基于深度学习的人脸识别系统设计与实现  

Design and implementation of face recognition system based on deep learning

作　　者：陈帅男 贾开吉 侯福鸽 CHEN Shuainan;JIA Kaiji;HOU Fuge(Hebei University of Science and Technology,Shijiazhuang 050018,China;No.95911 Unit,the PLA,Jiuquan 735018,China)

机构地区：[1]河北科技大学,石家庄050018 [2]95911部队,甘肃酒泉735018

出　　处：《空天预警研究学报》2024年第3期189-195,共7页JOURNAL OF AIR & SPACE EARLY WARNING RESEARCH

摘　　要：为了解决传统人脸识别方法在准确性和鲁棒性方面的不足,研究设计并实现了一种新型基于深度学习的人脸识别系统.该系统由人脸检测、预处理、特征表征和识别四个主要模块构成.采用深度卷积神经网络结合FaceNet模型和DRNet结构,优化了特征提取和空间嵌入;通过三元组损失函数,提高了识别的精度.实验结果表明,在公认的LFW数据集上,系统达到了99.71%的平均识别准确率.此外,通过量化和剪枝技术,系统在保持实时高准确性处理的同时,实现了较低的计算复杂度和内存需求.In order to address the shortcomings of traditional face recognition methods in terms of accuracy and robustness,a novel type of deep learning-based face recognition system is designed and implemented in this study.The system consists of four main modules:face detection,preprocessing,characterization and recognition.The deep convolutional neural network used combines FaceNet model and DRNet structure to optimize feature extraction and spatial embedding.The recognition accuracy is improved by using ternary loss function.Experimental results show that the system achieves 99.71%average recognition accuracy on the recognized LFW dataset.In addition,through quantification and pruning techniques,the system achieves low computational complexity and memory requirements while maintaining real-time high accuracy processing.

关 键 词：深度学习 人脸识别 人脸检测 FaceNet模型 DRNet架构 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]