PaperlayerView exact revision on arXiv

Source-linked AI summary

Infant Fingerprints: Quality Assessment & Recognition

João L. H. Dall Agnol, Dalcimar Casanova, Rodrigo Minetto, Luiz F. P. Southier, Marcelo Teixeira, Marcelo Filipak, Anil K. Jain

arXiv:2610.20432v1cs.CV

TL;DR

Infant-specific quality assessment is limited, although adult-derived structural criteria may not fit infant fingerprints. This paper evaluates quality methods and tunes parameters on longitudinal data across infant ages. Tuned configurations improve mean TMR over defaults across method-age combinations, while fusion and latent comparison substantially raise matching rates in the evaluated setting.

  • Problem

    Adult fingerprint structural bounds may not define acceptable infant samples, and infant-specific quality assessment remains largely unexplored.

    Source: §I · p. 1, §II · p. 2

  • Method

    The study searches age-specific parameters for fingerprint quality methods and evaluates tuned and default configurations through quality-guided gallery selection.

    Source: §V · p. 9, §VI · p. 9

  • Results

    At FMR=1.0%, tuned configurations obtained higher mean TMR than defaults in every method-age combination.

    Source: §V · p. 8

  • Takeaways & Limitations

    Existing biometric frameworks can be adapted to infant subjects through parameter search when representative data is available.

    Source: §VI · p. 10

Abstract

from arXiv · show

Infant fingerprint recognition is hindered by a combination of physiological characteristics, such as small ridge width and spacing, and acquisition challenges that include involuntary motion, skin residue, and uneven fingertip pressure. These factors undermine fingerprint quality standards originally designed for adults, which assume feature distributions that do not hold for infants. To address this gap, this paper evaluates ten fingerprint quality assessment methods on a longitudinal dataset that spans six age groups from newborns to three months, alongside a parameter tuning routine that uncovers age-specific fingerprint characteristics to recalibrate methods with domain-dependent assumptions. Both default and tuned configurations are compared through recognition experiments assessing the effectiveness of quality-guided gallery selection. Template-fusion strategies and latent fingerprint recognition technologies are further examined to identify the best possible matching rates achievable by the top-performing method. Results show that tuned parameter sets outperform default configurations for the majority of methods and age groups, indicating that targeted recalibration of existing frameworks improves gallery utility and contributes to more reliable infant fingerprint recognition.

I. INTRODUCTION

Infant fingerprint quality standards and methods face a gap because infant ridge structure differs from adult fingerprints, while infant-specific quality assessment remains limited. This study evaluates established methods and age-sensitive tuning on longitudinal infant data, then tests quality-guided gallery selection and matching with fusion and a latent comparator.

I. INTRODUCTION

  • Adult fingerprint criteria may not define acceptable infant samples because ridge and valley structural bounds need to be reconsidered by age.Infant ridges are thinner, spacing can be up to four times smaller, and softer, more elastic fingers are susceptible to capture deformation.Source: §I · p. 1
  • Quality assessment for infants remains relatively unexplored, with prior work largely applying adult-oriented NFIQ methods to infant data.Other infant-biometrics work has addressed high-resolution scanners and image enhancement, including binarization and super-resolution.Source: §I · p. 1
  • The experiments compare standard and optimized configurations through quality-ranked gallery selection across six age subsets from newborns to three months.The study also examines template fusion with a latent comparator to investigate matching rates for a top-performing quality method.Source: §I · pp. 1–2
  • The study evaluates ten quality assessment methods and introduces age- and method-specific parameter search to identify infant ridge and valley ranges that maximize gallery utility.Methods span classical signal-processing, structure-driven, and learning-based approaches, using local and global image features.Source: §I · p. 2

II. FINGERPRINT QUALITY ESTIMATION METHODS

Fingerprint quality methods differ in the features they measure, how they extract them, and their parameter assumptions. The review organizes these methods by adaptability to infant fingerprints while noting resolution constraints and the limited infant-specific evidence.

II. FINGERPRINT QUALITY ESTIMATION METHODS

  • The taxonomy distinguishes explicitly tunable methods from fixed, learned, or image-statistics approaches, and groups tunable methods by the strength of their fingerprint-specific assumptions.Strong-assumption examples include Gabor-based methods, LCS, and RPS; weaker-assumption examples include OCL, RVU, and DFB.Source: §II · p. 2
  • Most methods assume sampling near 500 PPI, while infant fingerprints require higher resolution; this study analyzes 5000-PPI images downsampled to 500 PPI for COTS compatibility.Pore-based quality indicators are an exception, but infant pores are rarely visible in the first months.Source: §II · p. 2
  • Methods are also classified as global, with one score per image, or local, with separate regional scores; local scoring can support enhancement and composition.Global approaches remain widely adopted, largely due to NFIQ2.Source: §II · p. 2
  • Infant fingerprint quality assessment remains largely unexplored, and no existing study specifically targets it, amid scarce infant datasets and limited synthetic fingerprint generation.Existing quality-estimation surveys focus on adult fingerprints.Source: §II · p. 2

III. PARAMETER SEARCH FOR INFANTS

The parameter search uses longitudinal infant fingerprint data to examine age-related patterns and adapt tunable quality methods. Its dataset protocol selects subjects with complete six-session coverage, and its search space targets parameters affecting ridge structure, spatial scale, and frequency.

III. PARAMETER SEARCH FOR INFANTS

  • The search investigates whether parameters derived from adult feature distributions capture infant fingerprints from hours after birth through three months.It uses the longitudinal UTFPR-IFD dataset and considers methods with tunable parameters.Source: §III · p. 3, B. Parameter Search Space · p. 3

A. Infant Dataset and Experimental Protocol

  • The curated dataset includes 45 infants, 3,780 video sequences, and 135,382 frames across six sessions, with subject-disjoint five-fold validation applied consistently across age groups.Subjects had complete coverage at every milestone and all 14 expected video samples per session; training frames were subsampled to reduce temporal redundancy.Source: A. Infant Dataset and Experimental… · p. 3

B. Parameter Search Space

  • The search tunes parameters affecting ridge structure, spatial scale, and frequency, excluding methods without tunable parameters.For frequency-tuned filters, f spans 2–20 pixels; Gaussian deviation σ is computed from the chosen frequency, and the number of orientations and angular-window deviation are also tuned.Source: B. Parameter Search Space · pp. 3–4

2) LCS:

The LCS and RPS search spaces encode infant-relevant ridge and valley dimensions and dominant frequency using bounds that extend adult reference ranges.

2) LCS:

  • LCS evaluates ridge–valley clarity from block-level average ridge and valley widths, tuning each width interval with a lower bound and displacement.Bounds range from 1 to 20 pixels, extending adult ranges to accommodate narrower valleys and relatively wider infant ridges after downsampling.Source: 2) LCS · p. 4

3) RPS:

  • RPS measures energy concentration around dominant ridge frequency, tuning frequency bounds plus radial distance and angular resolution.The frequency range extends from 0.01 to 0.70, compared with the adult reference rmin of approximately 0.077; radial distance spans 1–100 and angular resolution 60–180 degrees.Source: 3) RPS · p. 4

4) Grid-search methods:

Grid-search and Bayesian strategies tune fingerprint-quality parameters according to their search-space size, using FNMR-based utility to assess whether quality ranking selects more informative galleries.

4) Grid-search methods:

  • Grid-search methods OCL, RVU, and DFB vary local block size, as does LCS; larger blocks may stabilize estimates by including more ridges and valleys.The evaluated blocks ranged from 4 × 4 to 128 × 128 pixels.Source: 4) Grid-search methods · p. 4

1) Gabor:

  • Block sizes of 256 × 256 pixels and above were excluded because they would cover about one quarter of the fingerprint at 500 PPI, compromising local analysis.DFB also evaluates multiples of 2 from 4 to 128 because its base block is 4 × 4.Source: 1) Gabor · pp. 4–5

C. Parameter Optimization

  • Bayesian optimization handles methods with multiple parameters, while exhaustive grid search is reserved for block size.The search used 10 initial observations and 20 further iterations; block sizes ranged from 4 to 128 in powers of 2.Source: C. Parameter Optimization · p. 5
  • Lower FNMR on the retained genuine pairs indicates that a quality method better predicts biometric utility after the lowest-quality 20% of pairs are discarded.The threshold is fixed where 80% of initial training-set genuine pairs match.Source: C. Parameter Optimization · p. 5

IV. EXPERIMENTAL RESULTS

Across the cross-validation experiment, tuned configurations reduce FNMR relative to defaults, with the size of the gain varying by method. The parameter analysis also describes age-related ridge characteristics and illustrates how tuning can change frame selection.

A. Parameter Tuning

  • Table II reports tuned RPS and LCS parameter intervals by age group from a separate execution using all 450 digits, with adult ranges included for comparison.The parameters relate to ridge-valley features that influence matching performance across ages.Source: Table II · p. 5, A. Parameter Tuning · p. 5
  • The dominant ridge frequency is higher for infants and increases with age over the observed period, partly as poorly defined young-infant valleys make ridges appear wider.At 15 days, estimates already align with older age groups; at one month, the lower ridge-width bound decreases, and early-age fluctuations are noted.Source: A. Parameter Tuning · p. 5
  • Tuned configurations consistently reduce FNMR relative to defaults across methods, with LCS more responsive to tuning than Gabor and OCL.Figure 4 shows held-out-subject results, using parameters optimized on each fold’s training data.Source: A. Parameter Tuning · p. 5, Figure 4 · p. 6
  • For a newborn thumb, tuned LCS selects a more informative frame, whereas the default configuration selects a blurred frame captured mid-movement.Figure 5 presents the fingerprint core region for the best-scored frames from the test-only subset.Source: A. Parameter Tuning · p. 5, Figure 5 · p. 6

B. Recognition Results

Tuning generally improved quality-guided matching across age groups, while the strongest methods and recognition rates varied by age. Newborns and 7-day infants remained harder to recognize than older groups.

B. Recognition Results

  • RPS had the lowest mean rank (2.8), led 2- and 3-month groups at both FMRs, and led 15-day infants at FMR=0.1%.It was less competitive for newborns: 47.5% versus 53.6% for LCS at FMR=1.0%.Source: B. Recognition Results · p. 6, Table IV · p. 7
  • Tuned LCS significantly improved its mean rank by 5.6 and had the highest mean TMR at FMR=1.0% for infants aged 15 days or younger.Its margins over the next-best method were within fold standard deviation; at FMR=0.1%, the 1-month group was the sole exception to improvement over default.Source: B. Recognition Results · pp. 6–7
  • 36/36 tuned method-age combinations had higher mean TMR at FMR=1.0%, versus 32/36 at FMR=0.1%.The exceptions at FMR=0.1% were Gabor in three age groups and LCS at 1 month.Source: B. Recognition Results · p. 6, Table III · p. 7
  • Gabor tuning raised mean TMR in all six age groups at FMR=1.0%, but its rank difference of 2.3 was below the critical difference.At FMR=0.1%, default Gabor outperformed tuned Gabor in three age groups, by no more than 0.9 point.Source: B. Recognition Results · p. 7
  • Tuned DFB improved mean TMR over default in all six age groups at FMR=0.1% and ranked above NFIQ2, while RVU’s gains did not lift it above NFIQ2.RVU’s rank gain was significant; OCL’s TMR gains were capped at 2.1 points and its mean rank difference was below the critical difference.Source: B. Recognition Results · p. 7
  • LCS exceeded NFIQ2’s mean TMR at FMR=1.0% in all six age groups, supporting consideration of tuned features such as LCS, RPS, and Gabor for a child-focused composite.LCS’s rank gap to NFIQ2 was 3.9, slightly below the critical difference; the paper also suggests RPS and Gabor as potential additions.Source: B. Recognition Results · pp. 7–8
  • Newborn and 7-day samples had considerably lower recognition rates than older groups, while differences after 15 days were comparatively small.The 3-month group fluctuated to slightly lower rates despite having the oldest samples.Source: B. Recognition Results · p. 8, Figure 6 · p. 8

C. Improvements to Matching Rates

Frame-fusion and latent matching improve infant fingerprint matching rates, with the strongest reported mean at two months. Newborn matching remains difficult, and the experiment does not isolate how much tuning contributes to fusion gains.

C. Improvements to Matching Rates

  • Frame-fusion improved on single-frame results in every age group, and latent matching exceeded the VeriFinger baseline in all cases.Frame-fusion accepted a probe when it matched at least two frames of the same digit gallery.Source: C. Improvements to Matching Rates · p. 8
  • At 2 months, fusion and latent matching raised mean TMR from 90.5% (single-frame VeriFinger) to 97.6%; newborn MegaMatcher gained 29.7 points but remained the most challenging group.The 2-month result had a standard deviation of 0.9 points; the newborn example was RPS, rising from 47.5% to 77.2%.Source: Table V · p. 8, C. Improvements to Matching Rates · p. 8
  • From 15 days onward, MegaMatcher TMR exceeded 96% for all listed methods, supporting video sampling as a viable way to capture useful fingerprints from early life.The experiment did not include default- or NFIQ2-selected galleries, so it does not isolate tuning’s contribution to fusion gains.Source: C. Improvements to Matching Rates · p. 8
  • During the first week, reliable capture and evaluation remain challenging; 100% TMR may be unattainable at reasonable FMR without a Failure-To-Enroll rule.The authors suggest real-time quality evaluation and collector feedback could identify cases needing additional acquisitions.Source: C. Improvements to Matching Rates · p. 8
  • Involuntary movement, acquisition noise, and unsuitable pairings can prevent matching even for relatively clear frames; quality thresholds trade TMR gains against exclusions.The passage illustrates this with a false reject involving newborn little fingers.Source: C. Improvements to Matching Rates · p. 8

V. DISCUSSION

Tuning improves mean TMR over default configurations across every method-age combination, though gains and rankings vary by method. Matching is lower for newborns, while fusion and latent comparison raise rates; no single top-five method is statistically superior.

V. DISCUSSION

  • Tuned configurations achieved higher mean TMR than defaults in every method-age combination at FMR=1.0%.The comparison indicates that parameters derived from adult fingerprint norms limit the applicability of otherwise suitable methods to infants.Source: §V · p. 8
  • The illustrated genuine newborn thumb pair exceeded the MegaMatcher threshold, whereas the little-finger pair did not.Source: Figure 7 · p. 9
  • LCS and RPS rank gains exceeded the critical difference at both operating points, and together they had the highest mean TMR in five of six age groups.Their margins were within fold standard deviation, and neither differed significantly from the other top-five methods.Source: §V · p. 9
  • Gabor tuning gained consistently at FMR=1.0% but not FMR=0.1%, while untunable SIG and MQCDA ranked among the top five at FMR=1.0%.SIG and MQCDA together had the highest mean TMR at FMR=0.1% in three of six age groups; block-size tuning gave varying gains for RVU and OCL, neither surpassing standard NFIQ2.Source: §V · p. 9
  • Newborn matching rates were notably lower, while frame fusion with a latent comparator raised the 2-month frame-comparison TMR from 90.5% (RPS) to 97.6%.Rates for ages 15 days to three months were 97.4–97.5%, within fold standard deviation; the fusion rule requires two matching frames of the same finger.Source: §V · p. 9
  • Friedman–Nemenyi analysis found no single superior method among the top five, although they ranked consistently above most remaining methods and configurations.A composite quality index combining complementary properties such as orientation, width, and frequency is proposed as a promising direction.Source: §V · p. 9

VI. CONCLUSIONS

The conclusions identify adult-derived parameters as a poor fit for infant fingerprint characteristics and report that recognition-directed parameter search improves matching across evaluated ages. The resulting parameter sets offer a foundation for future infant recognition work, while broader age ranges and methods remain open research directions.

VI. CONCLUSIONS

  • Recognition-directed parameter search identified optimal ranges for features including ridge and valley width, and tuned parameters improved TMR@FMR across all evaluated age groups.The conclusions state that adult-standard parameters are insufficient for infant fingerprints.Source: §VI · p. 9
  • Existing biometric frameworks can be adapted to infant subjects when representative data is available, and the optimized parameter sets provide a foundation for future maternity-setting solutions.The study followed subjects from birth to three months using a longitudinal dataset.Source: §VI · pp. 9–10
  • Future work includes studying ridge and valley changes beyond the observed ages, age-aware standards, infant composite quality indexes, and high-resolution estimation and matching methods.Source: §VI · p. 10

APPENDIX A COMPLETE PARAMETER SET

Appendix A presents the tuned parameter sets in two parts.

APPENDIX A COMPLETE PARAMETER SET

Loading 2610.20432v1…