视觉图神经网络的人脸识别方法
Face recognition method based on visual graph neural network
魏永超 1朱泓超 2朱姿翰 3徐未其 2刘伟杰2
作者信息
- 1. 中国民用航空飞行学院学院科研处,德阳 618307;中国民用航空飞行学院民航安全工程学院,德阳 618307
- 2. 中国民用航空飞行学院民航安全工程学院,德阳 618307
- 3. 中国民用航空飞行学院航空电子电气学院,德阳 618307
- 折叠
摘要
目前,大部分人脸识别方法依赖CNN,通过级联融合局部特征实现特征提取,却忽视全局语义空间信息且训练代价巨大.基于Transformer的方法相较于CNN,参数更少且能有效表征全局特征信息,但对全局各特征区域相对空间依赖关系表征不足.针对以上问题,提出了一种视觉图神经网络的人脸识别方法,引入GCN作为特征提取网络,捕获邻近特征关系并建立全局特征区域的依赖性;结合ECA模块,提高模型对人脸特征感知能力.此外,基于Triplet Loss与Center Loss,构建联合损失函数作为目标函数,约束类内特征,提高模型泛化能力.本方法在LFW、CFP和AgeDB-30基准测试集上取得较好的效果,且模型参数量与计算复杂度更少.
Abstract
Currently,most face recognition methods rely on CNNs,which fuse local features through cascading to achieve fea-ture extraction,but ignore global semantic spatial information and are expensive to train.Compared with CNN,Transformer-based methods have fewer parameters and can effectively characterize global feature,but do not sufficiently characterize the relative spa-tial dependencies of each global region.To address these problems,a visual graph neural network approach for face recognition is proposed,introducing GCN as a backbone to capture the feature relationships and establish the dependencies of global feature re-gions;combined with the ECA module to improve the model's ability to perceive face features.In addition,based on Triplet Loss and Center Loss,a joint loss function is constructed as the objective function to constrain the intra-class features and improve the generalization ability of the model.Experimental results on LFW,CFP and AgeDB-30,show that the proposed method can achieve better performance,and the number of parameters and computational complexity are less.
关键词
人脸识别/图神经网络/空间多尺度/注意力机制Key words
face recognition/GNN/spatial multi-scale/attention mechanisms引用本文复制引用
基金项目
西藏科技厅重点研发项目(XZ202101ZY0017G)
四川省科技厅重点研发项目(2022YFG0356)
中国民用航空飞行学院科研项目(J2020-040)
中国民用航空飞行学院科研项目(CJ2020-01)
出版年
2024
NETL
NSTL
万方数据