基于多层次特征融合的Transformer人脸识别方法
Transformer face recognition method based on multi-level feature fusion
夏桂书 1朱姿翰 1魏永超 2朱泓超 3徐未其3
作者信息
- 1. 中国民用航空飞行学院航空电子电气学院,德阳 618307
- 2. 中国民用航空飞行学院科研处,德阳 618307
- 3. 中国民用航空飞行学院民航安全工程学院,德阳 618307
- 折叠
摘要
卷积神经网络中的卷积操作只能捕获局部信息,而Transformer能保留更多的空间信息且能建立图像的长距离连接.在视觉领域的应用中,Transformer缺乏灵活的图像尺寸及特征尺度适应能力,通过利用层级式网络增强不同尺度建模的灵活性,且引入多尺度特征融合模块丰富特征信息.本文提出了一种基于改进的Swin Transformer人脸模型——Swin Face模型.Swin Face以Swin Transformer为骨干网络,引入多层次特征融合模块,增强了模型对人脸的特征表达能力,并使用联合损失函数优化策略设计人脸识别分类器,实现人脸识别.实验结果表明,与多种人脸识别方法相比,Swin Face模型通过使用分级特征融合网络,在LFW、CALFW、AgeDB-30、CFP数据集上均取得最优的效果,验证了此模型具有良好的泛化性和鲁棒性.
Abstract
The convolutional operation in a convolutional neural network only captures local information,whereas the Transformer retains more spatial information and can create long-range connections of ima-ges.In the application of vision field,Transformer lacks flexible image size and feature scale adaptation capability.To solve this problems,the flexibility of modeling at different scales is enhanced by using hi-erarchical networks,and a multi-scale feature fusion module is introduced to enrich feature information.This paper propose an improved Swin Face model based on the Swin Transformer model.The model u-ses the Swin Transformer as the backbone network and a multi-level feature fusion model is introduced to enhance the feature representation capability of the Swin Face model for human faces.a joint loss function optimisation strategy is used to design a face recognition classifier to realize face recognition.The experimental results show that,compared with various face recognition methods,the Swin Face recognition method achieves best results on LFW,CALFW,AgeDB-30,and CFP datasets by using a hi-erarchical feature fusion network,and also has good generalization and robustness.
关键词
人脸识别/Transformer/多尺度特征/特征融合Key words
Face recognition/Transformer/Multi-scale features/Feature fusion引用本文复制引用
基金项目
西藏科技厅重点研发计划(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基金(J2020-126)
中国民用航空飞行学院科研基金(J2020-040)
中国民用航空飞行学院科研基金(J2021-056)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据