OFIQ community reviews early results of biometric quality assessment tool

The standardization of image quality for face biometrics is a major step towards making population-scale biometric systems functional, and as such organizations around the world are adopting the Open-Source Face Image Quality (OFIQ), the reference implementation for the standard.

The European Association of Biometrics hosted an OFIQ User Group Meeting, organized in cooperation with the German Federal Office for Information Security (BSI) and eu-LISA.

The meeting was chaired by Christoph Busch of Hochschule Darmstadt, and included three presenters from secunet who are part of the project team that developed OFIQ.

Busch noted that the OFIQ user group is growing, as indicated by the 300 registrations for the meeting.

Empowering large-scale biometrics applications

The ISO/IEC 29794-5 standard for face biometric sample quality was published in completed form in April, a year after the draft and assessment software library were released. The standard provides the quality metrics used in the requirements for ISO/IEC 19794-5 and ISO/IEC 39794-5 for biometric data interchange formats. As such, it applies to the use case of providing a reference image for a machine readable travel document (MRTD), a reference image for live enrollment at a kiosk, such as for the EU’s biometric Entry/Exit System (EES), and probe images, such as at an automated border control (ABC) gate.

In other words, compliance with the standard is how organizations ensure that when a person shows up at a border crossing, their image is suitable to be used for a face biometrics comparison.

ISO 29794-5 assesses quality through the unified quality score (UQS) and the component quality measures (CQM), the latter of which has capture and subject-related components. A low UQS indicates a facial recognition match using the image will likely have low accuracy. There are 27 factors which go into the unified score.

Anna Stratmann of BSI talked about the increasing use of facial images in public sector applications as one of the motivations for formulating the quality assessment standard, and the need for a common language and approach for those assessments. The quality requirements to ensure reliable performance with large scale databases are very high, she notes.

Stratmann also points out that transaction time must be kept in mind, as a system which delivers perfect quality but takes too long could incur unacceptable costs or fail to meet the operational requirements of the use case.

Javier Galbally of eu-LISA explained the importance of OFIQ for his organization, given its operation of large-scale biometric databases and collaboration with multiple stakeholders  and data providers in multiple business areas. He picked up on Stratmann’s point about speed considerations, noting efforts to build them into OFIQ.

Galbally also reviewed the results of an evaluation carried out by eu-LISA with the help of the European Commission’s Joint Research Center. It compared the predictions about accuracy of the agency’s shared biometric matching system made by OFIQ to those from the same contractor that provided the system. OFIQ significantly outperformed the biometric engine vendor’s own quality assessment.

eu-LISA is supporting OFIQ by overseeing the tool’s maintenance.

How it works and next steps

Johannes Merkle of secunet outlined how the tool works. That begins with pre-processing algorithms for face detection, landmark estimation, alignment and segmentations. Algorithms then assess the various characteristics or components of image quality.

Merkle notes that secunet’s quality assessment algorithm has scored the best result out of 53 tested in NIST’s FATE Quality.

An algorithm searches for artifacts caused by digital compression, in both JPEG and JPG2000 formats, and for detecting if the subject’s eyes are open and mouth is closed, which Merkle notes did not show any demographic bias. An algorithm determines the ratio of occluded space if face occlusion is detected during the segmentation phase. The head pose algorithm used stands out as an area for improvement in NIST testing, but face detection errors appear to be a major factor in those results, he says.

The OFIQ Demonstrator, which takes the form of a GUI wrapper for the assessment tool, was explained by secunet’s Benjamin Tams. The visualization of the CQM results enhances the overall explainability of OFIQ.

The Demonstrator supports Windows, with Linux and MacOS versions nearly complete.

Secunet’s Maxim Schaubert discussed the platform independence of OFIQ and how the conformance of the implementations to the standard was tested.

OFIQ 2.0 is already in development by secunet, Merkle said during his presentation, with a launch hoped for before the end of 2027. Computational performance, largely meaning speed, is an area of planned improvement, along with accuracy, especially for background uniformity and expression neutrality. The developers are also hoping for reductions in  demographic bias (in under-exposure prevention) and to include additional quality checks, such as for motion blurring and gaze direction.

Article Topics

biometric data quality  |  EAB  |  EAB 2025  |  European Association for Biometrics  |  face biometrics  |  German Federal Office for Information Security (BSI)  |  Open Source Face Image Quality (OFIQ)  |  secunet