CASIA OpenIR  > 毕业生  > 硕士学位论文
一个集成的无约束手写数字串识别系统
其他题名An Integrated System For Unconstrained Handwritten Numeral String Recognition
李世峰
学位类型工学硕士
导师刘昌平
2003-06-01
学位授予单位中国科学院研究生院
学位授予地点中国科学院自动化研究所
学位专业模式识别与智能系统
关键词手写数字识别 数字串分割 机器学习 聚类 分类器集成 Handwritten Numeral Recognition Machine Learning Clustering Multiple Classifier Fusion
摘要手写数字识别的目的是使数据信息能够自然、方便地输入计算机,以便于 进一步的处理。手写数字识别有着极为广泛的应用前景,在大规模数据统计、 财务、税务、金融领域以及邮件分拣中都有着成功的应用。同时手写数字识别 作为一个有难度的开放问题,可以为理论的深入分析、验证及评估提供具体的 实验平台,因而吸引了世界各地的研究工作者参与。 对于无约束的手写数字串识别来说,难点在于怎样使计算机具有鲁棒的识 别能力,不因为噪声和不同的个人书写方式而导致错误的结果。这就为学习理 论提出了新的挑战。 本文先就手写数字识别的相关研究发展作了综述,然后的一章介绍了机器 学习,包括非监督学习和监督学习,在数字识别中的应用。在非监督学习小节 中,除了比较各种成熟的聚类方法,还提出了自己的交迭聚类算法和基于RBF 核的升维聚类算法。在监督学习章节中,着重描述了将在实际工作中利用和展 开的各种方法,如贝叶斯理论、EM算法、学习矢量量化、支持向量机等。接下 来的一章着重讲述数字串识别领域的主要研究方向。最后一章描述了一个己实 现的具体的手写数字串识别系统。在该系统中,数字串图像首先被分割为单个 数字字符,然后经过预处理包括平滑去噪、倾斜校正、归一化,再使用多种方 法进行特征提取以便于多特征的分类器集成。在单分类器训练阶段,结合了无 监督聚类和有监督学习的方法,保证学习训练过程的快速和有效。在有监督学 习阶段,高斯混合模型和概率的学习矢量量化被统一在学习算法中并得到了不 错的效果。在多分类器集成阶段,分析了该集成框架对于识别率和识别速度的 影响,并引入决策轮廓矩阵以充分利用各分类器的识别结果信息。我们针对已 有的基于决策模板的算法提出了改进算法。最后给出了我们的集成方法的识别 效果,并和部分的其它集成方法作了比较。可以看出,我们的集成方法取得了 较高的识别效果,在中科院自动化所文字识别工程中心的大规模样本集上达到 99.22%的识别率。
其他摘要The purpose of handwritten numeral recognition is to input data into computers for further disposal naturally and conveniently. Recognition of handwritten numerals has been successfully used in fields such as large-scale statistics, finance, revenue and mails sorting. As an open and difficult challenge, the recognition problem of handwritten numerals also build up a testing platform for in-depth analysis, verification and evaluation of different kinds of theories. Researchers all over the world are appealed to take part in the work. For unconstrained handwritten numeral string recognition problem, the difficulty lies in how to endue the computer with robust recognition ability and erroneous outcome would not appear as a result of noises and different writing styles by person. Thus new challenges are brought forward to Machine Learning theories. In the thesis, the related researches in recognition of handwritten numeral string are overviewed for the first. The following chapter introduces the application of Machine Learning in numerals recognition, including unsupervised learning methods and supervised learning methods. In unsupervised learning chapter, we advance our overlapping clustering algorithm and RBF kernel based clustering algorithm aside from comparison of several mature clustering methods. In supervised learning chapter, we focus on introduction of methods that will be utilized and expanded in our application, such as Bayesian theory, EM algorithm, Learning Vector quantification and Support Vector Machine. The next chapter concentrates on main directions in the research field of numeral string recognition. The last chapter describes our concrete handwritten numeral string recognition system. In the system, images of numeral strings are segmented into separate digits for the first, then the separated numeral images are fed into pre-processing step, including smoothing, noise removal, slant correction and normalization. Then, several kinds of feature extraction methods are applied in facility of future fusion of classifiers. In training phase of single classifier, unsupervised clustering and supervised learning methods are integrated to ensure the rapidity and efficiency of training. In supervised learning phase, Gaussian Mixture Model and probabilistic Learning Vector Quantification are combined in our algorithm to achieve better results. In fusion phase of multiply classifiers, the influence of our framework in recognition accuracy and speed is analyzed, and the concept of Decision Profile Matrix is introduced to utilize the resulting recognition information of each classifier fully, also we advance an improved Decision Template based fusion algorithm. At last, the result of our system is presented to compare with other integration methods. It can be seen that our fusion strategy achieves better result, 99.22% accuracy rate on the large pattern set of Character Recognition Engineering Center in Institu
馆藏号XWLW680
其他标识符680
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/6787
专题毕业生_硕士学位论文
推荐引用方式
GB/T 7714
李世峰. 一个集成的无约束手写数字串识别系统[D]. 中国科学院自动化研究所. 中国科学院研究生院,2003.
条目包含的文件
条目无相关文件。
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[李世峰]的文章
百度学术
百度学术中相似的文章
[李世峰]的文章
必应学术
必应学术中相似的文章
[李世峰]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。