针对多视图三维人脸重建时,由于图像特征提取能力有限而导致三维人脸模型重建效果不佳的问题,本文提出了一种基于改进MobileNet V2网络的三维人脸重建方法.首先在MobileNet V2网络中引入多路径结构,使不同子路径的特征能够进行交互,从而增强网络的表达能力;然后在通道维度上加入压缩激励(Squeeze and Excitation,SE)模块,在空间维度上加入空间分组增强(Spatial Group-wise Enhance,SGE)模块,从而有效提取人脸形状参数和表情参数;最后使用三维可形变模型(3D Morphable Model,3DMM),并根据形状参数和表情参数完成三维人脸重建.在300W-LP数据集上,与基于卷积神经网络(Convolutional Neural Network,CNN)的重建方法以及基于Visual Geometry Group Batch Normalization(VGG-BN)改进网络的重建方法进行对比实验,实验结果表明平均重建精度分别提升约0.46％和0.31％.
Multi-View 3D Face Reconstruction Based on Improved MobileNet V2
Aiming at the problem that the3D face model reconstruction effect is poor due to the limited image feature extraction ability in multi-view 3D face reconstruction,a 3D face reconstruction method based on improved MobileNet V2 network is proposed. Firstly,the multipath structure is introduced into the MobileNet V2 network,so that the features of different subpaths can interact with each other,so as to enhance the expression capability of the network. Then Squeeze and Excitation ( SE) module is added to the channel dimension,and Spatial Group-wise Enhance ( SGE ) module is added to the spatial dimension,so as to extract the face shape parameters and expression parameters effectively. Finally,the 3D Morphable Model (3DMM) is used to reconstruct the 3D face according to the shape parameters and expression parameters. On the 300W-LP dataset,compared with the Convolutional Neural Network ( CNN) reconstruction method and the Visual Geometry Group Batch Normalization ( VGG-BN) improved network reconstruction method,the experimental results show that the average reconstruction accuracy is improved by about 0.46％ and 0.31％ respectively.
three-dimensional face reconstructionMobileNet V2multipath structureSE moduleSGE module3DMM
NETL
NSTL
万方数据
