EAB tackles future of fingerprint biometrics image quality assessment

The tool used around the world to judge the quality of fingerprint images has been updated, giving it more flexibility and capabilities, though work on its limitations remains. Expert speakers from around the world discussed the latest version of NFIQ, the open source image quality assessment tool, along with assessing synthetic data and contactless fingerprints in a recent European Association for Biometrics (EAB) event.

Elham Tabassi of NIST spoke about the history of NFIQ software, which came out of explorations of why fingerprint biometrics do not work well with a small percentage of people. The FBI adopted the quality assessment metric, which then required the code to be made publicly available.

Tabassi spoke about the uses of quality assessment and factors in fingerprint image quality, as well as challenges for the further development of NFIQ and other fingerprint quality assessment algorithms. She also went into the details of how NFIQ 1.0 and 2.0 compare.

Christoph Busch of NTNU and ATHENE considered the link between NFIQ 2 and ISO/IEC 29794-4:2017. 29794-4 is a set of explainable algorithms, Busch says, in contrast deep learning. He explains the role of the RandomForest machine learning algorithm it uses. He also discussed quality metrics for fingerprint images, such as orientation certainty level.

Greg Fiumara of NIST followed with a presentation on code changes in NFIQ 2.1, and gave credit to HID Global’s Ralph Lessmann and UMD student Andrew Figlarz for exceptional contributions to the open source development.

New elements include a new command-line interface, cross-platform support, and multithreading for faster results. API integration has been simplified, and new compliance tests introduced. Another significant change with version 2.1 is reduced tolerance for images that are not 500 PPI, or images without the resolution included.

Official support for Android (which was supported in NFIQ 2.0) and iOS, a C API, and code optimization are among the future projects for NFIQ, Fiumara says.

Lessman spoke about HID’s experience implementing NFIQ 2.0. These included platform and compiler compatibility, but also resource utilization and the affect of different canvas sizes on scores. He also discussed what specifically quality scores mean.

Ramon Blanco of euLISA spoke about how NFIQ 2 is used in the European Entry-Exit System, and Mickey Cohen of Shanit presented the findings from a comparison of scores generated by NIFQ 1 and NFIQ 2.1.

Contactless fingerprints in focus

Veridium CTO John Callahan, spoke about touchless fingerprint capture on mobile devices in the field. The technology performed well with worn prints in study by Anil Jain in India funded by Bill and Melinda Gates Foundation in 2018, Callahan says, and has been successful in commercial deployments by Mexico’s INE for financial inclusion, Peruvian police and a telecom operator, Hamburg police and a financial inclusion program in Pakistan. In the latter deployment, NADRA has paused the use of contactless fingerprinting pending answers around enrollment and NFIQ 2 evaluation, though the match results were highly satisfactory.

He also talked about the potential for real-time quality assessment, and the use of NFIQ 2 to produce fusion quality scores for multiple fingers.

Callahan recently reviewed NIST’s new guidance on contactless fingerprint biometrics in an interview with Biometric Update.

A panel concluded the first day’s activities, moderated by NIST’s Fiumara and including representatives of Veridum, euLISA, Shanit and Thales.

The panel discussed when customers should demand a mechanism like NFIQ to assess the quality of the biometrics they collect. For larger databases and interoperability between siloed systems, quality assessment is more important than in, for instance, an enterprise time and attendance use case.

On the second day, Christopher Schiel of Germany’s federal law enforcement BKA spoke about his agency’s use of NFIQ 2.0.

A series of presentations then focused on NFIQ 2 and contactless fingerprints.

Secunet’s Johannes Merkle presented a case study which showed that some NFIQ 2 features are quite predictive of contactless print quality, while others are likely to require tuning. Approximation of 500 PPI resolution is a particular challenge for contactless fingerprint biometrics, as does providing users with feedback for positioning guidance.

NIST’s John Libert discussed how NIST developed some of its guidance on contactless fingerprints and image assessment.

Areas of contactless fingerprint images with issues like saturation or shadows can sometimes result in the appearance of false minutiae, Libert notes. Ultimately, a new or adapted model trained with contactless images will be necessary to properly asses the image quality, he concludes.

Jannis Priesnitz of ATHENE’s da/sec Research Group spoke about what would be needed to apply NFIQ 2.0 to contactless prints. The groups’ research shows that for contact print datasets with consistent quality, NFIQ has relatively little predictive power for whether the image is good enough quality to match, but with datasets of heterogeneous quality, predictive power increases. For contactless, NFIQ 2.0 can be useful, depending on how much preprocessing has been applied.

He also presented findings from a self-captured database, which found differences based on the finger being analyzed, and that NFIQ 2.0’s predictive power was low for unoptimized contactless samples.

Andreas Uhl of Paris Lodron University of Salzburg talked about using NFIQ 2.0 with synthetic data, and research indicating that fingerprints can be generated that behave similarly to real samples in terms of data quality distributions.

Lessman reviewed tools that can be used to retrain (or more accurately, the experts say, “modify”) NFIQ 2, along with several caveats and recommendations.

Jean-Christophe Fondeur of Idemia concluded the solo presentations, addressing the acquisition of contactless fingerprint images. He discussed dedicated capture devices, calling them “controlled,” and contrasting them with uncontrolled capture systems, like smartphones. Controlled contactless fingerprint systems can deliver similar biometric performance to contact systems, Fondeur says, referring to NIST statistics, and both capture technologies are mature enough to begin delivering benefits today.

Article Topics

biometrics  |  biometrics research  |  contactless biometrics  |  EAB  |  European Association for Biometrics  |  fingerprint recognition  |  machine learning  |  NIST  |  open source  |  standards