基于注意力机制的轻量级双卷积手指静脉识别网络

赵冰艳; 梁义怀; 张中霞; 张文政

doi:10.11999/JEIT250380

基于注意力机制的轻量级双卷积手指静脉识别网络

doi: 10.11999/JEIT250380 cstr: 32379.14.JEIT250380

赵冰艳¹,
梁义怀^1, ,,
张中霞²,
张文政^{2, 3}

1.
西南交通大学信息科学与技术学院成都 611756
2.
中国电子科技网络信息安全有限公司成都 610041
3.
保密通信全国重点研究所成都 610041

基金项目: 中央高校基本科研业务费(2682024CX031)

详细信息

作者简介:
赵冰艳：女，硕士生，研究方向为手指静脉图像识别

梁义怀：男，讲师，研究方向为数据安全与隐私保护、群智感知、联邦学习

张中霞：女，博士，研究方向为信息安全、图像处理

张文政：男，高级工程师，研究方向为密码学理论和技术

通讯作者:
梁义怀　liangyh@swjtu.edu.cn

中图分类号: TP389
计量
- 文章访问数: 57
- HTML全文浏览量: 30
- PDF下载量: 2
- 被引次数: 0
出版历程
- 修回日期: 2026-01-06
- 录用日期: 2026-01-06
- 网络出版日期: 2026-01-15

Lightweight Dual Convolutional Finger Vein Recognition Network Based on Attention Mechanism

1.
School of Information Science and Technology, Southwest Jiaotong University, Chengdu 611756, China
2.
China Electronis Technology Cyber Security Co., Lt, Chengdu 610041, China
3.
Science and Technology on Communication Security Laboratory, Chengdu 610041, China

Funds: The Fundamental Research Funds for the Central Universities(2682024CX031)

摘要

摘要: 基于深度学习的指静脉识别方法已广泛应用于生物特征识别领域，然而现有模型普遍存在复杂度与分类性能失衡的问题，难以在内存受限和计算资源稀缺环境下高效完成识别任务。针对上述问题，该文提出了一种融合注意力机制的轻量化双通道卷积神经网络模型。此模型设计有双分支协同架构，旨在分别提取核心特征与辅助特征，从而丰富特征集合并增强网络对远程依赖特征的捕捉能力。通过设计一种并行双重注意力机制，以促进融合特征间的信息交互，引导模型聚焦于高价值信息，学习更具区分度的特征表示。实验结果显示，此模型在USM、HKPU和SDUMLA三个公开数据集上的识别准确率分别达到99.70%、98.33%和98.27%，比现有先进方法分别提升2.34%、1.79%和2.03%，而参数量减少11.35%-60.19%，表明提出的双卷积模型实现了网络规模与识别准确率之间的有效平衡。
- 指静脉识别 /
- 双卷积模型 /
- 特征表示 /
- 轻量化 /
- 并行双重注意力机制
Abstract: Objective Finger vein recognition, an emerging biometric authentication technology, has garnered considerable research attention due to its unique physiological characteristics and advantages in in vivo detection. However, the existing mainstream recognition frameworks based on deep learning still face significant challenges: on the one hand, high-precision recognition often relies on complex network structures, resulting in a sharp increase in model parameters, which makes deployment difficult in memory-constrained embedded devices and computing resource-scarce edge scenarios; On the other hand, although model compression technology can reduce the computational cost, it is often accompanied by the attenuation of feature expression ability, resulting in the inherent contradiction between recognition accuracy and efficiency. To address these challenges, a lightweight double convolutional model integrating an attention mechanism is proposed. By designing a parallel heterogeneous convolution module and an attention guidance mechanism, diversified features of images are deeply mined, and recognition accuracy is finally improved while the lightweight characteristics of the model are maintained. Methods The proposed network architecture adopts a three-level collaborative mechanism involving "feature extraction, dynamic calibration, and decision fusion". First, a dual convolution feature extraction module is constructed based on normalized ROI images. This module employs a strategy combining heterogeneous convolutional kernels, where rectangular convolutional branches with varying shapes and sizes are utilized to capture vein topological structure characteristics and track vein diameter directions; meanwhile, square convolution branches employ stacked square convolutions to extract local texture details and background intensity distribution characteristics. These two branches operate in parallel with a reduced channel count, forming complementary feature responses through kernel shape differences, which compresses parameter quantities while enhancing feature representation differentiation. Secondly, a parallel dual attention mechanism is designed to achieve two-dimensional calibration through joint optimization of channel attention and spatial attention. Channel dimensions adaptively assign weights to strengthen key discriminative features of vein textures; spatial dimensions construct pixel-level dependency models and dynamically focus on effective discriminative regions. This mechanism adopts a parallel feature concatenation fusion strategy, preserving input structural information while avoiding additional parameter burdens and improving model sensitivity to critical features. Finally, a three-level progressive feature optimization structure is constructed. Initially, multi-scale receptive field nesting is implemented through a convolutional compression module with a stride of 2, gradually purifying primary features during dimensionality reduction; subsequently, dual fully connected layers are employed for feature space transformation. The first layer utilizes ReLU activation to construct sparse feature representations, while the final layer employs Softmax for probability calibration. This structure effectively balances the risks of shallow underfitting and deep overfitting while maintaining forward inference efficiency. Results and Discussions The effectiveness and robustness of the proposed network are verified on three publicly available datasets: USM, HKPU and SDUMLA. The Acc metric is employed to evaluate detection accuracy. Experimental results on network recognition performance (Table 1) demonstrate that the proposed method achieves favorable outcomes in finger vein image recognition tasks. The feature visualization heatmaps (Fig. 4, Fig. 6) prove that the model can extract complete vein discrimination features. Visualization results (Fig. 7, Fig. 8) indicate that model loss and accuracy exhibit normal trends during training, with 100% classification performance achieved, thereby validating the reliability and robustness of the proposed approach. Quantitative comparisons (Tables 2 and 3) reveal that the proposed method effectively addresses the imbalance between model complexity and classification performance, demonstrating superior performance across three datasets. Furthermore, ablation studies (Table 4) confirm the efficacy of the proposed module, showing significant improvements in finger vein image recognition performance. Conclusions This paper proposes a lightweight dual-channel convolutional neural network architecture incorporating an attention mechanism, comprising three core innovative modules: a dual-convolution feature extraction module, a parallel dual-attention module, and a feature optimization classification module. During feature extraction, long-range venous features and background information are collaboratively encoded through a low-channel parallel architecture, significantly reducing parameter quantities while enhancing inter-individual discriminability. The attention module efficiently captures critical venous features, maintaining feature sensitivity while overcoming the parameter expansion bottleneck of traditional attention mechanisms. The feature optimization classification module employs a progressive feature recalibration mechanism, effectively mitigating the conflicts between underfitting and overfitting during stacked dimensionality reduction. Quantitative experiments demonstrate that the proposed method achieves recognition accuracies of 99.70%, 98.33% and 98.27% on the USM, HKPU and SDUMLA datasets respectively, representing an average improvement of 2.05% compared to existing state-of-the-art methods. When benchmarked against comparable lightweight finger vein recognition approaches, the proposed method reduces parameter scale by 11.35%-60.19%, successfully balancing model lightening and performance enhancement.
- Finger vein recognition /
- Double convolutional model /
- Feature representation /
- Lightweight /
- Parallel dual attention mechanism

HTML全文

图 1 手指静脉识别流程图和现有方法的问题

下载: 全尺寸图片幻灯片

图 2 ALDCNet算法整体结构

下载: 全尺寸图片幻灯片

图 3 DCFE模块结构

下载: 全尺寸图片幻灯片

图 4 矩形卷积和方形卷积作用下输出特征的热图

下载: 全尺寸图片幻灯片

图 5 PDA模块结构

下载: 全尺寸图片幻灯片

图 6 Grad-CAM热图

下载: 全尺寸图片幻灯片

图 7 两个公共数据集的损失曲线和准确率曲线

下载: 全尺寸图片幻灯片

图 8 混淆矩阵图

下载: 全尺寸图片幻灯片

表 1 不同卷积核组合在三个数据集上的识别准确率(%)

矩形卷积核	方形卷积核(USM)			方形卷积核(HKPU)			方形卷积核(SDUMLA)
矩形卷积核	5×5	7×7	9×9	5×5	7×7	9×9	5×5	7×7	9×9
1×3	99.49	99.59	99.39	97.14	97.62	98.33	97.64	97.48	97.80
1×5	99.70	99.39	99.49	97.38	97.62	98.10	97.32	97.16	98.27
1×7	99.29	99.29	98.78	97.86	97.86	97.62	97.48	97.64	97.80
注：粗体表示所有方案中性能指标的最高值

下载: 导出CSV

表 2 与其他方法性能指标的对比(%)

方法	Acc(USM)	EER(USM)	Acc(HKPU)	EER(HKPU)	Acc(SDUMLA)	EER(SDUMLA)
AlexNet^[12]	97.26	0.10	86.90	3.99	92.45	1.10
VGG16^[13]	95.93	0.31	76.67	4.78	78.14	1.73
Liu等人^[14]	97.97	0.20	77.14	4.24	69.65	4.56
LALBP^[15]	79.47	–	73.10	–	76.89	–
MMNBP^[16]	83.29	–	75.16	–	80.00	–
ViT-Cap^[17]	98.33	0.28	96.18	1.66	93.79	1.30
Lu等人^[18]	91.76	2.46	90.98	3.89	94.12	0.94
ALA Net^[19]	97.74	0.34	97.86	0.32	94.50	0.53
NLNet^[20]	98.98	0.48	96.22	0.55	94.30	1.13
ALDCNet	99.70	0.08	98.33	0.28	98.27	0.35

下载: 导出CSV

表 3 模型参数配置的对比

方法	准确率(%)	卷积核数	操作大小	卷积层数	参数量(M)	时间(ms)
AlexNet^[12]	97.26	64,192,384,256×2	11×11,5×5,3×3,3×3	5	2.454	1.01
VGG16^[13]	95.93	64×2,128×2,256×3,512×3,512×3	3×3,3×3,3×3,3×3,3×3	13	14.714	3.00
Liu等人^[14]	97.97	64,64,64,64,64	5×5,5×5,5×5,5×5,5×5	5	0.412	1.22
ALDCNet	99.70	32×2,32×3,32×3,32×3,64	3×3,1×5,5×5,3×3,3×3	12	0.164	1.66

下载: 导出CSV

表 4 消融实验(%)

模型	Acc (USM)	EER (USM)	Acc (HKPU)	EER (HKPU)	Acc (SDUMLA)	EER (SDUMLA)
无PDA	99.09	0.20	97.38	0.52	96.54	0.47
无DCFE-RC	99.29	0.10	97.38	0.48	96.69	0.64
无DCFE-SC	99.39	0.10	97.86	0.48	97.01	0.63
ALDCNet	99.70	0.08	98.33	0.28	98.27	0.35

下载: 导出CSV

参考文献(20)

[1]	LI Yantao, FAN Chao, QIN Huafeng, et al. Deep reinforcement learning-based feature extraction and encoding for finger-vein verification[J]. IEEE Transactions on Emerging Topics in Computational Intelligence, 2025, 9(1): 522–536. doi: 10.1109/TETCI.2024.3398022.
[2]	TAN Xiao. Advanced feature extraction algorithms for deep learning in image recognition[C]. 2024 International Conference on Power, Electrical Engineering, Electronics and Control (PEEEC), Athens, Greece, 2024: 1055–1060. doi: 10.1109/PEEEC63877.2024.00195.
[3]	HUANG Houjun, LIU Shilei, ZHENG He, et al. DeepVein: Novel finger vein verification methods based on deep convolutional neural networks[C]. 2017 IEEE International Conference on Identity, Security and Behavior Analysis (ISBA), New Delhi, India, 2017: 1–8. doi: 10.1109/ISBA.2017.7947683.
[4]	何鑫, 陈迅. 基于改进卷积神经网络的指静脉识别[J]. 计算机工程与设计, 2019, 40(2): 562–566. doi: 10.16208/j.issn1000-7024.2019.02.046. HE Xin and CHEN Xun. Finger vein recognition based on improved convolution neural network[J]. Computer Engineering and Design, 2019, 40(2): 562–566. doi: 10.16208/j.issn1000-7024.2019.02.046.
[5]	HU Hui, KANG Wenxiong, LU Yuting, et al. FV-Net: Learning a finger-vein feature representation based on a CNN[C]. 2018 24th International Conference on Pattern Recognition (ICPR), Beijing, China, 2018: 3489–3494. doi: 10.1109/ICPR.2018.8546007.
[6]	REN Hengyi, SUN Lijuan, REN Jinting, et al. FV-DDC: A novel finger-vein recognition model with deformation detection and correction[J]. Biomedical Signal Processing and Control, 2025, 100: 107098. doi: 10.1016/j.bspc.2024.107098.
[7]	BOUCHERIT I, ZMIRLI M O, HENTABLI H, et al. Finger vein identification using deeply-fused convolutional neural network[J]. Journal of King Saud University-Computer and Information Sciences, 2022, 34(3): 646–656. doi: 10.1016/j.jksuci.2020.04.002.
[8]	WANG Kaixuan, CHEN Guanghua, and CHU Hongjia. Finger vein recognition based on multi-receptive field bilinear convolutional neural network[J]. IEEE Signal Processing Letters, 2021, 28: 1590–1594. doi: 10.1109/LSP.2021.3094998.
[9]	CHAI Tingting, LI Jiahui, PRASAD S, et al. Shape-driven lightweight CNN for finger-vein biometrics[J]. Journal of Information Security and Applications, 2022, 67: 103211. doi: 10.1016/j.jisa.2022.103211.
[10]	MOHD ASAARI M S, SUANDI S A, and ROSDI B A. Fusion of band limited phase only correlation and width centroid contour distance for finger based biometrics[J]. Expert Systems with Applications, 2014, 41(7): 3367–3382. doi: 10.1016/j.eswa.2013.11.033.
[11]	KUMAR A and ZHOU Yingbo. Human identification using finger images[J]. IEEE Transactions on Image Processing, 2012, 21(4): 2228–2244. doi: 10.1109/TIP.2011.2171697.
[12]	FAIRUZ S, HABAEBI M H, and ELSHEIKH E M A. Finger vein identification based on transfer learning of AlexNet[C]. 2018 7th International Conference on Computer and Communication Engineering, Kuala Lumpur, Malaysia, 2018: 465–469. doi: 10.1109/ICCCE.2018.8539256.
[13]	NADIR C, ATTALLAH B, and BRIK Y. Finger vein based CNN algorithms for human recognition[C]. 2022 International Conference of Advanced Technology in Electronic and Electrical Engineering, M'sila, Algeria, 2022: 1–6. doi: 10.1109/ICATEEE57445.2022.10093105.
[14]	LIU Wenjie, LI Weijun, SUN Linjun, et al. Finger vein recognition based on deep learning[C]. 2017 12th IEEE Conference on Industrial Electronics and Applications (ICIEA), Siem Reap, Cambodia, 2017: 205–210. doi: 10.1109/ICIEA.2017.8282842.
[15]	黄艳国, 杨训根, 周满国. 基于感兴趣区域的改进型LBP手指静脉识别[J]. 传感器与微系统, 2023, 42(4): 143–147. doi: 10.13873/J.1000-9787(2023)04-0143-05. HUANG Yanguo, YANG Xungen, and ZHOU Manguo. Improved LBP finger vein recognition based on RoI[J]. Transducer and Microsystem Technologies, 2023, 42(4): 143–147. doi: 10.13873/J.1000-9787(2023)04-0143-05.
[16]	付华, 李涛, 司南楠. 基于MMNBP的手指静脉识别方法[J]. 传感器与微系统, 2019, 38(5): 45–48. doi: 10.13873/J.1000-9787(2019)05-0045-04. FU Hua, LI Tao, and SI Nannan. Finger vein recognition method based on MMNBP[J]. Transducer and Microsystem Technologies, 2019, 38(5): 45–48. doi: 10.13873/J.1000-9787(2019)05-0045-04.
[17]	LI Yupeng, LU Huimin, WANG Yifan, et al. ViT-Cap: A novel vision transformer-based capsule network model for finger vein recognition[J]. Applied Sciences, 2022, 12(20): 10364. doi: 10.3390/APP122010364.
[18]	LU Zhiying, WU Runhao, and ZHANG Jianfeng. Finger-vein feature extraction method based on vision transformer[J]. Journal of Electronic Imaging, 2022, 31(4): 043010. doi: 10.1117/1.JEI.31.4.043010.
[19]	HUANG Yiwei, MA Hui, and WANG Mingyang. Axially enhanced local attention network for finger vein recognition[J]. IEEE Transactions on Instrumentation and Measurement, 2023, 72: 5020210. doi: 10.1109/TIM.2023.3291785.
[20]	GUO Zishuo, MA Hui, and LIU Junbo. NLNet: A narrow-channel lightweight network for finger multimodal recognition[J]. Digital Signal Processing, 2024, 150: 104517. doi: 10.1016/j.dsp.2024.104517.