PaperlayerView exact revision on arXiv

Source-linked AI summary

AFID: A Unified Open Framework for Automated Fingermark Identification, Quality Assessment and Feature Extraction

Tim Oblak, Rudolf Haraksim, Peter Peer

arXiv:2609.07439v1cs.CV

TL;DR

Fingermark research faces scarce data and few open solutions, limiting reproducible comparison. AFID uses a public-data-trained recognition encoder as a shared basis for recognition, quality assessment, and feature extraction. It leads fixed-length identification across three NIST fingermark datasets and outperforms a commercial matcher, while its quality module exceeds compared baselines.

  • Problem

    Fingermark processing is difficult, and the field has limited open solutions, often trained on private or withdrawn data that impede reproduction and direct comparison.

    Source: §I · p. 1

  • Method

    AFID trains a fixed-length identity encoder on public data with heavy augmentation, then derives downstream tasks from the shared representation.

    Source: §I · p. 2, C. Input canonicalization and… · p. 5

  • Results

    AFID leads fixed-length fingermark recognition across NIST SD 27, SD 302, and SD 303, surpasses the commercial VeriFinger matcher, and AFID-Q predicts quality more accurately than every compared baseline.

    Source: §VII · p. 11

  • Takeaways & Limitations

    A single public-data-trained recognition backbone supports AFID’s downstream quality assessment and feature extraction alongside identification.

    Source: §I · p. 2, §VII · p. 11

    Segmentation is the weakest extraction task: trained only on manually annotated masks, its decoder tends to detect all ridge structure rather than isolate the central impression.

    Source: §VII · p. 11

Abstract

from arXiv · show

Automated fingermark identification is the foundation of forensic investigation, yet progress in the field is held back by fragmented, closed-source solutions trained on private or discontinued data. We present AFID, a unified open-source framework for friction ridge image processing that performs recognition, quality assessment, and feature extraction based on a single shared encoder, trained exclusively on publicly available data. At its core is a fixed-length representation learned for identity discrimination, trained under heavy augmentation. Despite applying essentially no preprocessing beyond resizing and padding at inference, AFID sets a new state of the art in fixed-length fingermark recognition, leading identification across NIST SD 27 (67.6% rank-1) , SD 302 (54.9% rank-1), and SD 303 (67.6% rank-1), surpassing a commercial matcher on fingermarks. From the same frozen backbone, a quality assessment module predicts recognition utility more accurately than any compared baseline and generalizes across independent matchers, while lightweight decoders recover minutiae, ridge orientation, and segmentation competitive with dedicated methods. The framework proves that a single, efficiently trained encoder can support the full fingermark processing pipeline, from recognition through quality assessment all the way to feature extraction. To accelerate research on fingermark analysis even further, we release the code, models, and annotations to the community.

I. INTRODUCTION

Fingermarks are difficult to process, and progress is constrained by scarce data, limited open solutions, and missing shared resources for quality assessment and feature extraction. AFID addresses these gaps with a public-data-trained recognition encoder that also supports quality assessment and feature extraction.

I. INTRODUCTION

  • AFID uses one recognition-trained encoder on publicly available data as the shared foundation for recognition, feature extraction, and quality assessment.It reverses conventional feature-first pipelines by learning the identity representation before deriving the downstream tasks.Source: §I · p. 2
  • The fixed-length representation is robust to rotation and scale without explicit alignment, needs essentially only resizing and padding, and leads fixed-length fingermark identification across major datasets.The introduction reports that it surpasses a commercial matcher.Source: §I · p. 2
  • AFID-Q derives a recognition-aligned quality metric from the shared representation that predicts recognition utility more accurately than existing methods and generalizes across independent matchers.It is not merely a matcher-specific confidence score.Source: §I · p. 2
  • Lightweight decoders recover minutiae, ridge orientation, and segmentation competitively with dedicated methods, while retaining local ridge structure in the learned representation.The introduction identifies these as the core fingermark feature set.Source: §I · pp. 2–3
  • The authors release the framework, pretrained models, and annotations as an open reference for fingermark analysis.They also identify it as a candidate reference implementation for the emerging ISO/IEC 29794-12 standard.Source: §I · p. 3

II. RELATED WORK

Related work spans recognition representations, feature extraction, and quality assessment, but these tasks share underlying ridge information. AFID builds on fixed-length recognition representations and derives quality and feature tasks from a shared encoder.

II. RELATED WORK

  • Recognition, feature extraction, and quality assessment operate on mostly overlapping feature sets derived from ridge structure.This shared basis motivates connecting the three processing tasks.Source: §II · p. 3

A. Representation Learning

  • Fixed-length global embeddings enable fast distance-based search, but existing approaches often rely on alignment or proprietary or discontinued training data; AFID instead uses public data, heavy augmentation, and synthetic data without inference-time alignment.Prior examples include DeepPrint, AFR-Net, Pan et al., and FLARE.Source: A. Representation Learning · p. 3

B. Feature Extraction

  • Feature-extraction methods use CNNs and U-Net designs to predict minutiae, orientation, and segmentation, including joint and multibranch approaches.Examples include FingerNet, MinutiaeNet, Finger-UNet, and Nguyen et al.’s SegFinNet.Source: B. Feature Extraction · p. 3

C. Quality Assessment

  • Fingermark quality methods range from NFIQ 2 and handcrafted metrics to expert-labeled models such as LQMetric and pAFQA; recognition-derived quality remains comparatively unexplored.AFID combines an embedding-norm signal with an expert-supervised signal to estimate utility.Source: C. Quality Assessment · p. 3

III. QUALITY-AWARE FIXED-LENGTH REPRESENTATION

  • AFID trains a rotation-, translation-, and scale-invariant identity encoder on public data, then uses it as the shared backbone for downstream tasks.The encoder requires no capture metadata at inference.Source: §III · p. 4

A. Architecture

AFID uses a ConvNeXt-T v2 encoder and recognition head to produce a fixed-length identity embedding, then freezes the encoder for downstream tasks. Its recognition objective combines identity discrimination with embedding-magnitude behavior that supplies a quality signal.

A. Architecture

  • The ConvNeXt-T v2 encoder maps images to multi-scale feature maps, while GeM pooling and batch normalization produce a 768-dimensional identity embedding.The encoder is frozen for downstream feature-extraction and quality-assessment modules.Source: A. Architecture · p. 4

B. Learning objective

  • MagFace makes the angular margin and an auxiliary regularizer depend on embedding magnitude, coupling identity discrimination with quality-awareness.The regularizer encourages larger norms for easy samples and smaller norms for hard ones.Source: B. Learning objective · p. 4
  • No-identity background inputs can fall outside the labeled training distribution, where their embedding norms may no longer track sample quality.The paper notes that recognition training with angular-margin losses requires an identity label for each image.Source: B. Learning objective · p. 4
  • A background regularizer pulls background-sample norms toward a fixed low target; the target is set to τ = 5, below MagFace’s lower norm bound of 10.Background inputs include blank images, public fingermarks without visible ridge content, and synthetically generated everyday surfaces.Source: B. Learning objective · p. 4

C. Input canonicalization and augmentation

AFID uses minimal inference preprocessing and heavy training augmentation to handle variation in friction-ridge images. The frozen encoder also supplies multi-scale features to task-specific decoders for feature extraction.

C. Input canonicalization and augmentation

  • At inference, images are resized to fit 768 × 768, zero-padded, and normalized; no alignment, resolution metadata, or auxiliary signals are required.Training augmentation varies physical area, translation, rotation, occlusion, intensity polarity, blur, brightness, and contrast.Source: C. Input canonicalization and… · pp. 4–5
  • Training samples are rescaled to areas from 1–25 cm2, spanning small partial marks to full rolled impressions; the corresponding inference assumption is a roughly 1–25 cm2 impression.Slap and rolled fingerprints may be occluded centrally or peripherally during training.Source: C. Input canonicalization and… · pp. 4–5

D. Training procedure

  • The recognition backbone is trained for up to 60 epochs, with the best validation-identification checkpoint retained and the MagFace classification head discarded afterward.Training uses AdamW on a single NVIDIA H100 GPU with batch size 48.Source: D. Training procedure · p. 5

IV. FEATURE EXTRACTION

  • AFID-X attaches task-specific decoders to the frozen encoder to predict foreground segmentation, minutiae locations and orientations, and dense ridge orientation.The shared U-Net-style decoder architecture uses multi-scale encoder features and outputs 768 × 768 predictions.Source: §IV · p. 5, A. Decoder architecture · p. 5

A. Decoder architecture

  • The ridge-orientation decoder learns from nearest-minutia Voronoi targets but receives supervision only at minutiae locations, where orientation confidence is high.The figure depicts annotated minutiae, the resulting orientation partition, and the locations used for supervision.Source: A. Decoder architecture · p. 5, Figure 2 · p. 5

B. Prediction heads and supervision

AFID’s frozen encoder supports three feature-prediction heads: segmentation, minutiae detection, and ridge orientation. The orientation head uses sparse, high-confidence minutiae orientations to train predictions across the image.

B. Prediction heads and supervision

  • The segmentation head predicts a single-channel 768 × 768 binary mask separating developed ridge regions from background, supervised by manual masks with per-pixel binary cross-entropy.The output logits are converted through a sigmoid for the mask prediction.Source: B. Prediction heads and supervision · p. 5
  • The minutiae head predicts a 12-channel 768 × 768 heatmap, encoding location and orientation but not minutia type.Each channel spans 30°, and a minutia is represented by a spatial Gaussian with σ = 0.2 mm.Source: B. Prediction heads and supervision · p. 5
  • The ridge-orientation head encodes each pixel’s 180°-symmetric angle as the unit vector (cos 2θ, sin 2θ).Reliable dense orientation labels are unavailable at scale, so training instead uses sparse minutiae orientations.Source: B. Prediction heads and supervision · p. 5
  • A nearest-minutia Voronoi field supplies approximate orientation targets, while a Gaussian confidence map concentrates masked mean-squared-error supervision near annotated minutiae.Despite sparse supervision, the decoder produces smooth, consistent orientation predictions across the image, including far from annotated minutiae; the authors describe this as the first such approach.Source: B. Prediction heads and supervision · pp. 5–6

C. Training procedure

  • Three independent decoders are jointly trained on NIST SD 302 images with a weighted sum of task losses and separate optimizers.The training set contains 5,050 images with minutiae templates and segmentation masks.Source: C. Training procedure · p. 6

V. QUALITY ASSESSMENT

AFID-Q combines an embedding-norm signal with expert-supervised quality estimates from the frozen encoder. Its regressor retains separate examiner scores and averages its ten predictions at inference.

V. QUALITY ASSESSMENT

  • AFID-Q produces a quality score from 0 to 100 by combining an unsupervised embedding-norm signal with a supervised expert-score regressor.The score is calibrated to the operational range specified by ISO/IEC 29794-1.Source: §V · p. 6

A. Quality from the embedding norm

  • The embedding norm serves as an unsupervised quality signal because MagFace training makes recognizable samples’ embeddings larger and background or content-free samples’ norms closer to the origin.The paper describes the norm as varying monotonically with sample quality across the operational range.Source: A. Quality from the embedding norm · p. 6

B. Quality from expert supervision

  • Because public training corpora provide limited fingermark coverage, AFID-Q supplements the norm signal with a regressor trained on expert quality annotations.The annotations cover 1,000 NIST SD 301 and SD 302 images scored independently by ten certified examiners, then rescaled from [1, 10] to [0, 100].Source: B. Quality from expert supervision · p. 6
  • The expert regressor pools features from all four encoder stages into a 512-dimensional descriptor and averages ten predicted expert scores for its final output.Training against the separate examiner scores preserves disagreement among examiners rather than targeting only an averaged consensus.Source: Figure 3 · p. 6, B. Quality from expert supervision · p. 7

C. Combining the signals

The combined quality score multiplies complementary norm-based and expert-supervised signals, with the product giving the best ordering among tested fusion rules. AFID’s experiments use public training data and held-out evaluation sets, while the quality regressor is trained with transformations that preserve visible ridge content.

C. Combining the signals

  • The product fusion gave the best ordering of samples by recognition utility among the tested rules, and either quality signal can veto a low final score.A high score requires agreement between the norm-based and expert-supervised signals; inference needs no capture-condition information.Source: C. Combining the signals · p. 7

D. Training procedure

  • The expert regressor is trained on a frozen backbone using content-preserving transformations, excluding augmentations that crop or discard ridge area.The regressor uses 1,000 annotated images and the checkpoint with the lowest validation loss.Source: D. Training procedure · p. 7

A. Datasets

  • Training uses approximately 68,000 real images across 12,520 fingers, while evaluation includes FVC benchmarks and three fingermark datasets with identities held out from training.The last 200 SD 302 fingers (455 fingermarks) are reserved for evaluation; training data includes LFIW, IIIT-D MOLF, and NIST SD 300–302.Source: A. Datasets · p. 7, §VI · p. 7
  • The dataset pool includes some licensed datasets accessible only upon request to their owners.Reported sample counts reflect cleanup, including removal of mislabeled or corrupted images and samples lacking required annotations.Source: A. Datasets · p. 7, Table I · p. 7

B. Metrics

AFID leads the reported fingermark identification comparisons and performs strongly in fingermark-to-fingerprint verification, while clean fingerprint verification is near saturation. The section evaluates recognition across verification, identification, and decoder and quality-assessment metrics, and describes orientation invariance as a distinction from alignment-based pipelines.

B. Metrics

  • Verification uses EER and TAR at specified FARs, identification uses CMC and rank-n rates, and decoder and quality assessment use task-specific metrics including F1, IoU, RMSD, and nAUC.Fingermark-to-fingerprint verification additionally reports TAR at 0.01% FAR.Source: B. Metrics · p. 8

D. Identification performance

  • 54.9% rank-1 on SD 302 puts AFID ahead of VeriFinger (48.1%), FLARE (26.8%), and FDD (22.0%), while AFID has the highest identification rate on all three datasets.The lead narrows on SD 303, whose upright partial marks involve less geometric variation; Figure 4 shows CMC curves through rank 100.Source: Table IV · p. 8, D. Identification performance · p. 9, Figure 4 · p. 9

C. Verification performance

  • 99.15% TAR@0.1% is AFID’s average across nine FVC databases, close to FLARE’s 99.16% and below VeriFinger’s 99.47%.All methods perform at or near saturation; AFID ranks highest among open methods on FVC 2002 DB3A.Source: C. Verification performance · p. 8
  • 54.51% TAR@0.1% on SD 302 is more than double FLARE’s 23.52% and FDD’s 21.32%, and AFID surpasses VeriFinger on SD 27 and SD 302.AFID has the highest TAR on SD 27 and SD 302 at both operating points and the highest TAR@0.01% on SD 303.Source: C. Verification performance · p. 8

D. Identification performance

  • A direct AFID–AFR-Net comparison is precluded by differing gallery sizes and VeriFinger versions across protocols.AFR-Net used a 100,000-image gallery, while AFID’s comparison used VeriFinger 2025.1 and AFR-Net’s reported reference used VeriFinger v12.3.Source: D. Identification performance · p. 9
  • AFID’s rotation augmentation learns an orientation-invariant representation instead of explicitly aligning each impression to a canonical orientation.The paper reports near-invariance across the full 360° at both embedding and local-descriptor levels.Source: D. Identification performance · p. 9

E. Identity encoder ablation

The ablations identify rotation and synthetic training data as major contributors to recognition performance, with effects varying by dataset. Resolution and loss choice also affect results.

E. Identity encoder ablation

  • Removing rotation augmentation drops mean rank-1 from 63.5% to 50.7%, chiefly through SD 302 falling from 54.9% to 29.9%.SD 303 is essentially unaffected, while the rotation-invariant representation is designed for the rotated marks that dominate SD 302.Source: E. Identity encoder ablation · pp. 9–10
  • Scale augmentation contributes most visibly on SD 27, whereas SD 303 is largely insensitive to augmentation and scale choices.SD 303 consists of upright partial slap marks with limited geometric variation; intensity inversion contributes less overall.Source: E. Identity encoder ablation · p. 10
  • Removing synthetic data lowers mean rank-1 to 50.7% and SD 27 rank-1 from 67.6% to 36.9%.The synthetic marks include substantial background clutter, training the model to learn invariance to background-specific features.Source: E. Identity encoder ablation · p. 10
  • At 768^2 input resolution performance is optimal: 512^2 costs over four mean points, while 1024^2 offers no gain.The 512^2 reduction is attributed to lost ridge detail.Source: E. Identity encoder ablation · p. 10
  • MagFace outperforms ArcFace and AdaFace, while the background regularizer has little effect on recognition.The regularizer is described as stabilizing norms on out-of-distribution inputs rather than aiding identification; AdaFace suppresses hard, low-quality samples.Source: E. Identity encoder ablation · p. 10

F. Minutiae extraction

The frozen recognition encoder supports minutiae, segmentation, orientation, and quality outputs. Results range from competitive feature extraction to quality estimates that lead across tested matchers.

F. Minutiae extraction

  • On SD 27 fingermark probes, AFID-X places third; at the looser threshold its F1 is 68.7%, versus 73.4% for the reported MinutiaeNet result.On SD 27 rolled exemplars, AFID-X reaches an F1 of 85.3%, the highest of tested methods, ahead of VeriFinger and FDD.Source: F. Minutiae extraction · pp. 10–11

G. Segmentation and Local Orientation

  • AFID-X segmentation achieves 59.43% IoU, with a 14.97% missing-detection rate and a 6.53% false-detection rate.The reference masks omit secondary and partial ridge impressions that AFID-X recovers, so the reported IoU understates qualitative mask accuracy.Source: G. Segmentation and Local Orientation · p. 11
  • AFID-X orientation reaches 12.55° average RMSD, on par with RefNet at 12.10° despite receiving no dense orientation-field supervision.It is trained using sparse, high-confidence angle annotations carried by minutiae points and recovers a smooth field across the impression.Source: G. Segmentation and Local Orientation · p. 11

H. Quality assessment

  • AFID-Q obtains the lowest nAUC for all three matchers and leads in 8 of 9 matcher/dataset combinations.On SD 27 with AFID, discarding the worst 20% of marks lowers rank-1 error from 32.4% to 22.0%, closing 61% of the gap to the 15.5% ideal floor.Source: H. Quality assessment · p. 11, Table VII · p. 11

VII. CONCLUSION

AFID combines recognition, quality assessment, and feature extraction on a shared frozen encoder trained on public data. The paper also describes synthetic fingermark generation and identifies remaining limitations in segmentation and quality validation.

VII. CONCLUSION

  • AFID derives quality assessment and feature extraction from a shared frozen recognition encoder trained on public data, achieving leading fingermark identification with essentially no inference preprocessing.Its decoders recover segmentation, orientation, and minutiae at levels comparable to other methods despite no feature-level supervision during recognition training.Source: §VII · p. 11
  • Segmentation is the weakest extraction task: trained on manual masks, its decoder tends to detect all ridge structure rather than isolate the central impression.The authors also call for further AFID-Q analysis across capture conditions and operating points before a validated standard contribution.Source: §VII · pp. 11–12

VIII. SYNTHETIC DATA GENERATION AND AUGMENTATION

  • The authors generate approximately 84,000 synthetic fingermarks from clean rolled or plain prints, masks, minutiae annotations, and sampled surface backgrounds.They create 50 surface photographs for background regularization and synthetic mark generation.Source: §VIII · p. 14
  • Synthetic marks combine transformed ridges with pressure variation, a partial-contact mask, and a randomly cropped and rotated background patch.The deposition opacity combines these effects with the eroded segmentation mask.Source: §VIII · p. 14, Figure 8 · p. 14
  • Image transforms are propagated to minutiae and segmentation labels, and training augmentation adds rescaling, full-range rotation, translation, occlusion, and photometric jitter.Labels falling outside the canvas or below the visibility threshold are discarded.Source: §VIII · p. 14

IX. ADDITIONAL MINUTIAE EXPERIMENTS

Additional experiments find that AFID-X minutiae perform strongly on rolled prints and downstream identification. These results concern specific SD 27 evaluation settings.

IX. ADDITIONAL MINUTIAE EXPERIMENTS

  • AFID-X reaches the highest rolled-print minutiae F1 among evaluated methods: 85.3% at the looser threshold.The original MinutiaeNet paper did not report this benchmark, so its result applies only to the publicly released implementation.Source: Table VIII · p. 15, §IX · p. 15
  • With MCC, AFID-X minutiae achieve 58.1% rank-1 identification versus 38.4% for VeriFinger, the highest rate of any automated extractor.This experiment uses the native SD 27 gallery and 258 probes; its absolute rank-k rates are not comparable to the main-text extended-gallery results.Source: Table IX · p. 15, §IX · p. 15

X. ROTATION INVARIANCE OF AFID EMBEDDINGS

AFID’s embeddings remain similar across large rotations, including in local descriptors; its norm regularizer suppresses no-ridge inputs while preserving real-impression quality ordering. On FVC fingerprints, AFID-Q has the lowest nAUC across all tested matchers.

X. ROTATION INVARIANCE OF AFID EMBEDDINGS

  • At the worst-case 180° offset, AFID’s mean cosine similarity stays above 0.93 on fingermark probes and 0.97 on rolled gallery prints; FLARE collapses beyond roughly ±45°.Images are rotated in 30° increments and compared with their unrotated embeddings.Source: §X · p. 16, Figure 10 · p. 16
  • Local descriptors retain high similarity across large print–fingermark rotations; RANSAC selects correspondences for a visualization-only transform, while recognition uses no prior alignment.The transform aligns the fingermark to the exemplar frame and is computed post-hoc purely for visualization.Source: §X · p. 16, Figure 11 · p. 16

XI. EFFECT OF THE BACKGROUND REGULARIZER ON EMBEDDING NORMS

  • The background regularizer pulls degenerate and background norms toward τ=5 while leaving genuine mark and print norms essentially unchanged and preserving their quality ordering.The resulting norm gap also lets ∥y∥ serve as a simple detector of visible friction ridges.Source: §XI · p. 17

XII. QUALITY PREDICTION ON FINGERPRINTS

  • 0.321 mean nAUC: AFID-Q is lowest under every matcher across nine FVC databases, versus 0.525 for VeriFinger’s quality value and 0.754 for NFIQ 2.The EDC curves show AFID-Q at or below the baselines in nearly every database–matcher panel.Source: §XII · p. 17, Figure 13 · p. 18
Loading 2609.07439v1…