摘要
针对现有基于深度学习的人脸对齐算法难以实现真正意义上“端对端”、浅层特征表征能力及鲁棒性差的问题,提出一种基于深度卷积神经网络的端对端人脸对齐算法。该算法网络含5个模块,使用堆叠卷积层提取人脸特征,通过增加隐含层的宽度达到丰富特征的目的,在前3个模块引入监督约束层提取更有效的人脸特征,采用两次迭代训练过程获取更佳的网络模型,该网络能够学习高维空间特征并预测人脸特征点的坐标位置。实验结果表明,即使在遮挡、复杂姿势、光照等情况下,该算法依然可以取得较优的人脸对齐效果。
For the existing face alignment algorithms based on deep learning,it is difficult to achieve an end-to-end training in the real sense,the shallow network has limited capacity to represent features and is poor in robustness.To solve these problems,an end-to-end face alignment method based on deep convolution neutral network was proposed.The network of algorithm contained five modules and stacked convolutional layer was used to extract features.The width in the hidden layers was increased to achieve rich features.Supervisory signal was added to the first three modules to get more effective features,two iterative procedures were designed to get a better network model.The network can learn global high-level features and directly predict the coordinates of facial landmarks.Extensive experiments demonstrate that the propsed method possesses superior capability of handling occlusions and complex variations of pose,illumination and so on.
作者
杨晓青
莫建文
YANG Xiao-qing;MO Jian-wen(School of Information and Communication,Guilin University of Electronic Technology,Guilin 541004,China)
出处
《计算机工程与设计》
北大核心
2019年第9期2666-2671,2717,共7页
Computer Engineering and Design
基金
国家自然科学基金项目(61661017、61362021)
中国博士后科学基金项目(2016M602923XB)
广西自然科学基金项目(2014GXNSFDA118035、2016GXNSFAA380149)
桂林电子科技大学研究生教育创新计划基金项目(2016YJCXB02)
广西科技创新能力与条件建设计划基金项目(桂科能1598025-21)
桂林科技开发基金项目(20150103-6)
关键词
深度卷积神经网络
特征提取
监督信号
端对端
人脸对齐
deep convolution neural network
feature extraction
supervisory signal
end-to-end
face alignment