Open-Set Vein Biometric Recognition with Deep Metric Learning

Anna G\'orska; Marcel Musia{\l}ek; Pawe{\l} Pilarek

arxiv: 2604.14874 · v1 · submitted 2026-04-16 · 💻 cs.CV

Open-Set Vein Biometric Recognition with Deep Metric Learning

Pawe{\l} Pilarek , Marcel Musia{\l}ek , Anna G\'orska This is my paper

Pith reviewed 2026-05-10 11:47 UTC · model grok-4.3

classification 💻 cs.CV

keywords vein recognitionopen-set recognitiondeep metric learningL2-normalized embeddingsprototype matchingbiometric identificationfinger veinhand vein

0 comments

The pith

Deep metric learning enables open-set vein recognition by learning embeddings that reject unseen subjects without retraining.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper demonstrates that vein biometric systems can move beyond closed-set classification, which forces full model retraining whenever new users enroll, to an open-set approach. It trains networks to produce L2-normalized embeddings, then matches them against prototypes using one fixed similarity threshold to accept known users and reject unknowns. Experiments on four datasets covering finger, wrist, and hand veins, including the large MMCBNU 6000 collection, report 99.6 percent rank-1 accuracy, 0.9945 open-set classification rate, and 1.57 percent equal error rate. A reader would care because this removes a major scalability barrier for real-world biometric deployments where user populations change over time.

Core claim

By training deep networks under metric-learning objectives to yield L2-normalized embeddings and applying prototype-based matching with a single calibrated similarity threshold, the framework performs open-set vein recognition: it identifies enrolled subjects at high accuracy while rejecting unseen impostors. On the MMCBNU 6000 benchmark the ResNet50-CBAM model reaches an OSCR of 0.9945, AUROC of 0.9974, and EER of 1.57 percent with 99.6 percent rank-1 accuracy. The same pipeline generalizes across four datasets but remains sensitive to domain shifts when data are scarce, and ablation results show that triplet losses paired with a simple 1-NN classifier give the best accuracy-efficiency mix.

What carries the argument

L2-normalized embeddings produced by deep metric learning, combined with prototype-based matching and one calibrated similarity threshold for open-set decisions.

If this is right

High identification accuracy (99.6 percent rank-1) is achieved simultaneously with strong rejection of unknown subjects on large-scale data.
Triplet-based metric learning plus a 1-NN classifier yields an efficient trade-off that supports real-time operation on ordinary hardware.
The system handles cross-dataset shifts to a useful degree but shows clear performance drops in low-data regimes.
Prototype matching removes the need to retrain the entire network when adding new enrolled users.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same embedding-plus-threshold recipe could be tried on other biometric modalities such as fingerprints or iris images to achieve open-set operation.
In deployment the threshold may still require occasional recalibration when the imaging environment changes, even if the paper does not test this explicitly.
Because new users can be enrolled by simply storing their embeddings, the approach lowers the computational cost of maintaining large biometric databases over time.
Further tests could check whether few-shot fine-tuning of the embeddings improves robustness when only a handful of samples are available for a new subject.

Load-bearing premise

A single fixed similarity threshold on the L2-normalized embeddings can reliably separate enrolled users from unseen impostors across different acquisition setups and data regimes.

What would settle it

If the equal error rate rises well above 1.57 percent or the open-set classification rate falls below 0.99 when the trained model is tested on a new vein dataset collected with different hardware, without any threshold adjustment.

Figures

Figures reproduced from arXiv: 2604.14874 by Anna G\'orska, Marcel Musia{\l}ek, Pawe{\l} Pilarek.

**Figure 1.** Figure 1: Overview of the proposed open-set vein recognition framework. The input vein image is mapped to a compact, ℓ2-normalised embedding by the ResNet50-CBAM backbone. During inference, the system computes cosine similarity scores against enrolled class prototypes. If the maximum similarity Smax falls below the calibrated threshold τ , the sample is rejected as an unknown identity; otherwise, it is assigned to … view at source ↗

read the original abstract

Most state-of-the-art vein recognition methods rely on closed-set classification, which inherently limits their scalability and prevents the adaptive enrollment of new users without complete model retraining. We rigorously evaluate the computational boundaries of Deep Metric Learning (DML) under strict open-set constraints. Unlike standard closed-set approaches, we analyze the impact of data scarcity and domain shift on recognition performance. Our approach learns discriminative L2-normalised embeddings and employs prototype-based matching with a calibrated similarity threshold to effectively distinguish between enrolled users and unseen impostors. We evaluate the framework under a strict subject-disjoint protocol across four diverse datasets covering finger, wrist, and dorsal hand veins (MMCBNU 6000, UTFVP, FYO, and a dorsal hand-vein dataset). On the large-scale MMCBNU 6000 benchmark, our best model (ResNet50-CBAM) achieves an OSCR of 0.9945, AUROC of 0.9974, and EER of 1.57%, maintaining high identification accuracy (99.6% Rank-1) while robustly rejecting unknown subjects. Cross-dataset experiments evaluate the framework's generalisation across different acquisition setups, confirming that while the model handles large-scale data robustly, performance remains sensitive to domain shifts in low-data regimes. Ablation studies demonstrate that triplet-based objectives combined with a simple 1-NN classifier offer an optimal trade-off between accuracy and efficiency, enabling real-time deployment on commodity hardware.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Applies deep metric learning to open-set vein biometrics with good results on large data but limited transfer across domains.

read the letter

This paper shows that deep metric learning can support open-set vein biometric recognition with high accuracy on a large benchmark, but the results also reveal that performance is sensitive to changes in data acquisition and scale. The new part is the evaluation of DML embeddings under strict subject-disjoint open-set conditions on vein datasets. They use L2-normalized features from models like ResNet50 with CBAM, match to prototypes, and apply a calibrated similarity threshold to reject unknowns. This setup is tested on MMCBNU 6000, UTFVP, FYO, and a dorsal hand dataset. It does well on the main task. On MMCBNU 6000 the best model reaches 0.9945 OSCR, 0.9974 AUROC, 1.57% EER, and 99.6% rank-1 identification. Ablation studies indicate triplet loss with a 1-NN classifier balances accuracy and speed for real-time use on standard hardware. The cross-dataset tests add value by showing where the method holds up and where it does not. The weaker area is generalization. The high numbers are tied to one large dataset, and the paper notes drops in low-data regimes under domain shift. This suggests the single calibrated threshold may need adjustment for new setups, which limits how plug-and-play the approach is. More detail on how the threshold is selected and the exact splits would strengthen the claims. Biometrics researchers focused on vein patterns or open-set identification problems would find this useful for the reported metrics and practical considerations. It is an incremental but honest extension of existing metric learning techniques to this domain. I recommend sending it for peer review. The concrete results and acknowledged limitations make it worth a closer look by referees.

Referee Report

3 major / 3 minor

Summary. The manuscript proposes a deep metric learning (DML) framework for open-set vein biometric recognition. It learns L2-normalized embeddings via triplet loss, uses prototype-based 1-NN matching, and applies a single calibrated similarity threshold to distinguish enrolled subjects from unseen impostors. Evaluations follow subject-disjoint protocols on four datasets (MMCBNU 6000, UTFVP, FYO, dorsal hand-vein), reporting OSCR 0.9945, AUROC 0.9974, EER 1.57%, and 99.6% Rank-1 on MMCBNU 6000 with ResNet50-CBAM; ablations favor triplet objectives with 1-NN for accuracy-efficiency trade-off, while cross-dataset tests note sensitivity to domain shift in low-data regimes.

Significance. If the open-set claims hold under fully specified protocols, the work advances scalable vein biometrics by removing closed-set retraining requirements for new enrollments. The multi-dataset evaluation, explicit handling of data scarcity, and ablation studies on loss functions and classifiers provide concrete evidence of practical trade-offs. The domain-shift analysis is a strength, as it identifies real limitations rather than overstating generalization.

major comments (3)

[§3 and §4.1] §3 (Methodology) and §4.1 (Experimental Setup): The procedure for calibrating the similarity threshold on L2-normalized embeddings is not described, including whether it is selected via validation on held-out data from the same acquisition distribution or fixed a priori. This is load-bearing for the central claim, as the reported OSCR/EER on MMCBNU 6000 and the assertion of robust rejection of unknown subjects both depend on this threshold separating known from unseen subjects without post-hoc adjustment.
[§4.2 and Table 2] §4.2 (Results on MMCBNU 6000) and Table 2: Exact subject-disjoint split details (e.g., number of subjects and samples allocated to training, validation for threshold calibration, and testing) are omitted, as is any error analysis of false rejects or impostor acceptance cases. Without these, the headline metrics (OSCR 0.9945, EER 1.57%) cannot be fully assessed for independence from distribution-specific tuning, especially given the paper's own cross-dataset drops.
[§5.3] §5.3 (Cross-dataset experiments): While performance sensitivity to domain shift in low-data regimes is reported, no quantitative analysis of embedding distribution shift (e.g., via cosine similarity histograms or t-SNE) or alternative threshold strategies is provided. This weakens the claim that prototype-based 1-NN with a single threshold offers reliable open-set separation across acquisition setups.

minor comments (3)

[Abstract] Abstract: The phrase 'rigorously evaluate the computational boundaries' is vague; replace with a concrete statement of the open-set protocol and metrics used.
[§4] §4: Ensure all tables report the exact number of trials or runs for the reported means and standard deviations; currently some ablation rows appear to omit variance.
[§3] Notation: Consistently use 'cosine similarity' rather than mixing 'similarity threshold' and 'L2-normalized embeddings' without clarifying the distance metric in the matching step.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which have helped us improve the clarity and reproducibility of the manuscript. We address each major comment point by point below, providing the strongest honest defense of our work while incorporating necessary clarifications and additions in the revised version.

read point-by-point responses

Referee: [§3 and §4.1] The procedure for calibrating the similarity threshold on L2-normalized embeddings is not described, including whether it is selected via validation on held-out data from the same acquisition distribution or fixed a priori. This is load-bearing for the central claim.

Authors: We thank the referee for highlighting this omission. The similarity threshold is calibrated exclusively on a held-out validation set drawn from training subjects (subject-disjoint from the test set) by selecting the value that maximizes OSCR on that validation data. No test-set information is used. We have now added a dedicated subsection in §3 describing the full calibration procedure, its optimization criterion, and confirmation that it remains fixed during testing. revision: yes
Referee: [§4.2 and Table 2] Exact subject-disjoint split details (e.g., number of subjects and samples allocated to training, validation for threshold calibration, and testing) are omitted, as is any error analysis of false rejects or impostor acceptance cases.

Authors: We agree these details are required for full assessment. The revised manuscript now includes explicit subject-disjoint split specifications (subjects and sample counts for training, validation used for threshold calibration, and testing) for every dataset in an expanded Table 2 and §4.2. We have also added a concise error analysis characterizing the main sources of false rejects (primarily acquisition artifacts) and rare impostor acceptances (inter-subject vein similarities), supported by representative examples. revision: yes
Referee: [§5.3] While performance sensitivity to domain shift in low-data regimes is reported, no quantitative analysis of embedding distribution shift (e.g., via cosine similarity histograms or t-SNE) or alternative threshold strategies is provided.

Authors: We accept that quantitative support strengthens the domain-shift discussion. The revised §5.3 now incorporates cosine-similarity histograms comparing intra- and cross-dataset embedding pairs together with t-SNE visualizations of the embedding distributions. We additionally compare the single global threshold against dataset-specific calibration alternatives and report the resulting trade-offs, confirming that the single-threshold approach remains practical within-domain while highlighting its sensitivity under low-data cross-dataset conditions. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical results rest on external benchmarks and standard open-set metrics

full rationale

The paper presents an empirical DML framework for open-set vein recognition. It learns L2-normalized embeddings, applies prototype-based 1-NN matching, and reports OSCR/AUROC/EER on subject-disjoint splits of external datasets (MMCBNU 6000, UTFVP, FYO). No equations, derivations, or self-citations are shown that reduce the reported performance numbers to quantities defined by fitted parameters inside the paper. Threshold calibration is a standard post-training step; the metrics are computed on held-out test data and remain falsifiable against the cited benchmarks. This is a normal non-circular empirical evaluation.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on standard assumptions of deep metric learning rather than new postulates; the similarity threshold is the main fitted element.

free parameters (1)

similarity threshold
Calibrated on data to separate known from unknown subjects; exact value and selection procedure not stated in abstract.

axioms (1)

domain assumption L2-normalized embeddings from triplet loss are sufficiently discriminative for vein patterns across subjects
Invoked when claiming prototype matching works for open-set rejection.

pith-pipeline@v0.9.0 · 5575 in / 1224 out tokens · 25094 ms · 2026-05-10T11:47:52.557828+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages

[1]

Intelligent Systems with Applications19, 200256 (2023)

Abdullahi, S.B., et al.: Sequence-wise multimodal biometric fingerprint and finger- vein recognition network (stmfpfv-net). Intelligent Systems with Applications19, 200256 (2023). https://doi.org/10.1016/j.iswa.2023.200256

work page doi:10.1016/j.iswa.2023.200256 2023
[2]

Advances in Science and Technology – Research Journal16, 36–46 (2022)

Al-Khafaji, R., Al-Tamimi, M.: Vein biometric recognition methods and systems: A review. Advances in Science and Technology – Research Journal16, 36–46 (2022). https://doi.org/10.12913/22998624/144495

work page doi:10.12913/22998624/144495 2022
[3]

Barcina-Blanco,M.,etal.:Managingtheunknown:asurveyonopensetrecognition and tangential areas (2024)

work page 2024
[4]

In: Biometric Recognition

Chen, Z., Yu, W., Bai, H., Li, Y.: An arcloss-based and openset-test-oriented finger vein recognition system. In: Biometric Recognition. pp. 287–294. Springer (2021)

work page 2021
[5]

Geng, S.-j

Geng, C., Huang, S.J., Chen, S.: Recent advances in open set recognition: A sur- vey. IEEE Transactions on Pattern Analysis and Machine Intelligence43(10), 3614–3631 (2021). https://doi.org/10.1109/tpami.2020.2981604

work page doi:10.1109/tpami.2020.2981604 2021
[6]

Hemis, M., Kheddar, H., Bourouis, S., Saleem, N.: Deep learning techniques for hand vein biometrics: A comprehensive review (2024), https://arxiv.org/abs/2409.07128

work page arXiv 2024
[7]

IEEE Transactions on Instrumentation and Measurement71, 1–26 (2022)

Hou, B., Zhang, H., Yan, R.: Finger-vein biometric recognition: A review. IEEE Transactions on Instrumentation and Measurement71, 1–26 (2022). https://doi.org/10.1109/TIM.2022.3200087

work page doi:10.1109/tim.2022.3200087 2022
[8]

Pattern Recognition Letters79, 80–105 (2016)

Jain, A.K., Nandakumar, K., Ross, A.: 50 years of biometric research: Accom- plishments, challenges, and opportunities. Pattern Recognition Letters79, 80–105 (2016). https://doi.org/10.1016/j.patrec.2015.12.013 14 P. Pilarek, M. Musiałek, A. Górska

work page doi:10.1016/j.patrec.2015.12.013 2016
[9]

Academic Platform Journal of Engineering and Smart Systems13, 71–93 (2025)

Kocakulak, M., Avci, A., Acır, N.: A survey of finger-vein recognition using deep learning: Concepts, challenges, and opportunities. Academic Platform Journal of Engineering and Smart Systems13, 71–93 (2025). https://doi.org/10.21541/apjess.1672743

work page doi:10.21541/apjess.1672743 2025
[10]

Sensors (Basel, Switzerland)13, 14339 – 14366 (2013)

Lu, Y., Xie, S.J., Yoon, S., Yang, J.C., Park, D.S.: Robust finger vein roi local- ization based on flexible segmentation. Sensors (Basel, Switzerland)13, 14339 – 14366 (2013)

work page 2013
[11]

In: 2013 6th International Congress on Image and Signal Processing (CISP) (2014)

Lu, Y., Xie, S., Yoon, S., Wang, Z., Park, D.S.: An available database for the research of finger vein recognition. In: 2013 6th International Congress on Image and Signal Processing (CISP) (2014). https://doi.org/10.1109/CISP.2013.6744030

work page doi:10.1109/cisp.2013.6744030 2013
[12]

In: 2021 IEEE FourthInternationalConferenceonArtificialIntelligenceandKnowledgeEngineer- ing (AIKE)

Mahdavi, A., Carvalho, M.: A survey on open set recognition. In: 2021 IEEE FourthInternationalConferenceonArtificialIntelligenceandKnowledgeEngineer- ing (AIKE). p. 37–44. IEEE (2021). https://doi.org/10.1109/aike52691.2021.00013

work page doi:10.1109/aike52691.2021.00013 2021
[13]

Sensors23(6) (2023)

Marattukalam, F., Abdulla, W., Cole, D., Gulati, P.: Deep learning- based wrist vascular biometric recognition. Sensors23(6) (2023). https://doi.org/10.3390/s23063132

work page doi:10.3390/s23063132 2023
[14]

Turkish Journal of Electrical Engineering and Com- puter Sciences24(3), 1863–1878 (2016)

Radzi, S.A., Hani, M.K., Bakhteri, R.: Finger-vein biometric identification using convolutional neural network. Turkish Journal of Electrical Engineering and Com- puter Sciences24(3), 1863–1878 (2016). https://doi.org/10.3906/elk-1311-43

work page doi:10.3906/elk-1311-43 2016
[15]

Knowledge-Based Sys- tems227, 107159 (2021)

Ren, H., Sun, L., Guo, J., Han, C., Wu, F.: Finger vein recognition system with template protection based on convolutional neural network. Knowledge-Based Sys- tems227, 107159 (2021). https://doi.org/10.1016/j.knosys.2021.107159

work page doi:10.1016/j.knosys.2021.107159 2021
[16]

Adabi, H

Singh, M., Singh, R., Ross, A.: A comprehensive overview of biometric recogni- tion: approaches, challenges and future directions. Journal of Ambient Intelligence and Humanized Computing12, 1–23 (2021). https://doi.org/10.1007/s12652-021- 03123-w

work page doi:10.1007/s12652-021- 2021
[17]

Dataset (2013), https://www.utwente.nl/en/eemcs/dmb/downloads/utfvp/

Ton, B.T., Veldhuis, R.: University of twente finger vascular pattern (utfvp). Dataset (2013), https://www.utwente.nl/en/eemcs/dmb/downloads/utfvp/

work page 2013
[18]

IEEE Access8, 82461–82470 (2020)

Toygar, Ö., Babalola, F.O., Bitirim, Y.: Fyo: A novel multimodal vein database with palmar, dorsal and wrist biometrics. IEEE Access8, 82461–82470 (2020). https://doi.org/10.1109/ACCESS.2020.2991475

work page doi:10.1109/access.2020.2991475 2020
[19]

IEEE Access7, 183118–183132 (2019)

Zhang, J., Lu, Z., Li, M., Wu, H.: Gan-based image augmentation for finger-vein biometric recognition. IEEE Access7, 183118–183132 (2019). https://doi.org/10.1109/ACCESS.2019.2960411

work page doi:10.1109/access.2019.2960411 2019
[20]

Concurrency and Computation: Practice and Experience34(12), e5697 (2022)

Zhang, Y., Li, W., Zhang, L., Ning, X., Sun, L., Lu, Y.: Agcnn: Adaptive gabor convolutional neural networks with receptive fields for vein biometric recognition. Concurrency and Computation: Practice and Experience34(12), e5697 (2022). https://doi.org/10.1002/cpe.5697

work page doi:10.1002/cpe.5697 2022

[1] [1]

Intelligent Systems with Applications19, 200256 (2023)

Abdullahi, S.B., et al.: Sequence-wise multimodal biometric fingerprint and finger- vein recognition network (stmfpfv-net). Intelligent Systems with Applications19, 200256 (2023). https://doi.org/10.1016/j.iswa.2023.200256

work page doi:10.1016/j.iswa.2023.200256 2023

[2] [2]

Advances in Science and Technology – Research Journal16, 36–46 (2022)

Al-Khafaji, R., Al-Tamimi, M.: Vein biometric recognition methods and systems: A review. Advances in Science and Technology – Research Journal16, 36–46 (2022). https://doi.org/10.12913/22998624/144495

work page doi:10.12913/22998624/144495 2022

[3] [3]

Barcina-Blanco,M.,etal.:Managingtheunknown:asurveyonopensetrecognition and tangential areas (2024)

work page 2024

[4] [4]

In: Biometric Recognition

Chen, Z., Yu, W., Bai, H., Li, Y.: An arcloss-based and openset-test-oriented finger vein recognition system. In: Biometric Recognition. pp. 287–294. Springer (2021)

work page 2021

[5] [5]

Geng, S.-j

Geng, C., Huang, S.J., Chen, S.: Recent advances in open set recognition: A sur- vey. IEEE Transactions on Pattern Analysis and Machine Intelligence43(10), 3614–3631 (2021). https://doi.org/10.1109/tpami.2020.2981604

work page doi:10.1109/tpami.2020.2981604 2021

[6] [6]

Hemis, M., Kheddar, H., Bourouis, S., Saleem, N.: Deep learning techniques for hand vein biometrics: A comprehensive review (2024), https://arxiv.org/abs/2409.07128

work page arXiv 2024

[7] [7]

IEEE Transactions on Instrumentation and Measurement71, 1–26 (2022)

Hou, B., Zhang, H., Yan, R.: Finger-vein biometric recognition: A review. IEEE Transactions on Instrumentation and Measurement71, 1–26 (2022). https://doi.org/10.1109/TIM.2022.3200087

work page doi:10.1109/tim.2022.3200087 2022

[8] [8]

Pattern Recognition Letters79, 80–105 (2016)

Jain, A.K., Nandakumar, K., Ross, A.: 50 years of biometric research: Accom- plishments, challenges, and opportunities. Pattern Recognition Letters79, 80–105 (2016). https://doi.org/10.1016/j.patrec.2015.12.013 14 P. Pilarek, M. Musiałek, A. Górska

work page doi:10.1016/j.patrec.2015.12.013 2016

[9] [9]

Academic Platform Journal of Engineering and Smart Systems13, 71–93 (2025)

Kocakulak, M., Avci, A., Acır, N.: A survey of finger-vein recognition using deep learning: Concepts, challenges, and opportunities. Academic Platform Journal of Engineering and Smart Systems13, 71–93 (2025). https://doi.org/10.21541/apjess.1672743

work page doi:10.21541/apjess.1672743 2025

[10] [10]

Sensors (Basel, Switzerland)13, 14339 – 14366 (2013)

Lu, Y., Xie, S.J., Yoon, S., Yang, J.C., Park, D.S.: Robust finger vein roi local- ization based on flexible segmentation. Sensors (Basel, Switzerland)13, 14339 – 14366 (2013)

work page 2013

[11] [11]

In: 2013 6th International Congress on Image and Signal Processing (CISP) (2014)

Lu, Y., Xie, S., Yoon, S., Wang, Z., Park, D.S.: An available database for the research of finger vein recognition. In: 2013 6th International Congress on Image and Signal Processing (CISP) (2014). https://doi.org/10.1109/CISP.2013.6744030

work page doi:10.1109/cisp.2013.6744030 2013

[12] [12]

In: 2021 IEEE FourthInternationalConferenceonArtificialIntelligenceandKnowledgeEngineer- ing (AIKE)

Mahdavi, A., Carvalho, M.: A survey on open set recognition. In: 2021 IEEE FourthInternationalConferenceonArtificialIntelligenceandKnowledgeEngineer- ing (AIKE). p. 37–44. IEEE (2021). https://doi.org/10.1109/aike52691.2021.00013

work page doi:10.1109/aike52691.2021.00013 2021

[13] [13]

Sensors23(6) (2023)

Marattukalam, F., Abdulla, W., Cole, D., Gulati, P.: Deep learning- based wrist vascular biometric recognition. Sensors23(6) (2023). https://doi.org/10.3390/s23063132

work page doi:10.3390/s23063132 2023

[14] [14]

Turkish Journal of Electrical Engineering and Com- puter Sciences24(3), 1863–1878 (2016)

Radzi, S.A., Hani, M.K., Bakhteri, R.: Finger-vein biometric identification using convolutional neural network. Turkish Journal of Electrical Engineering and Com- puter Sciences24(3), 1863–1878 (2016). https://doi.org/10.3906/elk-1311-43

work page doi:10.3906/elk-1311-43 2016

[15] [15]

Knowledge-Based Sys- tems227, 107159 (2021)

Ren, H., Sun, L., Guo, J., Han, C., Wu, F.: Finger vein recognition system with template protection based on convolutional neural network. Knowledge-Based Sys- tems227, 107159 (2021). https://doi.org/10.1016/j.knosys.2021.107159

work page doi:10.1016/j.knosys.2021.107159 2021

[16] [16]

Adabi, H

Singh, M., Singh, R., Ross, A.: A comprehensive overview of biometric recogni- tion: approaches, challenges and future directions. Journal of Ambient Intelligence and Humanized Computing12, 1–23 (2021). https://doi.org/10.1007/s12652-021- 03123-w

work page doi:10.1007/s12652-021- 2021

[17] [17]

Dataset (2013), https://www.utwente.nl/en/eemcs/dmb/downloads/utfvp/

Ton, B.T., Veldhuis, R.: University of twente finger vascular pattern (utfvp). Dataset (2013), https://www.utwente.nl/en/eemcs/dmb/downloads/utfvp/

work page 2013

[18] [18]

IEEE Access8, 82461–82470 (2020)

Toygar, Ö., Babalola, F.O., Bitirim, Y.: Fyo: A novel multimodal vein database with palmar, dorsal and wrist biometrics. IEEE Access8, 82461–82470 (2020). https://doi.org/10.1109/ACCESS.2020.2991475

work page doi:10.1109/access.2020.2991475 2020

[19] [19]

IEEE Access7, 183118–183132 (2019)

Zhang, J., Lu, Z., Li, M., Wu, H.: Gan-based image augmentation for finger-vein biometric recognition. IEEE Access7, 183118–183132 (2019). https://doi.org/10.1109/ACCESS.2019.2960411

work page doi:10.1109/access.2019.2960411 2019

[20] [20]

Concurrency and Computation: Practice and Experience34(12), e5697 (2022)

Zhang, Y., Li, W., Zhang, L., Ning, X., Sun, L., Lu, Y.: Agcnn: Adaptive gabor convolutional neural networks with receptive fields for vein biometric recognition. Concurrency and Computation: Practice and Experience34(12), e5697 (2022). https://doi.org/10.1002/cpe.5697

work page doi:10.1002/cpe.5697 2022