Development and Evaluation of Robust Statistical Method for Adaptive Handwritten Character Recognition System

自适应手写字符识别系统鲁棒统计方法的开发与评估

• 批准号: 11680324
• 负责人: HARA Takahiko
• 金额: $ 0.19万
• 依托单位: Oita University
• 依托单位国家: 日本
• 项目类别: Grant-in-Aid for Scientific Research (C)
• 财政年份: 1999
• 资助国家: 日本
• 起止时间: 1999 至 2000
• 项目状态: 已结题
• 来源: https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-11680324/
• 关键词: Canonical Discriminant Analysis Tree; CDAT; Decision Tree; Canonical Discriminant Analysis; 筆者適応型手書き文字認識; 頑健性

As non-parametric methods of classifying based on decision tree, CART and C4. 5 are typical. In this research, an import of such a non-parametric approach based on decision tree into canonical discriminant analysis is discussed. We call the decision tree based approach which has been imported into canonical discriminant analysis CDAT (Canonical Discriminant Analysis Tree). In CDAT, the index of improvement by dividing a region into two is defined as the difference between before-division sum of eigenvalues for between-groups variance-covariance matrix against within-groups variance-covariance matrix and after-division one. By using programming language C on work station, a routine of CDAT was implemented. For the sake of variety of handwritings, 3 writers prepared learning and testing data sets which consisted of 5 kinds of handwritten characters selected out of 109 cursive kana letters respectively, by using X-window system and programming language C on work staton. Then, CDAT was compared with canonical discriminant analysis with the above data sets, by using cross-validation. As a result, in the case that there was a variety of handwritings between writers, CDAT surpassed canonical discriminant analysis, where a threshold was needed so that subregion divided by decision tree might not become too small. In nonlinear structure of data sets, CDAT is effective to pile strata of decision trees recursively.

作为基于决策树的非参数分类方法，有CART和C4。 5个是典型的。在本研究中，讨论了这种基于决策树的非参数方法在典型判别分析中的重要性。我们将引入规范判别分析的基于决策树的方法称为 CDAT（规范判别分析树）。在CDAT中，将区域一分为二的改进指数定义为组间方差-协方差矩阵的特征值之和在划分前与组内方差-协方差矩阵的特征值之和与划分后的差值。在工作站上采用C语言编程，实现了CDAT例程。为了实现笔迹的多样性，3位作者在工作站上使用X-window系统和编程语言C编写了分别从109个草书假名字母中选出的5种手写文字组成的学习和测试数据集。然后，通过交叉验证，将 CDAT 与上述数据集的典型判别分析进行比较。结果，在作者之间存在多种笔迹的情况下，CDAT超越了规范判别分析，其中需要一个阈值，以便决策树划分的子区域不会变得太小。在数据集的非线性结构中，CDAT对于递归地堆积决策树层是有效的。