摘要
针对沙漏网络应用于人脸对齐中存在网络结构复杂、时间开销大的问题,提出一种带有深度可分离的轻量级沙漏网络。通过知识蒸馏的思想构造轻量级沙漏网络,解决网络结构复杂的问题;在叠层沙漏网络中使用深度可分离卷积,通过深度卷积和逐点卷积共同作用简化复杂网络,解决时间开销大的问题。实验结果表明,在300w数据集和WFLW数据集上,该方法与主流的人脸对齐方法相比,对齐精度在保持基本不变的情况下,对齐速度具有明显的优势。
Aiming at the problems that the hourglass network has complex network structure and large time overhead in face alignment,a lightweight hourglass network with deep separability was proposed.A lightweight hourglass network was constructed through the idea of knowledge distillation to solve the problem of complex network structure.Deep separable convolution was used in a stacked hourglass network,both deep convolution and pointwise convolution were used.The function simplified complex networks and solved the problem of large time complexity.Experimental results show that,on the 300w dataset and the WFLW dataset,compared with the mainstream face alignment method,this method has obvious advantages in alignment speed while the alignment accuracy maintains the same.
作者
贺怀清
陈琴
惠康华
HE Huai-qing;CHEN Qin;HUI Kang-hua(Department of Computer Science and Technology,Civil Aviation University of China,Tianjin 300300,China)
出处
《计算机工程与设计》
北大核心
2021年第8期2316-2323,共8页
Computer Engineering and Design
基金
国家重点研发计划基金项目(2020YFB1600101)
天津市教委科研基金项目(2020KJ024)。
关键词
沙漏网络
人脸对齐
轻量级
知识蒸馏
深度可分离卷积
hourglass network
face alignment
lightweight
knowledge distillation
deep separable convolution