基于压缩子结构特征的手写体数字识别

在对字符结构进行分析的基础上 ,提出了一种用于自由手写体数字识别的子结构特征 .由于绝对位置、笔画长度等特征因人而异 ,文中利用字符的拓扑信息来增强特征的稳定性 ,并将字符模式表达为一个矩阵 ,矩阵的每一列即为字符的一个子结构特征矢量 .由于子结构特征表达的模式可分性强 ,可通过矩阵运算对模式进行特征压缩 ,同时将不同模式等维化 ,利用一变结构神经网络构造分类器 ,避免了传统子结构特征规则匹配的缺点 ,提高了模式匹配速度 .利用信函分拣机提供的数字进行测试 ,识别率可达 97.58%.

Recognition of Handwritten Numerals Based on Compression of Sub-Structural Features

A setof sub- structural features was proposed for handwritten numeral recognition by analysing numeral structure.Some features which are changeable for different person were substituted for others who only rely on relative location among the strokes.The features in this paperare less changeable in con- trast with others by stressing on the numeral topology information.To avoid matching the troublesome rule,every numeral was described as a matrix which every row stands for a sub- structural feature vector. As the pattern described by sub- structural feature is of strong classification,the feature can be compressed by the matrix operation.Consequently,there exists the same dimension in all patterns.This allows an easy pattern matching being conducted by using a structure- changeable neural network.The experiments show thatthe correct recognition rate is about97.58%.