三维动态人脸定位研究

发布时间：2018-03-17 19:02

本文选题：三维人脸识别　切入点：三维人脸定位　出处：《吉林大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着社会的不断进步,生物识别技术逐渐受到各界研究学者的青睐.而人脸识别因其自然性和不易被被测个体察觉的特点以及在个人信息安全、自助服务、电子商务、安全防务等领域的巨大应用前景日益受到重视.在研究人员几十年的钻研下,二维人脸识别算法取得了显著地进展.然而不同于传统的二维人脸识别容易受到光线、化妆等外界因素的影响,三维人脸识别因其处理对象是三维人脸空间模型而更加安全、鲁棒.人脸定位作为人脸识别中的关键基础步骤,其结果的准确性对于之后的表情提取与分析、矫正人脸等处理举足轻重.因此,快速准确地定位人脸成为重中之重.本文针对三维动态人脸的空间模型,深入探讨了三维人脸识别中的数据处理部分:人脸的定位.主要分为人脸模型的读入、特征点标记和人脸区域获取三个阶段.针对采集到的三维动态人脸模型数据,首先要借助环境VS2010,将数据读入并且显示人脸模型.文中先简要说明了数据获取的三维设备:三维照相机.然后针对三维照相机采集到的人脸模型数据,简要描述了将其读入电脑显示的方式.因为扫描仪采集的数据范围较大,采集到的人脸模型数据存在耳朵、脖子、肩膀等无效部分,因此首先需要去除无用部分,将感兴趣的用于人脸识别的有效部分进行切割和定位.为了能够准确提取人脸特征并且减少冗余信息的干扰,文中选取球体对人脸进行定位.本文先在三维人脸模型上寻找到特征点.由于提取特征的鉴别力决定了识别性能,所以特征提取在三维人脸识别中占据举足轻重的地位.基于特征提取策略,根据需要由简到繁的给出手工定位特征点、根据深度信息定义特征点以及经典的ASM算法三种获取方法.然后简要地比较了三种方法的优劣性,以及各种方法所适应的场景.最后,以定位的鼻尖点为球心构造一个球体对三维模型进行切割,得到的球体内区域即为切割的人脸区域,从而实现人脸区域的定位.人脸获取部分的关键之处在于计算球体的半径,这里主要给出了测地距离的方法.通过将其与最为常用的欧氏距离进行比较,能够突出测地距离的优势.上机实验结果可以看出,本文提出三维人脸的定位方法具有较好的实用性.
[Abstract]:With the development of society, biometric recognition technology has been favored by researchers from all walks of life, and face recognition, because of its naturalness and characteristics that are not easy to be detected by tested individuals, can be used in personal information security, self-service, electronic commerce, etc. Great application prospects in the field of security and defense have been paid more and more attention. After decades of research by researchers, two-dimensional face recognition algorithms have made remarkable progress. However, unlike traditional two-dimensional face recognition, two-dimensional face recognition is vulnerable to light. Due to the influence of external factors, such as makeup, 3D face recognition is more secure and robust because its object is 3D face space model. Face location is a key step in face recognition. The accuracy of the result is very important for facial expression extraction and analysis and face correction. Therefore, it is very important to locate the face quickly and accurately. This paper focuses on the spatial model of 3D dynamic face. In this paper, the data processing in 3D face recognition is deeply discussed, which is divided into three stages: face model reading, feature point marking and face region acquisition. First of all, we should read the data into and display the face model with the help of the environment VS2010. Firstly, this paper briefly describes the 3D device for data acquisition: 3D camera, and then aims at the face model data collected by 3D camera. This paper briefly describes the way to read it into the computer display. Because the data collected by the scanner has a large range of data, the collected face model data has invalid parts, such as ears, neck, shoulders, etc., so it is necessary to remove the useless parts first. In order to extract face features accurately and reduce the interference of redundant information, the valid parts of face recognition of interest are cut and located. In this paper, the ball is selected to locate the face. Firstly, the feature points are found on the 3D face model. The ability of extracting features determines the recognition performance. Therefore, feature extraction plays an important role in 3D face recognition. Based on feature extraction strategy, handmade feature points are given from simple to complex according to the need of feature extraction. According to the depth information, the feature points and the classical ASM algorithm are defined. Then, the advantages and disadvantages of the three methods are briefly compared, as well as the scenarios adapted by the various methods. Finally, A sphere is constructed to cut the 3D model with the nasal tip as the center. The region inside the sphere is the area of the cut face, so the location of the face region is realized. The key point of the face acquisition part is to calculate the radius of the sphere. This paper mainly gives the method of geodesic distance. By comparing it with the most commonly used Euclidean distance, the advantage of geodesic distance can be highlighted. In this paper, a 3D face localization method is presented, which has good practicability.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】