少数民族面部特征自动语义刻画及族群识别方法研究

Automatic facial semantic extraction and ethnic groups recognition are two of the most challenging research interests in the area of computer vision. This project aims to study an automatic semantic characterization of multi-ethnic facial features and Chinese ethnic groups recognition technique, in order to promote the intelligent development of ethnic groups recognition. Firstly, a face landmark detector will be used to locate sufficient landmark points, and then some important low-level features based on these landmarks are extracted. Secondly, the dynamic curves and statistical theory are utilized to describe the shapes of facial components. And the clustering approaches will be built to achieve the automatic semantic extraction for the size and the geometric features of facial components. Thirdly, the above facial semantic extraction will be applied into the semantic characterization for multi-ethnic facial features. Meanwhile, a new feature selection algorithm will be proposed in order to extract the salient semantic characteristics of multi-ethnic facial features. Consequently, an recognition method for ethnic groups will be designed based on deep learning technology and the salient semantic characteristics of multi-ethnic facial features. The continuous research and improvement of this project will effectively solve the key issues in computer vision (manually labeling attributes and limitation of many features in discrete representation), which has significant theoretical and practical value for the intelligent development of the semantic extraction of multi-ethnic facial features and ethnic groups recognition techniques.

人脸语义全自动提取以及异族人脸识别困难问题是计算机领域非常具有挑战性的研究方向。本课题旨在研究面向少数民族面部语义全自动提取方法以及族群识别技术，进而设计出一套少数民族面部语义特征刻画及族群识别的系统方法，促使族群识别技术向着智能化的方向发展。首先，通过检测出的人脸特征点提取面部关键区域的底层特征；其次，用动态曲线及统计理论来刻画人脸关键区域的形状，并结合聚类技术实现人脸关键区域大小和几何特征的语义自动提取；将上述人脸语义自动提取技术进一步应用到少数民族面部特征刻画中，并设计新的特征选择方法，实现少数民族面部显著特征的提取；最后，基于深度学习和民族面部显著语义特征来设计族群识别方法。此项研究的成功推进将有效解决人脸语义提取中存在的手动标记属性、属性的二进制或离散型受限表示及异族人脸识别困难问题，这对于少数民族面部语义提取及族群识别技术的智能化发展有着重要的理论意义及实用价值。

本课题针对人脸语义全自动提取以及异族人脸识别困难问题，研究面向少数民族面部语义全自动提取方法以及族群识别技术。在公理化模糊集理论的相关研究基础上，基于信息粒理论建立了人脸特征刻画与语义提取方法、基于方向型三角形面积曲线聚类思想设计了人脸主要区域形状语义提取方法。并在此基础上，建立了少数民族面部显著特征提取和面部特征刻画方法。同时建立族群标准脸，借鉴SFDAE深度网络模型，再结合上一部分少数民族面部显著特征刻画的研究成果，利用训练图像和每个民族的重要特征及相对应的脸的部位共同训练深度学习网络，得到基于深度学习和语义特征的族群识别方法。本项目的成果将为民族特征刻画和族群识别智能化发展提供技术支持和保障，对于我国非物质文化遗产的保护与传承、民族间和谐稳定发展都具有十分重要的意义。