一种新的手写体字符识别算法  

A new algorithm on handwriting character recognition

作　　者：孙权森[1] 程显毅[1] 张长温[2] 夏德深[1] 

机构地区：[1]南京理工大学计算机系,江苏南京210094 [2]济南大学理学院,山东济南250022

出　　处：《江苏大学学报（自然科学版）》2005年第6期517-520,共4页Journal of Jiangsu University：Natural Science Edition

基　　金：国家自然科学基金资助项目(60473039)

摘　　要：研究模式识别的核心问题——特征抽取.基于偏最小二乘(PartialLeastSquares,简称PLS)回归和特征融合的思想,提出了一种组合特征抽取的新方法并将之用于手写体字符识别中.在PLS建模阶段,为了提高PLS成分(特征)的抽取速度,提出了一种非迭代PLS算法.在特征融合阶段,用所抽取的PLS成分特征组成模式的相关特征矩阵,并依此相关特征矩阵进行分类.在ConcordiaUniversityCENPARMI手写体阿拉伯数字数据库上的试验结果证实了该方法的有效性和鲁棒性,其分类结果优于基于单一特征的FSLDA方法的分类结果.另外,与已有的迭代PLS算法相比,所提出的非迭代PLS算法的复杂度和特征抽取的速度均占有优势.The feature extraction is the core problem in pattern recognition. Based on the ideas of Partial Least Squares （PLS） model and feature fusion, a new method of combined feature extraction is proposed. In PLS modeling, in order to enhance the extracted speed of the PLS component（ feature vectors）, a noniterativePLS （NIPLS） algorithm is proposed. In feature fusion, two sets of PLS components are extracted and the correlative feature matrix of the same pattern sample is made for classification. Experimental results on the Concordia University CENPARMI database of handwritten Arabic numerals show that classification results of the proposed method is better than that of FSLDA method adopting the single feature, and this algorithm is efficient and robust. The proposed NIPLS algorithm superior to other iterative PLS （IPLS） algorithm in computational cost： and speed of feature extracted.

关 键 词：手写体字符识别 偏最小二乘回归 特征抽取 特征融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]