基于卷积神经网络的单幅图像三维人脸重建  

3D face reconstruction of single image based on CNN

作　　者：王育坚[1] 李深圳 韩静园 谭卫雄 WANG Yujian;LI Shenzhen;HAN Jingyuan;TAN Weixiong(School of Information,Beijing Union University,Beijing 100101,China)

机构地区：[1]北京联合大学信息学院,北京100101

出　　处：《传感器与微系统》2021年第6期52-56,共5页Transducer and Microsystem Technologies

基　　金：国家自然科学基金资助项目(61572077)。

摘　　要：利用单幅二维图像进行三维人脸重建是图像处理研究领域的热点问题。受深度卷积神经网络(CNN)和三维形变模型(3DMM)的启发,提出一种采用CNN回归3DMM形状和表情参数的方法,进行三维人脸重建。在CNN模型VGG-16的基础上设计一种VGG-BN的改进网络模型,通过在每个卷积层后加入批归一化层,优化网络模型性能;并采用迁移学习方法,将预训练模型引入到VGG-BN网络的训练中。将改进的网络模型在300W-LP数据集上训练,在AFLW2000-3D数据集上测试,并和现有方法进行了对比分析。实验结果表明:改进的网络模型在人脸重建的准确性和泛化性方面都有一定的改善,重建人脸的形状和表情效果较好。Three-dimensional face reconstruction using a single two-dimensional image is a hot topic in the field of image processing.Inspired by deep convolutional neural network(CNN)and 3 D morphable model(3 DMM),a method of 3 D face reconstruction using convolution neural network to regress the shape and expression parameters of 3 DMM is proposed.Based on the convolution neural network model VGG-16,an improved network model of VGG-BN is designed.By adding batch normalization layer after each convolution layer, the performance of network model is optimized.The pre-training model is introduced into the training of VGG-BN network by transfer learning method.The improved network model is trained on 300 W-LP dataset, tested on AFLW2000-3 D dataset, and compared with existing methods.Experimental results show that the improved network model can improve the accuracy and generalization of face reconstruction, and the effect of face reconstruction is better.

关 键 词：三维人脸重建 三维形变模型(3DMM) 卷积神经网络(CNN) 单幅图像 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]