基于四元数的彩色人脸识别  

Quaternion Based Color Face Recognition

作　　者：郎方年[1] 田丽媛[2] 刘莉莉[3] 

机构地区：[1]成都大学信息科学与技术学院,四川成都610106 [2]江苏大学附属医院信息科,江苏镇江212013 [3]营口职业技术学院成人教育分院,辽宁营口115000

出　　处：《成都大学学报（自然科学版）》2013年第4期359-367,382,共10页Journal of Chengdu University（Natural Science Edition）

基　　金：国家自然科学基金(69572027);中国博士后基金(20100471665)资助项目

摘　　要：为充分挖掘人脸模式样本之间的鉴别信息、强化不同样本之间的区分性,以利于增强识别系统鲁棒性、提高人脸正确识别率,提出一种新颖的基于四元数的彩色人脸识别算法.将定义于实数域的PCA方法以及Fisher鉴别分析法向四元数体作合理推广,得到定义于体上的广义主成分分析方法及广义线性鉴别分析法,将这2种方法用于彩色人脸识别,从而得到全新的识别算法.该算法巧妙地将彩色像素的R、G、B 3个分量结合在一起,从数学上有机融合具有丰富鉴别信息的肤色成分以及反映人脸轮廓形状信息的灰度成分,较传统仅利用灰度信息的识别方法,具有更稳定的性能以及更高的正确识别率.提出的关于共轭四元数矩阵正交特征矢量集的获取方法,数学上有详细的推导证明,该方法在理论上合理,同时在自己建立的彩色人脸库上进行的实验表明,该方法可行且实用.In order to dig discriminant information among face samples fully and intensify the differentiating feature between different samples,a novel quatemion based color face recognition algorithm is proposed inthis paper for toning up the stability of recognition system and enhancing right recognition rate. Firstly, quatemion based principal component analysis（QPCA） and general linear disefiminant analysis（QLDA） are obtained through generalizing the principal component analysis （PCA） algorithm and Fisher pattem classification algorithm defined in real domain to quatemion domain. Then, by using QPCA and QLDA in our special pattern classification field--color face recognition, a novel recognition method is obtained. By combining Red, Green and Blue three components of color pixels artfully, our method integrates organically the color component of face which contains plentiful discriminant information with the gray component which reflects the face contour shape information. What＇ s more, our algorithm has more stable performance and higher right recognition rate than the conventional algorithm which uses gray information merely. The detailed derivation in math shows that our method has considerable rationality in theory. The experimental result in our color face database indicates that our method is feasible and practical.

关 键 词：特征抽取 四元数体 自共轭矩阵 模式分类 彩色人脸识别 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]