«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

26

期数:

2016年07期

页码:

11-13

栏目:

智能、算法、系统工程

出版日期:

2016-07-10

文章信息/Info

Title:

 Face Pose Classification Method Based on Deep Learning

文章编号:

1673-629X（2016)07-0011-03

作者:

 邓宗平; 赵启军; 陈虎

 四川大学 计算机学院 视觉合成图形图像技术国防重点学科实验室

Author(s):

 DEND Zong-ping; ZHAO Qi-jun; CHEN Hu

关键词:

 ; 姿态分类; 级联; 深度学习 ; 卷积神经网络

Keywords:

 ; pose classification; cascade; deep learning; convolutional neural network

分类号:

TP301

文献标志码:

A

摘要:

 人脸姿态通常表达着有用的信息，准确地把握人脸的姿态，往往在人脸对齐、人类行为分析以及司机疲劳驾驶监控等方面有着重要的作用。文中方法与以往姿态估计方法不一样，是一种基于卷积神经网络，应用深度学习做人脸姿态分类的方法。首先，第一次网络对姿态在yaw方向上进行5分类，同时在roll方向具有鲁棒性。之后，将第一次输出正脸的结果进入第二次网络，对姿态在pitch方向进行3分类。所有的输出结果对光照都具有鲁棒性。文中采用级联的方法在公开库上做测试，准确率高达95%以上。在实际监控视频中，姿态估计不仅有较高的准确率，而且有惊人的速度。由于本身实验的设计独特性，只做了自身实验对比。结果充分展示了用合理的神经网络与网络级联的方法在姿态估计上面的发展潜力。

Abstract:

 Face pose usually contains useful information,so detecting it accurately plays an important role in face alignment,human be-havior analysis and drivers’ fatigue driving monitoring. A novel method is proposed in this paper which applies deep learning to human face pose classification based on convolutional neural networks. It can be divided into two steps mainly. First,layer one classifies pose into 5 categories at direction yaw,and it’s of robustness at direction roll. Then layer two takes the result of step one as input to classify pose into 3 categories at direction pitch. All outputs are robust to illumination. The cascade connection is used to test on public benchmark,and the result shows that its accuracy is 95%. In real surveillance video,it has both high accuracy and fast estimating speed. Due to the partic-ularity of experiment,it only contrasts the result to itself. Experimental results show that well-designed cascade connection of neural net-work can estimate pose well.

常用功能

更新日期/Last Update: 2016-09-28