Talk to usTry the demo

Face Image Quality: What the 29794-5 Standard, the ICAO Portrait Report and the US Evaluation Measure, and Why Capture Changes Match Results

Evaluation and benchmarks

For: Product owner, machine learning engineer or procurement reviewer who has to decide what a verification flow does with a poor selfie or a poor document portrait

Which defects of a face photograph the public sources can measure, what each does to match errors as reported, and how to read those figures before you set a quality gate.

A person holds up a phone, the selfie looks fine to a human, and the match against the document portrait fails. Another selfie, taken in a dim room at arm's length with the head turned, passes. Before changing the matching threshold, ask what the two photographs actually contained. Face recognition systems depend on the quality of the image they are given, and there is now a public, standardised vocabulary for describing that quality, a reference implementation, an international travel-document report on how portraits should be captured, and an ongoing public evaluation of software that detects image defects.

This article reads those four sources side by side. It separates two ideas that product pages merge, utility for recognition and conformance to a requirement; lists the defects the sources can measure; reports what the German Federal Office for Information Security found each defect does to false non-match errors; and then lists the limits of those figures.

The article closest to this one is how to read a face recognition evaluation, which explains the error rates of the match itself. This one is about the input to the match. The article on face age estimation accuracy is a sibling in method: it reads a public evaluation of a different face analysis task.

Two meanings of "quality"

The word covers two different things, and the sources keep them apart.

The first is utility: how well an image supports recognition. The Federal Office for Information Security describes face image quality assessment algorithms as designed to calculate a unified quality score for a captured image that indicates its utility for recognition purposes. The second is conformance: whether the image meets a specific requirement, such as a frontal pose or open eyes. The same document explains that the unified score aims to predict utility, while the quality components each assess the conformance of the image to a certain requirement. It adds that the components make the decisions of quality assessment more transparent for users and operators, which is essential in biometric systems such as the European Entry/Exit System.

The US National Institute of Standards and Technology draws the same line as a scalar and a vector. A quality scalar converts an image into one value, and low values might trigger the collection of a new image or a switch to another modality. A quality vector reports properties related to recognition failure, such as focus, illumination, distortion and noise, and subject-related properties such as head pose, expression and eyeglasses, which makes it useful for actionable feedback to a user or photographer.

The consequence for a verification flow is practical. A single score can say "retake", but only the components can say what to retake: hold the phone level, remove the cap, move into the light. A supplier that reports only one number leaves the user guessing and leaves you unable to tell which defect is driving rejections.

What the international standard says

The standard for face image data quality is part 5 of the series on biometric sample quality and was published jointly by ISO and IEC in April 2025 as Edition 1, number 29794-5. Its public abstract says it establishes requirements on implementations that quantify how a face image's properties conform with those of canonical face images, for three use cases. The first use case is the collection of reference samples for identity documents, the second is sample system enrolment, and the third is probes for instantaneous response. The abstract also establishes terms and definitions for quantifying face image quality and specifies methods for it.

It is as important to read what the abstract excludes. The standard does not establish requirements on assessing the quality of pairs or sequences of images, on assessing the quality of three-dimensional captures, on encodings of face image quality data, or on performance evaluation of face image quality assessment algorithms. It adds that the software it covers inspects exactly one image and that the computations can be applied separately to each image of a pair. The use cases primarily address images of subjects who consent to the processing of their biometric data or for whom capture is operationally authorised.

Three consequences follow, and they are our reading of the abstract rather than a statement in it. First, a claim of alignment with the standard says what is computed, not how well it is computed: the standard leaves the performance evaluation of the algorithms to other work, which is where the public evaluation in a later section comes in. Second, the public abstract supplies no product-specific acceptance threshold, and the full text was not available to us; the threshold for your product is a decision you make from your own error costs. Third, a selfie and a document portrait are two images in a pair, and the standard scores each alone.

The reference implementation

The German Federal Office for Information Security develops the Open Source Face Image Quality framework, abbreviated OFIQ, as the software-side reference implementation of the algorithms in the standard. The office says the standard's revision and the software were developed in parallel, with findings flowing both ways, and that OFIQ is applicable to biometric applications such as border control. The EU agency that runs the large-scale IT systems for border and visa management has contributed to the maintenance of OFIQ since August 2024. It describes OFIQ as the reference face image quality assessment tool of the standard, and says the tool also follows the Commission implementing decision (EU) 2019/329 for the Entry/Exit System, which requires that the quality of facial images comply with the frontal image type of the 2011 edition of the earlier face image format standard. The agency also notes that quality measures are valuable as self-auditing tools to check that the quality of biometric data held in large datasets meets the minimum levels needed for the expected recognition accuracy.

OFIQ outputs a vector of quality measures, each an integer in the range from 0 to 100, where higher values signify higher quality. Its components fall into two groups.

  • Capture-related components depend on the environment of image acquisition: background uniformity, illumination uniformity, moments of the luminance distribution, over-exposure prevention, under-exposure prevention, dynamic range, sharpness, no compression artifacts and natural colour.
  • Subject-related components depend on the presentation of the face: single face present, eyes open, mouth closed, eyes visible, mouth occlusion prevention, face occlusion prevention, inter-eye distance, head size, crop of the face, head pose, expression neutrality and no head coverings.

What the ICAO portrait report asks for

The International Civil Aviation Organization published a technical report on portrait quality for reference facial images in travel documents, Version 1.0, in April 2018. Its scope is the enrolment of biometric reference data for the electronic travel document, as an application profile. It says image capturing for verification and identification, such as automated border control, is outside its scope, even if many of its requirements apply to such images too. The report notes that the portrait is one of the most important information carriers binding the document to the holder. For some criteria the report gives two levels: a minimum requirement, and a recommended value that will result in better overall performance or quality.

This is the document to read to understand what the portrait on a passport or identity card was supposed to look like when it was issued, which is the reference your selfie is compared with. Some of the specific values:

  • Resolution. The inter-eye distance in the captured photograph shall be at least 90 pixels for legacy applications, and for the design of a new process it should be at least 240 pixels.
  • Chip image. For the image stored in the chip, the inter-eye distance should be at least 90 pixels, preferably 120.
  • Focus. The image shall be in focus from nose to ears, and the report says a depth of field of 15 cm will be sufficient in most cases.
  • Pose. The pitch of the head shall be less than 5 degrees from frontal, the yaw less than 5 degrees, and the roll less than 8 degrees, with 5 degrees recommended for roll. The report adds that any stronger pose deviation may have a negative impact on facial recognition error rates.
  • Lighting. Illumination shall be adequate and uniform, with no difference between the brightness of the right and left side of the face. Its uniformity check measures the mean intensity in four squares on the forehead, the cheeks and the chin, and requires the maximum to be at most twice the minimum in each colour channel.
  • Expression and eyes. The face shall have a neutral expression with the mouth closed, and both eyes shall be open naturally, with pupils and irises completely visible.
  • Compression. The JPEG compression ratio shall not exceed 15 to 1, and lossy compression may be applied once per step: by the camera, by the photographer or citizen, and by the issuer.
  • Retouching. Images shall not be modified locally, and background removal shall not be implemented.

Two points matter for remote capture. The report states in a note that selfie-style portraits are likely not to maintain the minimal camera-subject distance requirement. And its annex on camera-subject distance reports a study of a database of excellent-quality images of 435 test subjects, in which camera-subject distance did not have a great influence on face verification performance over the range investigated, while the enrolment and verification images should be captured from a similar distance whenever possible. So the distance requirement is stricter than the measured effect on those excellent images; whether distance or blur or pose matters more in a remote flow is an untested hypothesis here, and the pilot at the end of this article is a way to test it.

What can be measured: the US evaluation

The US National Institute of Standards and Technology runs an ongoing evaluation of software that detects image defects, called Face Analysis Technology Evaluation Quality, with a track named Specific Image Defect Detection. In September 2023 it released its first report on the track, Internal Report 8485, which summarises accuracy for seven submissions from five developers across 20 quality measures. The report says the measures implemented most frequently were the total number of faces present, pitch, yaw, roll, eyes open and inter-eye distance. It also marks which measures must be checked for an image to be used as a reference photograph in a machine readable travel document, which is the first use case of the standard.

The test method is worth knowing, because it decides what the results can show. The test images are sequestered at the institute, so developers have no access to them. Ground truth comes from three approaches: camera placement at a known angle at the time of capture, manual labelling by human inspection, and synthetic degradation, in which varying amounts of blur, overexposure and underexposure are applied to a natural image. The report states a limitation of the second approach: where ground truth is measured by inspection, for example for inter-eye distance and mouth aperture, it has some measurement error, so the software can never be perfect. The test sets range in size from 30 images for face occlusion to 12000 for roll.

The institute's page also explains how the unified score is judged, and the method is the one that connects image quality to match errors. A quality assessment algorithm can make two kinds of error: false rejection, saying an image is poor when it is not, which drives costs, and false acceptance, saying an image is good when it is not, which drives future recognition errors. The page then sets a recognition threshold so that the false non-match rate of a set of accurate verification algorithms is 1 percent, discards a fraction of the images ranked last by the quality algorithm, and recomputes the false non-match rate. If the quality score predicted failure perfectly, discarding that fraction would remove every false non-match. In practice the quality algorithm does not always assign low values to the images that cause false non-matches.

The same page records that the institute has added a demographic analysis of the unified quality score, with plots by region of birth and sex, and a section on thresholds on quality measures and on demographic variability in quality measures. That is a reminder that a quality gate is itself a classifier whose rejection rate can differ between groups.

What each defect does to match errors

The German office's short report on OFIQ is the most direct public source on the question the title asks. It evaluates the quality measures in several ways, including how much the false non-match rate of face recognition algorithms falls when images are filtered by a quality measure. It uses error-versus-discard curves, several open-source and commercial recognition algorithms, and a large test set derived from a public face database. Read every figure below against one baseline: the report starts from a false non-match rate of 10 percent, and then gives the rate after discarding the 10 percent of images that the quality measure ranks as worst.

  • Unified score. Algorithms based on the MagFace approach, which encodes quality in the magnitude of the face feature vector, achieved the strongest results, and one such model decreased the false non-match rate from 10 percent to 6 percent.
  • Sharpness and compression artifacts. These were found to have the highest impact among the capture-related components; discarding the 10 percent of images ranked last on them reduced the false non-match error from 10 percent down to 8 percent.
  • Illumination, exposure, dynamic range and natural colour. Several components, including illumination uniformity, the moments of the luminance distribution, under-exposure prevention, over-exposure prevention, dynamic range and natural colour, had only a marginal impact: from 10 percent down to 9.5 percent. The office concludes that state-of-the-art face recognition systems are largely robust to non-severe defects caused by these components.
  • Background. Background uniformity was found to have negligible effects on face recognition performance.
  • Occlusion. Discarding images by face occlusion prevention, eyes visible and mouth occlusion prevention reduced the error by up to 2 percentage points, from 10 percent down to 8 percent. The office says this confirms the well-known fact that facial occlusions affect recognition negatively, in particular in the ocular region.
  • Eyes and face size. Eyes open gave a reduction of up to 2 percentage points, and inter-eye distance reduced the error from 10 percent to around 8 percent.
  • Pose. Roll and pitch had marginal impact, with reductions of 1 to 2 percentage points depending on the recognition system. Yaw had a higher impact: for some systems the error fell by more than 4 percentage points, from 10 percent to below 6 percent, and for others the reduction was less pronounced.
  • Motion blur. Discarding images by motion blur prevention reduced the error from 10 percent to 8 percent, but the office assumes the algorithm detects ordinary blur as well as motion blur.
  • Mouth, head coverings and expression. These had only marginal impact, reducing the error by up to 0.5 percentage points. The office notes that recognition systems are largely robust against variation in facial expression, and that head coverings rarely affect accuracy unless they occlude more relevant facial regions.

Put plainly, the ranking in that evaluation is: sharpness and compression, yaw, occlusion of the eye region and face size at the top, and illumination, background and expression at the bottom. That ranking is what the office measured on one test set with the algorithms it used, and it is a ranking of impact on recognition, not a statement that the lower items are unimportant to a human reviewer or to a regulator.

Reading these figures with care

Seven cautions apply, and the sources state most of them themselves.

The baseline is chosen. The 10 percent starting error is a feature of the evaluation design, not of an operational system, and an operational threshold may start from a much lower error. The institute's page starts its own design from a false non-match rate of 1 percent, and it reports that the quality algorithms are more effective when the initial rate is set low, and that if they are asked to detect the worst 5 percent of images, the relative gains are reduced.

Detection accuracy and recognition impact are different questions. The German report evaluates each algorithm twice, for how accurately it detects the defect on labelled images and for how much filtering by it helps recognition. It reports detection accuracy generally above 90 percent for capture-related components and above 95 percent for most subject-related ones, with exceptions, in particular significantly worse results for motion blur, where results reported in the scientific literature could not be reproduced.

Synthetic defects are not field defects. The German report says results obtained on synthetic test data should be treated with care, and that the institute's test sets were seen to include images with irrelevant defects, for example slight degradations of contrast and brightness that are not expected to affect recognition. For motion blur it says that simulating it is difficult.

Defects interact. The evaluation measures one component at a time. A dark, slightly blurred, off-axis selfie of a person wearing a cap combines several marginal effects, and no source in this article reports the combined effect.

The matcher matters. The report says some recognition systems were more robust than others to variation in the roll and pitch angle, and that the yaw reduction was less pronounced for other systems. A figure for one matcher is not a figure for yours.

Cooperative capture is assumed. The standard's use cases address consenting subjects, and the German report refers to cooperative capture scenarios when it discusses strong yaw. A remote flow has users who are cooperative but uncoached, which sits between a photo booth and a border gate.

Quality is not liveness. A high-quality photograph of a photograph has excellent sharpness and lighting. The ICAO report lists presentation attack detection and morphing prevention among the topics outside its scope. Quality scoring and attack detection are different layers, and the article on presentation attacks, deepfakes and injection attacks places them.

From the sources to a remote flow

The sources do not describe a remote flow, so what follows is our reasoning from them, not a requirement.

  1. Treat the document portrait and the selfie as two different quality problems. The portrait is a small image, printed or stored, produced once under issuer rules and then photographed or read; the ICAO report's values apply to its issuance, not to your capture of it.
  2. Give feedback by component. The German office states the purpose of components as actionable feedback for users and operators when a re-capture is initiated.
  3. Order the checks by measured impact, but keep the cheap ones. Face presence, a single face, size and sharpness before pose, and pose before background, which cost little to check.
  4. Set thresholds from your own data. The public abstract of the standard gives none, and the figures above are conditional on a baseline you do not have. Measure the false rejection that the gate causes, the false non-match rate with and without it, and the split by device and region.
  5. Log the components with the decision. A pass or fail without the underlying measures cannot be audited later or compared between suppliers.
  6. Limit retries by design. Each retry costs a user; the gate should stop asking at some point and hand over to a fallback path.
  7. Check the gate for group differences. The institute added a demographic analysis of the unified quality score to its report.

A quality-gate pilot

None of the public sources ranks distance, blur and pose for a remote onboarding flow, so a team that wants the ranking has to measure it on its own capture conditions. This is a design, with no result attached.

  1. Collect genuine pairs, a selfie and a document portrait of the same person, from the real device mix and the real lighting, with the consent that the collection needs.
  2. For every pair store the quality components with the decision: inter-eye distance, a sharpness measure, head pose angles and the other components of the reference implementation.
  3. Fix the match threshold in advance and do not change it during the pilot.
  4. Stratify the pairs by one component at a time, for example sharp against blurred with distance and pose held in a narrow band, and compute the false non-match rate in each stratum.
  5. Rank the components by how much each stratum moves the false non-match rate, and report the intervals, not just the point values, so that a small stratum is not read as a finding.
  6. Replay the pairs through a candidate quality gate and record two numbers: the genuine sessions the gate would have rejected, and the false non-matches it would have prevented.
  7. Split every figure by device class and by the other groups the team monitors, as in the last item of the list above.

Questions to put to a supplier

  • Which quality measures does the product compute, and are they the components of the standard or its own?
  • Is the answer a single number or a vector, and what does the user see when a capture is rejected?
  • Which defects block a capture and which only lower a score?
  • How were thresholds chosen, on whose data, and what are the false rejections caused by the gate?
  • Was the quality software submitted to the public defect-detection evaluation, and for which measures?
  • Is the quality check separate from the liveness check, and what does each report?
  • How does the gate behave on low-end cameras and on glasses in poor light, and what is the split of rejections by group?
  • Are quality measures stored with the session result, and for how long?

Where this sits in a verification flow

A NeuroVision KYC session places document, selfie, liveness and face match in one session. The face match step compares the selfie with the document portrait in a one-to-one comparison. The liveness step is active: the user moves the face closer through two ovals. The platform states input rules for images: JPEG or PNG only, with minimum sizes for the document and for the face. To discuss your device mix and capture conditions, contact the team.

Sources

All sources were fetched on 28 October 2026.

Ready to talk?

Tell us what you are building and we will help you plan the integration.

Talk to us