PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Sushrut Patwardhan, Raghavendra Ramachandra

arXiv:2608.25515v1cs.CV

TL;DR

Open-set finger vein verification must handle cross-dataset shifts and reject non-enrolled identities, not merely match enrolled users. OpenVeinNet combines adaptive tubular feature extraction, graph-based relational modelling, and centroid-angular embedding supervision, achieving strong and stable performance across five leave-one-dataset-out settings and both evaluation protocols.

  • Problem

    Cross-dataset sensor and acquisition shifts reduce robustness, while open-set verification requires rejecting non-enrolled probes at fixed operating thresholds.

  • Method

    OpenVeinNet combines DSConv-based local tubular feature extraction, graph-based long-range relational modelling, cosine-similarity verification, and Centroid Angular Hybrid Loss.

  • Results

    OpenVeinNet achieves strong and stable performance across five leave-one-dataset-out settings under enrollment-based unknown-rejection and full-subject verification, with consistent gains in EER and threshold-based TAR.

  • Takeaways & Limitations

    Explicitly modelling local vein geometry, global vascular relationships, and angularly constrained embeddings is effective for open-set finger vein verification.

  • Takeaways & Limitations

    FV-300 reported counts are approximate because 124 low-quality images were removed before partitioning, although subject-disjoint splitting was applied consistently.

Abstract

from arXiv · show

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.

1 INTRODUCTION

Finger vein verification offers a relatively spoof-resistant biometric modality, but cross-dataset shifts and open-set unknown rejection remain central practical challenges. OpenVeinNet addresses these challenges by combining adaptive tubular feature extraction, graph-based relational modelling, and angularly constrained embeddings.

  • Motivation: Finger vein patterns are difficult to observe externally and relatively resistant to presentation attacks, supporting their use in secure access control.They are acquired with cost-effective near-infrared imaging systems.
  • Challenges: Cross-dataset shifts can reduce robustness because sensor characteristics, image quality, and acquisition protocols differ between training and testing data.Open-set deployment additionally requires rejecting non-enrolled probes at fixed operating thresholds.
  • Proposed framework: OpenVeinNet integrates Dynamic Snake Convolution for local tubular vein features with graph-based modelling of long-range vascular relationships.The framework is designed for robust cross-dataset and open-set evaluation.
  • Proposed framework: Centroid Angular Hybrid Loss promotes intra-class compactness and inter-class angular separation for cosine-similarity-based verification.The loss is intended to improve discriminative representation learning for open-set verification.
  • Evaluation: The evaluation covers five public datasets using leave-one-dataset-out training, enrollment-based unknown-rejection, and full-subject verification protocols.Reported measures include AUC, EER, and TAR@FAR operating points, alongside comparisons with handcrafted and deep learning-based methods.
  • Analysis: Ablation, computational-cost, and Grad-CAM++ analyses examine the architecture, loss formulation, and decision behaviour.The component analysis reports that graph-based relational modelling is effective when local node features are geometrically aligned with vein structures.

2 PROPOSED METHOD: OPENVEINNET FOR OPEN-SET FINGER VEIN VERIFICATION

OpenVeinNet combines adaptive local vein extraction, graph-based relational modelling, and angularly structured embeddings for open-set finger vein verification. Its pipeline uses DSConv stem blocks, Grapher Blocks, positional information, and a hybrid loss for cosine-similarity verification.

  • Architecture: OpenVeinNet extracts local tubular vein features with DSConv and models local and long-range relationships between vein regions using Grapher Blocks.The stem uses adaptive sampling, while the graph backbone constructs k-nearest-neighbour graphs and aggregates neighbouring-node information.
  • Stem: Dynamic Snake Convolution: DSConv learns deformable sampling offsets and bilinearly samples directional feature maps to align receptive fields with curved and irregular vein structures.The offsets contain vertical and horizontal components, and directional convolutions use K×1 or 1×K kernels depending on the operating direction.
  • Stem: Dynamic Snake Convolution: The stem uses three DSConv layers with kernel sizes 9, 7, and 3, stride 2, and GELU activation for processing 224×224 input images.Each DSConv block is followed by batch normalisation and GELU, except the final block, where GELU is omitted to retain a wider range of fine-grained responses.
  • Feature representation: A learnable positional embedding is added before graph construction so the backbone retains the spatial arrangement and topology of vein structures.The stem output is then processed by graph blocks and adaptive average pooling to produce the final global representation.
  • Centroid Angular Hybrid Loss: Centroid Angular Hybrid Loss augments softmax with an angular term to compact same-identity embeddings and separate different identities angularly for threshold-based open-set verification.The balancing factor β controls angular supervision, with overly large values potentially reducing optimisation stability.

3 EXPERIMENTS AND RESULTS

Across five leave-one-dataset-out settings, OpenVeinNet shows strong and stable cross-dataset performance under both unknown-rejection and full-subject verification protocols. Its results indicate reliable generalisation across acquisition conditions, although dataset quality and sample density materially affect operating-point performance.

  • Overall evaluation: OpenVeinNet achieves strong and stable performance across both enrollment-based unknown-rejection and full-subject verification protocols on five leave-one-dataset-out settings.The learned embeddings generalise to unseen identities and different acquisition conditions.
  • Enrollment-based unknown-rejection: 97.72% AUC and 7.48% EER are achieved in the challenging ABDE →C unknown-rejection setting, with 53.52% TAR at FAR=1% and 95.47% TAR at FAR=10%.This protocol enrolls half the held-out subjects and treats the remainder as unknown identities.
  • Enrollment-based unknown-rejection: 98.67% AUC and 5.34% EER are obtained for ABCE →D, alongside the strongest reported TAR values across operating points.The setting directly tests rejection of non-enrolled identities at fixed thresholds.
  • Full-subject verification: OpenVeinNet records the lowest EER in 4 of 5 full-subject splits, including 3.17% in BCDE →A, 7.44% in ABDE →C, and 5.21% in ABCE →D.The results indicate that the embedding space remains discriminative under substantial dataset shift.
  • Operating points: OpenVeinNet generally achieves higher TAR than deep learning baselines at FAR=0.1% and FAR=1%, especially in challenging ABDE →C and ABCE →D protocols.Handcrafted MCP and WLD methods can perform strongly on clean data but degrade under stronger domain shift.
  • Operating points: VeinAttNet exceeds OpenVeinNet at FAR=0.1% in ACDE→B, achieving 39.09% versus 28.13% TAR, while OpenVeinNet has lower EER: 3.99% versus 4.35%.The paper attributes this trade-off to stricter-threshold score concentration versus broader operating-point calibration.
  • Challenging dataset analysis: The ABCD →E protocol is especially difficult because VERA has two images per identity and higher acquisition variability; at FAR=0.1%, TAR is zero for all methods.The paper relates this outcome to coarse score resolution from few genuine comparisons, while OpenVeinNet remains competitive at FAR=1% and FAR=10%.
  • Interpretation and significance: Performance variation is primarily associated with image quality, sample density, and acquisition variability, while OpenVeinNet maintains strong generalisation across diverse conditions.The reported improvements over competing deep learning baselines are statistically significant after Holm correction, with p < 10^-3.

4 ABLATION STUDY

The ablations identify effective configurations for DSConv, graph depth, neighbourhood size, and CAH loss, while showing complementary and synergistic gains from the major components.

  • DSConv Stem: The progressively decreasing DSConv kernel configuration (9, 7, 3) achieves the best EER of 10.82%.Larger shallow-layer kernels capture broader vascular structures, while smaller deeper-layer kernels refine local details.
  • Grapher Backbone: Four Grapher Blocks in Stage I and six in Stage II achieve the best EER of 10.595%.This moderate depth balances long-range connectivity against over-smoothing and redundancy.
  • Graph Neighbourhood: The asymmetric graph neighbourhood configuration k=18 in Stage I and k=9 in Stage II outperforms symmetric alternatives.Using a large neighbourhood in the later stage can introduce noise as spatial resolution decreases.
  • Loss Function: CAH achieves the lowest mean EER at 7.44%, compared with 7.83% for ArcFace, 8.20% for MagFace, and 8.22% for AdaFace.The loss also outperforms standard MSE and NLL objectives in the reported ablation.
  • Loss Function: β = 0.3 gives the best EER of 7.44%, whereas smaller or larger values underweight or overemphasise the angular term.The result supports balancing classification supervision with angular regularisation.
  • Component-Level Ablation: The full model reaches an EER of 7.44%, versus 13.59% for the standard baseline, 10.96% for DSConv alone, and 11.64% for graph modelling alone.The combined improvement is reported as larger than the sum of the individual gains, linking geometric alignment with effective graph reasoning.

5 COMPUTATIONAL COST AND INFERENCE LA-

OpenVeinNet prioritises verification robustness through a more expressive architecture, creating a measurable accuracy–efficiency trade-off while remaining practical for GPU-enabled deployment.

  • Evaluation Protocol: The study characterises the accuracy–efficiency trade-off by measuring parameters, FLOPs, latency, and throughput under a unified batch-size-one protocol.Learned methods are evaluated on held-out FV-300 checkpoints across CPU and GPU platforms.
  • Comparison: FV-ViT has 0.36M parameters and 0.03 GFLOPs, while ArcVein achieves 0.94 ms GPU latency and the highest throughput.These lightweight models provide faster inference but are associated with reduced verification robustness.
  • Computational Cost: OpenVeinNet uses 4.76M parameters and 5.07 GFLOPs, with latency of 55.82 ms on CPU and 13.00 ms on GPU.The model remains more compact than several CNN-based baselines in parameter count.
  • Cost Justification: DSConv adaptive sampling and Grapher Block graph construction increase computation while supporting curvilinear feature alignment and long-range topological modelling.The added operations are described as contributing to embedding discriminativeness and generalisability.
  • Verification Trade-off: CAH improves angular embedding separability, producing consistently lower EER and higher TAR under strict low-FAR operating conditions.These operating points are identified as important for real-world biometric systems.

6 INTERPRETATION OF THE PROPOSED METHOD

Grad-CAM++ analysis indicates that OpenVeinNet bases verification decisions on vascular structures and their broader structural relationships, with distinct roles for the stem and backbone.

  • Interpretation Setup: Grad-CAM++ examines genuine and impostor comparisons, including high- and low-similarity pairs, across the stem and backbone.The analysis is designed to identify regions contributing to embedding responses.
  • Stem Activations: Stem activations align strongly with visible vein structures, concentrating on consistent vascular regions in genuine high-similarity cases.Low-similarity or impostor cases produce more diffuse or shifted responses.
  • Backbone Activations: Backbone activations highlight broader regions associated with higher-level identity-specific information rather than only local vein texture.This behaviour is consistent with relational aggregation across neighbouring and non-local vein regions.
  • Interpretation Findings: The interpretation results provide qualitative evidence that decisions rely on meaningful vascular patterns and higher-level structural relationships.The analysis supports the intended division between local tubular extraction and graph-based feature refinement.

7 CONCLUSION

OpenVeinNet is presented as an open-set finger vein verification framework that combines local vascular morphology, global relational modelling, and angularly constrained embeddings.

  • Conclusion: OpenVeinNet integrates a DSConv stem for adaptive curvilinear vein extraction with a graph backbone for long-range dependency modelling.The architecture explicitly represents both local vascular morphology and global structural relationships.
  • Conclusion: CAH loss jointly enforces intra-class compactness and inter-class angular separability in the embedding space.This design is described as particularly beneficial for open-set scenarios.
  • Conclusion: Cross-dataset experiments under unknown-rejection and full-subject protocols demonstrate strong generalisation across multiple benchmarks.The conclusion reports improvements in EER and TAR, especially under strict low-FAR operating conditions.

18 Supplementary Results OpenVeinNet: Robust Open-Set

The supplementary material adds experimental evidence and implementation details for OpenVeinNet’s finger-vein verification framework.

  • The supplement reports additional experimental evidence for OpenVeinNet.The supplied passage identifies the supplement as providing further evidence for the framework.
  • OpenVeinNet is presented as using dynamic snake convolution and graph-based representation learning.
  • The supplementary material also includes implementation details.

8 STATISTICAL SIGNIFICANCE ANALYSIS

The study formally tests whether OpenVeinNet’s full-subject verification improvements over competing methods are statistically reliable across matched evaluation runs.

  • Paired exact two-sided sign-flip permutation tests compare OpenVeinNet with competing approaches on full-subject verification EER.Pairing uses seed-matched leave-one-dataset-out runs across five evaluation protocols.
  • Holm correction is applied to account for multiple comparisons.
  • All pairwise comparisons remain statistically significant after Holm correction.The result is reported in Table 15 and is described as consistent across matched evaluation runs.

9 INTRA-DATABASE OPEN-SET EVALUATION

Intra-database open-set experiments on FV-300, FV-USM, and MMCBNU show that OpenVeinNet remains competitive within individual databases, complementing its cross-dataset evaluation.

  • FV-300 results show a 0.39% EER for OpenVeinNet, the lowest among the evaluated methods.OpenVeinNet also achieves the highest TAR at FAR= 0.1% and matches the best TAR at FAR= 1% and FAR= 10%.
  • FV-USM results show 98.26% AUC and 7.06% EER for OpenVeinNet, versus 11.25% EER for LGFIN.OpenVeinNet also records the highest TAR across the reported operating points, including 53.38% at FAR= 0.1% and 73.66% at FAR= 1%.
  • MMCBNU results show 99.31% AUC and 4.07% EER for OpenVeinNet.LGFIN has slightly higher TAR at FAR= 0.1%, while OpenVeinNet has the best TAR at FAR= 1% and FAR= 10%.
  • The intra-database experiments support OpenVeinNet’s effectiveness for conventional open-set verification within the same database.The passage distinguishes this setting from the stronger domain-shift conditions of cross-dataset evaluation.
  • The learned baselines and proposed method were retrained under a unified evaluation protocol, while handcrafted methods were executed as fixed pipelines.ArcVein, LGFIN, FV-ViT, and OpenVeinNet were implemented in PyTorch; VeinAttNet was retrained in MATLAB.

11 SHARED EXPERIMENTAL PROTOCOL

The shared experimental protocol aligns data splits, seeds, metrics, and evaluation procedures across learned methods while retaining architecture-specific training choices and avoiding test-set tuning.

  • Each leave-one-dataset-out experiment trains on four datasets and tests on one held-out dataset.
  • All methods use the same statistical seeds and a common pipeline for computing AUC, EER, and TAR@FAR.
  • PyTorch-based models share grayscale-to-three-channel conversion, random flips, and resizing to method-specific input resolutions.VeinAttNet instead follows its original MATLAB augmentation strategy, including reflections, translations, rotations, and scaling.
  • No dataset-specific tuning is performed on held-out test sets, and fixed training configurations are used across splits except for source-class count.
  • The comparison aligns protocols, metrics, code, and matched runs without forcing identical hyperparameters across architectures.Baselines retain architecture-specific choices such as optimizer, learning-rate schedule, input resolution, and loss formulation.

12 SUBJECT-DISJOINT TRAIN–VALIDATION SPLIT

The intra-database open-set experiments enforce identity-level separation between model development and final testing, while embedding analyses show dataset-dependent cluster quality and persistent discriminative structure.

  • Subject-disjoint partitioning: The first 80% of sorted identities support model development, while the remaining 20% are held out entirely for open-set testing.This subject-disjoint split prevents final test subjects or finger identities from appearing during model development.
  • Subject-disjoint partitioning: Dataset-provided train and test image folders are used for optimisation and validation within development identities, not to establish open-set subject disjointness.Final evaluation uses only held-out identities and computes genuine and impostor scores from within-identity and cross-identity image pairs.
  • Experimental scope: FV-300 counts are approximate because 124 low-quality images were removed before subject partitioning.The same subject-disjoint development and testing strategy is nevertheless applied across the intra-database experiments and learned methods.
  • Embedding-space analysis: FV-300 produces compact, well-separated t-SNE clusters, whereas FV-USM, MMCBNU, and PolyU show greater dispersion, overlap, or reduced boundary sharpness.These visual differences are associated with image quality, acquisition conditions, cross-session variability, and subject diversity.
  • Embedding-space analysis: VERA has the most challenging embedding structure, with sparse clusters, substantial overlap, and minimal global cluster organisation.The passage attributes this pattern to only two images per identity and higher acquisition variability.
  • Embedding-space analysis: Across datasets, higher image quality and more samples per identity correspond to more compact clusters, while challenging datasets still retain partial cluster structure.The analysis interprets these patterns as evidence that OpenVeinNet embeddings remain partially discriminative under low-sample and high-variability conditions.

14 DET CURVE ANALYSIS

DET analyses compare OpenVeinNet across five datasets and two evaluation protocols, showing favourable false-match and false-non-match trade-offs, especially in difficult and low-FMR settings.

  • Overall DET behaviour: Across all datasets, OpenVeinNet's DET curves lie closer to the origin under both half and full protocols, indicating a favourable FMR–FNMR trade-off.The full protocol shows slight improvements attributed to increased enrolment data.
  • Dataset-specific comparisons: On FV-300, OpenVeinNet maintains lower FNMR, particularly in the low-FMR regime, despite strong performance from all methods.Low-FMR operation is identified as important for secure applications.
  • Dataset-specific comparisons: On FV-USM and MMCBNU, OpenVeinNet shows consistently lower FNMR across most operating regions than competing methods.The passage relates this pattern to robustness against intra-class variation and acquisition differences.
  • Dataset-specific comparisons: On PolyU, OpenVeinNet retains a clear DET advantage, especially at moderate FMR values, despite greater dispersion across methods.The dataset exhibits more variable method performance than FV-300.
  • Dataset-specific comparisons: On VERA, all methods degrade, but OpenVeinNet performs comparatively better and degrades more gradually under low-sample and high-variability conditions.This indicates stronger generalisation on the most challenging dataset in the DET analysis.
Loading 2608.25515v1…