Source-linked AI summary
Automated Latent Fingerprint Recognition
Kai Cao, Anil K. Jain
TL;DR
Latent fingerprint recognition remains difficult because crime-scene marks are noisy and incomplete, while automated systems achieve limited identification accuracy. The paper combines ConvNets with complementary minutiae and texture templates and fuses their comparison scores. On two latent databases searched against 100K rolled prints, the system achieves strong rank-1 performance, and fusion with a COTS AFIS improves it further.
Problem
Automated latent recognition remains limited because latent images have poor ridge quality and severe background noise, despite the need to search large reference collections accurately.
Method
The system uses ConvNets for ridge-flow estimation and multi-scale minutiae descriptors, represents each latent with two minutiae templates and one texture template, and fuses comparison scores.
Results
73.3% (74.4%) and 76.6% (78.4%) rank-1 accuracy were achieved after score-level (rank-level) fusion with a leading COTS latent AFIS on NIST SD27 and WVU, respectively.
Takeaways & Limitations
The proposed method performs significantly better than published algorithms on NIST SD27 and WVU and complements the COTS latent AFIS.
Takeaways & Limitations
Future work identifies improving recognition effectiveness, scalability, training data, and processing speed as remaining needs.
Abstract
from arXiv · showhide
Latent fingerprints are one of the most important and widely used evidence in law enforcement and forensic agencies worldwide. Yet, NIST evaluations show that the performance of state-of-the-art latent recognition systems is far from satisfactory. An automated latent fingerprint recognition system with high accuracy is essential to compare latents found at crime scenes to a large collection of reference prints to generate a candidate list of possible mates. In this paper, we propose an automated latent fingerprint recognition algorithm that utilizes Convolutional Neural Networks (ConvNets) for ridge flow estimation and minutiae descriptor extraction, and extract complementary templates (two minutiae templates and one texture template) to represent the latent. The comparison scores between the latent and a reference print based on the three templates are fused to retrieve a short candidate list from the reference database. Experimental results show that the rank-1 identification accuracies (query latent is matched with its true mate in the reference database) are 64.7% for the NIST SD27 and 75.3% for the WVU latent databases, against a reference database of 100K rolled prints. These results are the best among published papers on latent recognition and competitive with the performance (66.7% and 70.8% rank-1 accuracies on NIST SD27 and WVU DB, respectively) of a leading COTS latent Automated Fingerprint Identification System (AFIS). By score-level (rank-level) fusion of our system with the commercial off-the-shelf (COTS) latent AFIS, the overall rank-1 identification performance can be improved from 64.7% and 75.3% to 73.3% (74.4%) and 76.6% (78.4%) on NIST SD27 and WVU latent databases, respectively.
1 INTRODUCTION
Latent fingerprint recognition is difficult because crime-scene marks often have poor ridge quality, noise, small friction-ridge areas, and distortion, while current automated performance remains limited. The paper develops an automated, ConvNet-based system intended to produce accurate and efficient candidate lists with reduced examiner dependence.
- Recognition gap: Latent prints are difficult to recognize because crime-scene deposition produces poor ridge clarity and large background noise, unlike supervised reference-print acquisition.Poor-quality latents are also not repeatably deposited for later collection.
- Recognition gap: 67.2% rank-1 identification was achieved by the best automated latent system in NIST ELFT-EFS, searching 1,114 latents against 100,000 reference prints.Adding expert feature markup improved the rate to 70.2%.
- Current practice: Current latent processing faces reproducibility, throughput, and bias concerns, including approximately 15 minutes of manual markup per latent and only 46% median markup reproducibility.Examiner decision repeatability was reported as 90.0% for mated pairs and 85.9% for nonmated pairs.
- Proposed approach: The proposed representation uses two minutiae templates and one texture template to capture complementary latent information.The minutiae templates use ConvNet-learned ridge flow and a dictionary of ridge-structure elements.
- Proposed approach: ConvNets learn minutiae descriptors from multi-scale, multi-location fingerprint patches, while graph-based correspondence methods reduce false minutiae matches.Fourteen patch types are investigated, and second- and third-order graph matching address minutiae pairs and triplets.
- Evaluation: 73.3% (74.4%) and 76.6% (78.4%) rank-1 accuracy were obtained after score-level (rank-level) fusion with a leading COTS latent AFIS on NIST SD27 and WVU, respectively.The system is evaluated against a leading COTS AFIS, with most preprocessing automated except ROI handling and related feature-processing stages.
2 RELATED LITERATURE
Prior work addresses individual latent-processing modules, but poor latent quality limits automated recognition and published systems have not yet formed a fully evaluated end-to-end algorithm.
- Latent processing must address background noise, low ridge contrast, and small friction-ridge areas.
- ROI segmentation: ROI segmentation methods perform poorly on poor-quality latents, so examiners commonly mark the ROI when impressions overlap.
- Ridge flow estimation: ConvNet-based ridge-flow estimates generally outperform dictionary-based methods against manually marked ridge flow.
- Feature extraction and comparison: Published latent comparison methods rely on manually marked minutiae because robust automated latent minutiae extraction remains unavailable.
- To the authors’ knowledge, no published automated latent recognition algorithm had evaluated ConvNets in an end-to-end system.
3 PREPROCESSING AND FEATURE EXTRACTION
The proposed system represents each latent with two complementary minutiae templates and a texture template, using learned ridge flow and multi-scale ConvNet descriptors alongside dictionary-based processing.
- Each latent is represented by two minutiae templates and one texture template.
- Minutiae templates: The two minutiae templates use different ridge-flow and ridge-enhancement methods to provide complementary minutiae sets.
- Texture template: The texture template represents each non-overlapping latent block with a pair of virtual minutiae when minutiae are insufficient or unreliable.
- Minutiae descriptor: Fourteen ConvNets learn minutiae descriptors from multi-scale, multi-location patches around minutiae.
- Minutiae descriptor: Each descriptor ConvNet produces a 128-dimensional feature vector that can be concatenated into a minutia descriptor.
- Reference-print features: Reference prints use one minutiae template and a texture template, with minutiae extracted by a COTS tenprint AFIS and ridge flow obtained by STFT.
4 LATENT TO ROLLED COMPARISON
The comparison framework establishes latent–reference minutiae correspondences using descriptor similarity and second- and third-order graph matching, then combines minutiae and ridge-flow similarity. ConvNet-derived descriptors and complementary patch representations support correspondence selection and template scoring.
- Comparison framework: Two latent-to-reference comparison algorithms are designed: minutiae template comparison and texture template comparison.
- Minutiae correspondence: Minutiae correspondence is formulated as an assignment between latent and reference minutiae, with no minutia matched more than once.
- Descriptor extraction and evaluation: Fourteen ConvNets use multi-scale and multi-location patches to produce minutiae descriptors, while Fig. 10 illustrates fewer false correspondences for the proposed method on an impostor comparison.
- Graph matching: Second- and third-order graph matching compare compatibility among minutiae pairs and triplets using location and orientation distances, with power iteration providing approximate solutions.
- Minutiae correspondence: Top N = 120 candidate correspondences are selected by descriptor similarity, followed by second-order and third-order graph matching to remove false correspondences.
- Template similarity: The minutiae template similarity multiplies descriptor-based minutiae similarity by ridge-flow similarity after alignment and comparison of overlapping block orientations.
5 EXPERIMENTAL RESULTS
Experiments evaluate complementary latent templates, benchmark the proposed method against a COTS latent AFIS, and test their fusion on two databases against 100,000 rolled prints.
- 258 NIST SD27 latents and 449 WVU latents are evaluated against a 100,000-print reference database following the NIST ELFT-EFS protocol.
- Individual latent templates: The three-template system fuses two minutiae templates and one texture template designed to provide complementary latent information.
- Individual latent templates: 64.7% rank-1 accuracy on NIST SD27 and 75.3% on WVU DB result when all three templates are fused, versus 58.5% and 70.6% without the texture template.
- Benchmarking against COTS Latent AFIS: 75.3% rank-1 accuracy on WVU DB exceeds the COTS AFIS result of 70.8%, while 64.7% on NIST SD27 is below the COTS result of 66.7%.
- Benchmarking against COTS Latent AFIS: 74.4% and 78.4% rank-1 accuracies result from rank-level fusion with the COTS AFIS on NIST SD27 and WVU DB, respectively.Score-level fusion reaches 73.3% on NIST SD27 and 76.6% on WVU DB.
- Benchmarking against COTS Latent AFIS: The proposed method succeeds on some latents missed by the COTS AFIS, including cases where the commercial system assigns a zero comparison score.
- Benchmarking against COTS Latent AFIS: 51.2% and 65.7% rank-1 accuracies for ConvNet-learned minutiae descriptors exceed 21.3% and 35.2% obtained with MCC descriptors on NIST SD27 and WVU DB.
6 CONCLUSIONS AND FUTURE WORK
The paper addresses the difficulty of accurate, robust, fully automated latent recognition by presenting and benchmarking a ConvNet-based system with complementary templates. It concludes that fusion with a leading COTS AFIS improves rank-1 recognition, while identifying several avenues for further improvement.
- Latent recognition remains difficult because latent images contain poor ridge quality, severe background noise, small friction ridge areas, and distortion.
- The paper presents an automated latent fingerprint recognition algorithm and benchmarks it against a leading COTS latent AFIS.
- Three templates—two minutiae templates and one texture template—extract complementary information for latent recognition.
- Fusion of the proposed method with the COTS latent AFIS boosts rank-1 recognition accuracy on both benchmark latent databases.
- Future work: Future work includes stronger ConvNet architectures, additional latent features, cascaded recognition engines, larger training collections, and faster extraction and comparison.