基于SURF特征的栈式自编码网络人脸对齐算法  

Stacked auto-encoder networks face alignment algorithm based on SURF features

作　　者：崔凯 才华[1] 刘广文[1] 刘智[1] CUI Kai;CAI Hua;LIU Guang-wen;LIU Zhi(School of Electronic Information Engineering,Changchun University of Science and Technology,Changchun 130022,China)

机构地区：[1]长春理工大学电子信息工程学院,吉林长春130022

出　　处：《液晶与显示》2018年第3期254-260,共7页Chinese Journal of Liquid Crystals and Displays

基　　金：国家自然科学基金(No.61172111);吉林省科技发展计划(No.20160101260JC);吉林省教育厅"十三五"科学技术项目(No.JJKH20170625KJ)~~

摘　　要：人脸对齐是人脸识别系统中的一个核心部分,定位的准确性和定位速度直接影响到人脸识别的效果。人脸图像存在不同姿态、不同表情、不同光照条件等因素的影响,真实场景下的人脸对齐成为一个难题。本文提出了一种基于SURF特征的栈式自编码网络人脸对齐方法,首先通过粗糙定位网络找到近似人脸特征点,并提取局部的SURF特征,输入到局部细化网络,通过级联结构,进一步精确化人脸特征点的具体位置。最后,在人脸数据集AFLW和HELEN上与近几年的对齐方法进行对比实验,平均错误率8.80%,i5四核CPU,2.3Hz主频硬件平台下计算时间7.6ms。我们的人脸对齐方法在真实场景下(包括单人和多人)具有较好的鲁棒性,可以实现准确定位。Face alignment is an important part of face recognition system.The accuracy and speed of alignment directly affect the effect of face recognition.Face images have different postures,different expressions,different lighting conditions and other factors.Face alignment becomes a difficult problem.This paper proposes a auto-encoder network face alignment algorithm based on SURF features.The approximate face feature points are found by the rough location network.The local SURF features are extracted and input into the local thinning network.The exact locations of the feature points of the face are further refined by the cascade structure.In the face data sets AFLW and HELEN,we compared some new alignment method.The average error rate is 8.80%,the computation time is 7.6 ms under i5 quad core CPU,2.3 Hz dominant frequency.Our face alignment method has good robustness in real scene(including single and multiple faces),and it can achieve accurate positioning.

关 键 词：SURF特征 人脸对齐 自编码器网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]