【作者】 胡元奎；

【导师】 汪增福；

【作者基本信息】 中国科学技术大学 ， 模式识别与智能系统， 2006， 博士

【摘要】 自动人脸识别在公共安全、智能监控、数字身份认证、电子商务、多媒体和数字娱乐等领域具有巨大的应用价值，同时，人脸识别的研究涉及多个学科，具有重要的理论研究价值，受到各国政府、科研单位以及军事、安全、情报部门的广泛关注和高度重视。经过几十年的研究，人脸识别取得了长足的发展与进步，目前在控制和配合条件下，人脸图像识别可以取得比较高的识别率，但是在非控制条件和非配合条件下的人脸图像识别仍然是一个极具挑战性的课题，当人脸图像中光照和人脸姿态变化时，识别率急剧下降。 本文对人脸识别中光照变化的影响进行研究，在对当前人脸识别技术中解决光照变化问题的方法进行分析的基础上，提出了一种人脸图像的低维光照空间表示方法，在此基础上进行可变光照下的人脸识别。通过实验发现，9个基本点光源可以近似表示人脸识别应用中几乎所有的光照条件，在这9个基本光源照射下的9幅人脸基图像构成了低维人脸光照空间，它可以表示不同光照条件下的人脸图像，结合光照比图像方法，可以生成不同光照下的虚拟人脸图像。将这些虚拟人脸图像作为模板图像，可以进行不同光照下的人脸图像识别。本文提出的低维光照空间的主要优点是：通过这个光照空间，不仅能够由输入图像估计其光照参数，而且能够由给定的光照条件生成虚拟的人脸图像；利用某个特定人脸的图像建立的光照空间，可以应用于任意一张人脸，生成其在不同光照下的虚拟人脸图像。低维光照空间表示的思想虽然是以人脸为对象提出的，但可以推广应用于其他的对象。 本文对人脸识别中姿态变化的影响进行研究，提出基于单张正面照片的三维人脸模型合成方法，利用合成的三维人脸模型，生成不同姿态下的虚拟人脸图像，从而进行可变姿态下的人脸识别。由二维图像恢复对象的三维模型是计算机视觉领域的一个基本问题，传统的由二维人脸图像恢复人脸三维模型的方法需要多幅人脸图像、图像序列，或者需要限定条件下的立体图像对、正面和侧面图像对等，不利于实际应用。本文提出的算法只需要一幅正面人脸图像，降低了对使用条件的要求，便于实际应用，具有广阔的应用前景，该方法合成的三维人脸模型可以满足人脸识别、表情动画、人机交互的需要。 本文提出一种基于人脸对称性的快速人脸姿态估计方法，该算法能够仅由一幅输入图像快速、准确地估计出人脸的3D姿态。和其它检测方法相比，该方法具有模型简单、计算速度快等优点，利用面积作为输入信息，降低了估计结果对特征更多还原

【Abstract】 Automatic face recognition has great potential applications in public security, intelligent surveillance, digital personal identify, electronic commerce, multimedia, digital entertainment, etc. and has great theory value in many subjects, so face recognition has attracted much research attention from the research institutes, governments, military and security departments. Over the past 30 years, great progress and developments have been made in face recognition. Now, under the controlled and cooperative conditions, face recognition systems perform very well, but under the uncontrolled and uncooperative conditions, especially when the illumination in face images and facial poses are variant, the recognition rate degrade quickly and face recognition is also a great challenge.In this paper, the effect of illumination in face recognition is addressed. On the analysis of the present methods for this problem, a low dimensional illumination space representation (LDISR) of human faces for arbitrary lighting conditions is proposed in this paper. The LDISR is based on the observation that 9 basis point light resources can represent almost arbitrary lighting conditions for face recognition application and different human faces have the similar LDISR. The principal component analysis (PCA) and nearest neighbor clustering method are adopted to obtain the 9 basis point lights. The 9 basis images under the 9 basis point light sources construct a LDISR which can represent almost all face images under arbitrary lighting conditions. Illumination ratio image (IRI) is adopted to generate virtual face images under different illuminations. Based on the virtual face images, face recognition with variant illumination can be performed. The advantage of LDISR is that it can not only synthesize a virtual face image when given lighting condition but also estimate lighting conditions when given a face image, and it can be trained on images of one human face and can be used for all human faces. The LDISR is proposed for human faces, but it can be expanded for other objects.This paper proposes a new method of modeling 3D face based on single frontal face image. Using the 3D face model, virtual face images under different poses can be generated and then be used as template images for face recognition with variant poses. Recovering the 3D shape from 2D images is the basic problem in computer vision. The traditional methods of recovering 3D face model from 2D face images need multi images, image sequence, stereo images, or front and profile images, and the qualifications are difficult to fulfill in many practical applications. The proposed method needs only one frontal face image, facilitates the use in practical applications, and has great potential applications. As the experiment shows, face model synthesized by our method can fit for the applications such as face recognition, expression animation and human computer interaction.In order to estimate the facial 3D pose quickly and accurately, Area Model and更多还原