[Paper Review] A Universal Latent Fingerprint Enhancer Using Transformers

André Brasil Vieira Wyzykowski, Anil K. Jain|arXiv (Cornell University)|May 31, 2023

Biometric Identification and Security4 citations

TL;DR

This paper proposes ULPrint, a universal latent fingerprint enhancer using a two-stage Transformer-based method combining UNet with Mix Vision Transformer (SegFormer-B5) and dilated convolutions to improve ridge segmentation and image quality. It achieves state-of-the-art performance, boosting MSU-AFIS identification accuracy from 61.56% to 75.19% on NIST SD27, 67.63% to 77.02% on MSP Latent, and 46.90% to 52.12% on NIST SD302, outperforming Verifinger SDK 12.4.

ABSTRACT

Forensic science heavily relies on analyzing latent fingerprints, which are crucial for criminal investigations. However, various challenges, such as background noise, overlapping prints, and contamination, make the identification process difficult. Moreover, limited access to real crime scene and laboratory-generated databases hinders the development of efficient recognition algorithms. This study aims to develop a fast method, which we call ULPrint, to enhance various latent fingerprint types, including those obtained from real crime scenes and laboratory-created samples, to boost fingerprint recognition system performance. In closed-set identification accuracy experiments, the enhanced image was able to improve the performance of the MSU-AFIS from 61.56\% to 75.19\% in the NIST SD27 database, from 67.63\% to 77.02\% in the MSP Latent database, and from 46.90\% to 52.12\% in the NIST SD302 database. Our contributions include (1) the development of a two-step latent fingerprint enhancement method that combines Ridge Segmentation with UNet and Mix Visual Transformer (MiT) SegFormer-B5 encoder architecture, (2) the implementation of multiple dilated convolutions in the UNet architecture to capture intricate, non-local patterns better and enhance ridge segmentation, and (3) the guided blending of the predicted ridge mask with the latent fingerprint. This novel approach, ULPrint, streamlines the enhancement process, addressing challenges across diverse latent fingerprint types to improve forensic investigations and criminal justice outcomes.

Motivation & Objective

• To address the challenge of low-quality and variable latent fingerprints in forensic investigations, which hinder accurate identification due to noise, contamination, and partial impressions.
• To overcome the scarcity and limited accessibility of real crime scene fingerprint databases, which restrict algorithm development and benchmarking.
• To reduce human bias and workload in latent fingerprint analysis by developing an automated, universal enhancement method that improves recognition accuracy across diverse fingerprint types.
• To create a robust, generalizable enhancement framework that performs well on both real-world and laboratory-generated latent fingerprints, ensuring broader forensic applicability.
• To improve the reliability and efficiency of fingerprint recognition systems by integrating advanced deep learning architectures like Vision Transformers and dilated convolutions for superior ridge structure recovery.

Proposed method

• The method employs a two-stage pipeline: first, ridge segmentation using a modified UNet with multiple dilated convolutions to capture non-local, intricate ridge patterns.
• Second, the ridge mask is generated using a Mix Vision Transformer (MiT) encoder, specifically SegFormer-B5, to enhance feature representation and segmentation accuracy.
• The predicted ridge mask is then guidedly blended with the original latent fingerprint image to preserve texture and structural details while enhancing ridge clarity.
• The architecture integrates dilated convolutions in the UNet encoder to expand感受 receptive fields and improve segmentation of fine, discontinuous ridge structures.
• The model is trained end-to-end on a combination of real crime scene (NIST SD27, MSP Latent) and lab-generated (NIST SD302) databases to ensure generalization across fingerprint types.
• Evaluation follows NIST ELFT-EFS protocols, using closed-set 1:N identification with 100,000 rolled fingerprints as gallery, and NFIQ 2 for image quality assessment.

Experimental results

Research questions

• RQ1 Can a deep learning-based enhancement method universally improve the quality and recognizability of diverse latent fingerprint types from real crime scenes and lab settings?
• RQ2 How does the integration of dilated convolutions and Vision Transformers improve ridge segmentation compared to standard UNet or CNN-based approaches?
• RQ3 To what extent does the guided blending of the ridge mask with the original image enhance recognition accuracy without introducing artifacts?
• RQ4 Does the proposed method outperform existing commercial and research-grade fingerprint enhancement tools like Verifinger SDK 12.4 in closed-set identification tasks?
• RQ5 Can the method achieve consistent performance gains across multiple benchmark databases with varying image quality and complexity?

Key findings

• ULPrint improved MSU-AFIS identification accuracy from 61.56% to 75.19% on the NIST SD27 database, representing a 13.63 percentage point increase.
• On the MSP Latent database, recognition accuracy rose from 67.63% to 77.02%, a 9.39 percentage point improvement.
• For the NIST SD302 database, the method increased accuracy from 46.90% to 52.12%, a 5.22 percentage point gain.
• The NFIQ 2 quality score increased significantly after enhancement: mean rose from 20.04 to 26.62 on NIST SD27, from 9.47 to 15.91 on MSP Latent, and from 4.67 to 16.31 on NIST SD302.
• The method outperformed the Verifinger V12.4 SDK, which achieved only 27.51% on NIST SD27 and 55.06% on MSP Latent, confirming its superiority.
• Visual comparisons showed that ULPrint-generated enhancements preserved image details better than FingerNet and produced clearer, more consistent ridge structures across diverse latent types.