基金项目:国家自然科学基金资助项目(62106206)
通信作者 E-mail:wangmw@swjtu.edu.cn
基于双分支深度图卷积网络的指静脉识别研究
程俊军,王明文
(西南交通大学,数学学院,四川成都 610000)
摘要:基于图卷积神经网络的手指静脉识别方法不仅可以解决传统指静脉识别方法识别率较低的问题,还
可以解决其计算量大的问题。针对目前指静脉图模型在图结构不稳定性和匹配效率因模型增大而下降的问
题,采用了 SLIC 超像素分割算法来构建加权图并改变了图卷积神经网络提取加权图的图级特征。为了有
效抓取图数据中的高阶特征并避免过平滑,研究了一种双分支多交互的深度图卷积网络,旨在提升节点对
高阶特征的掌握能力。该方法首先根据节点特征对图结构进行调整,然后通过结合原始和重构后的图结构,
构建了双分支网络架构以充分挖掘高阶特征。此外,通过设计一种通道信息互动机制,以促进不同分支间
的信息交流,从而提高特征的多样性。实验结果显示,该网络结构在多个标准数据集上进行指静脉识别任
务时,能提高识别精度,减少单张图片识别时间,提高效率,并有效减轻过平滑现象。
关键词:指静脉识别;图像分割算法;图卷积神经网络;交叉熵函数;通道信息交互
Finger-vein Recognition Research Based on Deep Graph
Convolutional Network With Dual-Branch
CHENG Junjun,WANG Mingwen
(School of Mathematics,Southwest Jiaotong University,Chengdu 610000,Si Chuan,China)
[Abstract] The finger-vein recognition method based on graph convolutional neural network can not only solve the
problem of low recognition rate of traditional finger vein recognition method, but also solve the problem of its large
computational volume. To address the current finger vein graph model's problems of graph structure instability and
matching efficiency decreasing due to model increase, SLIC super-pixel segmentation algorithm is used to construct
the weighted graph and change the graph convolutional neural network to extract the graph-level features of the
weighted graph. In order to effectively capture the higher-order features in the graph data and avoid over-smoothing,
a two-branch multi-interaction deep graph convolutional network is investigated, aiming to improve the node's
ability to grasp the higher-order features. The method first adapts the graph structure according to the node features,
and then by combining the original and reconstructed graph structures, a two-branch network architecture is
constructed to fully mine the higher-order features. In addition, the diversity of features is improved by designing a
channel information interaction mechanism to facilitate information exchange between different branches.
Experimental results show that this network architecture can improve recognition accuracy, reduce single-image
recognition time, improve efficiency, and effectively mitigate oversmoothing when performing finger vein
recognition tasks on multiple standard datasets.
[keywords]Finger-vein Recognition;Image Segmentation Algorithm;Graph Convolutional Neural Network;Cross-
entropy function;Channel Information Interaction
0 概述