基于MARS图的人脸人耳多模态识别研究

发布时间：2018-01-30 21:39

本文关键词： 人脸识别人耳识别多模态识别点云配准稀疏表示　出处：《北京科技大学》2015年博士论文　论文类型：学位论文

【摘要】：随着信息化社会的不断发展,信息安全成为社会关注的热点,基于生物特征的身份识别在社会生活中的需求越来越强烈。近年开展的生物特征识别研究工作已经表明,单一模态的生物特征识别在实际应用中的准确性和鲁棒性难以满足需要,多模态识别可以融合多种生物特征体,丰富个体的鉴别信息,提高识别的准确性和鲁棒性。以人脸和人耳两种生物模态进行融合的识别具有友好性和非打扰性等特点,成为多模态生物识别研究的热点之一。受益于三维数据采集技术的发展,生物特征识别领域中的相当一部分研究延伸到使用三维信息进行识别。相对于二维识别,三维识别对光照和姿态变化的鲁棒性有所提高,但仍有受表情变化的影响明显、三维数据存储和计算开销大等不足,另外,三维识别同样面临着遮挡和数据缺失的问题。在非受控识别场景中,姿态、遮挡等带来的影响使得获取的个体生物特征数据在多数情况下是部分的,存在不可控的变化、缺失,因此实际场景下的识别往往是利用部分数据所进行的识别,如何利用部分数据来进行身份识别是生物特征识别要解决的典型核心问题之一。为实现更为鲁棒的身份识别,克服单一模态识别的不足,本文通过球面变换将采集到的人脸人耳三维数据转换为以识别对象为中心进行表达,进而生成多模态人脸人耳球面深度图与球面纹理图(MARS图)。MARS图自然融合了人脸人耳两种模态,包含了更完整的结构信息和纹理信息,有助于克服人脸、人耳单模态识别中姿态、遮挡、表情等问题带来的影响。MARS图消除了平面外旋转,能够实现无需对准的识别,其二维表达形式可减少数据存储开销,降低识别过程的计算复杂度。鉴于非受控场景下中的身份识别往往是利用部分数据所进行的识别,因此本文重点研究非受控场景下基于部分数据来进行识别的方法。在注册阶段,通过多视角三维人脸人耳数据的融合,构建注册时相对更为完整的全景MARS图原型库来表达身份信息;在识别阶段,构建单视角MARS图并提取单视角MARS图的局部特征与原型库中的全景MARS图局部特征匹配,进行多任务稀疏表示识别。本文的主要研究内容和创新点包括：第一,研究把人脸人耳三维数据由以采集设备为中心的表达转换为以识别对象为中心进行表达的方法,提出了MARS图的数据表达方法,降低了存储和计算开销,有助于实现非受控场景下无需数据对准的身份识别。第二,研究三维人脸人耳定位提取方法以及非刚性部分重合的多视角数据融合方法,提出了基于肤色检测的纯人脸人耳提取算法和基于BANICP的点云配准方法,实现人脸人耳的自动提取和非刚性人脸人耳点云的多视角数据配准和融合。第三,针对非受控场景下基于部分数据的身份识别问题,提出了基于MARS图仿射SIFT特征的多任务稀疏表示识别算法(ASMSRC:Affine-Sift based Multitask Sparse Represent Classifica-tion),通过多任务稀疏表示字典的构建和多任务最优稀疏表示系数求解,对测试样本的局部特征进行重构,依据平均重构误差进行分类和识别。本文提出的基于MARS图的人脸人耳多模态识别方法同时融合了结构特征和纹理特征,对光线变化、姿态变化、部分遮挡和表情变化具有较强的鲁棒性,很大程度上解决了非受控场景下基于部分数据匹配的身份识别问题。本文的研究不仅对基于人脸人耳的身份识别,而且对更广泛领域中的应用基础和理论研究都是有意义的。
[Abstract]:With the continuous development of information society, information security has become the focus of the society, based on biometric identification needs in the social life of the increasingly strong. Biometrics research work carried out in recent years have shown that the single modal biometric recognition in practical application, the accuracy and robustness of the difficult to meet the needs of multi modal identification can the integration of a variety of biological characteristics, rich individual identification information, to improve the recognition accuracy and robustness. The face and ear of two biological modal fusion recognition has the characteristics of friendly and non intrusive, becomes one of the hot research of multimodal biometrics.
Benefiting from the development of three-dimensional data acquisition technology, biometric identification technology is part of the study is extended to identify the use of three-dimensional information. Compared with two-dimensional recognition, robust 3D recognition of illumination and pose changes has increased, but there are still affected by the expression was affected obviously, lack of three-dimensional data storage and computing cost etc. in addition, 3D recognition also faces occlusion and the problem of missing data. The attitude in non controlled recognition in the scene, and the occlusion caused by making the individual biometric data obtained in most cases is part of the existence, change, uncontrollable loss, therefore the recognition scenario is often identified using part of the data, how to use the data for identification is one of the core issues of typical biometric identification to solve.
In order to achieve more robust identification, to overcome the shortcomings of single modal identification, the spherical transform 3D face data acquisition to the human ear to convert to the recognition object as the center of expression, and then generate the multimodal face and ear spherical depth map and spherical texture map (MARS map).MARS map of natural fusion of face two kinds of ear mode, contains the structure information and the texture information is more complete, helps to overcome the human face, gesture, ear recognition of single mode occlusion, bring expression and so on.MARS map to eliminate the plane rotation, to achieve recognition without the alignment, the two-dimensional expression can reduce data storage overhead. To reduce the computational complexity of the recognition process. In view of the identification of uncontrolled scenarios is often in recognition by using part of the data, so this paper focuses on the research of non controlled scenarios of data based on. The method for identification. In the registration phase, through the fusion of multi view 3D face and ear data, construct the registration relatively more complete panoramic MARS prototype library to express identity information; at the recognition stage, construction of single view MARS map and MARS map extraction panoramic local features of the local feature and the prototype of single view MARS map in matching, multi task sparse representation recognition.
The main research contents and innovations include: first, the research method of face and ear by expression of 3D data acquisition equipment for conversion to the center is to identify the object as the center of the expression of the proposed MARS map data expression method, reduce the storage and computation overhead, and identity recognition helps to realize the non controlled scene without data alignment. Second fusion localization method based on 3D face extraction method of human ear and multi view data coincide with non rigid part of the proposed facial skin detection pure ear extraction algorithm and registration method based on point cloud BANICP based on the realization of the human ear face automatic extraction and non rigid face and ear point cloud multi view data registration and fusion. In third, for the non controlled scene based on the identification problem of data in the proposed MARS map affine SIFT feature recognition based on multi task sparse representation Don't (ASMSRC:Affine-Sift based Multitask Sparse Represent algorithm, Classifica-tion) by multi task sparse dictionary construction and multi task optimal sparse representation coefficients, the local characteristics of the test samples to reconstruct, classification and identification based on the average reconstruction error.
The proposed multimodal face recognition method based on MARS graph and combines ear structure feature and texture feature of light change, attitude change, has strong robustness to occlusion and facial expression changes, largely solves the identification problem based on partial data, non controlled scene. This study not only the identity recognition based on face and ear, and the application of a broader base in the field of theory and research are meaningful.

【学位授予单位】：北京科技大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP391.41

【参考文献】