[Paper Review] A Fully Automated Latent Fingerprint Matcher with Embedded Self-learning Segmentation Module

Jinwei Xu, Jiankun Hu+1 more|arXiv (Cornell University)|Jun 26, 2014

Biometric Identification and Security11 references3 citations

TL;DR

This paper proposes a fully automated latent fingerprint matcher with an embedded self-learning dictionary learning-based ROI segmentation module and a genetic algorithm (GA)-based minutiae matching unit. It achieves superior performance on the NIST SD27 database, outperforming state-of-the-art public methods, with a mean penetration rate of 38.159% in Scenario 1 and improved results in Scenario 2 using the proposed segmentation, demonstrating robustness in challenging latent fingerprint conditions without manual intervention.

ABSTRACT

Latent fingerprint has the practical value to identify the suspects who have unintentionally left a trace of fingerprint in the crime scenes. However, designing a fully automated latent fingerprint matcher is a very challenging task as it needs to address many challenging issues including the separation of overlapping structured patterns over the partial and poor quality latent fingerprint image, and finding a match against a large background database that would have different resolutions. Currently there is no fully automated latent fingerprint matcher available to the public and most literature reports have utilized a specialized latent fingerprint matcher COTS3 which is not accessible to the public. This will make it infeasible to assess and compare the relevant research work which is vital for this research community. In this study, we target to develop a fully automated latent matcher for adaptive detection of the region of interest and robust matching of latent prints. Unlike the manually conducted matching procedure, the proposed latent matcher can run like a sealed black box without any manual intervention. This matcher consists of the following two modules: (i) the dictionary learning-based region of interest (ROI) segmentation scheme; and (ii) the genetic algorithm-based minutiae set matching unit. Experimental results on NIST SD27 latent fingerprint database demonstrates that the proposed matcher outperforms the currently public state-of-art latent fingerprint matcher.

Motivation & Objective

• To develop a fully automated latent fingerprint matcher that eliminates manual intervention in ROI detection and minutiae extraction.
• To address the challenges of poor image quality, overlapping structured noise, and partial prints in latent fingerprint images.
• To improve matching accuracy by integrating a self-learning dictionary learning-based ROI segmentation module with a genetic algorithm-based minutiae matching unit.
• To provide a publicly available, reproducible system for benchmarking and advancing latent fingerprint recognition research.

Proposed method

• A dictionary learning-based ROI segmentation module is used to automatically detect and extract the region of interest from latent fingerprint images.
• The segmentation module learns from training data to adaptively identify foreground regions while suppressing structured noise such as stains, lines, and graphics.
• Minutiae are automatically extracted only from the segmented ROI to improve feature reliability and reduce false matches.
• A genetic algorithm-based matching unit performs set-wise comparison of minutiae between the query latent print and database prints, optimizing matching scores through evolutionary search.
• The system is evaluated in three scenarios: baseline (no segmentation), proposed segmentation + GA matching, and Cao’s method + GA matching, to isolate the impact of segmentation.
• The fitness function in the GA unit is designed to maximize matching score based on minutiae correspondence, orientation, and spatial distribution.

Experimental results

Research questions

• RQ1 Can a fully automated latent fingerprint matcher achieve high performance without manual ROI labeling or feature extraction?
• RQ2 How does a self-learning dictionary learning-based segmentation module improve the reliability of minutiae extraction in low-quality latent prints?
• RQ3 What is the performance gain of integrating a dedicated ROI segmentation module with a genetic algorithm-based matching unit compared to existing methods?
• RQ4 How does the proposed matcher perform across different image quality categories (Good, Bad, Ugly) in the NIST SD27 database?
• RQ5 To what extent does the proposed system outperform public state-of-the-art latent fingerprint matchers in terms of mean penetration rate and CMC?

Key findings

• The proposed matcher achieved a mean penetration rate (pr) of 38.159% in Scenario 1, where no segmentation was applied, establishing a strong baseline.
• In Scenario 2, using the proposed self-learning segmentation module, the mean penetration rate improved to 34.496%, demonstrating the effectiveness of the ROI segmentation in enhancing matching reliability.
• Scenario 3, using Cao’s segmentation method, yielded a mean penetration rate of 36.434%, confirming that the proposed segmentation method outperforms existing alternatives.
• The system achieved the highest performance on 'Good' quality latent images, with a mean penetration rate of 28.40%, while performance degraded on 'Bad' (36.17%) and 'Ugly' (39.11%) classes due to poor image quality.
• The CMC (Cumulative Matching Characteristic) curves show that the proposed matcher consistently outperforms the baseline and other segmentation-based approaches, especially in the early ranks.
• The study confirms that the integration of a robust, self-learning segmentation module significantly enhances the accuracy of automated minutiae-based matching in latent fingerprint recognition.