基于双目立体视觉的三维人脸重构及其识别-电路与系统专业论文.docxVIP

摘要 由于三维人脸识别在人脸识别方面表现优异，其应用变得越来越广泛。作为 三维人脸识别的前提，三维人脸重构也引起了研究者越来越多的关注。 仿照人类视觉系统，双目立体视觉是三维人脸重构的一个比较重要的方法。 本文使用两个普通的摄像机搭建了双目视觉系统，并利用 Matlab 工具对两个摄像 机进行校正，计算出摄像机内部参数和外部参数。立体匹配是双目立体视觉的难 点，本文选择了稠密匹配方法来实现，并在 Texas 3DFRD 数据库中成功进行了人 脸匹配。作为方法的改进，本文提出利用主动形状模型(ASM)提取人脸特征点， 并根据这些特征点的视差范围来约束匹配算法的搜索范围，减少匹配的时间，并 降低误匹配率。 本文提出了的人脸识别改进方法，来降低人脸姿态对人脸识别率的影响。该 方法是先估计待识别人脸的姿态，并将 3D 人脸数据库旋转到相同的姿态，再投 影到二维图像作为训练数据库，最后利用经典的二维识别算法进行人脸识别。经 实验验证，在待识别人脸与数据库中的人脸有 10°的差别时，传统的二维人脸识 别率开始急剧下降，而本文提出的算法识别率基本保持不变。 经过大量的数据库验证，本文提出的人脸稠密匹配的改进算法缩短了人脸重 构的时间，并提高了其精确度；人脸识别的改进方法降低了人脸姿态对识别率的 影响。 关键词：摄像机标定 稠密匹配 视差范围 人脸姿态估计 人脸识别 Abstract 3D face recognition has become more and more popular as its extraordinary performance. As the pre-condition of the 3D face recognition, 3D face reconstruction attracts more and more researches’ attention. Imitating the human visual system, the binocular stereo visual system is regarded as a very effective method for 3D face reconstruction. Using two web cameras, this paper builds a binocular stereo visual system and gets internal and external parameters with the Matlab camera calibration toolbox. As a difficult part of the binocular stereo visual, stereo match is realized by the dense stereo correspondence algorithm, which has a good performance in Texas 3DFRD. In order to improve the algorithm, this paper uses the disparity of facial feature points to restrain the search range of the stereo correspondence algorithm. Facial feature points are got by the Active Shape Model (ASM). The improved algorithms can reduce the search time and false matching probability. With the 3D face database, the paper presents a novel face recognition system to weaken the bad effect of the pose on face recognition. The rotate matrix of the new face is estimated to make 3D faces have the same pose. And then these 3D faces are projected on 2D images which are trained to recognize the identity of the new face. The experiment shows that the performance of the traditional 2D face recognition algorithm deterio