基于MobieNetV2轻量级人脸识别算法  

Lightweight face recognition algorithm based on MobieNetV2

作　　者：蔡啟军 彭程[1,2] 石向文 CAI Qijun;PENG Cheng;SHI Xiangwen(Chengdu Institute of Computer Application,Chinese Academy of Sciences,Chengdu Sichuan 610041,China;School of Computer and Control,University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院成都计算机应用研究所,成都610041 [2]中国科学院大学计算机与控制学院,北京100049

出　　处：《计算机应用》2020年第S01期65-68,共4页journal of Computer Applications

基　　金：四川省科技厅项目(2018GZ0231)。

摘　　要：针对射频识别(RFID)会议报到系统出现漏签、替签等问题,采用一种人脸识别算法辅助报到系统解决上述人证不匹配的问题。由于会议报到系统硬件设备的内存有限,需要一个轻量级的网络模型,所以考虑将MobileNetV2网络作为人脸识别算法的研究对象。由于报到系统对实时性和准确率要求比较高,所以需要提高该算法的识别速度和识别准确率,于是对MobileNetV2网络进行改进。为了进一步提高人脸识别的泛化能力和算法的收敛速度,采用自适应缩放余弦损失函数AdaCos替换Arcface损失函数进行网络训练。实验结果表明,所设计的轻量级人脸识别算法在识别准确率、识别速度、算法模型大小等方面取得了很好的效果,该算法在不失精度的前提下,完全可以满足报到系统的需求。In view of the problems of missing signatures and signing for someone else in the Radio Frequency IDentification(RFID)conference registration system,a face recognition algorithm was used to assist the registration system to solve the above-mentioned problem of mismatching of person and identification. Due to the limited memory of the hardware equipment of the conference registration system,a lightweight network model was needed,so that the MobileNetV2 network was considered as the research object of the face recognition algorithm. Since the high real-time performance and accuracy required by the registration system,it is necessary to improve the recognition speed and accuracy of the algorithm,so the MobileNetV2 network was improved. In order to further improve the generalization ability of face recognition and the convergence speed of algorithm,the adaptive scaling cosine loss function AdaCos was used to replace the Arcface loss function to perform network training. Experimental results show that the designed lightweight face recognition algorithm achieves good results in terms of recognition accuracy,recognition speed and algorithm model size. The designed algorithm can fully meet the requirements of registration system without losing accuracy.

关 键 词：图像处理 深度学习 卷积神经网络 人脸检测 人脸识别 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP389.1[自动化与计算机技术—控制科学与工程]