters in the lexicon entry and segment(s) of the input word image is used to ranking the lexicon entries in order to get best match. As the result the performance for lexicons of size100,500,1000,10000are84%、78%、68%and 47%respectively.更多还原