Assessments for image quality in face biometrics improving, new standard TBD
Algorithms for assessing the quality of facial images used in biometric matching are increasingly effective, but a modern standard that supports interoperability and can be applied to different use cases may yet be some way off.
The European Association for Biometrics (EAB) held the three-day workshop on face image quality in cooperation with the U.S. Department of Homeland Security’s Office of Biometric Identity Management, National Institute of Standards and Technology (NIST), European Union Agency for the Operational Management of Large-Scale IT Systems in the Area of Freedom, Security and Justice (eu-LISA), the ISO/IEC JTC1 SC37 working group 3 and project iMARS to explore key topics and themes relating to facial recognition.
Expert speakers on the first day of the event included experts from DHS-OBIM, and DHS’ Maryland Testing Facility (MdTF), and universities Hochscule Darmstadt and NTNU.
Day two featured representatives of Innovatrics, Biometix, Neurotechnology, secunet, Rank One and Idemia, among a series of presentations on the implications of image quality for presentation attack detection (PAD) and industry experiences addressing facial image challenges.
A panel representing law enforcement agencies from around the world discussed operational aspects of face biometrics systems.
Moderated by Patrick Grother of NIST, Frøy Løvåsdal of Norway, Johanna Morley of the UK, Istvan Racz of eu-LISA, Bill Perry from Australia, and Arun Vemury of the U.S. each spoke about how face image quality is assessed in their systems, and the steps they have taken to ensure good-quality probe and reference images.
Morley noted the necessity of getting image capture stations right “because it’s the one thing that we have control over, so we just need to get that bit right.” She also said that the use of commercial image quality assessment software has made a significant difference in delivering consistent quality of reference images. For probe images, “all bets are off,” despite guidance on CCTV operation from Home Office.
One of the issues that has improved since the adoption of the image quality software, according to Morley, is the standardized ordering of profile and frontal images.
The panelists describe several significantly different systems and approaches to ensuring quality images, from physical photography setup efforts in Norway to common standards for EU member states to the service contract clauses for “service delivery partners” (STPs) in Australia, who deliver data before people arrive at the country’s borders.
Grother asked if an effective quality assessment algorithm would improve the situation in production. Løvåsdal responded that it would, but that the ISO/IEC 19794-5 standard is ambiguous, and compliance does not necessarily ensure good image quality. Worse, biometrics vendors and forensics experts may not adequately agree on what constitutes a good quality image.
The relative advantages of open-source and closed-source quality assessment technologies, such as in terms of interoperability and security, were also discussed by the panel.
Operational challenges come up frequently for many agencies, however, where image quality assessment will not necessarily deliver any benefit.
Perry described how Australia stipulates requirements in contracts to avoid vendor lock-in and ensure consistent results.
Vemury expressed interest in an equivalent for the NFIQ algorithm for other biometric modalities including face, though noted there are limitations to its use.
Panelists struggled, however, to explain the extent to which low-quality images degrade the effectiveness of matching. Grother referred to a situation in which a hair-style that occluded part of a subject’s face resulted in false positives, and noted that the situation was highly dependent on the algorithm used.
Supporting standard development
A panel moderated by Ramon Blanco of eu-LISA and including Pierre Gacon of Idemia, Grother, Ralph Lessmann of EAB and HID Global, and secunet’s Benjamin Tams closed the event.
The panel discussed the importance of the development of face image quality standards, and the benefit of such standards to biometric algorithm and system developers.
The use case behind the ISO/IEC 19794 standard, which is based on passport issuance and similar applications, is not entirely aligned with the needs of other use cases, Gacon points out. He also notes that NFIQ was also developed for a very specific use case, however.
Defining image quality in the first place can pose challenges, and given the use of capture devices from legacy mobile phones to dedicated cameras in controlled environments, any standard cannot set requirements for image quality uniformly across different applications.
The importance of image size to image quality has become significantly less important than it once was, as an example of a consideration that has evolved over time.
biometric liveness detection | biometrics | biometrics research | DHS | EAB | eu-LISA | European Association for Biometrics | face biometrics | facial recognition | forensics | Maryland Test Facility (MdTF) | NIST | OBIM | standards