一种结合结构和统计特征的脱机数字识别方法

2016-12-21张玉叶姜彬李开端王春歆

微型电脑应用 2016年8期

张玉叶，姜彬，李开端，王春歆

张玉叶，姜彬，李开端，王春歆

脱机手写数字识别归根结底是数字的图像特征匹配识别问题。为了提高识别效率，需要降低数字的特征维数；同时要提高数字识别的准确性，必须考虑手写数字的笔画结构不稳定的特点。提出了一种结合字符统计特征和结构特征的识别方法。首先，利用主分量分析法抽取数字字符图像的统计特征，通过对主分量重建模型的误差分析进行数字识别；为了进一步提高数字识别的准确度，再加入数字的宽高比结构特征进行比对识别。最后自制训练样本及测试样本库进行识别实验，实验结果表明本方法识别准确率较高。

脱机手写数字识别；主分量分析；结构特征；统计特征

0 引言

脱机数字识别不能利用联机识别可以得到的时间、笔顺等动态信息，系统实现比较困难[1]。手写数字识别最重要的环节是数字字符的特征提取[2]。目前，手写数字的特征可分为两类：统计特征和结构特征。统计特征是利用字符样本库，找出的0到9中每类字符空间分布的统计规律。结构特征包括数字的构造如端点、交叉点、轮廓等。两类特征各有优势，统计特征可以描述数字的本质特征，适用于给定训练集差别不大的情况；利用结构特征能够精确描述数字的细节特征，对书写较规范的数字有较高的识别率。可以将两类特征结合运用，以便更好地进行数字识别。

本文提出了一种能够结合字符统计特征和结构特征的识别方法，利用主分量分析法抽取数字字符样本的统计特征，通过对主分量重建模型的误差分析进行字符识别；为了进一步提高字符识别的准确度，加入宽高比结构特征进行字符比对识别。

1 脱机手写数字识别预处理

脱机手写数字识别，首要任务就是将纸质载体扫描为图像信息以便计算机处理。得到的数字图像在进行识别之前需要进行图像的预处理，以消除图像中无关的信息，从而改进特征提取、图像分割、匹配和识别的可靠

An Off-line Handwritten Numeral Recognition Method Combined With the Statistical Characteristics and Structural Features

Zhang Yuye1, Jiang Bin2,Li Kaiduan1,Wang Chunxin3
(1.Naval Aeronautical and Astronantical Universing, Qingdao 266041,China; 2.Qingdao University,Qingdao 266071,China; 3. North China Sea Fleet,Qingdao 266041,China))

Off-line handwritten numeral recognition is a pattern recognition problem of the images of ten numbers. In order to improve the recognition efficiency， the character dimension of number’s image should be decreased. As well, in order to improve the recognition veracity, the character mode instability which resulted from different writing styles and habits should be considered. The article proposed a numbers recognition method combined with the statistical characteristics and structural features of numbers. Firstly, the principal component analysis (PCA) method is adopted to extract statistical characteristics of numeral image. The numeral recognition will be realized through analysis of the reconstruction error of model which is reconstructed by the principal components. In order to further determine the type of numeral, the structural features of width and height rate should be added. Finally, through experiments on the identification of numeral image, the reliability and accuracy of this method of digital recognition is verified, and the deficiency of this method in real-time recognition is analyzed.

Offline Handwritten Numeral Recognition; Principal Component Analysis; Structural Features; Statistical Characteristics