基于流形结构的人脸民族特征研究  

Research of Face Ethnic Features from Manifold Structure

作　　者：王存睿[1,2] 张庆灵 段晓东[2] 王元刚 李泽东 

机构地区：[1]东北大学系统科学研究所,沈阳110004 [2]大连民族大学民族文化数字技术大连市重点实验室,大连116600 [3]大连理工大学电子信息与电气工程学部,大连116024

出　　处：《自动化学报》2018年第1期140-159,共20页Acta Automatica Sinica

基　　金：国家自然科学基金(61370146;61672132)资助~~

摘　　要：人脸民族特征选取与分析是人脸识别与人类学重要研究方向之一.本文建立了中国三个民族人脸数据库,通过流形结构来研究和分析人脸的民族特征.首先,在体质人类学定义的人脸几何特征指标进行流形分析,未形成按语义分布的子流形.因此本文将人脸特征扩至全部组合的长度、角度和比例特征进行分析,利用m RMR算法对2 926个长度特征、21万余个角度特征、427万个比例特征中冗余特征进行筛选,加上人类学指标及混合筛选的数据集共形成5个数据集.利用LPP、Isomap、LE、PCA和LDA等流形方法分析5数据集,其中的4个数据集都形成了民族语义的子流形分布.为验证筛选特征指标的有效性,本文利用分类算法J48、SVM、RBF network、Naive Bayes、Bayes network在Weka平台对数据集以族群语义作为类别进行交叉验证实验,实验结果表明混合特征的人脸数据集族群分类平均准确率最高,且比例特征分类指标优于其他特征数据集.本文通过大量实验揭示了民族人脸数据可在子空间内形成按民族语义分布的子流形结构.中国三个民族人脸特征在低维空间存在不同民族语义的子流形,通过流形分析和特征筛选构建的人脸测量指标不仅可为人脸族群分析提供方法,同时也将丰富和补充体质人类学的相关研究工作.Facial ethnic feature selection and analysis is one of the most significant research focuses in face recognition and anthropology. In this paper, we build a Chinese ethnic face database including three ethnic groups. Manifold learning is used to analyze facial ethnic features. Firstly, we conduct manifold analysis on the basis of facial geometric indicators proposed by anthropologist, which, however, does not formulate sub-manifold distributed by semantics. Therefore, we intend to expand the scope of facial features by calculating the complete distances, angles and indexes with landmarks. Then, we adopt mRMR to filter 2 926 distance indicators, more than 219 450 angle indicators and more than 4 279 275 index indicators. Finally, we can obtain 5 datasets with features of distance, angle, index, anthropology and mixing. Several popular manifold learning methods including LPP, ISOMAP, LE, PCA and LDA are utilized to study the above mentioned datasets, and we get the distinguishable manifold structure of facial ethic feature and clusters in 4 of the 5 datasets. To evaluate the validity of filtered features, we make use of classification algorithms including J48, SVM, RBF network, Naive Bayes, and Bayes network implemented in Weka for cross validation experiments by ethnic semantics. Experimental results indicate that the average of classification accuracy on the dataset with mixing features is higher than that of other datasets, and that the index is more salient than other geometric features. Moreover, by full experimental investigation, we find that ethnic facial data can generate sub-manifold structure distributed by semantics. Facial features of three Chinese ethnic groups exhibit different ethnic semantic sub-manifolds in the low-dimensional space. Facial measurement indicators obtained by manifold analysis and feature selection not only provide a method for facial ethnic groups analysis, but also enrich and improve the related research work in anthropology.

关 键 词：人脸民族特征 生物特征识别 人脸识别 流形学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]