结合多尺度融合特征和残差注意力机制的联合三维人脸重建及密集对齐算法  

Joint 3D face reconstruction and dense alignment algorithm combining multi-scalefusion features and residual attention mechanism

作　　者：黄有达 周大可[1] 杨欣[1] Huang Youda;Zhou Dake;Yang Xin(College of Automation Engineering,Nanjing University of Aeronautics&Astronautics,Nanjing 211100,China)

机构地区：[1]南京航空航天大学自动化学院,南京211100

出　　处：《计算机应用研究》2021年第7期2175-2178,2187,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61573182);南京航空航天大学研究生创新基地(实验室)开放基金资助项目(kfjj20190322)。

摘　　要：针对三维人脸重建和密集对齐算法精度不足的问题,引入密集连接的多尺度特征融合模块和残差注意力机制设计了一种性能强大的网络。在编码器结构前,引入密集连接的多尺度特征融合模块获得多尺度融合特征,使编码器获得更丰富的信息;在解码器模块中引入残差注意力机制,强化网络对重要特征的关注同时抑制不必要的噪声。实验结果表明,相较其他算法,该算法取得了较显著的改进;相对PRNet,该算法以更少的参数量在各项指标上取得7.7%~12.1%的性能提升。Aiming at the problem of insufficient accuracy of 3D face reconstruction and dense alignment algorithm,this paper introduced densely connected multi-scale feature fusion module and residual attention mechanism to design a powerful network.For the encoder structure,this paper introduced a densely connected multi-scale feature fusion module to obtain multi-scale fusion features,which made the encoder obtain richer information.The decoder module introduced residual attention mechanism to strengthen the network’s attention to important features while suppressing unnecessary noise.According to the experimental results,compared with other algorithms,the proposed algorithm achieves a significant improvement.Compared with the PRNet,the algorithm uses fewer parameters to achieve a performance improvement of 7.7%to 12.1%in various indicators.

关 键 词：三维人脸重建和密集对齐 密集连接 空洞卷积 残差注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]