一种通过卷积─池化提升SVM人脸识别率的研究  

Research on Improving SVM Face Recognition Rate by Convolution–Pooling

作　　者：叶晓波[1] 秦海菲 吕永林 YE Xiaobo;QIN Haifei;LV Yonglin(Institute of Network&Information Systems,Chuxiong Normal University,Chuxiong,Yunnan Province 675000;School of Information Sciences&Technology,Chuxiong Normal University,Chuxiong,Yunnan Province 675000;School of Economics&Management,Chuxiong Normal University,Chuxiong,Yunnan Province 675000)

机构地区：[1]楚雄师范学院网络与信息系统研究所,云南楚雄675000 [2]楚雄师范学院信息科学与技术学院,云南楚雄675000 [3]楚雄师范学院经济与管理学院,云南楚雄675000

出　　处：《楚雄师范学院学报》2020年第3期106-114,共9页Journal of Chuxiong Normal University

摘　　要：以剑桥大学计算机实验室的ORL Faces数据库作为实验数据,通过卷积神经网络中的"卷积-池化"层对实验数据进行处理,选择LIBSVM集成软件为工具,对原始数据和经"卷积-池化"处理后的数据进行了分类识别研究,SVM参数选用C-SVC模型、nu-SVC模型与线性核函数、多项式核函数、径向基核函数、Sigmoid核函数进行组合。实验结论:增加训练数据可提高人脸识别率,卷积-池化处理可实现数据降维,"高斯平滑卷积核卷积-池化"处理可提高SVM人脸识别率,SVM在人脸识别中更适合选用C-SVC模型+线性核函数。Taking the ORL Faces database of the computer laboratory of Cambridge University as the experimental data,the experimental data are processed through the"convolution-pooling"layer in the convolution neural network,and the LIBSVM integration software is selected as the tool to classify and identify the original data and the data processed by"convolution-pooling".SVM parameters are combined with C-SVC model,nu-SVC model and linear kernel function,polynomial kernel function,radial basis kernel function and Sigmoid kernel function.The convolution-pooling processing can achieve data dimensionality reduction and"gaussian smooth convolution kernel convolution-pooling"processing can improve SVM face recognition rate.SVM is more suitable to choose C-SVC model linear kernel function in face recognition.

关 键 词：人脸识别 卷积 池化 SVM(支持向量机) 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]