Computer Enginee ring a nd Applications 计算机工程与应用
2019,55(24)
1 引言
人脸检测作为人脸识别系统的重要环节,它的作用
是把图像或者视频里的人脸检测出来并框出。被检测
出的人脸用于人脸验证或者人脸识别,所以投入实际运
级联的卷积神经网络人脸检测方法
李亚可
1
,玉振明
1,2
1. 桂林电子科技大学 信息与通信学院,广西 桂林 541004
2. 梧州学院 电子信息工程学院,广西 梧州 543002
摘 要:针对由于光照、分辨率、姿态和表情等因素变化引起的人脸检测准确性不高的问题和大多人脸检测算法使
用单一的卷积神经网络去提取特征引起的算法的泛化能力变弱的问题,提出了三层由浅及深的级联的卷积神经网
络结构。通过全卷积神经网络快速定位人脸候选区域,采用深度神经网络提取人脸鲁棒性特征,对候选区域进一步
分类验证,并用联合回归的方法确定最终人脸位置,提高检测精度。同时通过加权降低得分改进常用的非极大值抑
制的方法,解决了由于相邻人脸高度重叠引起的漏检问题。实验结果表明,该模型对上述引起人脸检测准确率不高
的因素具有较好的鲁棒性,并且在 FD DB 数据集上有着较高的准确率和运行速度。改进后的非极大值抑制算法对
在 FDDB 的测试准确率也有一定的提升。
关键词:人脸检测 ;全卷积网络 ;联合回归
文献标志码:A 中图分类号:TP 391 doi:10.3778/j.issn.1002-8331.1809-0212
李亚可,玉振明 . 级联的卷积神经网络人脸检测方法 . 计算机工程与应用,2019,55(24):184-1 89.
LI Yake, YU Zhenmi ng. Concatenated convolutional neural network face detection method. Computer Engineering and
Applications, 2019, 55(24):184-189.
Concaten ated Convolutio nal Neural Network F ace Detection Method
LI Yak e
1
, YU Zhenming
1,2
1.School of Information and Communication, Guilin University of Electronic Technology, G uilin, Guangxi 54100 4, C hina
2.School of Electronics and Information Engineering, Wuzh ou University, Wuzhou, Guangxi 5430 02, Chin a
Ab stract:Aiming at the problem of low face detection accuracy cau sed by changes in lighting, low resolution, posture
and expression, and the generalization of algorithms caused by most face detection algorithms using a single convolutional
neural network to extrac t featur es, a three-layer convolutional neural network structure consisting of shallow and deep cas-
cade is proposed. The fac e candidate region is q uickly locat ed by the full convolutional neural network. Then the depth
neural network is used to extract the face robust ness feature, and the candidate region is further classified and verified.
The joint regression face method is used to determine the final face position and improv e the detection accuracy. At the
same time, the commonly use d non-maxim um value suppression method is improved by weighting the reduction score,
and the mi ssed detection problem caused by the overla pping of adjacent faces is solved. The experimental results show
that th e model is robust to the above-mentioned fac tors that cause low face detection accuracy, and it has high accur acy
and r unning speed in FDDB dataset. The improved non-maximum suppres sion al gorithm also has a certain improvement
on the test accurac y of FDDB.
Key w ords:face detection; full convolutional network; join t regression
基金项目:广西重点研发计划(No. 桂科 AB16380273);国家自然科学基金(No.61562074)。
作者简介:李亚可(1990—),男,硕士生,主要研究领域为计算机视觉、机器学习、深度学习,E-mail:473243180@qq.com;玉振明
(1963—),男,博士,教授,主要研究领域为信号处理、视频大数据。
收稿日期:2018-09-17 修回日期:2018-11-30 文章编号:10 02-8331(2019)24-0184-06
CN KI网络出版:2019-01-24, http://kns.cnki.net/kcms/de tail/11.2127.tp.20190122.1701.009.html
184
评论0
最新资源