基于深度卷积网络的3D人脸重构算法

基于单张人脸图片的3D人脸模型重构，无论是在计算机图形领域还是可见光成像领域都是一个极具挑战性的研究方向，对于人脸识别、人脸成像、人脸动画等实际应用更是具有重要意义。针对目前算法复杂度较高、运算量较大且存在局部最优解和初始化不良等问题，本文提出了一种基于深度卷积神经网络的单张图片向3D人脸自动重构算法。该算法首先基于3D转换模型来提取2D人脸图像的密集信息，然后构建深度卷积神经网络架构、设计总体损失函数，直接学习2D人脸图像从像素到3D坐标的映射，从而实现了3D人脸模型的自动构建。算法对比与仿真实验表明，该算法在3D人脸重建上的归一化平均误差更低，且仅需一张2D人脸图像便可自动重构生成3D人脸模型。所生成的3D人脸模型鲁棒性好，重构准确，完整保留表情细节，并且对不同姿态的人脸也具有较好的重建效果，能够在三维空间中无死角自由呈现，将满足更多实际应用需求。

3D face reconstruction algorithm based on deepconvolutional networks

The reconstruction of 3D facial model based on single face image is a challenging research direction both in the field of computer graphics and visible light imaging area,and for face recognition,facial imaging,facial animation and other practical applications are of great significance.In view of the present high computational complexity,large operation quantity,local optimal solution and poor initialization and other problems,a convolutional neural network based on the depth of single image to 3D face reconstruction algorithm is proposed in this paper.Firstly,the intensive information of the 2D face image is extracted based on the 3D transformation model,then a depth convolution neural network architecture is constructed to design an overall loss function,and to directly study of 2D face image from the pixel to the 3D coordinate mapping.Thus,the automatic construction of the 3D face model is achieved.Comparison of algorithms and simulation experiments show that the proposed algorithm in 3D face reconstruction on the normalized average error is lower,and only a 2D image face can be automatic refactoring generated 3D model face.The generated 3D model face has good robustness,accurate reconstruction,complete preservation of facial details,and good reconstruction effect for faces with different postures.It can be freely presented in three dimensional space,which will meet more practical application needs.