【作者】 梁永涛；

【导师】 孙艳丰；

【作者基本信息】 北京工业大学 ， 计算机应用技术， 2006， 硕士

【摘要】 人脸识别是当前计算机视觉、模式识别、计算机图形学等领域的研究热点,具有重要的理论研究意义和巨大的应用价值。经过几十年的发展,人脸识别研究取得很大成就,在条件可控或者理想情况下基本达到实用水平。但是目前的人脸识别技术水平在非理想情况下与实用要求有很大距离,仍有许多关键性难题需要解决,特别是光照问题、姿态问题、表情问题等。在实验室对人脸识别多年研究的基础上并结合作者在研究生阶段参与的科研课题,本文围绕着人脸识别中的光照问题以及其他问题做了以下几方面的研究工作:1.对人脸识别研究做了概述。一些国内外研究者对人脸识别做了很好的综述,但是作为当前的研究热点,新的方法和技术不断出现,每年都有大量关于人脸识别的研究成果。本文从发展历程、国内外研究现状、经典算法、性能评价和面临问题五个方面对人脸识别研究进行了系统全面地介绍。2.分析了三维形状和二维纹理对人脸识别的影响。在本实验室建立的大规模中国人的三维人脸库和基于重采样三维人脸对齐算法的基础上,本文设计合理的实验,采用平均三维人脸来改变原始三维人脸的形状和纹理信息,系统地分析了在这种变换模式下的人脸三维形状和二维纹理对人脸图像识别的影响,在Eigenface算法上的试验结果表明,形状和纹理信息对人脸识别都有很大影响,并且形状信息的影响远远大于纹理信息的影响。3.在深入分析光照问题的基础上,提出了一种基于人脸形状预测模型合成虚拟图像的人脸识别方法。本文选用支持向量回归来训练和学习人脸图像和人脸三维形状之间的内在关系,并把这种内在关系作为先验知识用于预测新输入人脸图像的形状信息,避免了非人脸的产生,并保证预测的准确性。利用预测得到的人脸三维形状和输入图像合成不同光照条件下的虚拟图像,从而通过丰富人脸样本来提高了人脸识别对光照的鲁棒性,在合成虚拟图像过程中结合光照比例图方法,增强了合成效果。实验结果表明通过该方法增加虚拟人脸可以显著改进多光照条件下的人脸识别性能。更多还原

【Abstract】 Face recognition has been the research focus in computer graphics, computer vision and pattern recognition as its wide range of applications and scientific value. There are many face recognition techniques have been proposed and shown significant promise, and many commercial systems are available for various applications under well-controlled environment in return for the long time research effort, but robust face recognition is still difficult as many unresolved challenges, such as illumination and pose variation problems, expression problem and so on. This thesis is based on the foundation of researching on face recognition in my laboratory and the projects that I have taken part in, the main contributions of it are as following:1. Provides a review of face recognition literature. Although there are some good surveys of face recognition to be published in the past few years, we think it is necessary to give a new overview of face recognition from different viewpoints because there are many new algorithms and technologies to be presented very year. This survey describes the face recognition technology from viewpoints of historical development, state-of-the-arts in the world, the classical algorithms, performance evaluation and key issues in face recognition.2. Measures quantificationally the three-dimensional shape and two-dimensional surface reflectance contributions to face images recognition. Based on the BJPU-3D large-scale Chinese face database and the correspondence between these faces by though mesh resampling, we have altered the three-dimensional shape and two-dimensional surface reflectance of the original 3D faces by making use of the average face, then we have implemented a series of experiments to measure the contributions of three-dimensional shape and two-dimensional surface reflectance to face image recognition under this change pattern. From the experiment results based on the Eigenface algorithm, we concludes that the two types of information are both important for face recognition but the 3D shape is much more important than the 2D surface reflectance,this conclusion gives good suggestions and clues to the design of face recognition algorithms in future.3. Proposes a face recognition method by synthesizing virtual face images based on face shape prediction model by through the deep analysis of the illumination problem. We train the support vector regression (SVR) model through examples of two-dimensional (2D) and更多还原