Abstract: Abstract: Tubular tree-like veins were regarded as intrinsic structures within finger vein images and had been verified as highly discriminative for personal identification. However, prevailing finger vein recognition methodologies predominantly focused on extracting spatial domain features, while their expression in the frequency domain and the utilization of multi-scale details were often overlooked. A cascaded finger vein recognition network based on a multi-domain feature fusion module (MDFF-FVR) was designed. The core component, the Multi-Domain Feature Fusion (MDFF) module, consisted of four key parts: a depthwise separable wavelet convolution layer that enhanced structural information representation; a frequency–spatial coupling block that extracted global frequency-domain features to compensate for the limitations in spatial representation; an improved semi-wavelet spatial attention block that refined local frequency-domain details; and an adaptive fusion module that integrated the features to produce the final enhanced feature map.Built upon this MDFF module, a cascaded network architecture was further constructed to improve recognition performance. Evaluations on the HKPU and SDUMLA databases demonstrated state-of-the-art performance, achieving recognition rates (RR) of 99.68% and 99.53%, and equal error rates (EER) of 0.35% and 0.41%, respectively. In comparison with existing spatial- or frequency-only methods, the proposed multi-domain feature fusion cascaded network exhibited significant advantages in both recognition accuracy and robustness.The extraction of spatial-domain features was predominantly emphasized in existing finger vein recognition methods, while the frequency-domain representation of tubular tree-like veins and the embedded multi-scale details were often neglected. To overcome this limitation, a cascaded finger vein recognition approach based on multi-domain feature fusion was proposed. A hierarchical cascaded network architecture was designed, in which each stage was composed of a multi-domain feature fusion module and basic operations, enabling progressive feature extraction and enhancement. The multi-domain feature fusion module was employed as the core component. Structural information was first enhanced by a depthwise separable wavelet convolution block. Global frequency-domain features were then extracted through a frequency–spatial coupling block to compensate for insufficient spatial representation, and local frequency-domain details were further refined by an improved semi-wavelet spatial attention block. Finally, multi-domain features were integrated through an adaptive fusion module to generate the enhanced feature representation. Superior performance is demonstrated on the HKPU and SDUMLA databases, achieving recognition accuracies of 99.68% and 99.53%, with equal error rates of 0.35% and 0.41%, respectively. Compared with approaches that relied solely on spatial or frequency features, significant improvements in both recognition accuracy and robustness are achieved by the proposed cascaded network with multi-domain feature fusion.

Key words: finger vein recognition, deep neural network, multi-domain feature fusion, frequency domain analysis, spatial-wavelet attention, cascaded network architecture

摘要： 针对现有手指静脉识别方法主要关注空间域特征的提取，忽略了手指静脉血管作为管状树形结构在频率域的表达及其中蕴含的多尺度细节的问题，提出了基于多域特征融合的级联手指静脉识别方法。该方法采用层次化的级联网络架构，每一层次由多域融合模块与基础操作构成，实现了特征的逐步提取和增强。其中，多域特征融合模块作为核心模块：首先，通过深度可分离小波卷积块来增强结构信息；其次，利用空间–频域耦合块提取频域全局特征，从而弥补空间域表达的不足，并结合改进的半小波–空间注意力块进一步细化频域细节的表征；最后，通过自适应融合模块将多域特征进行整合，得到最终的增强特征表示。在HKPU与SDUMLA数据集上，所提方法识别准确率(Accuracy, ACC)分别达到了99.68%和99.53%，且等错误率(Equal Error Rate, EER)仅为0.35%和0.41%。与现有的仅在空间域或频域提取特征的方法相比，基于多域特征融合级联网络的方法在识别准确性与鲁棒性方面均表现出显著的优势。

关键词: 指静脉识别, 深度神经网络, 多域特征融合, 频率域分析, 空间域-小波注意力, 级联网络架构