单张照片的三维人脸重建方法研究

发布时间：2018-05-12 18:50

本文选题：三维人脸建模 + 单张图片　；参考：《南京理工大学》2007年硕士论文

【摘要】： 三维人脸模型的应用广泛存在于安全认证、影视动漫、医学科学等领域。近年来，以详尽的脸部信息进行三维人脸重建取得了许多成果。然而详尽的脸部信息获取不仅成本昂贵，更有诸如监控视频的特定任务的检索等应用却因对象原因无法采集更多的脸部信息。鉴此，本文将进行基于单张正面照片信息进行脸部三维重建方法的讨论。首先从大量的正侧面脸部照片采集着手构建了正、侧面脸部信息的人脸库，依据人体测量学、人体解剖学等脸部关键特征的原则定义了脸部测量点及测量项目。对提取的正侧面脸部数据进行神经网络训练，通过训练的权值得到个体的正面数据点的脸部深度数据，拟合及仿真的实验结果表明了方法的可行性。其次针对神经网络训练过程中收敛速度慢的不足，探讨了寻找最优形状因子、最优学习率以及两者最优组合的加速算法，结合基于epsilon向量外推的加速方法得到一种新的加速算法，数值试验表明能使速度和精度显著提高。第三，基于建立的人脸数据库中测量点定义模型的正侧面特征点，采用径向基函数插值的方法对模型进行调整，，生成特定人脸模型。对于粗糙的原始网格给出了一种改进的Loop细分方法，使细分后的模型更符合原有形状，并纹理映射生成具有真实感的人脸模型。最后用Visual C++.NET和OpenGL实现了从脸部数据库数据提取到三维人脸重建系统。
[Abstract]:Three-dimensional face models are widely used in the fields of security authentication, video animation, medical science and so on. In recent years, many achievements have been made in 3D face reconstruction with detailed facial information. However, detailed facial information acquisition is not only expensive, but also can not collect more facial information because of object reasons. In view of this, this paper will discuss the method of facial 3D reconstruction based on single front photo information. Firstly, the face database of face information is constructed from the collection of a large number of face photographs on the front and side sides. According to the principles of anthropometry, human anatomy and other key features of the face, the measurement points and measurement items of the face are defined. Neural network training is carried out on the extracted facial data, and the depth data of the face of the positive data points are obtained by the weights of the training. The experimental results of fitting and simulation show the feasibility of the method. Secondly, aiming at the shortage of slow convergence in the training process of neural network, the paper discusses the accelerated algorithm to find the optimal shape factor, the optimal learning rate and the optimal combination of the two. A new acceleration algorithm based on epsilon vector extrapolation is presented. The numerical results show that the speed and accuracy can be improved significantly. Thirdly, the radial basis function (RBF) interpolation method is used to adjust the model to generate a specific face model based on the positive side feature points of the measurement point definition model in the established face database. An improved Loop subdivision method is proposed for rough original meshes to make the subdivided models more consistent with the original shape and texture mapping to generate realistic face models. Finally, Visual C. Net and OpenGL are used to realize the face data extraction and 3D face reconstruction system.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2007
【分类号】：TP391.41

【引证文献】