基于卷积神经网络的人脸识别算法  

Research of face recognition algorithm based on CNN

作　　者：宋强 张颖[1] SONG Qiang;ZHANG Ying(School of Electronic and Information Engineering,University of Science and Technology Liaoning,Anshan 114051,China)

机构地区：[1]辽宁科技大学电子与信息工程学院,辽宁鞍山114002

出　　处：《辽宁科技大学学报》2020年第5期363-367,376,共6页Journal of University of Science and Technology Liaoning

摘　　要：人脸识别作为人工智能信息化的分支,可以通过卷积神经网络自主学习,快速实现人脸识别,但准确率及训练效率上仍存在不足。本文在传统模型的基础上改进模型结构,采用双对称LeNet并行连接的网络结构,运用DCT-LBP联合处理的方法对输入图像进行预处理,在输出层Softmax回归分类中的特征和权重向量间加上一个类内余弦矫正,增强了泛化能力,提高了识别效率。Face recognition,as a branch of artificial intelligence informatization,can learn by convolution neural network and realize face recognition quickly,but there are still some shortcomings in accuracy and training efficiency.In this paper,the traditional model structure is improved by adopting the network structure with double symmetric LeNet parallel connection.The input image is preprocessed by DCT-LBP joint processing method.An intra-class cosine correction is added between the features and weight vectors in Softmax regression classification of the output layer,which enhances the generalization ability and improves the recognition efficiency.

关 键 词：卷积神经网络 双对称LeNet 人脸识别 Softmax回归分类 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]