PaperlayerView exact revision on arXiv

Source-linked AI summary

FVeinSyn: Synthetic Finger Vein Image Generator

Yifan Wang, Jie Gui, Adams Wai Kin Kong, Baosheng Yu, Changsheng Chen, Qi Li, Zhenan Sun, James Tin-Yau Kwok, Alex Kot

arXiv:2608.27527v1cs.CVcs.AI

TL;DR

Finger vein recognition lacks large-scale, diverse public datasets, limiting the evidence available for learning broad identity and appearance distributions. FVeinSyn decouples vascular-topology synthesis from near-infrared rendering and adds controllable intra-class variation. The framework generates 500,000 images across 10,000 identities and is reported to improve realism, diversity, consistency, and downstream recognition performance.

  • Problem

    Finger vein datasets have limited identities, samples, and variation, while real data are insufficient to model broad vascular topologies and near-infrared appearances.

    Source: §I · p. 2, A. Finger Vein Recognition · p. 3

  • Method

    FVeinSyn combines a constrained stochastic L-System identity generator, a region-aware cascaded renderer, and an intra-class diversity generator.

    Source: §I · pp. 2–3, B. Finger Vein Identity Generator · p. 6

  • Results

    500,000 images across 10,000 vein identities were generated, with evaluations reporting high realism and diversity plus benefits for downstream recognition.

    Source: C. Image Quality Assessment · p. 9, §V · pp. 15–16

  • Takeaways & Limitations

    FVeinSyn provides large-scale, controllable synthetic finger vein data with identity-consistent patterns, realistic rendering, and modeled intra-class variation.

    Source: §I · p. 3, §V · p. 15

Abstract

from arXiv · show

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.

I. INTRODUCTION

Finger vein recognition is constrained by scarce, low-diversity public data and difficult synthesis requirements. FVeinSyn addresses these challenges by decoupling topology generation, near-infrared rendering, and intra-class variation modeling in a controllable framework.

Motivation

  • Public finger vein datasets contain fewer than 1K identities, typically 2–12 impressions per finger, and minimal intra-class variation.These limitations hinder evaluation of large-scale retrieval, cross-device generalization, and robustness.Source: §I · p. 2
  • Limited real samples cannot cover broad vascular-topology variability or jointly model identity distribution and near-infrared appearance.Adding vein patterns, imaging characteristics, and finger geometry further increases generation complexity.Source: §I · p. 2

Framework

  • FVeinSyn decouples identity topology generation from near-infrared appearance rendering while using controllable structural and degradation priors.The framework targets realistic complete images with rich inter-class and intra-class variation.Source: §I · p. 2
  • A stochastic L-System vascular identity generator uses physiological and topological constraints to produce anatomically plausible, identity-distinctive vascular patterns.The design is based on probabilistic production rules inspired by vascular bifurcation and constrained by branching angles, diameter decay, and curvature smoothness.Source: §I · pp. 2–3
  • A region-aware cascaded renderer maps binary vascular topologies to realistic near-infrared images through multi-region adversarial generation.Assigning anatomical structural priors such as finger shape and joint-cavity regions strengthens identity distinctiveness.Source: §I · pp. 2–3
  • An intra-class diversity generator simulates realistic variability using controllable geometric, photometric, and imaging transformations.Unlike methods focused on localized vein regions, FVeinSyn generates complete finger vein images and explicitly models intra-class diversity.Source: §I · p. 3

Evaluation

  • Extensive experiments report improved downstream recognition performance and greater identity scale and diversity than existing state-of-the-art generation methods.The introduction presents this as the principal empirical benefit of FVeinSyn-generated datasets.Source: §I · p. 3

II. RELATED WORK

Prior finger vein recognition methods progressed from handcrafted descriptors to deep learning, but public data remain too small and homogeneous for broad generalization. Synthetic generation addresses this gap, although existing methods struggle with realistic, anatomically coherent, full-finger images and diverse identities.

Finger Vein Recognition

  • Handcrafted methods use descriptors and matching algorithms, but are sensitive to lighting, finger misalignment, and imaging noise.Representative approaches include repeated line tracking, maximum curvature, Gabor filtering, IUWT, and wide line detection.Source: A. Finger Vein Recognition · p. 3
  • Deep learning improved recognition accuracy and robustness through architectures, loss functions, prior knowledge, preprocessing, domain adaptation, and robustness methods.Architectural trends include deeper backbones, feature pyramids, full-view processing, and attention mechanisms.Source: A. Finger Vein Recognition · p. 3
  • Public datasets remain insufficient in identities, samples per identity, and acquisition sessions, while posture, scattering, occlusion, and device differences create complex intra-class variation.These challenges limit generalization in cross-domain, few-shot, and large-scale retrieval scenarios.Source: A. Finger Vein Recognition · p. 3

Synthetic Finger Vein Generation

  • Synthetic biometric research includes intra-class generation for known identities and inter-class generation for new identities with corresponding multi-modal samples.Most finger vein studies still focus on intra-class generation, with few addressing inter-class generation.Source: B. Synthetic Finger Vein Generation · pp. 3–4
  • Analytical modeling established synthetic finger vein generation but showed limitations in visual realism, geometric continuity, and intra-class diversity.Its variations were mainly introduced through limited parameter perturbations.Source: B. Synthetic Finger Vein Generation · p. 4
  • Finger vein synthesis requires fine anatomical reconstruction and modeling of near-infrared optical effects, making high-fidelity full-finger generation more difficult than palm or dorsal-hand synthesis.Existing research mostly focuses on local ROI-level synthesis, whereas full-finger images offer greater potential for identity modeling and multi-task recognition.Source: B. Synthetic Finger Vein Generation · p. 4

C. Biometrics with Synthetic Dataset

Synthetic data can expand biometric training while reducing privacy and distributional-balance concerns, but synthetic-only recognition remains limited by realism and domain gaps. FVeinSyn addresses these challenges through constrained topology generation, region-aware rendering, and controllable synthesis stages.

  • Synthetic data can be expanded at scale without violating personal privacy, supporting ethical compliance and distributional balance.Source: C. Biometrics with Synthetic Dataset · p. 4
  • Synthetic-only recognition systems remain limited because texture fidelity and domain-distribution gaps restrict generalization to real-world test sets.Source: C. Biometrics with Synthetic Dataset · p. 4
  • FVeinSyn generates a binary vascular topology under physiological and geometric constraints before rendering it into a near-infrared image with structural priors.Source: §III · p. 4
  • The vascular model enforces flow conservation and Murray-type power-law scaling while using bifurcation angles, lengths, diameters, and flows as geometric parameters.Source: A. Preliminary · pp. 4–5
  • Its framework separates topology generation, near-infrared appearance rendering, and intra-class variation enhancement into three controllable stages.Source: Figure 4 · p. 5
  • Stochastic L-systems introduce probabilistic production rules and parameterized growth variations, while adversarial image transfer learns mappings between image domains with cycle consistency.Source: A. Preliminary · p. 5

B. Finger Vein Identity Generator

The Finger Vein Identity Generator models vascular formation as a stochastic, biologically constrained growth grammar. Recursive bifurcation, smooth curve construction, and rasterization produce controllable topology maps with anatomical plausibility and identity diversity.

  • The generator balances anatomical plausibility and topological diversity so each identity represents a physiologically valid, distinctive vascular pattern.Source: B. Finger Vein Identity Generator · p. 6
  • A probabilistic growth grammar implemented with a symbolic L-system uses anatomical and geometric priors to generate identity-distinctive vascular topologies.Source: B. Finger Vein Identity Generator · p. 6
  • Bezier interpolation smooths turtle trajectories before rasterization into a binary topology map whose vessel thickness follows segment diameters.Source: B. Finger Vein Identity Generator · p. 6
  • Recursive bifurcation begins from an initial trunk and applies a branching rule with probability p, while non-branching growth occurs with probability 1−p.Source: B. Finger Vein Identity Generator · p. 6
  • Inter-level length attenuation and diameter scaling constrain branch growth, with diameter attenuation chosen to approximately satisfy the Murray-type law.Source: B. Finger Vein Identity Generator · p. 6
  • Generated networks typically contain three to four vertically distributed primary trunks and secondary branches with physiologically consistent asymmetry and density.Source: B. Finger Vein Identity Generator · p. 6

C. Finger Vein Image Renderer: CascadedRA-GAN

CascadedRA-GAN renders binary vein identity patterns into realistic near-infrared images while preserving structural constraints. Its cascaded regional supervision combines global distributional alignment with local shape, brightness, and vein-pattern fidelity.

D. Intra-Class Diversity Generator

The intra-class diversity generator applies geometric perturbations to topology maps and optical degradations to rendered grayscale images, producing controllable variation across synthetic samples.

D. Intra-Class Diversity Generator

  • The two-stage ICDG applies geometric perturbations in the binary topology domain before adding optical degradations during grayscale rendering.This separates pose and viewpoint simulation from near-infrared acquisition effects.Source: D. Intra-Class Diversity Generator · p. 7
  • Composite transformations are ordered compositions of selected elementary operators from either the geometric or optical transformation family.Activated subsets and their parameters define each composite mapping.Source: D. Intra-Class Diversity Generator · p. 7
  • The ICDG pipeline transforms an identity topology map, renders it with structural priors, and then applies optical perturbations to produce the final image.The formulation is ˜Ip = TG(Ip; θG), Isyn = G(˜Ip, S), ˆI = TO(Isyn; θO).Source: D. Intra-Class Diversity Generator · p. 7

1) Geometric Perturbations (Binary/Topology Stage):

2) Optical Degradation (Grayscale/Render Stage):

The grayscale-stage operators simulate optical and acquisition variability through exposure changes, skin scattering, motion blur, and optical blur.

2) Optical Degradation (Grayscale/Render Stage):

  • The exposure operator applies a centrally symmetric, exponentially decaying field centered at a random location within the finger region.Its maximum exposure exponent controls the strength of the local exposure effect.Source: 2) Optical Degradation… · p. 8
  • Exposure modifies image intensity through a spatially varying multiplicative gain applied to the grayscale image.The gain is defined from the exposure field and the result is constrained to the image intensity range.Source: 2) Optical Degradation… · p. 8
  • The skin-scattering operator uses an atmospheric scattering model to simulate subsurface scattering and haze in near-infrared imagery.Its formulation includes spatially varying transmittance and a scattered-image computation.Source: 2) Optical Degradation… · p. 8
  • Motion blur is simulated by convolution with a normalized direction-dependent linear kernel, while optical blur uses a Gaussian point spread function.The two operators model relative motion and imaging-system defocus or residual blur, respectively.Source: 2) Optical Degradation… · p. 8

3) Triggering Strategy:

ICDG samples a variable number of perturbation operators and selects them without replacement, creating mostly mild variations alongside fewer strongly distorted samples.

3) Triggering Strategy:

  • The ICDG uses a two-stage probabilistic triggering strategy to select operators for each sample.It first samples the number of active operators and then selects a subset using normalized weights.Source: 3) Triggering Strategy · p. 8
  • Each operator’s non-negative base weight is normalized into a selection probability within its transformation family.The normalized probabilities determine the relative tendency of operators to be activated.Source: 3) Triggering Strategy · p. 8
  • A truncated Poisson variable controls the number of active operators, yielding mostly lightly perturbed samples and a smaller number with stronger distortions.The authors state that this long-tailed distribution better reflects intra-class variability in real finger vein imagery.Source: 3) Triggering Strategy · p. 8
  • The selected operators are sampled without replacement to form the activated transformation set.This prevents the same operator from being selected repeatedly within one sample’s activated set.Source: 3) Triggering Strategy · p. 8
  • The evaluation covers realism, image quality, recognition, identity uniqueness, vascular-pattern consistency, and intra-class diversity.Recognition experiments include open-set, cross-domain, and sample- or identity-limited settings.Source: §IV · p. 8

B. Visualization Analysis

Visualization and quality analyses compare FVeinSyn with real datasets and synthesis methods across appearance, vascular-pattern extraction, image-quality distributions, and human judgments of realism.

B. Visualization Analysis

  • FVeinSyn samples show anatomically consistent main veins, coherent branching, and plausible overall topology in visual comparisons with real datasets.The comparison includes eight real datasets and two columns of FVeinSyn-generated images.Source: B. Visualization Analysis · p. 9
  • Classical extraction methods produce synthetic pattern distributions closely matching real images in pattern locations, line connectivity, and local response intensity.The methods include RLT, maximum curvature, Gabor filtering, and IUWT.Source: B. Visualization Analysis · p. 9

C. Image Quality Assessment

  • FVeinSyn’s GCF distribution is closer to high-quality real datasets such as UTFVP and MMCBNU 6000, while its entropy distribution is also highly consistent with them.GCF reflects vein-to-tissue contrast, and entropy reflects information representation and statistical complexity.Source: C. Image Quality Assessment · pp. 9–10, Figure 7 · p. 10

B. Visualization Analysis

E. Recognition Performance Evaluation

Across open-set, cross-domain, and sample-limited evaluations, FVeinSyn consistently improves finger vein recognition over real-data-only and other synthetic-data baselines. Its synthetic data also shows strong identity distinctiveness, intra-class diversity, and alignment with real matching-score distributions.

Open-Set Finger Vein Recognition

Cross-Domain Recognition

Sample-Limited Recognition

Synthetic Dataset Evaluation

G. Ablation Study on Intra-Class Diversity Generator

The ablation study finds that geometric perturbations and optical degradations each improve recognition, while their combination models complementary structural and appearance-level variation. Generated intra-class variation counts also follow a long-tail distribution consistent with real-world data.

Transformation Contributions

  • All individual geometric and optical transformations improve recognition performance over the real-data baseline.The study evaluates isolated transformations using synthetic pretraining datasets under the open-set recognition setting.Source: G. Ablation Study on Intra-Class… · p. 14
  • Geometric perturbations introduce structural variability, while optical degradations model appearance-level changes.The two transformation categories contribute in a complementary manner.Source: G. Ablation Study on Intra-Class… · p. 14
  • The mixed setting supports empirically weighting structural and appearance-level variations jointly.Shift, rotation, and skin scattering receive higher weights, while exposure receives a lower weight.Source: G. Ablation Study on Intra-Class… · p. 14

Variation Distribution

  • The number of intra-class variations per generated sample follows a long-tail distribution consistent with real-world data.Most samples contain only a few variations, while samples with more variations gradually decrease.Source: G. Ablation Study on Intra-Class… · p. 14

H. Demographic Bias Analysis

The demographic analysis examines whether FVeinSyn pretraining affects recognition differently across demographic groups. Reported subgroup performance remains close to overall performance, but the renderer’s dependence on real data leaves potential demographic-bias inheritance or amplification as a limitation.

Bias Limitation

  • Potential demographic biases in real renderer-training data may be inherited or amplified by generated finger-vein data.The paper identifies this as important because such bias may affect downstream generalization across demographic groups.Source: H. Demographic Bias Analysis · p. 14

Subgroup Results

  • Subgroup recognition performance is generally close to overall performance across gender, age, race, and blood-type groups.UTFVP is evaluated by gender and age; MMCBNU 6000 is additionally evaluated by race and blood type.Source: H. Demographic Bias Analysis · p. 15, Table XI · p. 15
  • FVeinSyn pretraining produces relatively stable performance gains across the analyzed demographic groups.No specific group shows a persistent and significant deviation from the overall trend in the reported results.Source: H. Demographic Bias Analysis · p. 15

Evaluation Scope

  • The framework generates 500,000 images across 10,000 vein identities for recognition evaluation.The conclusion reports improved recognition, cross-domain robustness, and data-limited performance relative to real-data baselines.Source: §V · p. 15

Related papers

This paper cites

Loading 2608.27527v1…