Pith. sign in

REVIEW 5 major objections 6 minor 55 references

FingerVeinSyn-5M: A Million-Scale Dataset and Benchmark for Finger Vein Recognition

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read FingerVeinSyn-5M, a five-million-image synthetic dataset, transfers to real finger-vein recognition and lifts fine-tuned performance by 53.91 percent on average.

desk verdict A genuinely large synthetic finger-vein dataset with useful annotations, but the headline 53.91% gain is an absolute TAR jump measured on a 500k-image subset, not the full 5M dataset. read the letter →

arxiv 2506.03635 v1 pith:R6ZABVXT submitted 2025-06-04 cs.CV

classification cs.CV
keywords fingerveinrecognitionsyntheticdatasetbiometricgenerativeadversarialnetworkL-systempretrainingfew-shotannotations
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Finger vein recognition is held back by the small size of public datasets, which have under a thousand identities and only a handful of samples per finger. This paper argues that the bottleneck can be bypassed with a synthetic corpus: FVeinSyn generates 50,000 unique vein identities with an L-system, renders them into realistic images with a cascaded region-aware GAN, and then applies controlled transformations to produce 100 samples per identity. The released dataset, FingerVeinSyn-5M, contains 5 million fully annotated images, including vein-pattern masks, finger-shape masks, joint-cavity coordinates, ROI boxes, and per-sample transformation parameters. The central claim is that pretraining on this fully synthetic data and fine-tuning with minimal real data improves recognition by an average of 53.91% across six public benchmarks. If the claim holds, it means a dataset with no real identities can supply the scale and supervision that real finger-vein collections cannot.

What carries the argument

The load-bearing object is FVeinSyn, a three-stage generator. The identity stage uses a stochastic L-system, a grammar that grows branching structures, to synthesize physiologically plausible vein networks unique to each virtual finger. The rendering stage uses a cascaded region-aware GAN, trained with cycle-consistency losses, region-specific adversarial losses, and margin losses that keep joint cavities brighter than surrounding tissue and veins darker, to turn binary vein patterns into realistic images. The variation stage applies controlled shifts, rotations, scalings, rolls, exposures, and three kinds of blur to create 100 samples per identity. These components create the scale, annotation, and controlled intra-class variation that the transfer results depend on.

What would settle it

Evaluate the same pretrain-and-fine-tune protocol on a finger vein dataset acquired with a sensor not among HKPU-FV, FV-USM, or PLUS-FV3. If the 53.91% average gain collapses on that unseen hardware, the effect is caused by memorized supervision statistics rather than by realistic vein-pattern synthesis.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that synthetic finger vein imagery transfers to real recognition tasks. FingerVeinSyn-5M is the largest finger vein dataset, with 5,000,000 images from 50,000 distinct generated identities and 100 controlled variations per identity. The paper reports that synthetic-only training outperforms matched-size real-data training by 15.17% on average, that larger synthetic training sets improve open-set recognition further, and that pretraining plus minimal fine-tuning yields an average 53.91% gain over real-data training. It also reports that the synthetic identities are 99.83% unique in feature space and that the first comprehensive annotations enable supervised learning for tasks other than recognition.

Load-bearing premise

The transfer results assume the image generator learns general vein realism rather than memorizing the finger shapes, crop regions, and brightness patterns of the three real datasets (HKPU-FV, FV-USM, PLUS-FV3) used to train it.

Editorial extensions

If this is right

  • Pretraining on FingerVeinSyn-5M and fine-tuning on one real enrollment sample raises true-accept rate at a 1e-6 false-accept rate from roughly 0.04–0.71 to 0.38–0.86 across the six datasets, so single-sample registration becomes practical.
  • A model trained only on the synthetic set and tested on six real sensors obtains true-accept rates above 0.6 on five of six domains, indicating synthetic pretraining supplies domain-general features.
  • Because the dataset includes vein-pattern masks, finger-shape masks, joint-cavity coordinates, ROI boxes, and per-sample transformation parameters, supervised multi-task models can be trained at a scale impossible with real finger-vein collections.
  • The open-set results show synthetic-only training beating matched-size real training by 15.17% on average, rising to 29.15% when the full 100-variation training set is used, so controlled synthetic variation is a substitute for raw data volume.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The most direct extension is to treat the FVeinSyn pipeline as a general vascular-biometric generator: palm veins share the same two-vessel-plus-network anatomy and lighting physics, so the released recipe should transfer if the renderer's supervision data is replaced.
  • The renderer was supervised only with HKPU-FV, FV-USM, and PLUS-FV3, so the cleanest check of the paper's generality is an unseen-sensor evaluation; until that is run, the 53.91% gain should be understood as demonstrated on sensors statistically similar to those three.
  • The per-sample transformation annotations make it possible to learn disentangled representations of pose, roll, scale, and illumination, a direction the recognition experiments only evaluate indirectly.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes FVeinSyn, a synthetic finger-vein generator that combines stochastic L-systems for vein pattern generation, a cascaded region-aware GAN for rendering, and a set of intra-class augmentations (shift, rotation, scaling, rolling, exposure, and blur effects). Using this generator, the authors construct and release FingerVeinSyn-5M, a dataset of 50,000 synthetic identities with 100 samples each plus annotations, and evaluate recognition performance on six public finger-vein datasets. The central claim is that models pretrained on this synthetic data and fine-tuned with minimal real data achieve an average 53.91% performance gain, and that the dataset enables one/few-shot recognition and cross-domain transfer.

Significance. If validated, this would be the largest public finger-vein dataset and the first with full annotations, making a substantial contribution to supervised multi-task learning and to synthetic-to-real pretraining for biometric recognition. The paper has concrete strengths: it evaluates on real, disjoint test sets from six public datasets; it releases the dataset and annotations; and the generation pipeline contains plausible, controllable components. However, the headline 53.91% gain is not directly demonstrated for the released 5M, 50k-times-100 configuration, and several protocol details are under-specified. The core idea is promising and likely useful to the biometrics community, but the evidence as presented is not yet sufficient to support the paper's strongest quantitative claims.

major comments (5)
  1. [§5.3, Table 3] The 53.91% figure is not tied to a clearly defined baseline. It equals the absolute difference between the Avg. TAR of the 'Syn. data 50k 10 500k' row with fine-tuning (0.9013) and the 'Real data' row (0.3622). The abstract and §5.3 describe this as a 'performance gain' or 'further accuracy improvement,' which readers will naturally read as a relative gain. The same ambiguity affects the '15.17%' and '29.15%' claims in §5.3, which also appear to be absolute percentage-point differences from the real-data baseline. Please state explicitly whether these are absolute percentage points or relative percentages, and identify the exact rows being compared.
  2. [§5.3, §5.4] The central claim is not tested on the released dataset configuration. The best synthetic row in Table 3 uses 50,000 identities times 10 samples (500k images), not the full 5M images from 50,000 identities times 100 samples. In Table 4, the 'Syn. data' row repeats the values from the 'Syn. data 1.2k 100 120k' row of Table 3 (0.7324, 0.6831, 0.6443, 0.7310, 0.8450, 0.2861), and no fine-tuning or full-scale pretraining run appears. Consequently, the experiments do not demonstrate that pretraining on FingerVeinSyn-5M at its released scale yields the reported 53.91% gain. Please either run the full recommended configuration or explicitly qualify the claim as applying only to the tested subsets.
  3. [§5.2] The dataset-quality metrics U_class, C_intra, and D_intra depend on a recognition model F_eval, but the paper does not state what data F_eval was trained on. If F_eval was trained on synthetic data, U_class and C_intra largely measure self-consistency rather than realism; if it was trained on real data, that should be stated and justified. The threshold r is also not specified, aside from the remark 'r < 0.2' in the uniqueness discussion, and Eq. (5) for D_intra uses an unspecified distance threshold. Please provide the exact values and a real-data-trained F_eval, or explain why the chosen F_eval is appropriate.
  4. [§5.5] The one/few-shot protocol is under-specified. The paper does not state how the N registration samples are selected per identity, whether the same identities are used for all values of N, or how the fine-tuning and test sets are partitioned for each dataset. In addition, §5.1 says both training and fine-tuning are run for 20 epochs, while §5.5 says both are set to 60 epochs; this discrepancy must be reconciled. Without this information, the one-shot results in Table 5 cannot be reproduced.
  5. [§5.1, §5.5] No error bars or multiple-seed results are reported. Recognition metrics, especially one/few-shot results such as the N=1 row in Table 5, are point estimates that can vary substantially across random initializations and data splits. Please report mean and standard deviation over at least three seeds, or justify why a single run is sufficient for the conclusions.
minor comments (6)
  1. [§1, author block] The affiliations contain typos: 'Nnajing' should be 'Nanjing' and 'Sccience' should be 'Science'.
  2. [§5.1] The verification protocol is incomplete; please specify whether TAR@FAR=1e-6 is computed via 1:1 verification or 1:N identification, and how many genuine and impostor comparisons are used.
  3. [§3.3, Eq. (4)] The terms MeanA and MeanB are not defined; please specify which image regions they average and provide some justification or sensitivity analysis for the chosen margin values of 0.1 and 0.3.
  4. [§4] The sample count description is ambiguous: the listed variations sum to 50 samples, and the text says the 'final 50 samples include different combinations'; please clarify how the 100 samples per identity are exactly allocated and whether combinations are applied in addition to or instead of individual variations.
  5. [§6, Conclusion] The conclusion says the transformations generate 'inter-class variations'; since the variations are applied within a single finger identity, this should be 'intra-class variations'.
  6. [Table 5] The '-' entries for UTFVP at N=3 are not explained; please add a footnote describing why those results are absent.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central synthetic-to-real transfer claim is evaluated on disjoint real test sets, and the cited priors are external. The reported gaps are protocol or reporting issues, not derivation-by-construction circularity.

full rationale

The paper's central claim—that models pretrained on FingerVeinSyn-5M and fine-tuned on real data improve finger vein recognition—is tested against real, disjoint test sets (Section 5.1, Table 3), so the reported TAR numbers are not equivalent to the training inputs by construction. The synthetic generator does use real finger shapes and brightness priors (Section 3.3), but that is a conditioning input, not a fit to the evaluation labels; the test identities are non-overlapping. No 'uniqueness theorem' or load-bearing self-citation is invoked: self-citations [41,42,43] support only background statements, and the anatomical prior [10,36] is external. The reviewer-flagged issues are real but are not circularity: Table 4 reuses the 'Syn. data 1.2k 100 120k' row from Table 3, the headline 53.91% is an absolute TAR difference (0.9013−0.3622) described as a 'performance gain', and the reported experiments use at most 50k×10=500k images rather than the full 5M dataset. Section 5.2 also leaves the training data of F_eval unspecified, which weakens the dataset-quality metrics if F_eval is synthetic-trained, but the paper does not use those metrics as the load-bearing evidence for the transfer claim. These are reporting/protocol omissions that should be corrected, but they do not make any equation or result reduce to its own input. Therefore the derivation chain is not circular.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central transfer claim rests on the realism of the generated images and on the evaluation protocol. The generator uses several hand-set constants and domain assumptions (anatomical model, GAN transferability, SAM annotation quality). No new physical entities such as particles, forces, or conserved quantities are introduced.

free parameters (4)
  • margin_j in Eq. (4) = 0.1
    Hand-set to enforce that joint cavities appear brighter than surrounding tissue; based on empirical observations of real finger vein images; no sensitivity analysis reported.
  • margin_p in Eq. (4) = 0.3
    Hand-set to enforce that vein patterns appear darker than background; an empirical constant with no sensitivity analysis.
  • L-system branching probabilities (P_u, P_m, P_b) = not specified
    Stochastic grammar probabilities control vein branching directions and initial points; values are not reported, making the identity generator hard to reproduce.
  • Intra-class variation ranges = shift +/-20, scale +/-0.15, roll +/-20, rotation +/-20
    Hand-chosen augmentation ranges that determine the intra-class diversity score (68.30%); no ablation shows how these choices affect downstream recognition.
assumptions (3)
  • domain assumption Finger veins primarily consist of two main vessels on each side, connected by a venous network (from reference [10]).
    Used in Section 3.2 as the anatomical constraint for the L-system vein pattern generator; if this simplified anatomy is wrong, generated identities may not match real finger vein topology.
  • domain assumption 2D L-system generated vein patterns, rendered by a CycleGAN-style cascade, are sufficiently realistic that representations learned on them transfer to real sensor images.
    Core premise of the whole benchmark; Section 5.6 provides only a coarse matching-score distribution check.
  • domain assumption The Segment Anything Model provides accurate finger shape, joint cavity, and ROI annotations on the HaGRID gesture database and real finger datasets, and these annotations are anatomically correct for finger vein capture.
    Used in Section 3.3 to generate shape and region supervision; errors would propagate into the synthetic dataset's annotations and realism.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FingerVeinSyn-5M: A Million-Scale Dataset and Benchmark for Finger Vein Recognition." pith.science (2026). https://pith.science/paper/R6ZABVXT

@misc{pith2026250603635,
  author       = {Pith},
  title        = {Pith review of: FingerVeinSyn-5M: A Million-Scale Dataset and Benchmark for Finger Vein Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R6ZABVXT}},
  note         = {Machine review of arXiv:2506.03635}
}
read the original abstract

A major challenge in finger vein recognition is the lack of large-scale public datasets. Existing datasets contain few identities and limited samples per finger, restricting the advancement of deep learning-based methods. To address this, we introduce FVeinSyn, a synthetic generator capable of producing diverse finger vein patterns with rich intra-class variations. Using FVeinSyn, we created FingerVeinSyn-5M -- the largest available finger vein dataset -- containing 5 million samples from 50,000 unique fingers, each with 100 variations including shift, rotation, scale, roll, varying exposure levels, skin scattering blur, optical blur, and motion blur. FingerVeinSyn-5M is also the first to offer fully annotated finger vein images, supporting deep learning applications in this field. Models pretrained on FingerVeinSyn-5M and fine-tuned with minimal real data achieve an average 53.91\% performance gain across multiple benchmarks. The dataset is publicly available at: https://github.com/EvanWang98/FingerVeinSyn-5M.

Figures

Figures reproduced from arXiv: 2506.03635 by the authors.

Figure 1
Figure 1. The main FVeinSyn framework for large-scale fin [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the matching score distributions between real and synthetic finger vein images. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 42 canonical work pages

  1. [1]

    Mohd Shahrimie Mohd Asaari, Shahrel A Suandi, and Bakhtiar Affendi Rosdi

  2. [2]

    Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. 2018. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018) . IEEE, 67–74

  3. [3]

    Chinese Academy of Sciences Institute of Automation. 2025. CASIA Fin- gerprint Image Database. http://www.idealtest.org/dbDetailForUser.do?id=7# /datasetDetail/7. Accessed: 2025-05-26

  4. [4]

    Rig Das, Emanuela Piciucco, Emanuele Maiorana, and Patrizio Campisi. 2019. Convolutional Neural Network for Finger-Vein-Based Biometric Identification. IEEE Transactions on Information Forensics and Security 14, 2 (2019), 360–373. https://doi.org/10.1109/TIFS.2018.2850320

  5. [5]

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. 2019. Arcface: Additive angular margin loss for deep face recognition. In CVPR

  6. [6]

    Joshua James Engelsma, Steven Grosz, and Anil K. Jain. 2023. PrintsGAN: Syn- thetic Fingerprint Generator. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 5 (2023), 6111–6124. https://doi.org/10.1109/TPAMI.2022.3204591

  7. [7]

    Gregory P Fiumara, Patricia A Flanagan, John D Grantham, Bruce Bandini, Ken- neth Ko, and John M Libert. 2018. NIST special database 300: uncompressed plain and rolled images from fingerprint cards. (2018)

  8. [8]

    Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. 2016. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In Com- puter Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14 . Springer, 87–102

Show all 55 references
  1. [9]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition . 770–778

  2. [10]

    Fieke Hillerström, Ajay Kumar, and Raymond Veldhuis. 2014. Generating and analyzing synthetic finger vein images. In 2014 International Conference of the Biometrics Special Interest Group (BIOSIG) . IEEE, 1–9

  3. [11]

    Borui Hou and Ruqiang Yan. 2021. ArcVein-Arccosine Center Loss for Finger Vein Verification. IEEE Transactions on Instrumentation and Measurement 70 (2021), 1–11. https://doi.org/10.1109/TIM.2021.3062164

  4. [12]

    Saad Shakeel, Weili Yang, and Wenxiong Kang

    Junduan Huang, An Zheng, M. Saad Shakeel, Weili Yang, and Wenxiong Kang

  5. [13]

    Jianlong Jin, Lei Shen, Ruixin Zhang, Chenglong Zhao, Ge Jin, Jingyun Zhang, Shouhong Ding, Yang Zhao, and Wei Jia. 2024. Pce-palm: Palm crease energy based two-stage realistic pseudo-palmprint generation. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 3...

  6. [14]

    Jianlong Jin, Chenglong Zhao, Ruixin Zhang, Sheng Shang, Jianqing Xu, Jingyun Zhang, ShaoMing Wang, Yang Zhao, Shouhong Ding, Wei Jia, and Yunsheng Wu

  7. [15]

    Alexander Kapitanov, Karina Kvanchiani, Alexander Nagaev, Roman Kraynov, and Andrei Makhliarchuk. 2024. HaGRID – HAnd Gesture Recognition Image Dataset. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV). 4572–4581

  8. [16]

    Christof Kauba, Bernhard Prommegger, and Andreas Uhl. 2018. Focussing the Beam - A New Laser Illumination Based Data Set Providing Insights to Finger- Vein Recognition. In 2018 IEEE 9th International Conference on Biometrics Theory, Applications and Systems (BTAS). 1–9. https:...

  9. [17]

    Minchul Kim, Feng Liu, Anil Jain, and Xiaoming Liu. 2023. DCFace: Synthetic Face Generation with Dual Condition Diffusion Model. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . 12715–12725. https://doi. org/10.1109/CVPR52729.2023.01223

  10. [18]

    Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. 2023. Segment Anything. arXiv:2304.02643 (2023)

  11. [19]

    Ajay Kumar and Yingbo Zhou. 2011. Human identification using finger images. IEEE Transactions on image processing 21, 4 (2011), 2228–2244

  12. [20]

    Ridvan Salih Kuzu, Emanuela Piciucco, Emanuele Maiorana, and Patrizio Campisi

  13. [21]

    Shuyi Li and Bob Zhang. 2021. Joint Discriminative Sparse Coding for Robust Hand-Based Multimodal Recognition. IEEE Transactions on Information Forensics and Security 16 (2021), 3186–3198. https://doi.org/10.1109/TIFS.2021.3074315

  14. [22]

    Yantao Li, Congyi Tang, Shaojiang Deng, Huafeng Qin, and Hongyu Huang

  15. [23]

    Jingyu Lin, Guiqin Zhao, Jing Xu, Guoli Wang, Zejin Wang, Antitza Dantcheva, Lan Du, and Cunjian Chen. 2024. DiffTV: Identity-Preserved Thermal-to-Visible Face Translation via Feature Alignment and Dual-Stage Conditions. In Proceed- ings of the 32nd ACM International Conferenc...

  16. [24]

    Aristid Lindenmayer. 1968. Mathematical models for cellular interactions in development I. Filaments with one-sided inputs. Journal of theoretical biology 18, 3 (1968), 280–299

  17. [25]

    Yu Lu, Shan Juan Xie, Sook Yoon, Zhihui Wang, and Dong Sun Park. 2013. An available database for the research of finger vein recognition. In 2013 6th Interna- tional congress on image and signal processing (CISP) , Vol. 1. IEEE, 410–415

  18. [26]

    Naoto Miura, Akio Nagasaka, and Takafumi Miyatake. 2004. Feature extraction of finger-vein patterns based on repeated line tracking and its application to personal identification. Machine vision and applications 15 (2004), 194–203

  19. [27]

    Naoto Miura, Akio Nagasaka, and Takafumi Miyatake. 2007. Extraction of finger- vein patterns using maximum curvature points in image profiles. IEICE TRANS- ACTIONS on Information and Systems 90, 8 (2007), 1185–1194

  20. [28]

    National Institute of Standards and Technology. 2025. NIST Special Database 302. https://www.nist.gov/itl/iad/image-group/nist-special-database-302. Accessed: 2025-05-26

  21. [29]

    Shree K Nayar and Srinivasa G Narasimhan. 1999. Vision in bad weather. In Proceedings of the seventh IEEE international conference on computer vision , Vol. 2. IEEE, 820–827

  22. [30]

    Wei-Feng Ou, Lai-Man Po, Chang Zhou, Peng-Fei Xian, and Jing-Jing Xiong. 2022. GAN-based inter-class sample generation for contrastive learning of vein image representations. IEEE Transactions on Biometrics, Behavior, and Identity Science 4, 2 (2022), 249–262

  23. [31]

    El-Yacoubi, and Jun Wang

    Huafeng Qin, Zhipeng Xiong, Yantao Li, Mounim A. El-Yacoubi, and Jun Wang

  24. [32]

    Hengyi Ren, Lijuan Sun, Jian Guo, and Chong Han. 2022. A dataset and benchmark for multimodal biometric recognition based on fingerprint and finger vein. IEEE Transactions on Information Forensics and Security 17 (2022), 2030–2043

  25. [33]

    Edwin H Salazar-Jurado, Ruber Hernández-García, Karina Vilches-Ponce, Ri- cardo J Barrientos, Marco Mora, and Gaurav Jaswal. 2023. Towards the generation of synthetic images of palm vein patterns: A review.Information Fusion 89 (2023), 66–90

  26. [34]

    Sheng Shang, Chenglong Zhao, Ruixin Zhang, Jianlong Jin, Jingyun Zhang, Rizen Guo, Shouhong Ding, Yunsheng Wu, Yang Zhao, and Wei Jia. 2025. PVTree: Real- istic and Controllable Palm Vein Generation for Recognition Tasks. InProceedings of the AAAI Conference on Artificial Inte...

  27. [35]

    Lei Shen, Jianlong Jin, Ruixin Zhang, Huaen Li, Kai Zhao, Yingyi Zhang, Jingyun Zhang, Shouhong Ding, Yang Zhao, and Wei Jia. 2023. RPG-Palm: Realistic Pseudo-data Generation for Palmprint Recognition. InProceedings of the IEEE/CVF International Conference on Computer Vision (...

  28. [36]

    Paul W Sweeney, Lina Hacker, Thierry L Lefebvre, Emma L Brown, Janek Gröhl, and Sarah E Bohndiek. 2024. Unsupervised segmentation of 3D microvascular photoacoustic images using deep generative learning. Advanced Science 11, 32 (2024), 2402195

  29. [37]

    IEEE Transactions on Information Forensics and Security 19 (2024), 9330–9343

    Attention BLSTM-Based Temporal-Spatial Vein Transformer for Multi- View Finger-Vein Recognition. IEEE Transactions on Information Forensics and Security 19 (2024), 9330–9343. https://doi.org/10.1109/TIFS.2024.3468898

  30. [38]

    B. T. Ton and R. N. J. Veldhuis. 2013. A high quality finger vascular pattern dataset collected using a custom designed capturing device. In Proceedings of the 2013 International Conference on Biometrics (ICB), Madrid, Spain (Madrid, Spain). 1–5

  31. [39]

    Matthias Vanoni, Pedro Tome, Laurent El Shafey, and Sébastien Marcel. 2014. Cross-database evaluation using an open finger vein sensor. In2014 IEEE workshop on biometric measurements and systems for security and medical applications (BIOMS) proceedings. IEEE, 30–35

  32. [40]

    Tao Wang, Yushu Zhang, Xiangli Xiao, Lin Yuan, Zhihua Xia, and Jian Weng

  33. [41]

    Yifan Wang, Jie Gui, Xinli Shi, Linqing Gui, Yuan Yan Tang, and James Tin-Yau Kwok. 2025. ColorVein: Colorful Cancelable Vein Biometrics. IEEE Transactions on Information Forensics and Security (2025)

  34. [42]

    Yifan Wang, Jie Gui, Yuan Yan Tang, and James Tin-Yau Kwok. 2024. CFVNet: An End-to-End Cancelable Finger Vein Network for Recognition. IEEE Transactions on Information Forensics and Security 19 (2024), 7810–7823. https://doi.org/10. 1109/TIFS.2024.3436528

  35. [43]

    Su Tang, Shan Zhou, Wenxiong Kang, Qiuxia Wu, and Feiqi Deng. 2019. Finger vein verification using a Siamese CNN. IET biometrics 8, 5 (2019), 306–315

  36. [44]

    Lu Yang, Gongping Yang, Kuikui Wang, Fanchang Hao, and Yilong Yin. 2021. Finger Vein Recognition via Sparse Reconstruction Error Constrained Low-Rank Representation. IEEE Transactions on Information Forensics and Security 16 (2021), 4869–4881. https://doi.org/10.1109/TIFS.2021.3118894

  37. [45]

    Yilong Yin, Lili Liu, and Xiwei Sun. 2011. SDUMLA-HMT: A multimodal biometric database. In Biometric Recognition: 6th Chinese Conference, CCBR 2011, Beijing, China, December 3-4, 2011. Proceedings 6 . Springer, 260–268

  38. [46]

    Pengyang Zhao, Zhiquan Chen, Jing-Hao Xue, Jianjiang Feng, Wenming Yang, Qingmin Liao, and Jie Zhou. 2023. Single-Sample Finger Vein Recognition via Competitive and Progressive Sparse Representation. IEEE Transactions on Bio- metrics, Behavior, and Identity Science 5, 2 (2023)...

  39. [47]

    In Proceedings of the 32nd ACM International Conference on Multimedia (Melbourne VIC, Australia) (MM ’24)

    Make Privacy Renewable! Generating Privacy-Preserving Faces Supporting Cancelable Biometric Recognition. In Proceedings of the 32nd ACM International Conference on Multimedia (Melbourne VIC, Australia) (MM ’24). Association for Computing Machinery, New York, NY, USA, 10268–102...

  40. [48]

    Zheng Zhu, Guan Huang, Jiankang Deng, Yun Ye, Junjie Huang, Xinze Chen, Jiagang Zhu, Tian Yang, Dalong Du, Jiwen Lu, and Jie Zhou. 2023. WebFace260M: A Benchmark for Million-Scale Deep Face Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 2 (2023...

  41. [50]

    Yifan Wang, Huimin Lu, Xiwen Qin, and Jianwei Guo. 2023. Residual Gabor convolutional network and FV-Mix exponential level data augmentation strategy for finger vein recognition. Expert Systems with Applications 223 (2023), 119874. MM ’25, October 27–31, 2025, Dublin, Ireland ...

  42. [54]

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. 2017. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Pro- ceedings of the IEEE international conference on computer vision . 2223–2232

  43. [2014]

    Expert Systems with Applications 41, 7 (2014), 3367–3382

    Fusion of band limited phase only correlation and width centroid contour distance for finger based biometrics. Expert Systems with Applications 41, 7 (2014), 3367–3382

  44. [2020]

    IEEE Transactions on Information Forensics and Security 15 (2020), 2641–2654

    On-the-Fly Finger-Vein-Based Biometric Recognition Using Deep Neural Networks. IEEE Transactions on Information Forensics and Security 15 (2020), 2641–2654. https://doi.org/10.1109/TIFS.2020.2971144

  45. [2023]

    IEEE Transactions on Information Forensics and Security 18 (2023), 1322–1334

    FVFSNet: Frequency-Spatial Coupling Network for Finger Vein Authentica- tion. IEEE Transactions on Information Forensics and Security 18 (2023), 1322–1334. https://doi.org/10.1109/TIFS.2023.3238546

  46. [2024]

    IEEE Transactions on Instrumentation and Measurement 73 (2024), 1–10

    Mixed Automatic Adversarial Augmentation Network for Finger-Vein Recognition. IEEE Transactions on Instrumentation and Measurement 73 (2024), 1–10. https://doi.org/10.1109/TIM.2024.3400355

  47. [2025]

    arXiv:2503.18312 [cs.CV] https://arxiv.org/abs/2503.18312

    Diff-Palm: Realistic Palmprint Generation with Polynomial Creases and Intra-Class Variation Controllable Diffusion Models. arXiv:2503.18312 [cs.CV] https://arxiv.org/abs/2503.18312

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.