文档介绍：华中科技大学
硕士学位论文
基于统计法与结构法的手写数字识别研究
姓名:张勇
申请学位级别:硕士
专业:计算机系统结构
指导教师:鲁宏伟
2011-01-17
华中科技大学硕士学位论文
摘要
脱机手写数字识别是一个至今仍没有完美解决的问题,是一个很有挑战的问题。
它有着广泛而重要的应用需求,如识别邮政编码、银行收据、税收票据、各种打印
手写的数字等等,这吸引了众多研究者持续不断的为之努力。
介绍了手写数字识别的研究现状与未来发展的趋势。讨论了识别所需的预处理
工作如图像二值化、图像分割、归一化与字符细化。针对目前特征的组合对提高识
别率效果明显,而多种分类器组合不仅识别速度慢且效果甚微的情况,研究了一种
将梯度特征、曲率特征、小波变换特征组合,直接使用分类性能优秀的支持向量机
做为分类器的统计法。详细说明了特征提取的方法,论述了支持向量机与使用其实
现工具进行识别的方法。
研究了一种简便实用的结构法:将数字图像按特征点分解,去除细小分支后再
重组,对重组后的线段进行简化的基元分类,最后根据字符所含基元与特征点来分
类识别。该方法避免了传统方法中的学习过程,减少了识别的时间。为提高识别可
靠性,尝试性对同一批样本分别用两种方法识别,再对两种识别结果进行仲裁得出
最后的识别结果。
实验表明,研究的统计法与结构法简便可行,错误率均在可接受的范围内,结
合识别虽然拒识率比较高,但可靠性有所提高。
关键词:手写数字识别支持向量机特征组合骨架分解树形分类器
I
华中科技大学硕士学位论文
Abstract
Off-line handwritten digit recognition is a problem not solved perfect so far, so it is
very challenging. There are so many potential applications such as bank check processing,
postal mail sorting, automatic reading of tax forms and various handwritten and printed
materials that a lot of researcher persistent to research it.
Previous researches and tendency of handwritten digit recognition are introduced.
Pre-processing such as image binaryzation, image segmentation, normalization and
thinning are discussed. bination do a good contribution to recognition rate
but bination not. I research a statistical method use Support vector classifier
which has excellent classified ability to bined features of gradient and
curvature and wavelet features. The method of extract features and using tool of support
vector classifier are illustrate in detail.
A structural method is researched. More specifically, based on three kinds of feature
points, first we extract ten kinds of primitive segments for each image. And then, a
tree-like classifier based on the extracted feature points and primitives is applied to
classify the numerals. This method needn’t a study procedure and run fast. Finally, to get
a good reliability rate, using the two methods to r