OpenVeinNet: Robust Open-Set Finger Vein Verification with Dynamic Snake Convolution and Graph Learning

Sushrut Patwardhan; Raghavendra Ramachandra

cs.CVAug 26, 2026

OpenVeinNet: Robust Open-Set Finger Vein Verification with Dynamic Snake Convolution and Graph Learning

Authors: Sushrut Patwardhan, Raghavendra Ramachandra

Organizations: SAFE Centre, Department of Information Security and Communication Technology (IIK), Norwegian University of Science and Technology (NTNU), Gjøvik, Norway.

Abstract

Finger vein verification is a promising biometric modality for secure authentication because vascular patterns are internal, difficult to observe externally, and relatively resistant to presentation attacks. However, reliable verification remains challenging in open-set settings, where test identities are unseen during training and non-enrolled probes must be rejected at inference. This paper presents OpenVeinNet, a finger vein verification framework designed for cross-dataset and open-set evaluation. The proposed model combines Dynamic Snake Convolution with graph-based feature modelling. Dynamic Snake Convolution extracts local curvilinear and tubular vein structures using adaptive sampling, while the graph convolutional backbone models long-range topological relationships between vein regions. To improve the discriminative quality of the embedding space, we introduce a Centroid Angular Hybrid Loss, which jointly encourages intra-class compactness and inter-class angular separation for cosinesimilaritybased verification. Experiments are conducted on five public finger vein datasets: FV-300, MMCBNU, FV-USM, PolyU, and VERA. The method is evaluated using leaveonedatasetout training under both enrolmentbased unknownrejection and fullsubject verification protocols, and is compared with handcrafted and recent deep learning-based baselines. The results show that OpenVeinNet achieves strong cross-dataset generalisation, consistently low equal error rates, and competitive true accept rates at fixed false accept rate operating points. Ablation studies further confirm the individual and combined contributions of adaptive tubular feature extraction, graph-based relational modelling, and the proposed loss function. These findings indicate that explicitly modelling local vein geometry, global vascular relationships, and angularly compact embeddings is effective for openset finger vein verification.

Figures & tables

Fig. 1 : A brief taxonomy of finger vein verification techniques based on deep learning approaches.

Authors (Year)	Technique Description
Huafeng Qin et al., [ 1 ] (2015)	3 Sequential convolution layers and 2 dense layers
Itqan et al., [ 2 ] (2016)	3 Sequential convolution layers and 1 dense layers
Syafeeza Radzi et al., [ 3 ] (2016)	2 Sequential convolution layers
Huafeng Qin et al., [ 4 ] (2017)	4 Sequential convolution layers with path-based training.
Hyung Gil Hong et al., [ 5 ] (2017)	12 Sequential convolution layers and 3 dense layers
Rig Das et al., [ 6 ] (2018)	5 Sequential convolution layers

TABLE I : Summary of existing deep learning-based finger vein verification techniques.

Fig. 2 : Block diagram of the proposed OpenVeinNet framework for open-set finger vein verification. (a) The training phase incorporates a stem module with Dynamic Snake Convolution, followed by a multi-layer Graph Convolutional Network (GCN) and the proposed Centroid Angular Hybrid Loss for robust representation learning. (b) During inference, the system performs open-set verification by comparing probe samples against enrolled templates using the learned graph-based embeddings.

Fig. 3 : DSBlock (In: input channels, Out: output channels, K: kernel size, x and y are the directions in which DSConv operates)

Component	Layer	Output Dimensions
Input	-	224,224,3
	DSC module 1 (S=2, F=3)	(112,112,16)
Stem	DSC module 2 (S=2, F=3)	(56,56,32)
	DSC module 3 (S=2, F=3)	(28,28,64)
	Stage I (K=18, S=2, F=3)	(14,14,128)
Backbone	Stage II (K=9, S=2, F=3)	(7,7,256)

TABLE II : OpenVeinNet architecture details indicating output dimension.

K

indicates the number of nearest neighbours,

S

indicates stride, and

F

indicates filter size.

Fig. 4 : Stem architecture of OpenVeinNet, featuring a sequence of three Dynamic Snake Convolution (DSConv) layers with kernel sizes $K\in\{9,7,3\}$ , stride $2$ , and GELU activation, processing an input image $\bm{I}\in\mathbb{R}^{B\times 3\times 224\times 224}$ to produce a feature map $\bm{X}_{\text{stem}}\in\mathbb{R}^{B\times 64\times 28\times 28}$ for the backbone.

Fig. 5 : GrapherBlock (In: input channels, Out: output channels, K: kernel size, N: nearest neighbours taken into account by graph convolution.)

Fig. 6 : Backbone structure of OpenVeinNet, with four sequential Grapher Blocks, a convolution layer with batch normalisation, six additional Grapher Blocks, another convolution layer with batch normalisation, and adaptive average pooling.

Dataset	Classes	Images per identity	Total Images
FV-300 (A) [ 31 ]	300	92*	27,600*
MMCBNU (B) [ 44 ]	600	10	6,000
FV-USM (C) [ 45 ]	492	12	5,904
PolyU (D) [ 46 ]	218	4	872
VERA (E) [ 47 ]	220	2	440

TABLE III : Summary of the five public finger vein datasets used in this study. A total of 124 low-quality FV-300 images were removed due to severe blur, poor illumination, or failed vein visibility.

Dataset	Genuine Scores		Impostor Scores
Dataset	Half	Full	Half	Full
FV-300 (A) [ 31 ]	598,279	1,198,364	181,709,824	362,208,956
MMCBNU (B) [ 44 ]	13,500	27,000	9,000,000	17,970,000
FV-USM (C) [ 45 ]	16,236	32,472	8,714,304	17,393,184
PolyU (D) [ 46 ]	1,562	3,107	405,765	807,794
VERA (E) [ 47 ]	110	220	48,400	96,360

TABLE IV : Number of genuine and impostor scores used for the half-subject and full-subject verification protocols.

$Train\rightarrow Test$ Dataset	Algorithm	AUC (%)	EER (%)	TAR@FAR $=0.1\%$ (%)	TAR@FAR $=1\%$ (%)	TAR@FAR $=10\%$ (%)
$BCDE\rightarrow A$	MCP [ 35 ]	99.80	0.89	40.19	99.47	100.00
	RLT [ 36 ]	99.78	0.95	30.00	99.24	100.00
	WLD [ 37 ]	99.82	0.84	43.82	99.59	100.00
	ArcVein [ 17 ]	89.39 (82.64–96.14)	19.49 (11.45–27.54)	6.67 (0.00–16.88)	22.81 (2.86–42.77)	61.22 (35.63–86.81)
	LGFIN [ 34 ]	85.52 (83.95–87.08)	23.35 (22.54–24.15)	4.23 (0.00–8.51)	15.12 (6.50–23.74)	50.09 (41.24–58.95)
	FV-ViT [ 30 ]	89.72 (88.31–91.13)	17.74 (16.16–19.33)	0.95 (0.00–2.12)	6.34 (3.56–9.12)	58.68 (51.23–66.13)

TABLE V : Half-subject threshold-based operating-point results.

$Train\rightarrow Test$ Dataset	Algorithm	AUC (%)	EER (%)	TAR@FAR $=0.1\%$ (%)	TAR@FAR $=1\%$ (%)	TAR@FAR $=10\%$ (%)
$BCDE\rightarrow A$	MCP [ 35 ]	99.80	0.89	40.19	99.47	100.00
	RLT [ 36 ]	99.78	0.95	30.00	99.24	100.00
	WLD [ 37 ]	99.82	0.84	43.82	99.59	100.00
	ArcVein [ 17 ]	89.77 (83.13–96.42)	19.38 (10.81–27.95)	6.63 (1.15–12.10)	24.80 (5.88–43.72)	63.72 (39.50–87.95)
	LGFIN [ 34 ]	85.65 (84.32–86.99)	23.01 (22.30–23.72)	3.28 (1.06–5.50)	14.68 (6.02–23.35)	50.09 (42.60–57.59)
	FV-ViT [ 30 ]	89.87 (88.51–91.23)	17.37 (15.84–18.89)	1.04 (0.00–2.11)	6.70 (3.57–9.84)	59.09 (51.73–66.46)

TABLE VI : Full-subject threshold-based operating-point results.

Dataset	Gradient	Grad. Coherence	Laplacian	Contrast	Images/class
FV-300 (A) [ 31 ]	0.0687	0.3675	0.0609	0.0311	89.57
MMCBNU (B) [ 44 ]	0.0781	0.5398	0.0026	0.0392	10.00
FV-USM (C) [ 45 ]	0.0312	0.2928	0.0009	0.0137	12.00
PolyU (D) [ 46 ]	0.0478	0.4605	0.0017	0.0249	5.84
VERA (E) [ 47 ]	0.0476	0.1429	0.0074	0.0196	2.00

TABLE VII : Mean fingervein image-quality indicators computed across datasets. The metrics quantify gradient strength, structural coherence, edge sharpness (Laplacian variance), and local contrast. Higher values indicate better visibility and continuity of vein structures. The number of images per subject is also reported to highlight dataset density, which impacts verification reliability.

Kernel set	EER (%)
(3, 3, 3)	12.3351
(5, 5, 5)	12.4453
(7, 7, 7)	11.8358
(9, 9, 9)	11.8376
(9, 7, 5)	11.0605
(7, 5, 3)	13.0697

TABLE VIII : Ablation on the kernel sizes used sequentially in the DSConv Blocks of the stem.

Fig. 7 : Ablation on the number of Grapher Blocks in the backbone. The Y-axis indicates the number of Grapher Blocks in the first stage, while the X-axis indicates the number of Grapher Blocks in the second stage.

Loss Function	EER (%)
MSE	26.68
NLL	33.98
Proposed CAH Loss	7.44

TABLE IX : Ablation study on standard loss functions under the

ABDE\rightarrow C

protocol. Lower EER is better.

Stage I $k$	Stage II $k$	EER (%)
9	9	10.37
9	18	9.42
18	18	10.48
18	9 (proposed)	7.44

TABLE X : Sensitivity of the graph neighbourhood size

k

in the two backbone stages on the ABDE

\rightarrow

C protocol. Lower EER is better.

Loss Function	Mean EER (%)
ArcFace [ 48 ]	7.83
MagFace [ 49 ]	8.20
AdaFace [ 50 ]	8.22
Proposed CAH Loss	7.44

TABLE XI : Loss ablation against strong angular-margin losses on the

ABDE\rightarrow C

protocol. Lower EER is better.

$\beta$	EER (%)
0.1	9.07
0.3 (default)	7.44
0.5	10.60
0.9	11.57

TABLE XII : Sensitivity of the proposed CAH loss to the balancing factor

\beta

on the

ABDE\rightarrow C

protocol. Lower EER is better.

Variant	Mean EER (%)
Standard stem + no graph	13.59
DSConv stem + no graph	10.96
Standard stem + graph backbone	11.64
Full model (DSConv + graph backbone)	7.44

TABLE XIII : Component ablation isolating the contributions of DSConv and graph modelling on the

ABDE\rightarrow C

protocol. Lower EER is better.

Method	Params (M)	FLOPs (G)	CPU Lat. (ms)	CPU Thr. (img/s)	GPU Lat. (ms)	GPU Thr. (img/s)
ArcVein [ 17 ]	7.83	2.45	7.09	141.10	0.94	1067.42
LGFIN [ 34 ]	6.79	7.51	45.48	21.99	3.84	260.42
FV-ViT [ 30 ]	0.36	0.03	5.02	199.37	1.74	574.79
VeinAttNet [ 31 ]	0.09	n/a	11.24	88.96	10.63	94.08
Proposed Method	4.76	5.07	55.82	17.92	13.00	76.94

TABLE XIV : Runtime and computational-cost comparison on FV-300 using batch size

1

. Lower latency is better; higher throughput is better.

Fig. 8 : Grad-CAM++ [ 51 ] -based interpretation of the proposed model for finger vein images from the same identity, representing a genuine comparison.

Fig. 9 : Grad-CAM++ [ 51 ] -based interpretation of the proposed model for finger vein images from different identities, representing an impostor comparison.

Comparison	$n$	Mean $\Delta$ EER (pp)	Holm-adjusted $p$	Wilcoxon $p$	Significant
Proposed vs ArcVein	20	15.17	$7.63\times 10^{-6}$	$1.91\times 10^{-6}$	Yes
Proposed vs LGFIN	20	13.05	$7.63\times 10^{-6}$	$1.91\times 10^{-6}$	Yes
Proposed vs FV-ViT	20	13.66	$7.63\times 10^{-6}$	$1.91\times 10^{-6}$	Yes
Proposed vs VeinAttNet	20	4.44	$9.54\times 10^{-5}$	$1.68\times 10^{-4}$	Yes

TABLE XV : Paired statistical significance testing on full-subject EER. A positive mean

\Delta

EER indicates that the proposed method achieves lower EER than the comparator. Holm-adjusted

p

values are computed from the exact paired sign-flip permutation test, while the Wilcoxon column reports the paired two-sided signed-rank test.

Dataset	Algorithm	AUC (%)	EER (%)	TAR@FAR $=0.1\%$ (%)	TAR@FAR $=1\%$ (%)	TAR@FAR $=10\%$ (%)
FV-300	MCP	99.96	0.42	95.32	100.00	100.00
	RLT	99.92	0.74	78.14	99.70	100.00
	ArcVein	96.36 (94.01–98.72)	11.03 (5.95–16.11)	54.26 (45.30–63.22)	67.25 (57.34–77.15)	89.13 (81.94–96.31)
	LGFIN	99.99 (99.98–100.00)	0.43 (0.09–0.78)	98.42 (96.98–99.86)	99.77 (99.44–100.00)	100.00 (100.00–100.00)
	FV-ViT	96.55 (95.04–98.06)	9.60 (7.59–11.61)	29.05 (7.24–50.86)	48.83 (28.26–69.40)	90.48 (85.25–95.70)
	VeinAttNet	99.73 (99.63–99.83)	2.43 (1.81–3.05)	63.32 (47.63–79.01)	93.71 (90.81–96.60)	99.90 (99.86–99.94)

TABLE XVI : Intra-database open-set verification results with 95% confidence intervals over five statistical seeds. The experiments are reported on FV-300, FV-USM, and MMCBNU, where sufficient samples per identity are available for reliable intra-database training and evaluation.

Method	Training setup	Optimizer	LR	Epochs	Batch	Input	Loss
MCP / RLT / WLD	Handcrafted pipelines	n/a	n/a	n/a	n/a	Native	No learning
ArcVein	Retrained (PyTorch)	SGD (Nesterov 0.9)	$1\times 10^{-2}$ ( $\div 10$ @30)	100	64	$224\times 224$	Arc-cosine + softmax ( $\lambda=0.01$ )
LGFIN	Retrained (PyTorch)	AdamW	$1\times 10^{-4}$	50	8	$147\times 147$	Cross-entropy
FV-ViT	Retrained (PyTorch)	AdamW	$1\times 10^{-4}$	50	64	$224\times 224$	Cross-entropy
VeinAttNet	Retrained (MATLAB)	Adam	$1\times 10^{-3}$	50	128	$224\times 224$	Softmax classification
Proposed	Retrained (PyTorch)	AdamW	$1\times 10^{-3}$	150	128	$224\times 224$	CAH ( $\beta=0.3$ )

TABLE XVII : Training configurations for the proposed method and all learned baselines.

Dataset	Total IDs	Train / Val. IDs	Train / Val. imgs	Held-out IDs
FV-300	300	240	16,536 / 4,954	60
MMCBNU	600	480	2,400 / 2,400	120
FV-USM	492	393	2,358 / 2,358	99

TABLE XVIII : Identity-level subject-disjoint train–test statistics for the intra-database open-set experiments. The first 80% of sorted identities are used for model development, and the remaining identities are held out entirely for open-set testing. The development image counts report the dataset-provided train/test folders used for optimisation and validation/model selection within the development identities.

Panel of Figure 30 — Fig. 10 : FV-300 Database t-SNE visualisation of OpenVeinNet embeddings: (a) best 50 classes, (b) worst 50 classes, and (c) all classes.

Panel of Figure 31 — Fig. 14 : FV-USM database t-SNE visualisation of OpenVeinNet embeddings: (a) best 50 classes, (b) worst 50 classes, and (c) all classes.

Panel of Figure 32 — Fig. 18 : MMCBNU database t-SNE visualisation of OpenVeinNet embeddings: (a) best 50 classes, (b) worst 50 classes, and (c) all classes.

Panel of Figure 33 — Fig. 22 : PolyU database t-SNE visualisation of OpenVeinNet embeddings: (a) best 50 classes, (b) worst 50 classes, and (c) all classes.

Panel of Figure 34 — Fig. 26 : VERA database t-SNE visualisation of OpenVeinNet embeddings: (a) best 50 classes, (b) worst 50 classes, and (c) all classes.

Panel of Figure 35 — Fig. 27 : DET curves for the half evaluation protocol across the five datasets: (a) FV-300, (b) FV-USM, (c) MMCBNU, (d) PolyU, and (e) VERA.

Panel of Figure 36 — Fig. 28 : DET curves for the full evaluation protocol across the five datasets: (a) FV-300, (b) FV-USM, (c) MMCBNU, (d) PolyU, and (e) VERA.

Explore similar work

Apr 16, 2026

Open-Set Vein Biometric Recognition with Deep Metric Learning

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.

Paweł Pilarek, Marcel Musiałek, Anna Górska

Wroclaw University of Science and Technology, Wroclaw, Poland

Biometric Identification Domain Generalization Open-Set Recognition Metric Learning

Ask PDF

Aug 27, 2026

FVeinSyn: Synthetic Finger Vein Image Generator

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 propose FVeinSyn, a large-scale controllable synthetic data generation framework for finger vein. It explicitly decouples synthesis of vascular topology and imaging appearance to mitigate the limitations caused by insufficient training samples, such as inadequate identity diversity and restricted realism. Specifically: first, a finger vein identity generator models vascular topology under physiological and geometric constraints using stochastic L-systems, producing anatomically valid and identity-distinctive vascular patterns. Then, a cascaded region-aware GAN renders the topological maps into realistic near-infrared images. Finally, an intra-class diversity generator introduces geometric and optical perturbations to simulate realistic intra-class variations. Using FVeinSyn, we generated 500,000 images (10,000 vein identities, 50 samples per identity) and conducted extensive evaluations. Results show that FVeinSyn holds significant advantages in realism, identity diversity, vascular pattern consistency, and intra-class diversity. Models trained with FVeinSyn outperform real-data-only baselines a cross eight public datasets, achieving an average accuracy improvement of 27.43%. The code is available at: https://github.com/EvanWang98/Synthetic-Finger-Vein-Generator.

Yifan Wang, Jie Gui, Adams Wai Kin Kong +6

School of Cyber Science and Engineering, Southeast University, Nanjing 210000, China · Engineering Research Center of Blockchain Application, Supervision and Management (Southeast University), Ministry of Education; Purple Mountain Laboratories, Nanjing 210000, China · School of Computer Science and Engineering, Nanyang Technological University, Singapore 639798 +5

Synthetic Data Generation Synthetic Data Augmentation Image Generation Biometric Identification

Ask PDF

Jul 23, 2026

MAGE-Vein: Multi-Instance Age and Gender Estimation from Finger Vein Images

Age estimation from finger vein images has been widely considered impractical due to severe demographic biases in public datasets and physiological confounding factors like gender. To overcome these limitations, we propose MAGE-Vein, a novel multi-instance, multi-task learning framework. Our approach extracts robust structural aging signs by employing a hybrid feature-level fusion of three fingers, effectively suppressing local imaging noise. Furthermore, simultaneous optimization of gender classification conditions the network to effectively eliminate gender-specific vascular variations. Evaluated on a demographically balanced dataset of 402 subjects, MAGE-Vein achieves a mean absolute error of 6.12 years and a correlation of 0.880. Our results not only overturn the conventional consensus regarding the limitations of the finger vein modality but also demonstrate that previous estimation failures were primarily artifacts of biased public datasets. Our code is available at https://github.com/gsisaoki/MAGE-Vein.

Katsuki Tanaka, Koichi Ito, Takafumi Aoki +4

Graduate School of Information Sciences, Tohoku University, Japan. · Hitachi, Ltd., Japan.

Topic pending

Ask PDF