基于深度学习的三维人脸稀疏重建研究

Research on 3D Face Sparse Reconstruction Based on Deep Learning

为提升三维人脸稀疏重建的精确性,论文参考深度学习领域的点云特征提取网络模型,搭建了全新的人脸关键点特征提取网络。该网络克服了传统投影重建法只能在二维平面进行拟合计算的缺点,可以直接在三维空间从输入的稀疏关键点中提取出关键点的全局特征,并转化为所需的人脸模型参数,从而实现了由稀疏关键点到三维人脸的重建过程。实验结果表明,该方法在BFM三维人脸库上的最终结果,重建精度明显优于传统投影重建法,具有更加出色的重建性能。

In order to improve the accuracy of 3D face sparse reconstruction,referring to the point cloud feature extraction network in the field of deep learning,a new face landmarks feature extraction network is established.This work overcomes the disadvantage that the traditional projection reconstruction method can only fit on 2D plane.It directly extracts global features from the input 3D sparse landmarks and converts them into the required face model parameters to realize the reconstruction process from sparse landmarks to 3D face.The experimental results show that the reconstruction accuracy of the method on BFM 3D face database is significantly better than the traditional projection reconstruction method,and has desirable performance.