基于LDA降维和BP神经网络的手写数字识别  

Recognition of Handwritten Digit Data Based on LDA Dimensionality Reduction and BP Neural Network

作　　者：刘佳悦 LIU Jiayue(School of Mathematical Sciences,South China Normal University,Guangzhou Guangdong 510631,China)

机构地区：[1]华南师范大学数学科学学院,广东广州510631

出　　处：《信息与电脑》2023年第14期187-189,193,共4页Information & Computer

摘　　要：手写数字数据集是机器学习分类领域的优质数据集,文章以反向传播(Back Propagation,BP)神经网络为基础,对手写数字进行分类识别。为减少BP神经网络的计算开支,实验前,对比了过滤卡方检验法、主成分分析(Principal Component Analysis,PCA)降维、线性判别式分析(Linear Discriminant Analysis,LDA)降维以及多维尺度变换(Multidimensional Scaling,MDS)降维对特征选取的训练测试效果,从而确定了神经网络拟合之前的最优特征提取方法。实验中,利用Bagging对BP神经网络进行集成处理,分类识别了手写数字。实验后,将文中方法与朴素贝叶斯、决策树、随机森林、LDA多分类进行对比。结果表明,采取LDA降维方法时,降到9维的特征提取方式最优,单个BP神经网络对手写数字数据识别的准确率为92%左右,而基于Bagging集成的BP神经网络在识别准确率方面高达95%。Handwritten digital data set is a high-quality data set in the field of machine learning classification.Based on Back Propagation(BP)neural network,this paper classifies and recognizes handwritten digital data.In order to reduce the calculation cost of BP neural network,before the experiment,the training and testing effects of filtering chi-square test,Principal Component Analysis(PCA)dimensionality reduction,Linear Discriminant Analysis(LDA)dimensionality reduction and Multidimensional Scaling(MDS)dimensionality reduction on feature selection were compared,so as to determine the optimal feature extraction method before neural network fitting.In the experiment,the BP neural network is integrated by Bagging,and the handwritten digital data is classified and recognized.After the experiment,the proposed method is compared with naive Bayes,decision tree,random forest and LDA multi-classification.The results show that when LDA dimension reduction method is adopted,the feature extraction method reduced to 9 dimensions is the best,and the recognition accuracy of handwritten digital data by single BP neural network is about 92%,while the recognition accuracy of BP neural network based on Bagging integration is as high as 95%.

关 键 词：线性判别式分析(LDA) 反向传播(BP) 数据识别 手写数字 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]