鲁棒人脸正面化方法研究

发布时间：2018-01-26 00:16

本文关键词： 人脸正面化正交Procrustes Shatten-p范数三维模型　出处：《南京理工大学》2016年硕士论文　论文类型：学位论文

【摘要】：一直以来,人脸识别都是模式识别和计算机视觉领域的热点问题,尽管在正面视图上人脸识别已经取得了很好的结果,但是存在不同姿态的人脸图像仍然会导致人脸识别系统性能的下降。为了将不同姿态下的人脸图像转换成正面视图,本文围绕人脸正面化问题,运用统计分析、三维建模、特征表示、回归表示等理论,归纳总结了当前主流的人脸正面化方法,并对现有算法加以改进,以提升人脸正面化方法的鲁棒性。本文的主要工作和研究成果如下:(1)从基于二维平面的人脸正面化方法出发,本文提出了结构化正交Procrustes回归。与正交Procrustes回归用Frobenius范数约束误差项相比,我们的方法用Shatten-p范数可以更好地刻画图像的结构信息,从而能够得到更加鲁棒的结果。此外我们也分别讨论使用l1范数和l2范数来约束表示系数的情况,并给出了相应的优化算法。(2)本文提出了一种基于三维模型的鲁棒人脸正面化方法,该方法首先应用SDM(Supervised Descent Method)方法对给定的二维图像进行特征点定位。同时,给定一个正面的三维模型并手动标定相应的特征点,通过二维图像和三维模型之间的特征点位置对应关系,我们可以计算出二维图像特征点和三维模型特征点间的投影矩阵,利用该投影矩阵初步将二维图像的姿态变化矫正为正面视图。然后,再利用人脸的局部对称性完成对不可视区域填充,同时进行遮挡检测,并利用泊松图像编辑和局部对称性去除遮挡。
[Abstract]:Face recognition has always been a hot topic in the field of pattern recognition and computer vision, although it has achieved good results in frontal view. But face images with different pose still lead to the deterioration of face recognition system performance. In order to convert face images under different poses into positive view, this paper uses statistical analysis around face frontal image. Three-dimensional modeling, feature representation, regression representation and other theories, summarized the current mainstream face frontal methods, and improved the existing algorithms. In order to improve the robustness of the face obverse method. The main work and research results are as follows: 1) based on the two-dimensional plane face obverse method. In this paper, a structured orthogonal Procrustes regression is proposed, which is compared with the Frobenius norm constraint error term used in orthogonal Procrustes regression. Our method can better depict the structure information of images by using Shatten-p norm. In addition, we also discuss the case of using l 1 norm and l 2 norm to constrain the coefficients respectively. In this paper, a robust face facade method based on 3D model is proposed. Firstly, the SDM(Supervised Descent method is used to locate the feature points of a given two-dimensional image. A frontal 3D model is given and the corresponding feature points are manually calibrated. We can calculate the projection matrix between 2D image feature points and 3D model feature points, using this projection matrix, we can preliminarily correct the attitude change of 2D image into a frontal view. Then the local symmetry of the face is used to fill the invisible region, and the occlusion detection is carried out at the same time, and the occlusion is removed by Poisson image editing and local symmetry.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】