电子通信论文编号:TX214   论文字数:18949,页数:42                     
目   录
第1章  绪  论 1
1.1  选题的背景及意义 1
1.2  脱机手写识别发展概况 3
1.3  手写体数字的特点 4
1.4  本文的组织结构 5
第2章 图像预处理 6
2.1  数据获取 6
2.2  预处理 6
2.2.1 图像的灰度化 6
2.2.2 图像的二值化 7
2.2.3 图像的平滑去噪 7
2.2.4 图像的字符分割 8
2.2.5 图像的归一化处理 9
第3章  特征提取与选择 10
3.1 样本特征库初步分析 12
3.2 样品筛选处理 13
3.3 特征筛选处理 13
3.4 手写数字特征提取与分析 14
第4章 聚类识别理论支持之线性判据 16
4.1 对象相似性的测度 16
4.1.1 对象间距离种类 16
4.1.2 距离测度算法简介 18
4.2  Bayes决策理论之于手写体字符聚类识别 18
4.2  几种聚类判据算法浅析 20
4.2.1 试探性未知类别聚类算法 20
4.2.2 模糊聚类算法 21
4.2.3 动态聚类算法 21
第5章 聚类实现及结果分析 23
5.1 聚类软件执行流程 23
5.2 软件执行结果分析 26
结 论 29
致 谢 31
参考文献 32
附 录1 33
附 录2 34

摘要：脱机手写体数字识别是模式识别领域一个极具挑战性的课题，它将在信函分拣、银行支票识别、统计报表处理以及手写ID的自动输入等诸多方面发挥巨大的作用。然而，手写体数字的书写随意性很大，相邻数字之间的位置关系也复杂多样。本系统基于最小风险的贝叶斯（bayes）匹配决策进行设计，将本系统的主要应用方向标定为脱机手写数字样本的自动录入及聚类，主要工作如下：
 1、预处理方面，实现了基本的图像平滑，并针对不同纸张背景制定了区别对待的图像二值化策略：对以空白纸张为背景的数字图像采用迭代最佳分割阈值算法，以稿纸为背景的数字图像采用双重阈值法。
 2、回顾和总结了历年手写数字的主要细化方法，在结合本系统主要适用于数字录入这一用途的基础上，提出了改进细化算法。
 3、介绍了几种主要的统计特征和笔划结构特征提取方法。针对本次设计采用了像素分割提取的方式。
 4、在识别阶段，本系统基于bayes理论采用了多种识别算法进行图形聚类。
 系统中测试样本共包含8张数字记录扫描扩大图片，其中有标准印刷体数字杂陈类型，同种类异尺度印刷体类型，不规则手写体类型，图形元件类型等。算法聚类正确率视具体样本类型而定。
 
关键词：手写数字； 聚类； 贝叶斯； 特征提取； 模糊集； 聚类中心

 Design of handwritten numeral matching and recognition system base on Bayes theory
Abstract： The off-line recognition of handwritten numeral is one of the challenging items in the field of pattern recognition.It will paly a important role in many respects such as automatic inputting for letter classification,check recognition,statistical returns and handwritten ID.However,as a result of the lack of strict supervison and the complex relation of position between numbers.The present system is designed base on Bayesian theory and is mainly used to automatic inputting and automatic clustering for off-line handwritten numeral sample.The main work is as follows:
 1,Pretreatment of handwritten numeral sample,having achieved basical smoothness of picture and generating different strategy for binaryzation based on different papery substrate,such as using iterating the best threshold for dividing against the background of plain paper ,using two-tier threshold for dividing against the background of draft paper.
 2,This paper reviewed and summarize the main ways over the years in which thining numeral,On the basis of the present system can realize entering for numeral function,one modified arithmetic used to thining sample is proposed.
 3,This paper introduced some primary ways for refining feature and statistical nature.In connection with the present design,I adopt the method of dividing picture pixels.
 4,In the period of identifying,author use a few kinds of arithmetics to achieve image clustering base on Bayesian theory.
 The set of all test samples for this present system is eight piece of pictures enlarged recording information about numeral.the set includes set of standard print digits and the same category but different size numeral and handwritten numeral and set of geometrical shapes etc.The accuracy rate of all clustering algorithms is depend on the certain sample.
Keywords: handwritten digits, clustering, bayes, feature extraction, fuzzy set, center of clustering