基于SVM和BP神经网络的手写数字识别.pdf

I

摘 要

科技发展日新月异，智能识别推陈出新。如今是信息化时期，数字识别在很多智能领

域上运用广泛，拥有普遍的使用远景，因此探索这项技术有其重要的实际意义。由于手写

数字在写法上千差万别，且数字间字形差别相对较小，使得识别系统的开发具有很大的挑

战性。

当前手写数字识别采用的技术有 Bayes判别法、决策树法、神经网络和支持向量机

（Support Vector Machines ，SVM）等。诞生于 20世纪 90年代的 SVM技术是机器学习研究

的热点，因其良好的泛化性能成为了数字识别领域的热门方法。

II

Technological development changes rapidly, and intelligent recognition innovates

constantly. In the information era, numeral recognition has broad application prospects in many

pattern areas with a common vision, so it is of great practical significance to explore this

technology. As a result of handwritten numeral vary widely in the wording and digital shape

difference is relatively small, making recognition system development is a great challenge.

At present, Bayes discriminant analysis, decision tree method, neural network and support

vector machine (referred to as SVM) are the main methods of recognition of handwriting. SVM

technology, which was born in 1990s, is a hot topic in machine learning research. Because of its

good generalization performance, it has become a popular method in the field of numeral

recognition.

Keywordshandwritten numeral recognition; neural network; SVM

III

目录

摘要

...................................................................................................................................................................... I

Abstract

............................................................................................................................................................ I

第 1 章绪论

....................................................................................................................................................... 1

第 2 章手写数字识别综述

2.1.1 图像二值化

....................................................................................... 错误！未定义书签。

2.1.2 图像去噪锐化

................................................................................... 错误！未定义书签。

2.1.3 图像分割细化

................................................................................... 错误！未定义书签。

2.3.1 决策树法

.............................................................................................................................. 4

2.3.2 贝叶斯判别法

...................................................................................................................... 5

2.3.3 神经网络

.............................................................................................................................. 7

2.3.4 支持向量机

......................................................................................................................... 9

3.2.1 二分类支持向量机

.......................................................................................................... 12

................................................................................................................................ 13

3.4.1 线性分划

............................................................................................................................ 13

3.4.2 非线性分划

....................................................................................................................... 15

3.6.2 遗传算法

............................................................................................................................ 18

IV

第 4 章 BP 神经网络

4.2.1 BP 神经网络算法流程

4.2.2 数据归一化

....................................................................................................................... 20

4.2.3 BP 神经网络训练 ............................................................................................................. 21

4.2.4 BP 神经网络参数 ............................................................................................................. 22

4.2.5 BP 神经网络函数 ............................................................................................................. 23

................................................................................................... 32

5.2.1 设计目标

............................................................................................................................ 32

5.2.2 样本图片选择

................................................................................................................... 33

5.2.3 设计与实现

....................................................................................................................... 34

5.2.4 结果分析

............................................................................................................................ 39

................................................................................................................... 39

5.3.4 结果分析

............................................................................................................................ 42

............................................................................................................................... 43

结论

................................................................................................................................................................... 44

1

第 1 章 绪论

数字识别主要有印刷体与手写体识别，由于存在人体差异和个人习惯的不同，即使是

同一个数字，每个人书写的情况都会千差万别。手写数字识别技术在不同的数字数据库实

验中已得到了较为满意的识别辨认效果，然而将其使用到手写文档中，其识别的精准度仍

难以达到实际要求。

手写数字是借助电子仪器去自动辨别人手写出来的数字。尽管由于笔画、字型等因素

还未能在计算机上做到全部都正确识别，但其在各方面应用早已十分广泛，多存在于办公

和教学自动化、银行票据自动识别、邮政自动分拣等技术领域当中。

在日常生活和工作中，每天都会接触大量的数字信息，尤其是处理各种各样的数字文