Source-linked AI summary
A Comparative Study of Fingerprint Image-Quality Estimation Methods
Fernando Alonso-Fernandez, Julian Fierrez, Javier Ortega-Garcia, Joaquin Gonzalez-Rodriguez, Hartwig Fronthaler, Klaus Kollreider, Josef Bigun
TL;DR
Fingerprint verification is vulnerable to image-quality degradation because poor images produce spurious or missing features, motivating quality and validity estimation. This paper surveys and visually illustrates existing measures, compares selected algorithms on a real multisession, multisensor corpus, and studies low-quality rejection. The measures are highly correlated in most cases, and rejecting only 5% of samples can substantially improve verification error rates in the best cases.
Problem
Fingerprint-quality degradation produces spurious and missing features that reduce verification performance, so systems need to estimate captured-image quality and validity.
Method
The paper surveys quality-estimation approaches, provides visual examples, compares selected measures, and evaluates low-quality rejection with a minutiae-based matcher on the BioSec corpus.
Results
High correlation is observed among most tested quality measures, while rejecting 5% of samples improves best-case FRR by 10%, 50%, and 30% for capacitive, optical, and thermal sensors, respectively.
Takeaways & Limitations
Quality estimation can support monitoring, recapture, quality-based adaptation, and rejection of low-quality samples in fingerprint recognition systems.
Abstract
from arXiv · showhide
One of the open issues in fingerprint verification is the lack of robustness against image-quality degradation. Poor-quality images result in spurious and missing features, thus degrading the performance of the overall system. Therefore, it is important for a fingerprint recognition system to estimate the quality and validity of the captured fingerprint images. In this work, we review existing approaches for fingerprint image-quality estimation, including the rationale behind the published measures and visual examples showing their behavior under different quality conditions. We have also tested a selection of fingerprint image-quality estimation algorithms. For the experiments, we employ the BioSec multimodal baseline corpus, which includes 19200 fingerprint images from 200 individuals acquired in two sessions with three different sensors. The behavior of the selected quality measures is compared, showing high correlation between them in most cases. The effect of low-quality samples in the verification performance is also studied for a widely available minutiae-based fingerprint matching system.
I. INTRODUCTION
Fingerprint image quality is treated primarily as utility: the expected impact of an individual sample on recognition performance. The paper reviews quality measures and evaluates their relationship to verification performance using real multisession, multisensor data.
- Quality assessment matters because captured-image quality is affected by factors including user conditions, cuts, dryness or wetness, temperature, dirt, and residual sensor prints.Many of these factors cannot be controlled or avoided.
- Quality measures can monitor acquisition, trigger recapture until satisfactory samples are obtained, or support quality-based adaptation of recognition steps.
- Fingerprint quality is commonly defined by ridge-and-valley clarity and the extractability of identification features such as minutiae, core points, and delta points.
- Experiments use a real multisession and multisensor database to evaluate quality measures through verification error rates after rejecting the lowest-quality samples.The reported metrics are equal error rate, false acceptance rate, and false rejection rate.
- The paper surveys published measures, explains their rationale, provides visual examples, compares correlations among selected measures, and studies rejecting low-quality samples.
II. ALGORITHMS FOR FINGERPRINT IMAGE-QUALITY ESTIMATION
Fingerprint-quality estimation methods are organized by the image information they exploit and by whether they use local or global representations. The reviewed examples derive quality from direction, intensity, symmetry, and filter-based structure.
- Existing approaches are divided into local-feature, global-feature, and classification-based methods.
- Local-feature methods: Local methods commonly partition fingerprints into nonoverlapping square blocks, classify block quality, and aggregate the resulting local scores.
- Visual examples: The figures visualize quality blockwise: brighter blocks indicate higher quality for OCL and symmetry, whereas brighter LOQ blocks indicate greater orientation difference and lower quality.
- Based on the local direction: The OCL measures energy concentration along the dominant ridge direction, while ridge frequency and thickness features detect abnormal ridge structure.
- Based on the local direction: LOQ uses average absolute orientation differences with surrounding blocks, and its global score reflects how smoothly ridge direction changes across the image.
- Based on the local direction: Symmetry-based methods decompose orientation tensors into linear symmetry for coherent ridge flow and parabolic symmetry for high-curvature structures such as minutiae, cores, and deltas.
2) Based on Gabor Filters:
Gabor-filter methods estimate block quality from directional filter responses. Strong ridge direction produces unequal responses, whereas poor-quality or background blocks produce similar responses.
- Gabor filters are applied to each block at different directions to capture directional structure.
- High-quality blocks produce one or more larger directional responses, while poor-quality or background blocks produce similar responses.
- The standard deviation of directional filter responses classifies each block as good or poor.
- The whole-image quality index is the percentage of foreground blocks marked good, and images below a predefined threshold are rejected.
- Poor-quality images can additionally be categorized as smudged or dry.
3) Based on Pixel Intensity:
Pixel-intensity methods assess fingerprint quality from grayscale variation, ridge–valley contrast, directional intensity structure, and clustering. Their block-level decisions are aggregated into an image-quality score or rejection decision.
- Directional block classification: Directional classification methods label foreground blocks directional when a dominant directional histogram exceeds a threshold; otherwise, blocks are nondirectional.
- Directional block classification: Overall scores can assign relative weights to foreground blocks based on distance from the foreground centroid before aggregating block quality.
- Directional block classification: An image is considered poor quality when the resulting score falls below a threshold, with smudginess and dryness also defined for poor-quality images.
- Pixel-intensity measures: Intensity-based methods use gray-level variance or ridge–valley contrast, with larger variance and higher contrast indicating clearer, higher-quality blocks.
- Pixel-intensity measures: One method binarizes blocks with Otsu’s method and uses ridge–valley pixel clustering as a clarity measure.More clustered ridge or valley gray values indicate higher structural clarity and quality.
- Pixel-intensity measures: Another method models ridges and valleys as a sinusoidal wave, segments their gray-level distributions, and uses distribution overlap to measure clarity.
4) Based on Power Spectrum:
These methods analyze a sinusoidal representation of fingerprint ridges and valleys in the direction normal to local ridge orientation. Quality is inferred from dominant-frequency behavior or used to classify blocks by recoverability.
- 5) Based on Power Spectrum:: The method in computes the discrete Fourier transform of a sinusoidal-shaped wave extracted normal to the local ridge direction.Low-quality blocks lack an obvious dominant frequency or have one outside the normal ridge-frequency range.
- 5) Based on Power Spectrum:: Hong et al. extract sinusoid amplitude, frequency, and variance to classify fingerprint blocks as recoverable or unrecoverable.
5) Based on a Combination of Local Features:
Combination-based methods aggregate several local image-quality indicators, while related global-feature methods analyze the fingerprint holistically. NFIS MINDTCT produces quality maps from multiple local structural cues.
- 5) Based on a Combination of Local Features:: MINDTCT assesses each block using direction, low-contrast, low-flow, and high-curve maps.These maps identify sufficient ridge structure, weak contrast, absent dominant ridge flow, and highly curved regions, respectively.
- 5) Based on a Combination of Local Features:: Global-feature methods analyze the image holistically and compute a global quality measure from extracted features.
- 5) Based on a Combination of Local Features:: The local clarity score compares gray-level distributions of segmented ridges and valleys, with lower overlap indicating clearer separation.Figure 5 reports overlap values of 0.22 for the low-quality block and 0.10 for the high-quality block.
- 5) Based on a Combination of Local Features:: NFIS quality maps are generated for fingerprints of different quality to visualize spatial differences in local image conditions.
- 5) Based on a Combination of Local Features:: Lim et al. use direction-field continuity, accumulating abrupt changes between blocks into a global direction score.High-quality fingerprints exhibit smoothly changing ridge directions across the image.
2) Based on Power Spectrum:
Power-spectrum methods quantify fingerprint quality from the distribution and concentration of spectral energy, while classifier-based methods estimate quality from match-versus-nonmatch separability.
- 2) Based on Power Spectrum:: The global quality index uses a 2-D DFT and measures energy concentration within an annular ROI bounded by typical minimum and maximum ridge frequencies.
- 2) Based on Power Spectrum:: High-quality images concentrate spectral energy in a few frequency bands, whereas poor-quality images distribute it more diffusely.Entropy is used to measure this energy concentration.
- 2) Based on Power Spectrum:: The example global quality index assigns values of 0.35 to the low-quality image and 0.88 to the high-quality image.
- 2) Based on Power Spectrum:: Classifier-based quality measures define quality as the degree of separation between genuine-match and impostor-nonmatch score distributions.This formulation treats quality as a prediction of matcher performance.
- 2) Based on Power Spectrum:: In the classifier formulation, the similarity scores represent genuine and impostor comparisons, and the quality measure is defined from their distributional statistics.The cited equation uses expectations and standard deviations to measure separation, which is expected to increase with image quality.
- 2) Based on Power Spectrum:: The BioSec corpus provides fingerprint examples across capacitive, optical, and thermal sensors and across three different fingers.
- 2) Based on Power Spectrum:: A neural-network method outputs five quality classes: 5 poor, 4 fair, 3 good, 2 very good, and 1 excellent.
III. EXPERIMENTS
The experiments compare representative fingerprint-quality measures by their mutual correlation and utility for minutiae-based verification. Testing uses the NFIS2 matcher and its MINDTCT and BOZORTH3 components across BioSec sensor conditions.
- III. EXPERIMENTS: The study compares selected quality measures through their correlation and their effect on a widely available minutiae-based matcher.
- III. EXPERIMENTS: NFIS2 supplies the experimental matcher, with MINDTCT used for minutia extraction and quality assessment and BOZORTH3 used for fingerprint matching.
- III. EXPERIMENTS: Figure 9 plots pairwise quality values and reports Pearson correlations separately for capacitive, optical, and thermal sensors.
- III. EXPERIMENTS: BOZORTH3 matches minutiae templates using minutiae location and direction in a translation- and rotation-invariant manner.
B. Selected Quality Measures
The study selects representative fingerprint-quality measures spanning direction, pixel-intensity, power-spectrum, and classifier-based approaches, and evaluates them on the BioSec corpus across three sensors.
- Selected measures: The selected measures cover direction information, pixel intensity, power spectrum, and classifier-based quality estimation.OCL, LCS, and energy concentration represent the first three feature sources, while NFIQ represents classifier-based assessment.
- Quality representation: All image-quality values are normalized to [0–1], where 0 denotes the worst quality and 1 the best quality.
- Database: The BioSec corpus contains 19 200 fingerprint images from 200 individuals collected in two sessions using capacitive, thermal, and optical sensors.
- Sensors: The capacitive, thermal, and optical sensors use image sizes of 96 × 96, 400 × 496, and a separately specified optical configuration, respectively.
- Evaluation protocol: The development set comprises 50 individuals for parameter tuning, while the remaining 150 individuals form the test set.
- Evaluation protocol: Verification evaluation defines genuine matchings across sessions and impostor matchings between users for each sensor.
D. Results and Discussion
The quality measures are highly correlated in most cases and correlate with genuine, but not usually impostor, similarity scores; rejecting low-quality samples generally improves verification performance, especially FRR.
- Correlation among measures: High correlation occurs among the tested quality measures except when NFIQ is involved, possibly because NFIQ uses a finite number of quality labels.
- Quality and similarity: Genuine similarity scores show some correlation with quality, whereas impostor scores show almost no correlation in most cases.
- Quality and similarity: Low-quality conditions preserve low impostor scores, which is desirable for verification security.
- Verification performance: After rejecting 5% of samples, the best FRR improvements are about 10% for capacitive, 50% for optical, and 30% for thermal sensors.
- Verification performance: After rejecting 5% of samples, the best EER improvements are 3.5% for capacitive, 45% for optical, and 21% for thermal sensors.
- Verification performance: The best FAR improvements are 2.73% for optical and 6.8% for thermal sensors, while some cases show no improvement.
- Sensor dependence: Performance variations are similar for capacitive and thermal sensors across most algorithms but differ for the optical sensor.
IV. CONCLUSION
The conclusion organizes fingerprint-quality estimation methods into local-feature, global-feature, and classification approaches, then relates their behavior to verification performance across sensors.
- Conclusion: Fingerprint-quality estimation methods are grouped into local-feature, global-feature, and classification-based approaches.
- Conclusion: The reviewed methods use direction fields, Gabor filter responses, power spectra, and pixel-intensity values as feature sources.
- Conclusion: The experiments use 19 200 BioSec images from 200 individuals, acquired in two sessions with capacitive, optical, and thermal sensors.
- Conclusion: Quality correlates highly with genuine scores but almost not at all with impostor scores.
- Conclusion: Rejecting low-quality samples produces the highest improvement in false rejection rate at a specified false acceptance rate.
- Future work and scope: Correlation values vary by sensor, so some quality measures may be unsuitable for particular sensors and combinations may provide complementary information.