基于单张照片的三维人脸重建研究

发布时间：2018-06-24 22:45

本文选题：三维人脸重建 + 明暗恢复形状　；参考：《武汉理工大学》2011年硕士论文

【摘要】：人脸的几何形状和表面材质非常复杂,现实生活中的很多问题在二维的环境下并不能很好地解决。比如在人脸识别、人脸监控等方面,在表情变化或者正侧面匹配上其识别率始终无法提高,如果能有三维模型,那么就可以在三维模型上进行表情合成或者利用三维信息进行匹配,在理论上会识别率和鲁棒性都将会有提高。目前,三维人脸模型在安全认证、影视动漫、医学科学等领域的应用非常广泛。近年来,三维人脸重建方面的研究也取得了很多不错的成果。然而获取真实详细的脸部信息成本非常昂贵,例如需要利用三维激光扫描仪,而且由于环境因素或者建模对象的配合程度的影响导致实施起来比较困难。而人脸的二维图像获取相对而言非常简单,因此目前很多研究都围绕着基于二维图像去重建三维的人脸模型,而其中,尤其是单张正面图像所需要的条件少而非常容易获取,因此,本文将对基于单张照片的二维信息进行人脸的三维模型重建的方法进行讨论。本文首先深入研究了当前主流的基于图像明暗信息来恢复形状的三维重建算法,分析对比了传统的四类典型解法,包括最小化方法、演化方法、局部方法和线性化方法,并且从中选取了演化方法进行求解研究,在此基础之上提出了一种基于水平集算法和能量最小化方法的三维重建算法。首先本文在分析介绍了传统的水平集的数值解法之后,提出了一种改进的快速行进法来求解水平集问题,通过实验对比,该算法能够使求解的精度显著性地提高。使用改进的快速行进法求解基于图像明暗恢复形状的问题得到了和原始二维图像比较一致的三维人脸模型,但是由于明暗恢复形状问题所固有的病态性,求解结果不可避免地出现了凹凸二义性的问题。为了解决这个问题,本文首先基于三角面片的方向性寻找凹凸分界线,从而确定凹陷的区域,实验结果证明该方法的精确度比较高。然后采取能量最小化的方法,基于凹凸边界线和相对中心线对凹陷区域进行重构,实验结果表明最终获得的三维人脸模型和原始二维图像相似度很高,具有非常好的真实性。
[Abstract]:The geometric shape and surface material of human face are very complex, and many problems in real life can not be solved well in two-dimensional environment. For example, in respect of face recognition, face monitoring and so on, the recognition rate of facial expression changes or positive side matching can never be improved. If there can be a three-dimensional model, Then we can combine facial expression on 3D model or match it with 3D information. In theory, the recognition rate and robustness will be improved. At present, 3D face models are widely used in the fields of security authentication, video animation, medical science and so on. In recent years, the research of 3D face reconstruction has also made a lot of good results. However, it is very expensive to obtain real and detailed facial information, such as using a 3D laser scanner, and it is difficult to implement because of the influence of environmental factors or the degree of cooperation of modeling objects. The 2D image acquisition of human face is relatively simple, so many researches focus on the reconstruction of 3D face model based on 2D image, and especially the single frontal image needs less conditions and is very easy to obtain. Therefore, the method of 3D face reconstruction based on two-dimensional information of single photograph is discussed in this paper. In this paper, the current mainstream 3D reconstruction algorithm based on image shading information to restore shape is deeply studied, and four typical methods are analyzed and compared, including minimization method, evolution method, local method and linearization method. The evolutionary method is selected to solve the problem. Based on this, a 3D reconstruction algorithm based on level set algorithm and energy minimization method is proposed. In this paper, after analyzing and introducing the traditional numerical method of horizontal set, an improved fast moving method is proposed to solve the level set problem. The experimental results show that the accuracy of the algorithm can be significantly improved. The improved fast moving method is used to solve the shape restoration problem based on image light and dark. A 3D face model which is consistent with the original 2D image is obtained. However, because of the inherent ill-condition of the shape restoration problem, The problem of concavity and convexity is inevitable. In order to solve this problem, firstly, the concave and convex boundaries are found based on the directionality of the triangles, and the region of the depression is determined. The experimental results show that the accuracy of the method is high. Then the method of energy minimization is adopted to reconstruct the concave region based on the concave boundary line and the relative center line. The experimental results show that the resulting 3D face model has a high similarity with the original two-dimensional image and has a very good authenticity.
【学位授予单位】：武汉理工大学
【学位级别】：硕士
【学位授予年份】：2011
【分类号】：TP391.41

【相似文献】