基于并行组合分类器的脱机手写体数字识别  

Off-line Handwritten Digit Recognition Based on Parallel Combined Classifiers

作　　者：楚浩宇 高萌 刘永生 

机构地区：[1]东北农业大学电气与信息学院,黑龙江哈尔滨150030

出　　处：《计算机技术与发展》2018年第3期105-108,共4页Computer Technology and Development

基　　金：国家"863"高技术发展计划项目(2013AA10230304)

摘　　要：为了提高脱机手写体数字识别的识别率和可靠性,并且考虑到传统的单一分类器对数字之间差异的敏感性不同,综合K-近邻算法、广义回归神经网络、支持向量机三种机器学习算法,提出了一种并行组织结构的组合分类器。并行组合分类器通过改进的投票机制来判定识别结果。以MNIST数据库为数据来源,在MATLAB平台上开展各种分类器的性能对比实验。组合后的识别率、拒识率、误识率、可靠性分别可达到97.48%、1.55%、0.97%、99.02%。实验结果表明,并行组合分类器在鲁棒性方面优于传统的单一分类器,在识别率、拒识率、算法的时间复杂度上均优于其他组合分类器。并行组合分类器以简易结构实现了脱机手写体数字的快速、高效识别。In order to improve the recognition rate and reliability of off-line handwritten digit recognition,considering that traditional single classifiers have different sensitivity to the differences between digital,we propose a combined classifier of parallel organizational structure combining three machine learning algorithms of K-nearest neighbor,general regression neural network and support vector machine.It uses the improved voting mechanism to determine the recognition result.Using MNIST database as data source,the comparable experiment on the performance of classifiers is carried out on MATLAB,whose results(recognition rate,rejection rate,false accept rate,reliability)are 97.48%,1.55%,0.97% and 99.02%.The experiments indicate that the parallel combined classifier is superior to the traditional single classifier in terms of robustness,and other combined classifiers in terms of recognition rate,rejection rate and time complexity.With a simple structure,it can achieve fast and efficient off-line handwritten digit recognition.

关 键 词：模式识别 组合分类器 LR 广义回归神经网络 支持向量机 手写体数字 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]