融合纹理与形状的人脸加权新特征  

Face Feature Weighted by Fusing Texture and Shape

作　　者：方天红[1,2] 陈庆虎[1] 廖海斌[3] 邱益鸣 

机构地区：[1]武汉大学电子信息学院,湖北武汉430072 [2]湖北工程学院物理与电子信息工程学院,湖北孝感432000 [3]湖北科技学院计算机科学与技术学院,湖北咸宁437100

出　　处：《武汉大学学报（信息科学版）》2015年第3期321-326,340,共7页Geomatics and Information Science of Wuhan University

基　　金：国家自然科学基金资助项目(61271256);郑州市重大科技攻关资助项目(072SGZS38042)~~

摘　　要：根据人类进行人脸识别的特点,提出一种纹理与几何形状相结合的人脸新特征。新特征提取的第一步是提取脸部5个关键点;然后,根据人脸图像每个像素点到5个关键点距离动态对每个像素进行加权计算。新特征在纹理特征的基础上,融合了人脸关键点和每个纹理点与关键点之间的位置几何距离信息。与传统的单一纹理特征相比,提高了抗干扰性;而且,由于定位了5个关键点,有利于后续的人脸分块识别。在YALE人脸库和XJTU人脸库上采用线性判别方法与稀疏表示人脸识别方法的实验研究表明：新特征与传统的纹理特征相比,识别率提高了5%~10%;新特征加人脸分块方法识别率接近100%。Textural and structural characteristics are two of the principal features of human faces, both of which have advantages and disadvantages for face recognition. In this paper, we propose a Gaussian weights based new feature combined with texture feature and structure feature approach. In this method, we first use a standard fiducial point detector to locating five face key-points （e. g. , the mid- point of the eye）, and then a new feature will be generated dynamically by weighting the gray value of the pixel according to the distance between the pixel and its relative key-point. The new feature, orig- inated from the texture feature, ties the geometric structural feature of the pixels and the key-point in- formation and delivers better performance than the original texture feature for pose variations. Mean- while, the five key-points are beneficial to part-based face recognition. Experimental results show that the new feature can improve recognition rates by about 5 % than the original texture feature, and im- proved the part--based method by combining the new feature and the face blocking~ achieving recogni- tion rates close to 100% for two different available databases.

关 键 词：人脸识别 特征加权 人脸分块 线性判别分析 稀疏表示 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]