摘要
为了提高人脸识别在嵌入式设备中的识别精度和速度,提出一种基于Ghostnet轻量级人脸识别算法—Ghostfacenet。首先,通过预设卷积生成固定数目的内在特征;针对卷积运算计算消耗大的问题,使用计算成本低廉的线性操作代替卷积运算,产生一系列与内在特征相关联的特征信息;其次,基于Ghostnet中的Ghost模块以及深度可分离卷积设计出Ghostfacenet-Bottleneck,并且由其构建出Ghostfacenet轻量级卷积神经网络;最后,联合Softmax损失函数和Arcface损失函数进一步增加人脸类内紧凑性以及类间差异,同时使得轻量级模型有更好的收敛性以及泛化能力。实验结果表明,Ghostfacenet在嵌入式设备的识别速度分别是Resnet50、Efficientnet、MobilenetV2和Mobilefacenet的11.08倍、8.57倍、2.75倍和2.82倍。在不显著降低识别性能同时能够显著提高运行效率,非常适用于资源有限的嵌入式设备中。
In order to improve the recognition accuracy and recognition speed of face recognition in embedded devices,a Ghostnet-based lightweight face recognition algorithm called Ghostfacenet is proposed.Firstly,a fixed number of intrinsic features are generated by pre-determined convolution.To address the problem of computationally intensive convolutional operations,linear operations with low computational cost are used instead of convolutional operations to generate a series of feature information associated with intrinsic features.Secondly,the Ghostfacenet-Bottleneck is designed based on the Ghost module in Ghostnet and the depthwise separable convolution.And the Ghostfacenet lightweight convolutional neural network is constructed from Ghostfacenet-Bottleneck.Finally,the Arcface loss function and the Airface loss function are combined to further increase the intra-class compactness of faces as well as inter-class differences.It also allows for better convergence and generalization capabilities of lightweight models.The experimental results show that Ghostfacenet is 11.08 times,8.57 times,2.75 times and 2.82 times faster than Resnet50,Efficientnet,MobilenetV2 and Mobilefacenet respectively in embedded devices.This is a significant improvement in operational efficiency without a significant reduction in recognition performance and is ideal for embedded devices with limited resources.
作者
赵锋
张鹏
张冉
Zhao Feng;Zhang Peng;Zhang Ran(School of Instrument and Electronic,North University of China,Taiyuan 030051,China;School of Date Science and Technology,North University of China,Taiyuan 030051,China)
出处
《电子测量技术》
北大核心
2022年第16期130-136,共7页
Electronic Measurement Technology
基金
技术领域基金(2021-JCJQ-JJ-0726)项目资助。
关键词
深度学习
损失函数
轻量级卷积神经网络
人脸识别
嵌入式系统
deep learning
loss function
lightweight convolutional neural networks
face recognition
embedded system
