改进加权投票的PCA-Net多特征融合SSFR  

A Fusion method of PCA-Net and Multi-Features with Improved Weighted Voting for Single-Sample Face Recognition

作　　者：赵淑欢 葛佳琦 梁晓林 刘帅奇[1,2] ZHAO Shu-huan;GE Jia-qi;LIANG Xiao-lin;LIU Shuai-qi(College of Electronic Information Engineering,Hebei University,Baoding Hebei 071000,China;Hebei Machine Vision Technology Innovation Center,Baoding Hebei 071000,China)

机构地区：[1]河北大学电子信息工程学院,河北保定071000 [2]河北省机器视觉技术创新中心,河北保定071000

出　　处：《计算机仿真》2023年第4期223-230,共8页Computer Simulation

基　　金：国家自然科学基金项目(61572063,61401308);河北省自然科学基金项目(F2019201151,F2019201362,F2018210148,F2020201025);河北省高等学校科学技术研究项目(QN2016085,QN2017306,BJ2020030);河北大学校长基金(XZJJ201909);河北大学高层次人才科研启动经费项目(2014-303,8012605);河北省创新技术中心开放课题(2018HBMV01,2018HBMV02);广东省数字信号与图象处理技术重点实验室开放基金资助(2020GDDSIPL-04)。

摘　　要：单样本人脸识别是人脸识别在实际应用中面临的挑战性问题之一,虽然深度学习在人脸识别方面取得突破性进展但其性能依赖海量标注性数据,故其在单样本上性能有限。而传统浅层特征对有标注的数据量需求不高,但因单样本数据缺少类内变化其性能有限,提出一种改进加权投票的PCA-Net多特征融合算法。在数据集方面,利用LU分解生成虚拟样本扩展数据集;根据PCA-Net特征下样本的相关性细化数据集,实现对数据集初步特征提取和筛选;在细化数据集上提取多LBP特征并与PCA-Net特征进行加权投票。在AR、Extended Yale B、CMU-PIE三个数据库上的实验结果表明,所提方法提高了单样本人脸识别性能。Single-sample face recognition is one of the challenging problems for face recognition in practical applications.Although deep learning methods have made breakthroughs in face recognition,its performance relies on massive labeled data,so its performance on single-sample is limited.However,traditional shallow features do not have a high demand for labeled data,but their performance is limited due to the lack of intra-class variation for single-sample training set.For this reason,this paper proposes a PCA-Net multi-feature fusion algorithm with improved weighted voting.First,in terms of data sets,LU decomposition was used to generate virtual samples;Secondly,the data set was refined according to the correlation values between training samples and testing sample in the PCA-Net features space to achieve preliminary feature extraction;Finally,in the refined dataset,multiple LBP features and perform weighted voting with PCA-Net features were extracted.Experimental results on AR,Extended Yale B,and CMU-PIE databases show that our method improves the performance of single-sample face recognition.

关 键 词：单样本人脸识别 局部二值模式 虚拟样本 特征融合 加权投票 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]