LEADER: Lightweight End-to-End Attention-Gated Dual Autoencoder for Robust Minutiae Extraction

Raffaele Cappelli; Matteo Ferrara

cs.CVFeb 17, 2026

LEADER: Lightweight End-to-End Attention-Gated Dual Autoencoder for Robust Minutiae Extraction

Authors: Raffaele Cappelli, Matteo Ferrara

Organizations: Department of Computer Science and Engineering of the University of Bologna, Italy

Abstract

Minutiae extraction, a fundamental stage in fingerprint recognition, is increasingly shifting toward deep learning. However, truly end-to-end methods that eliminate separate preprocessing and postprocessing steps remain scarce. This paper introduces LEADER (Lightweight End-to-end Attention-gated Dual autoencodER), a neural network that maps raw fingerprint images to minutiae descriptors, including location, direction, and type. The proposed architecture integrates non-maximum suppression and angular decoding to enable complete end-to-end inference using only 0.9M parameters. It employs a novel "Castle-Moat-Rampart" ground-truth encoding and a dual-autoencoder structure, interconnected through an attention-gating mechanism. Experimental evaluations demonstrate state-of-the-art accuracy on plain fingerprints and robust cross-domain generalization to latent impressions. Specifically, LEADER attains a 34% higher F1-score on the NIST SD27 dataset compared to specialized latent minutiae extractors. Sample-level analysis on this challenging benchmark reveals an average rank of 2.07 among all compared methods, with LEADER securing the first-place position in 47% of the samples-more than doubling the frequency of the second-best extractor. The internal representations learned by the model align with established fingerprint domain features, such as segmentation masks, orientation fields, frequency maps, and skeletons. Inference requires 15ms on GPU and 322ms on CPU, outperforming leading commercial software in computational efficiency. The source code and pre-trained weights are publicly released to facilitate reproducibility.

Figures & tables

Fig. 1: Overview of fingerprint patterns and extraction paradigms. (a) Local ridge structure highlighting a ridge ( $\mathrm{a}_{1}$ ), a valley ( $\mathrm{a}_{2}$ ), a ridge ending ( $\mathrm{a}_{3}$ ), and a bifurcation ( $\mathrm{a}_{4}$ ), with their corresponding directions $\theta$ . (b) Conventional deep learning-based pipeline: the input image ( $\mathrm{b}_{1}$ ) is typically preprocessed ( $\mathrm{b}_{2}$ ) and fed into a neural network to generate heatmaps ( $\mathrm{b}_{3}$ ), requiring off-graph postprocessing to obtain the final minutiae ( $\mathrm{b}_{4}$ ). (c) Traditional extraction workflow: the input image ( $\mathrm{c}_{1}$ ) is enhanced ( $\mathrm{c}_{2}$ ), binarized ( $\mathrm{c}_{3}$ ), and thinned to a ridge skeleton ( $\mathrm{c}_{4}$ ) before connectivity analysis yields the minutiae ( $\mathrm{c}_{5}$ ).

Method	Single-pass	End-to-end	Minutiae direction	Minutiae type	Parameters (M)	Open-source
Sankaran et al. (2014) [ 7 ]	No 1	No	No	No	$29.7$	No
Jiang et al. (2016) [ 8 ]	No 1	No	No	No	$11.2$	No
Darlow and Rosman (2017) [ 9 ]	No 1	No	No	No	$3.3$	No
Tang, Gao and Feng (2017) [ 10 ]	No 2	No	Yes	No	$108.3$	No
Tang et al. (2017) [ 11 ]	Yes	No 4	Yes	No	$4.7$	Yes
Nguyen, Cao, and Jain (2018) [ 12 ]	No 2	No	Yes	No	$74.4$	Yes

TABLE I: Evolution of deep learning-based minutiae extraction methods and comparison with LEADER.

Fig. 2: High-level diagram of the LEADER model. The network processes the fingerprint $\mathbf{F}$ through a pipeline composed of the Stem , the Context-Autoencoder , the Attention-Gate , the Refinement-Autoencoder , and the multi-task Head producing position ( $\widehat{\mathbf{P}},\widetilde{\mathbf{P}}$ ), direction ( $\widehat{\mathbf{D}}$ ), and type ( $\widehat{\mathbf{T}}$ ) maps.

Fig. 3: Ground-truth encoding and adaptive spatial weighting. (a) Detail of a fingerprint image with annotated minutiae. (b) Generated binary position map $\mathbf{P}$ : the morphology of the positive components is dynamically constrained in high-density regions to prevent spatial overlap. (c) Corresponding weight map $\mathbf{W}$ , derived from the CMR geometry. The penalty zones adaptively reshape to maintain separation between closely spaced minutiae; the castle regions in $\mathbf{W}$ correspond to positive pixels in $\mathbf{P}$ , while the ramparts provide localized penalty peaks to enforce sharp localization.

Fig. 4: Radial profile of the CMR weighting strategy. Left: 2D weight map

\mathbf{W}

for an isolated minutia. Right: Profile of the weighting function

\omega(\rho_{\min}(\cdot))

along a cross-section from the center (pixels are highlighted in red). The castle (

w_{ij}=1

) covers the central region—nominally of radius

\delta

, but dynamically adjusted in high-density areas—followed by a zero-gradient moat of width

\beta

, which effectively extends into the lower tail of the Gaussian slope. The rampart rises to a peak penalty (

w_{ij}=1

) to enforce sharp localization before descending to the background plateau

\lambda

Dataset	Sensor Type	Fingers	Impressions	Total	Ground-truth Source 1	Minutiae	Usage
FVC2002 DB2-A	Optical	$100$	1 2	$100$	Manual (Authors)	$5\,041$	Training
FVC2002 DB3-A	Capacitive	$100$	8	$800$	[ 19 ] [ 20 ]	$19\,032$
FVC2004 DB1-A	Optical	$100$	8	$800$	[ 19 ] [ 20 ]	$32\,771$
FVC2004 DB3-A	Thermal Sweeping	$100$	8	$800$	[ 19 ] [ 20 ]	$32\,608$
FFE Good	Optical	$10$	1	$10$	[ 21 ] 3	$525$
FFE Bad	Optical	$50$	1	$50$	[ 21 ] 3	$2\,107$

TABLE II: Characteristics of the datasets used for training, hyperparameter tuning, and testing.

Fig. 5: Representative fingerprint samples from the datasets. The collection covers diverse acquisition technologies: (a) FVC2002 DB2-A (Optical); (b) FVC2002 DB3-A (Capacitive); (c) FVC2004 DB1-A (Optical); (d) FVC2004 DB3-A (Thermal sweeping); (e) FFE Good (Optical); (f) FFE Bad (Optical); (g) FVC2000 DB2-A (Capacitive); (h) FVC2002 DB1-A (Optical); (i) NIST SD27 (Latent).

		FVC2002 DB1-A			NIST SD27
	Method	(16 px, $\pi/6$ rad)	(12 px, $\pi/8$ rad)	(8 px, $\pi/10$ rad)	(16 px, $\pi/6$ rad)	(12 px, $\pi/8$ rad)	(8 px, $\pi/10$ rad)
Type-agnostic	FingerNet	0.87	0.86	0.83	0.67	0.64	0.57
	LatentAFIS	0.84	0.83	0.81	0.62	0.59	0.54
	MinutiaeNet	0.47	0.43	0.38	0.48	0.45	0.40
	FingerJet	0.83	0.80	0.75	0.40	0.36	0.29
	FDx	0.86	0.85	0.81	0.45	0.41	0.34

TABLE III: Comparative performance (

F_{1}

-score). Results are reported for the three threshold levels under both type-agnostic and type-aware evaluation regimes. Bold and underlined values indicate the best and second-best scores, respectively.

Fig. 6: PR curves for FVC2002 DB1-A (top two rows) and NIST SD27 (bottom two rows). Each column represents a different threshold level, from the most permissive

(16\text{\,px},\pi/6\text{\,rad})

to the most demanding

(8\text{\,px},\pi/10\text{\,rad})

. Results are reported for both type-agnostic and type-aware regimes. For methods that do not support minutiae type (FingerNet, LatentAFIS, and MinutiaeNet), curves are omitted from type-aware plots to ensure a fair comparison while maintaining legend consistency.

Fig. 7: Sample-level visual comparisons. True Positives (green), False Positives (orange), and False Negatives (magenta) are overlaid for LEADER and the two top-performing competitors. Rows illustrate representative performance scenarios: (a–c) high-scoring sample on FVC2002 DB1-A; (d–f) LEADER’s lowest-performing case on FVC2002 DB1-A; (g–i) high-scoring latent sample on NIST SD27; (j–l) failure case on NIST SD27. First column (a, d, g, j) shows LEADER results ( $F_{1}$ -score: 1.0, 0.75, 0.92, 0.00). Second and third columns display top competitors: (b) FDx (0.95), (c) MINDTCT (0.93), (e) LatentAFIS (0.77), (f) FingerJet (0.71), (h) VeriFinger (0.83), (i) FingerJet (0.70), (k) LatentAFIS (0.25), and (l) FDx (0.14). Full-image comparisons for all extractors are provided in Figs. 19 – 21 (Supplementary Material).

	Method	Parameters (M)	Time (ms)
cpu 1	FingerNet	$4.7$	$3\,964$
	LatentAFIS	$5.5$	$5\,860$
	MinutiaeNet	$74.4$	$25\,744$
	FingerJet	—	$63$
	FDx	—	$397$
	VeriFinger	—	$1\,199$

TABLE IV: Average execution time (ms) on NIST SD27.

Configuration 1						Type-agnostic						Type-aware
						FVC2002 DB1-A			NIST SD27			FVC2002 DB1-A			NIST SD27
#	S	CA	AG	RA	GTE	(16, $\frac{\pi}{6}$ )	(12, $\frac{\pi}{8}$ )	(8, $\frac{\pi}{10}$ )	(16, $\frac{\pi}{6}$ )	(12, $\frac{\pi}{8}$ )	(8, $\frac{\pi}{10}$ )	(16, $\frac{\pi}{6}$ )	(12, $\frac{\pi}{8}$ )	(8, $\frac{\pi}{10}$ )	(16, $\frac{\pi}{6}$ )	(12, $\frac{\pi}{8}$ )	(8, $\frac{\pi}{10}$ )
0	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\boxplus$	0.92	0.91	0.90	0.71	0.69	0.62	0.85	0.84	0.83	0.52	0.51	0.47
1	–	$\checkmark$	$\checkmark$	$\checkmark$	$\boxplus$	0.91	0.90	0.88	0.66	0.63	0.56	0.82	0.81	0.80	0.47	0.45	0.41
2	$\checkmark$	–	$\checkmark$	$\checkmark$	$\boxplus$	0.91	0.91	0.90	0.64	0.61	0.54	0.83	0.83	0.82	0.46	0.44	0.41
3	$\checkmark$	$\checkmark$	–	$\checkmark$	$\boxplus$	0.91	0.91	0.90	0.64	0.61	0.55	0.84	0.83	0.82	0.46	0.45	0.41

TABLE V: Ablation Study. Performance (

F_{1}

-score) evaluated across the three threshold levels for both type-agnostic and type-aware regimes, comparing the full LEADER architecture against various degraded configurations.

Fig. 8: Visualization of output maps and decoding logic. (a) Input fingerprint region. (b) Predicted position map $\widehat{\mathbf{P}}$ . (c) Refined map $\widetilde{\mathbf{P}}$ showing local peaks. (d) Direction map $\widehat{\mathbf{D}}$ : the angle $\hat{d}_{ij}$ is mapped to the Hue component (HSV); the two blobs show distinct shades of magenta, corresponding to the directions of the two minutiae, approximately $-\pi/5$ (top) and $-\pi/3$ (bottom). (e) Type map $\widehat{\mathbf{T}}$ : Hue encodes the classification score $\hat{t}_{ij}$ on a gradient from red (bifurcation) to blue (ridge ending). In (d) and (e), Saturation and Value are modulated by $\hat{p}_{ij}$ to suppress background noise. (f) Final extracted set $\widehat{\mathbb{M}}$ .

Fig. 9: Visual analysis of internal features through PCA projection. (a, h) Input fingerprints from FVC2002 DB1-A. (b–g, i–n) RGB visualization of the first three principal components of the feature maps. All visualized activations are generated via GELU, except for (d, k), which correspond to the Sigmoid gating signal $\mathbf{X}_{gate}$ from the Attention-Gate . (b, i) First StemBlock . (c, j) Second SeparableConvBlock in the encoding path of the Context-Autoencoder . The remaining panels illustrate successive InvBottleneckConvBlock activations within the Refinement-Autoencoder : (e, l) first in encoding, (f, m) penultimate in decoding, and (g, n) final in decoding.

Fig. 10: Analysis of ridge cut robustness. The sequence focuses on the bottom-left region of the fingerprint in Fig. 9 h: (a) raw input; (b) PCA projection of Stem features; (c) PCA projection of the Attention-Gate signal $\mathbf{X}_{gate}$ ; (d, e) individual $\mathbf{X}_{gate}$ channels exhibiting noise reduction and localized noise focusing; (f) topological inpainting and neural thinning in the PCA projection of Refinement-Autoencoder features.

Fig. 11: Internal representation of synthetic geometric patterns. Features are visualized via PCA projection. Top row (a–d): Concentric circles (constant frequency). Bottom row (e–h): Horizontal ridges (varying frequency). Columns: (a, e) input; (b, f) orientation-sensitive features ( Context-Autoencoder ); (c, g) frequency/segmentation features ( Refinement-Autoencoder ); (d, h) final common representation.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

	Symbol	Description
Model & Inference	$\mathbf{F}$	Raw input fingerprint image
	$w,h$	Width and height of the input image
	$i,j$	Row and column indices of an element in the spatial grid
	$\mathbf{X},\mathbf{X}^{\prime}$	Generic intermediate feature maps
	$\mathbf{X}_{gate}$	Recalibration gating signal within the Attention-Gate
	$\varsigma(\cdot)$	Sigmoid activation function

Appendix

TABLE VI: Mathematical symbols and notation.

Fig. 12: Stem component. Detailed view of the dual-path input stage. It accepts the raw fingerprint $\mathbf{F}$ and processes it through two specialized StemBlocks (detailed in Fig. 16 d).

Fig. 13: Attention-Gate component. Internal flow of the recalibration mechanism. It computes the multi-scale gating signal $\mathbf{X}_{gate}$ through parallel dilated paths and modulates the input features using element-wise multiplication before the final concatenation.

Fig. 14: Head component. Diagram of the task-specific decoding logic. The shared features are branched into three HeadBlocks (detailed in Fig. 16 e) for position, direction, and type estimation. Task-specific activations are applied within each block: Sigmoid for position and type maps and Linear (identity) for the regression of Cartesian directional components. The pipeline concludes with on-graph NMS and CartesianToPolar angular decoding.

Fig. 15: Context-Autoencoder and Refinement-Autoencoder components. Side-by-side comparison: The Context-Autoencoder (left) employs SeparableConvBlocks for topological extraction (Fig. 16 c), while the Refinement-Autoencoder (right) uses InvBottleneckConvBlocks for high-level semantic refinement (Fig. 16 f). Both stages adopt DownsamplingBlocks and UpsamplingBlocks (Fig. 16 a, b).

Fig. 16: Computational block library. Schematic of the modular blocks in LEADER: (a) DownsamplingBlock and (b) UpsamplingBlock for resolution management; (c) SeparableConvBlock for efficient feature extraction; (d) StemBlock for initial feature extraction and spatial mapping; (e) HeadBlock for task-specific projection; and (f) InvBottleneckConvBlock , the core building block of the Refinement-Autoencoder , designed for high-capacity representation with minimal parameter overhead. For blocks employed at multiple stages, input dimensions are denoted as $w^{\prime}\times h^{\prime}\times c^{\prime}$ to indicate that the operation is resolution-agnostic and applicable to feature maps with varying spatial sizes and channel depths throughout the network.

Fig. 17: Optimization dynamics during LEADER training. The primary axis (left) shows the minimization of the multi-task loss function

\mathcal{L}

. The secondary axis (right) illustrates the evolution of the normalized weight magnitude

\mathcal{W}

, indicating stable convergence and controlled model capacity.

Method	Avg. Rank $\pm$ S.D.	T1 (%)	T3 (%)	BH (%)	Avg. Rank $\pm$ S.D.	T1 (%)	T3 (%)	BH (%)
	FVC2002 DB1-A				NIST SD27
FingerNet	3.48 $\pm$ 1.90	13	56	19	3.29 $\pm$ 2.16	22	64	15
LatentAFIS	4.98 $\pm$ 1.99	3	27	43	3.75 $\pm$ 1.84	11	43	14
MinutiaeNet	9.79 $\pm$ 0.41	0	0	100	5.36 $\pm$ 2.42	5	24	48
FingerJet	5.80 $\pm$ 1.73	1	10	62	7.36 $\pm$ 2.05	0	5	80
FDx	4.04 $\pm$ 1.85	8	43	22	6.18 $\pm$ 1.93	1	9	62

Appendix

TABLE VII: Sample-level ranking analysis. Results summarize multiple independent rankings (one per sample) based on

F_{1}

-score computed at the (

16\text{\,px},\pi/6\text{\,rad}

) threshold level in the type-agnostic regime. Bold and underlined values indicate the best and second-best results for each metric, respectively.

	FingerNet	LatentAFIS	MinutiaeNet	FingerJet	FDx	VeriFinger	MINDTCT	FingerFlow	SourceAFIS	LEADER
FingerNet	—	71	100	81	63	55	72	100	91	13
LatentAFIS	28	—	100	59	34	30	54	99	82	5
MinutiaeNet	0	0	—	1	0	0	0	19	0	0
FingerJet	19	40	99	—	26	17	37	99	72	2
FDx	36	63	100	73	—	44	66	100	93	9
VeriFinger	44	68	100	80	51	—	69	100	92	7

Appendix

TABLE VIII: Pairwise direct-win rates (%) on FVC2002 DB1-A. Each cell

(i,j)

represents the percentage of samples where the method in row

i

outperforms the method in column

j

in terms of

F_{1}

-score. Bold values highlight the performance of LEADER against all competitors.

	FingerNet	LatentAFIS	MinutiaeNet	FingerJet	FDx	VeriFinger	MINDTCT	FingerFlow	SourceAFIS	LEADER
FingerNet	—	59	71	88	83	53	91	73	91	29
LatentAFIS	36	—	68	90	82	41	94	71	95	20
MinutiaeNet	26	26	—	71	57	28	82	44	83	14
FingerJet	8	7	28	—	33	7	57	31	57	2
FDx	15	15	41	64	—	12	82	45	84	3
VeriFinger	44	55	69	90	87	—	95	72	96	23

Appendix

TABLE IX: Pairwise direct-win rates (%) on NIST SD27. Notation as in Table VIII .

Fingerprint Identification

See all 21 papers

Revisiting Handcrafted Minutiae Detection: A Simple and Effective Open Source Baseline for Modern Fingerprint Workflows
Oct 8, 2026Raffaele CappelliFingerprint Identification Biometric Identification
Department of Computer Science and Engineering, University of Bologna, Cesena, Italy
AFID: A Unified Open Framework for Automated Fingermark Identification, Quality Assessment and Feature Extraction
Sep 7, 2026Tim Oblak, Rudolf Haraksim, Peter PeerBiometric Identification Fingerprint Identification
Joint Research Centre of the European Commission in Ispra, Italy
Unleashing the Power of Simplicity: A Minimalist Strategy for State-of-the-Art Fingerprint Enhancement
Mar 19, 2026Raffaele CappelliFingerprint Identification Image Enhancement
Department of Computer Science and Engineering of the University of Bologna, Italy

Oct 8, 2026J/K move · Enter open · S save

Revisiting Handcrafted Minutiae Detection: A Simple and Effective Open Source Baseline for Modern Fingerprint Workflows

Raffaele Cappelli

Department of Computer Science and Engineering, University of Bologna, Cesena, Italy

Fingerprint Identification Biometric Identification

Open paper AskPDF arXiv

Handcrafted minutiae detection algorithms remain fundamental to biometric science and forensic practice due to their full auditability, adherence to international standards, and operational independence from training datasets or GPU hardware. However, current open-source traditional baselines are severely outdated, relying almost exclusively on legacy C/C++ codebases that lack seamless integration with modern scientific software ecosystems. To bridge this gap, the present work introduces SBMEX (Skeleton-Based Minutiae EXtraction), a fast and deterministic minutiae detection method integrated into the open source pyfing package. SBMEX achieves high computational throughput by employing a dual Look-Up Table architecture that replaces runtime neighborhood scanning during Crossing Number computation and skeleton tracking. Additionally, it incorporates a continuous quality scoring framework driven by tracking path length, dual ridge-valley skeleton fusion, and spatial density decay. Rigorous evaluation on NIST SD302 datasets demonstrates that SBMEX delivers feature extraction accuracy comparable to or outperforming traditional open-source baselines without fine-tuning, while achieving a drastic reduction in minutiae detection latency relative to classical Crossing Number Python implementations.

Abstract

Fingerprint Identification

Explore similar work

Revisiting Handcrafted Minutiae Detection: A Simple and Effective Open Source Baseline for Modern Fingerprint Workflows

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

Unleashing the Power of Simplicity: A Minimalist Strategy for State-of-the-Art Fingerprint Enhancement