基于加权均值人脸的多姿态人脸识别  

Multi-pose face recognition based on weighted mean face

作　　者：邹国锋[1] 傅桂霞[1] 申晋[1] 高明亮[1] 王科俊[2] Zou Guofeng;Fu Guixia;Shen Jin;Gao Mingliang;Wang Kejun(College of Electrical&Electronic Engineering,Shandong University of Technology,Zibo Shandong 255049,China;College of Automation,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]山东理工大学电气与电子工程学院,山东淄博255049 [2]哈尔滨工程大学自动化学院,哈尔滨150001

出　　处：《计算机应用研究》2018年第10期3164-3168,共5页Application Research of Computers

基　　金：山东省自然科研基金资助项目(ZR2016FL14;ZR2015FL029);国家自然科学基金资助项目(61601266);中国博士后科学基金资助项目(2017M612306)

摘　　要：针对姿态变化人脸识别的困难,提出一种基于加权均值人脸的识别思路。根据人脸姿态左右摇摆角度变化,定义每幅姿态变化人脸的权值计算方法,提出加权均值人脸的构建策略;基于姿态变化人脸的俯仰角度,将姿态变化人脸划分为俯视、平视和仰视三个层次,并在每个层次中构建加权均值人脸,形成加权均值人脸矩阵;最后,采用改进的局部保持投影算法对加权均值人脸矩阵进行深层特征提取,实现多姿态人脸识别。实验结果表明,所提方法能有效提取姿态变化人脸的关键信息,使识别效果得到较大改善。This paper proposed the concept of weighted mean face to deal with the problem of pose varied face recognition.Firstly,according to face vacillating change,this paper defined the calculation method of the weight of each pose varied face,and then proposed the construction method of weighted mean face.Then,based on the pitch angles of pose varied faces,it divided the pose varied faces into three levels:upward view,parallel view and downward view.In each view level,constructing weighted mean face,thus it formed a weighted mean face matrix.Finally,this paper used the improved locality preserving projection method to extract deep feature of weighted mean face matrix and realized pose varied face recognition.The experiments show that the proposed approach can effectively extract the key information of pose varied faces and improve recognition effect greatly.

关 键 词：人脸姿态变化 加权均值人脸 加权均值人脸矩阵 局部保持投影 深层特征提取 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]