摘要
现有的大多数人脸识别算法均采用深度学习中各种改进的卷积神经网络算法。但算法存在参数多,训练时间长等问题。因此,为了减少训练过程所消耗的时间和分类过程中的计算量,设计运用了改进的MobileNet算法来实现人脸识别,并将其移植到Jetson nano设备上构成完整的室外安防系统。将MobileNet模型中原本的Soft Max分类器进行了改进,通过对比实验,发现使用A-SoftMax分类器的效果要好于Soft Max。实验结果表明:本文提出的模型在LFW人脸数据库上达到97.4%的准确率,计算时间减少为传统卷积神经网络的1/9,计算参数减少为传统卷积神经网络的1/7。
Most existing face recognition algorithms adopt various improved convolutional neural network(CNN)algorithms in Deep Learning.But the algorithm has many problems,such as many parameters and long training time.Therefore,in order to reduce the time consumed in the training process and the amount of calculation in the classification process,an improved MobileNet algorithm is designed to realize face recognition,and transplants it to Jetson nano equipment to form a complete outdoor security system.The original SoftMax classifier in MobileNet model is improved.Through comparative experiments,it is found that the effect of using A-SoftMax classifier is better than that of SoftMax.The experimental results show that the proposed model achieves 97.4%accuracy on LFW face database,reduces computing time to 1/9 of traditional CNN,and reduces computing parameters to 1/7 of traditional CNN.
作者
胡佳玲
施一萍
谢思雅
陈藩
刘瑾
HU Jialing;SHI Yiping;XIE Siya;CHEN Fan;LIU Jin(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)
出处
《传感器与微系统》
CSCD
北大核心
2021年第3期102-105,共4页
Transducer and Microsystem Technologies
基金
国家自然科学基金资助项目(61701296)
上海工程技术大学学科建设项目(19KY0229)。