[1]叶继华,郭祺玥,江爱文,等.基于特征子空间直和的跨年龄人脸识别方法[J].郑州大学学报(工学版),2021,42(05):7-12.
点击复制

基于特征子空间直和的跨年龄人脸识别方法()
分享到:

《郑州大学学报(工学版)》[ISSN:1671-6833/CN:41-1339/T]

卷:
42
期数:
2021年05期
页码:
7-12
栏目:
出版日期:
2021-09-10

文章信息/Info

Title:
Cross-age Face Recognition Method ba<x>sed on Feature Subspace Direct Sum
作者:
叶继华郭祺玥江爱文黎欣
文献标志码:
A
摘要:
近年来,学术界提出了许多判别性方法用于解决跨年龄人脸识别任务,并取得了可观的成果。然而,这些方法都在一定程度上忽视了年龄因子与身份因子之间的相关性。因此,本文在同时进行人脸身份识别和年龄分类这两个任务的多任务卷积神经网络的基础上引入直和模块,提出了一种基于特征子空间直和的多任务卷积神经网络 (Feature Subspace with Direct Sum CNN, FSDS-CNN)。该网络利用两个并行子网分别从卷积单元共享的深度特征中提取出身份相关特征和年龄相关特征,并对这两个相关特征所对应的特征子空间施加直和约束,使得身份相关特征与年龄相关特征尽可能无关。通过多损失的联合监督学习,该网络可以获得随年龄变化鲁棒的年龄无关人脸身份特征。本文在三个公开的基准老化数据集上进行了实验并与近几年的10种具有代表性的方法做了对比,在Morph Album 2数据集中,本文方法在Rank-1识别率(Rank-1 Identification Rate)上结果为98.41%,取得了次优值;在CACD-VS数据集中,本文方法在精确度(Accuracy)上结果为99.2%,取得了次优值,在AUC(Area Under Curve)上结果为99.7%,取得了最优值,比性能第2的模型提高了0.1%;在Cross-Age LFW数据集中,本文方法在等错误率(Equal Error Rate, EER)上结果为10.1%,在错误匹配率为0.1时的错误非匹配率(false non-match rate when false match rate is 10%, FNMR@FMR=0.1)上结果为10.2%,均取得了最优值,比性能第2的模型分别下降了4.7%和11.6%。同时本文在三个数据集上的实验均进行了消融对比实验以验证直和模块的有效性,实验结果证明了直和模块的有效性和优越性。本文提出的FSDS-CNN模型利用直和模块有效降低了身份特征与年龄特征的相关性,能够有效提升跨年龄人脸识别的性能。
Abstract:
The present methods ignore the correlation between age factor and identity factor to some extent. Therefore, we introduce the direct sum module on the basis of the multi-task convolutional neural network that simultaneously performs two tasks of face recognition and age classification, and propose the feature subspace with direct sum multi-task convolutional neural network (FSDS-CNN). The network first extracts the facial deep feature from the input face image through a convolution unit stacked by multiple convolution la<x>yers, batch normalization la<x>yers, activation function la<x>yers and pooling la<x>yers, and then two parallel discriminative subnets, namely the identity subnet and the auxiliary age subnet, extract the identity-related feature and age-related feature from the deep feature shared by the convolution unit to pass to the corresponding discriminator for multi-task classification. At the same time, by leveraging the direct sum module proposed in our paper, the direct sum constraint is applied to the feature subspaces corresponding to these two related features, then the redundant information between the two feature subspaces is eliminated as much as possible, so that the correlation between identity-related feature and age-related feature is decreased as much as possible. Through the joint supervised learning of multiple loss functions, the network can effectively separate the age information from the face identity feature and obtain age-invariant face identity feature that is robust with age. We first conduct explore experiment for the setting of the two important hyperparameters (the number of eigenvectors and the weight of direct sum loss) of the direct sum module on the Morph Album 2 dataset, and finally determine that the number of eigenvectors is 25 and the weight of direct sum loss is 10. After that, we conduct cross-age face recognition and verification experiments on three public -domain benchmark aging datasets, and compare with 10 representative methods in recent years. In the Morph Album 2 dataset, our method achieves the result of 98.41% on the evaluation metric of Rank-1 Identification Rate ( higher is better), which is second along all of the methods in the CACD-VS dataset, our method achieves the result of 99.2% on the evaluation metric of Accuracy ( higher is better), which is the second, and the result on the evaluation metric of AUC (Area Under Curve, higher is better ) is 99.7%, which achieves an optimal value along all of the methods and increases by 0.1% compared with the second model in the Cross-Age LFW dataset, our method achieves the result of 10.1% on the evaluation metric of Equal Error Rate (EER, less is better), and achieves the result of 10.2% on the evaluation metric of the false non-match rate when false match rate is 0.1 (FNMR@FMR=0.1, less is better), all of which achieve the optimal value and decrease by 4.7% and 11.6% respectively compared with the second model. At the same time, the ablation comparison experiments are conducted on the three datasets to verify the effectiveness of the direct sum module and the results have proved the effectiveness and superiority of the direct sum module. Our FSDS-CNN model uses the direct sum module to effectively separate the age information from the identity feature, which reduces the correlation between the identity feature and the age feature, and then effectively improve the performance of cross-age face recognition.
更新日期/Last Update: 2021-10-11