基于深度残差网络的人脸关键点检测  

Facial Keypoints Detection Based on Deep Residual Network

作　　者：李昊璇[1] 王芬 LI Haoxuan;WANG Fen(College of Physics and Electronic Engineering,Shanxi University,Taiyuan 030006,China)

机构地区：[1]山西大学物理电子工程学院

出　　处：《测试技术学报》2019年第6期516-519,546,共5页Journal of Test and Measurement Technology

摘　　要：为了解决神经网络结构加深导致的梯度消失问题,本文提出一种将残差块应用于神经网络中的深度残差网络,使神经网络能更好地应用于人脸关键点检测.该深度残差网络共包含18层,残差块将网络的输出上叠加了y=x的恒等映射层,此方法可以使网络在深度增加的同时不退化.实验使用源自YouTube Faces的数据集进行68个人脸关键点检测,将使用残差块的网络与不使用残差块的网络进行比较,结果显示使用残差块的网络能够更快地收敛,并且提升了关键点检测精度.In order to address the problem of gradient disappearance caused by the deepening of the neural network,this paper proposes a method of applying residual blocks to the neural network,so that the neural network can be better applied to facial keypoints detection.The total number of weighted layers is 18 in this deep residual network and the residual block superimposes the output of the network with an identity mapping,y=x.This method can increase the depth of the network without vanishing gradients.We use a face detection dataset from YouTube Faces to detect 68 facial keypoints and compare the network using the residual block with the unused ones.The results show that the network using the residual block can converge faster and improve the accuracy of k-eypoints detection.

关 键 词：人脸关键点检测 卷积神经网络 残差网络 梯度消失 恒等映射 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]