Source-linked AI summary
FingerNet: An Unified Deep Network for Fingerprint Minutiae Extraction
Yao Tang, Fei Gao, Jufu Feng, Yuhang Liu
TL;DR
Fingerprint minutiae extraction is reliable on rolled/slap images but degrades on latent fingerprints with fuzzy ridges and complex background noise. FingerNet integrates traditional fingerprint-processing knowledge into a convolutional network, expands a fixed-weight shallow pipeline into a trainable end-to-end model, and reports reliable automatic extraction on rolled/slap and latent fingerprints while outperforming other minutiae-extraction algorithms.
Problem
Minutiae extraction accuracy degrades on latent fingerprints because fuzzy ridges and complex background noise challenge methods developed for rolled/slap fingerprints.
Method
FingerNet transforms traditional orientation estimation, segmentation, enhancement, and extraction methods into a fixed-weight network, then expands it with trainable layers while preserving end-to-end differentiability.
Results
FingerNet extracts reliable minutiae automatically on rolled/slap and latent fingerprints without fine tuning and outperforms other minutiae-extraction algorithms.
Takeaways & Limitations
Combining domain knowledge with deep learning yields an efficient and reliable unified network for minutiae extraction across rolled/slap and latent fingerprints.
Takeaways & Limitations
Future work includes integrating ridge frequency, improving segmentation, and extending FingerNet to fingerprint matching.
Abstract
from arXiv · showhide
Minutiae extraction is of critical importance in automated fingerprint recognition. Previous works on rolled/slap fingerprints failed on latent fingerprints due to noisy ridge patterns and complex background noises. In this paper, we propose a new way to design deep convolutional network combining domain knowledge and the representation ability of deep learning. In terms of orientation estimation, segmentation, enhancement and minutiae extraction, several typical traditional methods performed well on rolled/slap fingerprints are transformed into convolutional manners and integrated as an unified plain network. We demonstrate that this pipeline is equivalent to a shallow network with fixed weights. The network is then expanded to enhance its representation ability and the weights are released to learn complex background variance from data, while preserving end-to-end differentiability. Experimental results on NIST SD27 latent database and FVC 2004 slap database demonstrate that the proposed algorithm outperforms the state-of-the-art minutiae extraction algorithms. Code is made publicly available at: https://github.com/felixTY/FingerNet.
1. Introduction
Latent fingerprints are important but difficult for minutiae extraction because fuzzy ridges and complex background noise degrade methods developed for rolled/slap images. FingerNet combines domain knowledge with deep learning to extract fingerprint representations and reliable minutiae automatically.
- Motivation: Latent-fingerprint minutiae extraction is difficult because fuzzy ridges and complex background noise reduce accuracy.Minutiae are important features in most fingerprint matching systems, and latent prints are directly obtained from crime scenes.
- Approach: FingerNet combines domain knowledge with deep-learning representation ability through a specially designed network.The design guides network structure and weight initialization while preserving end-to-end differentiability.
- Approach: The unified network acquires orientation fields, segmentation, enhancement, and minutiae representations during extraction.These representations integrate typical fingerprint-processing stages into one network.
- Results: Reliable minutiae were extracted automatically from both rolled/slap and latent fingerprints without fine tuning.The paper also reports promising performance on good-quality FVC 2004 fingerprints.
- Training: FingerNet introduces weak labels generated by matching latent fingerprints with corresponding rolled/slap fingerprints to support modular training.This addresses the lack of orientation and segmentation training labels.
2. Proposed FingerNet for Minutiae Extraction
The proposed FingerNet starts from traditional fingerprint-processing operations represented as a shallow fixed-weight network, then expands it into a trainable convolutional model.
- Network construction: Traditional normalization, orientation estimation, segmentation, Gabor enhancement, and minutiae extraction are transformed into convolutional kernels.These operations are integrated into a shallow network whose weights are initially fixed.
- Network construction: The plain network is expanded into a complete trainable fingerprint network.The expansion is intended to provide greater representation ability while retaining the integrated processing pipeline.
2.1. Traditional Methods to Equivalent ConvNets
The paper recasts traditional fingerprint preprocessing and minutiae extraction stages as differentiable convolutional operations. These components form a shallow ConvNet that integrates normalization, orientation estimation, segmentation, enhancement, and extraction.
- Traditional fingerprint processing is organized as normalization, orientation estimation, segmentation, enhancement, and minutiae extraction.
- All operators are pixel-wise and differentiable, allowing the integrated traditional pipeline to be expressed as a convolutional network.
- Gradient-based orientation estimation becomes convolution with Sobel masks, windowed aggregation, and angle computation.
- Learning-based segmentation is represented as a shallow ConvNet using handcrafted local features and a linear classifier.
- Selective convolution handles Gabor filters whose weights are shared across blocks with matching ridge frequency and orientation.
- Template matching for minutiae scoring is equivalent to a ConvNet containing one convolution layer and one maxout layer.
2.2. Expand to FingerNet
Plain FingerNet works reasonably on rolled/slap fingerprints but fails on latent images because shallow structures have limited expressive power against complex background noise. The expanded network adds trainable capacity while retaining domain-informed components and supports end-to-end training.
- Plain FingerNet fails on latent fingerprints because shallow ConvNet structures have poor expressive power for complex background noise.
- The complete FingerNet expands the simple network with convolutional layers and releases weights to learn complex background variance from data.
- Orientation Estimation: Orientation estimation uses multiscale features, orientation regression, and discrete angle probabilities that can be selected or averaged.
- Segmentation: Segmentation shares the entire multiscale feature maps with orientation estimation and predicts a pixel-wise region-of-interest probability map.
- The expanded architecture is trainable end-to-end and is described as an extension of plain FingerNet.
- Minutiae Extraction: Minutiae extraction predicts score, X-offset, Y-offset, and angle maps, then filters scores, recovers precise locations and angles, and applies non-maximum suppression.
2.3. Label, Loss and Training
FingerNet uses weak, strong, and ground-truth labels to train a unified fingerprint network with domain-informed losses. Its total loss combines nine losses while balancing label classes and incorporating orientation coherence and segmentation smoothness.
- Label construction: Weak orientation labels come from matched rolled/slap fingerprints, while weak segmentation labels use dilated and smoothed minutiae convex hulls.
- Label construction: Unoriented manually marked minutiae directions provide strong orientation labels, and manually marked minutiae lists provide ground-truth maps.
- Label construction: An inverted Gaussian angle label handles angular discontinuity and assigns equal loss to angles with the same directional distance.
- Loss definition: The total loss is a weighted sum of 9 losses from orientation, segmentation, and minutiae extraction, organized into three loss types.
- Loss definition: Class-balancing weights λ+ and λ− compensate for imbalanced positive and negative labels in the cross entropy loss.
- Loss definition: Orientation coherence constrains orientation distributions, while Laplacian edge suppression smooths segmentation and reduces noise and outliers.
- Training procedure: Training first uses orientation and segmentation losses, then adds minutiae losses so the network learns ridge properties step by step.
3. Experiments
Experiments evaluate FingerNet on latent and rolled/slap fingerprint databases using precision-recall and identification performance. The reported results include reliable extraction across quality levels, quantitative localization and angle errors, and improved identification against other methods.
- Experimental setup: FingerNet calculates orientation fields, segmentation, enhanced fingerprints, and minutiae without fine-tuning across different-quality fingerprint databases.
- Experimental setup: Training uses about 8000 matched rolled–latent fingerprint pairs, while testing uses NIST SD27 and FVC 2004 database set A.
- Evaluation: Minutiae extraction is evaluated with precision-recall curves, counting one-to-one matches within 15 pixels and 30° of manually labeled minutiae.
- Minutiae extraction performance: On NIST SD27, FingerNet is compared with MINDTCT, Gabor-based, autoencoder-based, fully convolutional, and VeriFinger extractors.
- Minutiae extraction performance: 4.4 pixels and 5.0° are the mean location and angle errors on the reported NIST SD27 evaluation.
- Minutiae extraction performance: 3.4 pixels and 6.4° are the mean location and angle errors reported for FVC 2004 database set A.
- Identification performance: FingerNet outperforms other methods on NIST SD27 identification using CMC curves with an extended clique matching model and an approximately 40K-fingerprint gallery.
4. Conclusion and Future Work
FingerNet combines domain knowledge with deep learning by expanding a fixed-weight traditional pipeline into a trainable network for minutiae extraction. Future work targets ridge frequency, segmentation accuracy, and fingerprint matching.
- FingerNet guides network structure and weight initialization by combining traditional fingerprint methods with deep learning representation ability.
- The traditional pipeline is represented as a fixed-weight network, then expanded with released weights while preserving end-to-end differentiability.
- FingerNet is reported to outperform other minutiae extraction algorithms on rolled/slap and latent fingerprints.
- Future work will integrate ridge frequency, improve segmentation accuracy, and extend FingerNet to fingerprint matching.