基于稀疏化双向二维主成分分析的人脸识别  

Face Recognition Based on Sparse Two-Direction Two-Dimensional Principle Component Analysis

作　　者：张裕平 龚晓峰[1] 雒瑞森 ZHANG Yuping;GONG Xiaofeng;LUO Ruisen(College of Electrical Engineering,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学电气工程学院

出　　处：《计算机工程》2019年第12期232-236,共5页Computer Engineering

基　　金：中国博士后科学基金(2017M612958)

摘　　要：双向二维主成分分析((2D)2PCA)易受异常值影响,鲁棒性差,且所提取的特征向量是非稀疏的。针对上述不足,提出基于L1范数的稀疏双向二维主成分分析方法(2D)2PCA-L1S。在(2D)2PCA目标函数中加入L1范数约束,以提高算法的抗干扰能力,同时引入弹性网约束,通过Lasso与Ridge惩罚函数实现稀疏性。在Feret和Yale数据库中进行基于最近邻的人脸分类、人脸重构和基于粒子群优化SVM参数的人脸识别实验,结果表明,相较于2DPCA、(2D)2PCA、(2D)2PCA-L1等主成分分析方法,该方法能准确提取人脸主要信息,人脸识别和人脸重构效果较好。Two-Direction Two-Dimensional Principle Component Analysis((2 D)~2PCA)is an improved method of Principle Component Analysis(PCA)in the two-dimensional space.However,just like PCA,the(2 D)~2PCA is susceptible to abnormal values,its robustness is weak and the extracted feature vectors are non-sparse.So this paper proposes a sparse(2 D)~2PCA method based on L1 norm,the(2 D)~2PCA-L1 S,to tackle these problems.First,this paper adds the L1 norm constraint to the(2 D)~2PCA objective function to improve the anti-interference ability.Then,this paper introduces the elastic network constraint into the objective function,so the sparsity is realized by the Lasso and Ridge penalty functions.The following experiments are carried out on the Feret and Yale datasets:the face classification and face reconstruction based on nearest neighbors,and the face recognition based on Particle Swarm Optimization(PSO)-SVM Parameters.The results show that compared with other PCA method,such as the 2 DPCA,(2 D)~2PCA and(2 D)~2PCA-L1,the proposed method can accurately extract the main face information,and it has better effects on face recognition and face reconstruction.

关 键 词：双向二维主成分分析 稀疏化 粒子群优化 支持向量机 人脸识别 

分 类 号：TN971[电子电信—信号与信息处理]