基于中心偏差估计和自适应间隔的人脸识别算法

Face Recognition Based on Center Bias Estimation and Adaptive Margin

损失函数的设计在深度人脸识别中至关重要.常见做法是给所有类别添加固定的间隔项,以修改类别间的决策边界,压缩类内特征间距,提高模型分离不同类别特征的能力.然而,为所有类别添加相同的间隔项可能会忽略人脸识别数据集内类别间的不一致性.为进一步提升模型效果,模型应依据类别的学习难易程度,对不同类别样本特征给予不同程度的关注.文中设计了基于类均值中心与类权重中心之间的偏差挖掘难类的方法,称之为中心偏差估计.本文提出的方法会根据中心偏差估计的程度,为不同类别自适应分配不同大小的间隔项.同时,为解决训练前期中心偏差计算不稳定问题,提出了动态变化的收敛参数,调整中心偏差估计的可信度,开展相关实验验证收敛参数的有效性.在人脸验证基准数据集中,本文提出的方法比基准方法的平均准确率提高了0.26%,达到96.62%.在2个大型人脸验证测试数据集上,在FPR等于0.01%时,提出方法的TPR分数分别提高了0.58%和0.22%,获得88.47%和92.29%的实验结果,且多组实验结果表明提出的方法优于一般现有算法.实现代码参见https://github.com/TCCof-WANG/FR-Centers-Bias.

The design of the loss function is crucial in deep face recognition.A common practice is to add a fixed margin term to all classes to modify the decision boundary between classes,compress the distance between intra-class fea⁃tures,and improve the ability of the model to separate features of different classes.However,adding the same margin term for all classes may ignore the inconsistency between classes in the face recognition dataset.In order to further improve the effectiveness of the model,we argue that the model should pay different attention to the samples of different classes accord⁃ing to the learning difficulty of the class.In this paper,we introduce a method for hard class mining based on the bias be⁃tween the center of the class mean and the center of the class weight,called center bias estimation.The method proposed in this paper adaptively assigns margin terms of different sizes to different classes according to the value of center bias estima⁃tion.At the same time,to solve the problem of unstable calculation of center bias estimation in the early stage of training,we propose an adaptively changing convergence parameter to adjust the credibility of center bias estimation and design rele⁃vant experiments to prove the effectiveness of the convergence parameters.In the face verification baseline dataset,the pro⁃posed method in this paper is improved by 0.26%on average accuracy compared with the baseline method,reaching 96.62%.In two large face verification test datasets,when FPR is equal to 0.01%,the TPR scores of our method is improved by 0.58%and 0.22%,respectively,and the experimental results of 88.47%and 92.29%are obtained,and multiple experi⁃mental results show that our method is better than the general existing algorithms.The implementation code is published on https://github.com/TCCofWANG/FR-Centers-Bias.