轻量级Transformer的双向交互近红外手指静脉图像识别  

Bidirectional interactive near infrared finger vein image recognition of a lightweight transformer

作　　者：陶志勇[1] 高亚静 王萌 林森 TAO Zhi-yong;GAO Ya-jing;WANG Meng;LIN Sen(School of Electronic and Information Engineering,Liaoning Technical University,Huludao 125105,Liaoning,China;School of Electronic and Electrical Engineering,Zhengzhou University of Science and Technology,Zhengzhou 450064,China;School of Automation and Electrical Engineering,Shenyang Ligong University,Shenyang 110159,China)

机构地区：[1]辽宁工程技术大学电子与信息工程学院,辽宁葫芦岛125105 [2]郑州科技学院电子与电气工程学院,郑州450064 [3]沈阳理工大学自动化与电气工程学院,沈阳110159

出　　处：《兰州大学学报（自然科学版）》2024年第5期621-628,共8页Journal of Lanzhou University(Natural Sciences)

基　　金：辽宁省教育厅重点攻关项目(LJKZ0349);辽宁省高等学校基本科研项目(LJKMZ20220679)。

摘　　要：针对现有手指静脉识别算法速度慢、复杂度高以及Transformer架构在小数据集上效果不佳的问题,提出轻量级Transformer的双向交互识别方法 .利用轻量级卷积神经网络与改进的Transformer架构组成并行主干网络,用于近红外手指静脉图像的局部和全局特征提取;设计交互结构,在并行结构的基础上,以交互方式融合两条分支上不同尺度的特征.为最大程度地保留近红外图像的局部特征和全局表示,将两条分支提取的信息拼接融合,通过输出层得出识别结果 .结果表明,该算法在多个数据集上的最高识别率可达99.77%,参数量仅1.33 MB.相较于其他指静脉算法,以及改进的Transformer架构,在保持高准确率的同时进一步降低了算法的复杂度.To address the issues of slow recognition speed,high algorithm complexity and poor perfor-mance of transformer architecture on small datasets in existing finger vein recognition algorithms,a light-weight transformer based bidirectional interactive near-infrared finger vein recognition algorithm was proposed,with a parallel backbone network composed of a lightweight convolutional neural network and an improved transformer architecture for local and global feature extraction of near-infrared finger vein images.A up and down structure was designed that integrated features of different scales on two branches in an interactive manner on the basis of a parallel structure.In order to preserve the local fea-tures and global representation of the near-infrared image to the greatest extent possible,the information extracted from the two branches was concatenated and fused,and the recognition results obtained through the output layer.The experimental results showed that the algorithm had a maximum recognition rate of 99.77%on multiple datasets,with a parameter size of only 1.33 MB.Compared to other novel fin-ger vein algorithms and improved transformer architectures,it further reduced the complexity of the algo-rithm while maintaining a high accuracy.

关 键 词：卷积神经网络 指静脉识别 近红外图像 轻量级网络 特征提取 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]