【作者】 高彦宇；

【导师】 杨扬；

【作者基本信息】 北京科技大学 ， 控制理论与工程， 2004， 博士

【副题名】And Applications in Bank Cheques Automatic Processing

【摘要】 脱机手写体汉字识别是模式识别研究领域中的难点,但是由于脱机手写体汉字识别有广阔的应用背景(如金融表单自动处理、自动阅卷等),并且囊括了模式识别领域中的所有典型问题,如特征选择、分类器选择以及样本集选择,因此对于它的研究具有深刻的理论意义和实用价值。本论文的主要工作及创新点如下：①提出基于矩特征和弹性网格技术的串行和并行特征融合策略。矩特征提取的是汉字图像的全局特征,尽管这种特征具有优秀的特征表达能力,并且抗噪声、抗变形能力很强,但是它对相似字的识别率不高。弹性网格方法提取的是汉字图像的局部特征,尽管它能更有效地反映汉字的结构细节,但是抗噪能力差。因此,将这两种特征向量有机结合,不仅能同时兼顾手写体汉字的全局和局部特征,而且具有很强的分辨能力。②提出基于多小波正交外壳扩展和多分辨率匹配策略的手写体汉字识别系统。通过多小波变换和正交外壳扩展得到的特征向量对于手写体汉字图像的位移、大小和旋转变化不敏感,而且所采用的多分辨率匹配策略与人眼观察事物的方式相似,能够快速准确地识别手写体汉字样本。③提出一种基于两分法和GLVQ算法的动态单模板字典制作方法和一种基于特征向量分布的多模板字典制作方法。其中,多模板字典的自学习功能能够有效提高整个系统的泛化能力。④分别针对小样本集和海量样本集手写体汉字,提出基于支持向量机的多类分类策略,所得的分类器不仅识别率高、泛化能力强,而且有效解决了多类分类问题。为了进一步提高分类识别的速度,本文将神经网络多类分类方法与最小二乘支持向量机算法结合,对大样本集手写体汉字进行识别,取得了很好的识别率和识别速度。论文的研究成果为脱机手写体汉字识别提供了新的思路和方法,为今后的研究和实际应用奠定了良好的基础。更多还原

【Abstract】 Off-line handwritten Chinese characters recognition (OHCCR) is a formidable task inpattern recognition. It not only has broad applications, such as automatic processing offinancial forms and test papers, but also involves all of the typical problems in patternrecognition, such as features extraction, classifiers selection and sample-set selection.Therefore, the research of OHCCR has great academic and practical values.The main work done in the dissertation and the innovation points were as follows:Proposed a serial features fusion technique and a parallel features fusion technique basedon moment feature and elastic mesh technique. The moment feature was used to extractglobal features. Although it has excellent representation capabilities and robustness in thepresence of noise and distortion, it is incapable of distinguishing similar characters. Theelastic mesh method was used to extract local features of the handwritten Chinesecharacters. Although it can represent the image’s detail information effectively, it issensitive to noise. Therefore, the effective fusion of these features not only presentedhandwritten Chinese characters’ global and local features, but also could get the mostdiscriminative features. Proposed a recognition system based on multi-waveletorthonormal shell expansion and multi-resolution matching strategy. The feature vectorsgot by multi-wavelet transforming and orthonormal shell expansion were insensitive to theshift transformation, scale transformation and rotation transformation of handwrittenChinese characters. The multi-resolution matching approach, which was similar to humansimultaneous interpretation of visual information, could distinguish handwritten Chinesecharacters fast and accurately. Proposed a method for constructing dynamicsingle-template dictionary based on dichotomy and GLVQ algorithm and a method forconstructing multi-template dictionary based on feature vectors’ distribution. Theself-learning strategy of the multi-template dictionary could greatly improve the wholesystem’s generalization ability. Proposed two kinds of multi-classification strategiesbased on support vector machine for recognizing small-set handwritten Chinese charactersand huge-set handwritten Chinese characters respectively. These classifiers not only had excellent performance of good generalization and high accuracy, but also could solve themulti-classification problem effectively. As for improving recognition speed further, aneural network multi-classification method was combined with Least Square SupportVector Machine algorithm for large-set handwritten Chinese characters recognition, whichgot excellent recognition rate and recognition speed.The research results of this dissertation not only proposed novel ideas and methods forOHCCR, but also provided a good foundation for further research and practicalapplications.更多还原