基于自学习深度卷积神经网络的姿态变化人脸识别  

Pose Varied Face Recognition Based on Self Learning Deep Convolutional Neural Network

作　　者：邹国锋[1] 傅桂霞[1] 高明亮[1] 尹丽菊[1] 王科俊[2] ZOU Guo-feng;FU Gui-xia;GAO Ming-liang;YIN Li-ju;WANG Ke-jun(College of Electrical ＆ Electronic Engineering, Shandong University of Technology, Zibo 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：《小型微型计算机系统》2018年第6期1156-1162,共7页Journal of Chinese Computer Systems

基　　金：山东省自然科研基金联合项目(ZR2015FL029;ZR2016FL14;ZR2015FL034)资助;国家自然科学基金青年基金项目(61601266)资助

摘　　要：针对卷积神经网络结构设计依赖人为经验,网络深度、特征图个数设置缺乏理论依据,网络训练需大量训练样本支持,并结合姿态变化人脸识别存在的问题,提出姿态变化人脸底层特征图的样本扩充方法和深度卷积神经网络模型的自学习方法.首先,根据姿态人脸分布规律,将姿态人脸非线性流形空间划分为不同流形层和局部子空间,针对局部子空间内姿态人脸定义人脸底层特征构建方法,实现姿态变化人脸样本扩充.然后,通过网络结构初始化、网络结构全局和局部自适应扩展,获得自学习深度卷积神经网络,实现姿态变化人脸的深层非线性特征提取和识别.实验表明,本文所提方法丰富了卷积神经网络的理论研究,有效改善了姿态变化人脸识别的准确率.The construction of Convolutional Neural Network depends on human experience,the setting of network depth and number of feature maps lack theoretical basis,and the network training needs a large number of training samples. In this paper,combining with the difficulties of pose-varied face recognition,we propose a method of sample expansion based on pose-varied face low-level feature map and a novel self-learning method for deep Convolutional Neural Network model. First of all,according to the distribution of posevaried faces,the pose-varied face manifold space is divided into different manifold layer and local subspace. For pose-varied faces in the subspace,the construction method of face low-level feature is defined and the expansion of training samples are realizes. Then,the self-learning deep Convolutional Neural Network is obtained through the network initialization,the global and local adaptive extension of network structure,this neural network realizes the deep nonlinear feature extraction and recognition of pose-varied faces. The experiment results show that the proposed method enriches the theoretical study of Convolutional Neural Network,and improves the accuracy of pose-varied face recognition.

关 键 词：卷积神经网络 自学习 深度学习 人脸姿态变化 人脸底层特征图 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]