摘要
在实际应用场景中,低分辨率、高噪声的人脸图像普遍存在,这严重制约了传统人脸识别方法在识别准确率和鲁棒性方面的表现。针对这一问题,文章提出了一种基于全局-局部特征融合的低分辨率人脸识别方法——Global-Local Transformer(GLFormer)。该方法结合了Vision Transformer模块与Swin Transformer模块,分别用于提取图像的全局特征与局部细节特征。其中,Vision Transformer模块擅长建模图像中的长距离依赖关系,提升整体语义理解能力;Swin Transformer模块则增强了对局部纹理和边缘信息的感知能力,提高了细节表达效果。在此基础上,引入特征融合模块,实现全局与局部信息的有效整合,充分发挥2类特征的互补优势。实验结果表明,与传统卷积神经网络及单一Transformer架构相比,GLFormer在低分辨率、高噪声条件下的人脸识别准确率和鲁棒性均有显著提升,验证了所提方法的有效性与优越性。
In practical application scenarios,low resolution and high noise facial images are commonly present,which seriously restricts the performance of traditional facial recognition methods in terms of recognition accuracy and robustness.To address this issue,the article proposes a low resolution face recognition method based on global local feature fusion—Global Local Transformer(GLFormer).This method combines the Vision Transformer module and Swin Transformer module,which are used to extract global features and local detail features of images,respectively.Among them,the Vision Transformer module excels in modeling long-range dependencies in images,enhancing overall semantic understanding capabilities.The Swin Transformer module enhances the perception of local texture and edge information,improving the expression of details.On this basis,a feature fusion module is introduced to effectively integrate global and local information,fully leveraging the complementary advantages of the two types of features.The experimental results show that compared with traditional convolutional neural networks and single Transformer architectures,GLFormer significantly improves the accuracy and robustness of face recognition under low resolution and high noise conditions,verifying the effectiveness and superiority of the proposed method.
作者
李家瑞
陶丽
刘少烔
LI Jiarui;TAO Li;LIU Shaotong(Yingkou Institute of Technology,Yingkou,Liaoning 115014,China)
基金
营口理工学院校级科研项目:面向恶劣成像环境的鲁棒视觉智能感知技术研究(QNL202417)
营口市自动化工业控制技术创新中心项目第10号。
