BSI seeks background removal technologies for new biometric benchmark

Germany’s Federal Office for Information Security (BSI) is calling on developers to submit background removal technologies for a new benchmark focused on facial images, bringing standardized evaluation to an image-processing task that can affect biometric image quality.

The BGR-PROJECT, or Background Removal Benchmark, is accepting software implementations and device-based solutions. BSI says its goal is to obtain a “representative and technically sound selection” of current background removal algorithms that will subsequently be used for benchmarking. Submissions are open until November 27, 2026.

Background removal separates a person from the surroundings in a photograph so the subject can be recomposited onto a uniform background. BSI points to international requirements for identity and travel-document portraits as one reason the technology is relevant to biometrics.

The task sounds straightforward. The technical challenge is determining exactly where the person ends and the background begins without introducing changes that degrade the resulting facial image.

Why remove the background?

Portrait photographs used for identity documents need to meet specifications governing how the subject is presented.

BSI says an isolated subject can be recomposited onto a uniform, plain background as required by standards including ISO/IEC 39794-5 and ICAO Doc 9303 specifications for machine-readable travel documents, which set requirements intended to prevent distracting patterns, shadows or objects behind the head.

Background removal provides a way to separate the person from the original scene before placing the subject against the required background.

Doing that reliably, however, requires considerably more than simply drawing an outline around a face.

Finding where the person ends and the background begins

At its simplest, a background removal algorithm has to determine which parts of an image belong to the person and which belong to the surroundings.

That becomes difficult when there is little visual contrast between the two. Dark hair against a dark background, for example, can make it difficult to determine precisely where the hair ends and the background begins. Clothing can create a similar problem when its color or texture closely resembles the surrounding scene.

Facial portraits make the separation particularly challenging because the boundary around a person is not always a clean line. Individual hair strands, blurred edges and other fine structures can contain contributions from both the subject and the background.

Technically, BSI describes the task as estimating a per-pixel segmentation mask, or more precisely an alpha matte. The algorithm determines the contribution of the foreground and background at each pixel, including the transition regions where the two blend.

Errors in that process can produce visible artifacts.

BSI identifies halos around the subject, clipped hair and color spill from the original background among the problems background removal algorithms need to avoid.

In biometric applications, those errors may matter beyond the appearance of the resulting photograph. BSI says artifacts introduced during background removal may degrade the quality of the facial image and, consequently, the performance of downstream face recognition.

That makes preserving the subject while removing the surroundings an important part of the problem BSI wants the benchmark to examine.

BSI moves toward standardized evaluation

The BGR-PROJECT is intended to establish a standardized framework for evaluating and comparing background removal technologies in the context of facial images.

The project is currently in its submission stage.

The public project description does not yet detail the complete set of metrics that will be used in the benchmark or state that downstream face recognition performance itself will be measured.

That distinction is important. BSI identifies possible degradation of face recognition performance as a reason background removal quality matters, rather than describing face recognition testing as an explicit benchmark objective.

Software and device vendors invited to participate

BSI is accepting two types of submissions: software-based implementations and device-based solutions.

For software submissions, developers must first register with BSI and exchange PGP credentials. Implementations can be provided as scripts or binaries and must run on Ubuntu 24.04 or Windows 11 on x86-64 systems.

The software receives an RGB PNG image and must return a mask with the same dimensions. The mask represents background pixels, foreground pixels and pixels influenced by both foreground and background.

Software evaluations will be conducted by secunet Security Networks on behalf of BSI. According to the repository, submitted software will be used exclusively for evaluation and will not be provided to third parties other than secunet.

BSI is also accepting device-based technologies, with photo booths cited as an example. Vendors are asked to contact BSI to coordinate the evaluation process before sending a device.

The deadline for submissions is November 27, 2026.

The project puts a standardized evaluation effort around a processing task that can appear simple but becomes considerably harder at the boundaries of a biometric subject: removing what is behind a person without inadvertently changing the facial image that remains.

Article Topics

biometric data quality  |  biometric research  |  biometrics  |  German Federal Office for Information Security (BSI)  |  Germany  |  travel documents