基于样本均衡蒸馏的少样本人脸识别方法  

Face recognition method with few samples based on sample balance distillation

作　　者：钟锐[1,2] 王晨 宋亚锋 廖华江[3] ZHONG Rui;WANG Chen;SONG Ya-feng;LIAO Hua-jiang(School of Mathematics and Computer Science,Gannan Normal University,Ganzhou 341000,China;Jiangxi Provincial Key Laboratory of Numerical Simulation and Simulation Technology,Gannan Normal University,Ganzhou 341000,China;Information and Education Technology Center,Gannan Normal University,Ganzhou 341000,China)

机构地区：[1]赣南师范大学数学与计算机科学学院,江西赣州341000 [2]赣南师范大学江西省数值模拟与仿真技术重点实验室,江西赣州341000 [3]赣南师范大学信息与教育技术中心,江西赣州341000

出　　处：《计算机工程与设计》2024年第11期3457-3462,共6页Computer Engineering and Design

基　　金：国家自然科学基金项目(62266003);江西省自然科学基金项目(20232BAB202056);江西省教育厅科技基金项目(GJJ180771、GJJ211401、GJJ161676);江西省基础教育研究课题基金项目(SZUGSZH2021-1147)。

摘　　要：为解决实际应用场景中人脸识别模型复杂度高和训练样本数量不足的问题,提出一种基于样本均衡蒸馏(sample balance distillation,SBD)的轻量高效人脸识别方法。通过构造样本均衡Focalloss损失函数解决训练样本数量不足的问题,该损失函数能够在模型训练过程中增加稀少样本的权重,使模型能够对稀少样本进行准确分类,将蒸馏损失函数与样本均衡Focalloss损失函数进行加权融合,将教师网络强大的特征表达能力迁移至学生网络中,达到提高模型训练效率和分类精度的目标。为验证模型有效性,在多个样本数量分布不均衡数据集中进行大量实验,其结果表明,所提模型的训练时间和识别效率得到了显著提高,具有较高识别率。To address the problems of the high complexity of facial recognition models and the insufficient number of training samples in practical application scenarios,a lightweight and efficient facial recognition method based on sample balance distillation(SBD)was proposed.The problem of insufficient training samples was solved by constructing a sample-balanced Focalloss loss function.This loss function increased the weight of rare samples in model training,so that the model accurately classified rare samples.The distillation loss function and the sample-balance Focalloss loss function were weighted and fused to transfer the powerful feature expression ability of the teacher network to the student network to achieve the goal of improving the model training efficiency and classification accuracy.To verify the model’s effectiveness,many experiments were carried out on multiple datasets with imbalanced sample size distribution.Results show that the proposed model significantly improves the training time and recognition efficiency and has a high recognition rate.

关 键 词：人脸识别 知识蒸馏 样本均衡蒸馏 轻量高效 深度神经网络 少样本 Focalloss 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]