融合子区域局部二值模式特征与深层聚合网络的人脸识别  

Face Recognition Based on Fusion of Sub-region Local Binary Pattern Features and Deep Aggregation Network

作　　者：傅桂霞[1] 魏文辉 邹国锋[1] 尹丽菊[1] 高明亮[1] FU Gui-xia;WEI Wen-hui;ZOU Guo-feng;YIN Li-ju;GAO Ming-liang(College of Electrical and Electronic Engineering,Shandong University of Technology,Zibo 255049,China)

机构地区：[1]山东理工大学电气与电子工程学院,淄博255049

出　　处：《科学技术与工程》2018年第19期237-243,共7页Science Technology and Engineering

基　　金：山东省自然科学基金(ZR2016FL14);国家自然科学基金(61601266);中国博士后科学基金(2017M612306);淄博市校城融合发展计划(2016ZBXC142)资助

摘　　要：针对深度网络对人脸噪声敏感,且学习过程容易忽视人脸结构信息的问题,提出融合子区域局部二值模式(local binary pattern,LBP)特征和深层聚合网络的人脸识别算法。将人脸图像划分为不同子区域,并采用局部二值模式对人脸进行预处理,获取子区域人脸的LBP特征。不同子区域LBP特征输入不同的稀疏自动编码器,实现深层特征提取;然后不同稀疏自动编码器的输出特征通过全连接方式实现特征聚合,获得人脸特征向量用于分类。通过大量实验获取了最优的聚合网络模型架构和网络参数取值,改善了人脸识别效果。Aiming at the problem that the depth network is sensitive to human face noise and the learning process is easy to ignore the information of face structure,a face recognition algorithm was proposed based on LBP features and deep aggregation network.The face image is divided into different sub-regions,and the local binary image is used to pre-process the face image in order to obtain the LBP features of the sub-region face.LBP features of different sub-regions are input to different sparse auto encoders to achieve deep feature extraction.Then,the output features of different sparse auto encoders are fully aggregated to obtain the feature vectors for classification.Through a large number of experiments,the optimal aggregation network model structure and network parameter values were got,and improved the face recognition effect.

关 键 词：局部二值模式特征 稀疏自动编码器 子区域划分 深层聚合网络 人脸识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]