* The Nat iona l Natural Science Foundat ion of China under Grant No. 61462049 (国家自然科学基金).
Received 2019-07-16, Accepted 2019-10-12.
CN KI 网络出版: 2019-11-05, https://kns.cnki.net/KCM S/detail/11.5602.TP.20191104.1815.012 .html
计算机科学与探索
Journal of Frontiers of Computer Sc ienc e and Technology
基于轻量级网络的实时人脸识别算法研究
*
张 典,汪海涛
+
,姜 瑛,陈 星
昆明理工大学 信息工程与自动化学院,昆明 650500
+ 通信作者 E-mail: kmwht@163.c om
摘 要:为了在嵌入式和移动设备上实现高精度的实时人脸识别,对常见的网络在人脸识别方面的优缺点进
行了分析,提出了一种高效的深度卷积神经网络模型 Lightfacenet。在网络中结合深度可分离卷积、逐点卷积、
瓶颈结构和挤压与激励结构提出了轻量化神经网络单元,使网络在保证有一定准确率的情况下有效地解决深
层的神经网络带来的参数冗余和计算量大的问题,再通过改进的非线性激活函数进一步提高网络的准确性。
该神经 网络在保留卷积神经 网络部分优点的 同时也很好地平衡了 网络的缺点。在 同样的实验环境 下,
Lightfacenet 网络既实现了非常高的识别精度,也在模型推理速度上达到实时的效果。在使用 MS- Celeb- 1M 数
据集训练后,该模型在 LFW 数据集上达到了 99.50%的准确率,其效果已经可以与现在的大型卷积神经网络媲
美。对于面部识别,Lightfacenet 比目前最先进的移动卷积神经网络在保证准确率的情况下提高了效率。
关键词:人脸识别;轻量化神经网络单元;实时;非线性激活函数
文献标志码:A 中图分类号:TP391.41
张典, 汪海涛, 姜瑛, 等. 基于轻量级网络的实时人脸识别算法研究[J]. 计算机科学与探索, 2020, 14(2): 317-324.
ZHANG D, WANG H T, JIANG Y, et al. Research on re al-time face reco gnition algorithm based on lightweight net-
wo rk[J]. Journal of Frontiers of Computer Science and Technolog y, 2020, 14(2): 317-324.
Research on Real-Time Face Recognition Algorithm Based on Lightweight Network
*
ZHANG Dian, WANG Haitao
+
, JI ANG Ying, CHEN Xing
Faculty of Information Engineering and Au tomation, Kunming University of Science and Technology, Kunming
650500, China
Abstract: In order to achieve high-precision real-time face recognition on embedded and mobile devices, the advant-
ages and disadvantage s of com mon networks in face recogniti on are analyzed, and an efficient de ep convol ution
neural network model Lightfacenet is proposed. In the network, a lightweight neural network unit is proposed,
wh ich combines the deep separable convolution, point-by-p oint convolution, bottleneck structure and squeeze and
excitat ion structu re. The networ k can effectively so lve the problem of parameter redundancy and large computation
caused by the deep neural ne two rk with a certain accuracy, and then further improve the accuracy of the network
through improved non-linear activation. The neural network not only retains so me advantages of the convoluti ona l
neural network, but also balances the disadvantages o f the network . In the same experimental environ ment, the
Lightfacenet network not o nly achieves very high recognition a ccuracy, but also achieves real- time effect in the
1673-9418/2020/14(2)-0317 -08
doi: 10.3778/j.is sn.1673-9418.1907037