基于卷积神经网络的胡须及帽子特征检测  

Beard and hat detection based on CNN

作　　者：顾旸铖 邱卫东[1] 

机构地区：[1]上海交通大学电子信息与电气工程学院,上海200240

出　　处：《信息技术》2017年第9期121-124,129,共5页Information Technology

摘　　要：随着安防监控、视频行为分析等领域的发展,人脸检测技术已经从基础的检测出人脸进一步发展到了获取人脸的相关特征。文中给出了一个基于卷积神经网络的检测人脸胡须与帽子特征的方法。首先利用主动形状模型(ASM)标记人脸关键点,之后进行仿射变换对齐人脸,输入卷积神经网络分类器。大量的仿真实验结果表明,卷积神经网络较基于传统特征模型的支持向量机(SVM)分类器更能够准确地区分出胡须与帽子两个关键的人脸特征,运算速度快,对于稍有模糊的人脸也具有一定的鲁棒性。With the development of security monitoring and human behavior analysis,face recognition technology has expanded to facial attribution algorithm on the basis of face detection algorithm. This paper proposes a solution using deep neural networks to detect beard and hat on human faces. First,it used active shape model（ ASM） to detect key points on human faces,then carried out affine transform to have the faces aligned. It used convolutional neural network models to distinguish the aligned faces. The massive experiments show that deep neural networks work better on distinguishing key facial attributions like beard and hat than traditional feature models with SVM. Convolutional neural network runs fast and shows robustness on faces with slight illegibility.

关 键 词：人脸特征检测 卷积神经网络 传统特征模型 主动特征模型 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]