ISO/IEC 29794-5:2025

International
Standard
ISO/IEC 29794-5
First edition
Information technology —
2025-04
Biometric sample quality —
Part 5:
Face image data
Technologies de l'information — Qualité d'échantillon
biométrique —
Partie 5: Données d'image de face
Reference number
© ISO/IEC 2025
All rights reserved. Unless otherwise specified, or required in the context of its implementation, no part of this publication may
be reproduced or utilized otherwise in any form or by any means, electronic or mechanical, including photocopying, or posting on
the internet or an intranet, without prior written permission. Permission can be requested from either ISO at the address below
or ISO’s member body in the country of the requester.
ISO copyright office
CP 401 • Ch. de Blandonnet 8
CH-1214 Vernier, Geneva
Phone: +41 22 749 01 11
Email: copyright@iso.org
Website: www.iso.org
Published in Switzerland
© ISO/IEC 2025 – All rights reserved
ii
Contents Page
Foreword .v
Introduction .vi
1 Scope . 1
2 Normative references . 1
3 Terms and definitions . 2
4 Abbreviated terms . 4
5 Conformance . 4
6 Common computations . 6
6.1 Overview .6
6.2 Conversion of 16 bits per channel images to 8 bits per channel images .7
6.3 Conversion of high bit-depth images to 8 bit greyscale or 24 bit colour images .7
6.4 Face detection.7
6.5 Face landmark estimation .8
6.6 Landmarked region segmentation .10
6.7 Face alignment .11
6.8 Face parsing . 12
6.9 Face occlusion segmentation . 13
6.10 Computing eye centres and inter-eye distance .14
6.11 Head pose estimation . 15
6.12 Conversion of 8-bits-per-channel colour images to luminance .16
6.13 Conversion of 8-bits-per-channel colour images to CIELAB space .17
6.14 Handling of greyscale images . .18
6.15 Luminance histogram .18
6.16 Entropy .18
6.17 Expressing binary quantities as continuous values .18
6.18 Representation and arithmetic of real and integer numbers .18
6.19 Normalization of image colour values .19
7 Quality measures . 19
7.1 General .19
7.2 Quality score (unified) . . 20
7.2.1 Description . 20
7.2.2 Computation of the native quality measure . 20
7.2.3 Mapping the computation result to the target range of the quality component .21
7.3 Capture-related quality components .21
7.3.1 General .21
7.3.2 Background uniformity . .21
7.3.3 Illumination uniformity . 22
7.3.4 Moments of the luminance distribution . 23
7.3.5 Under-exposure prevention . 25
7.3.6 Over-exposure prevention . 25
7.3.7 Dynamic range . 26
7.3.8 Sharpness . . .27
7.3.9 No compression artefacts . 28
7.3.10 Natural colour . 29
7.4 Subject-related quality components . 30
7.4.1 General . 30
7.4.2 Single face present . 30
7.4.3 Eyes open .31
7.4.4 Mouth closed .32
7.4.5 Eyes visible .32
7.4.6 Mouth occlusion prevention . 33
7.4.7 Face occlusion prevention . 34

© ISO/IEC 2025 – All rights reserved
iii
7.4.8 Inter-eye distance . . 34
7.4.9 Head size . 35
7.4.10 Crop of the face image . 36
7.4.11 Head pose . 38
7.4.12 Expression neutrality . 39
7.4.13 No head covering . 40
8 Face image quality block . 41
8.1 Binary encoding .41
8.2 XML encoding .41
8.3 Organization identifiers .41
8.4 Algorithm identifiers .41
Annex A (normative) Conformance test assertions .44
Annex B (informative) Quantitative goal for face image QAAs .50
Annex C (informative) Applications of quality measures .53
Annex D (informative) Quality requirements with no quality measure .56
Annex E (informative) OFIQ testing reports .58
Annex F (informative) Guidance for sequential use of ISO/IEC 29794-5 quality components .59
Bibliography .60

© ISO/IEC 2025 – All rights reserved
iv
Foreword
ISO (the International Organization for Standardization) and IEC (the International Electrotechnical
Commission) form the specialized system for worldwide standardization. National bodies that are
members of ISO or IEC participate in the development of International Standards through technical
committees established by the respective organization to deal with particular fields of technical activity.
ISO and IEC technical committees collaborate in fields of mutual interest. Other international organizations,
governmental and non-governmental, in liaison with ISO and IEC, also take part in the work.
The procedures used to develop this document and those intended for its further maintenance are described
in the ISO/IEC Directives, Part 1. In particular, the different approval criteria needed for the different types
of document should be noted. This document was drafted in accordance with the editorial rules of the ISO/
IEC Directives, Part 2 (see www.iso.org/directives or www.iec.ch/members_experts/refdocs).
ISO and IEC draw attention to the possibility that the implementation of this document may involve the
use of (a) patent(s). ISO and IEC take no position concerning the evidence, validity or applicability of any
claimed patent rights in respect thereof. As of the date of publication of this document, ISO and IEC had not
received notice of (a) patent(s) which may be required to implement this document. However, implementers
are cautioned that this may not represent the latest information, which may be obtained from the patent
database available at www.iso.org/patents and https://patents.iec.ch. ISO and IEC shall not be held
responsible for identifying any or all such patent rights.
Any trade name used in this document is information given for the convenience of users and does not
constitute an endorsement.
For an explanation of the voluntary nature of standards, the meaning of ISO specific terms and expressions
related to conformity assessment, as well as information about ISO's adherence to the World Trade
Organization (WTO) principles in the Technical Barriers to Trade (TBT) see www.iso.org/iso/foreword.html.
In the IEC, see www.iec.ch/understanding-standards.
This document was prepared by Joint Technical Committee ISO/IEC JTC 1, Information technology,
Subcommittee SC 37, Biometrics.
This first edition cancels and replaces the first edition of ISO/IEC TR 29794-5:2010 which has been
technically revised.
The main changes are as follows:
— the document has been completely revised to become and International Standard;
— information on the role of quality measures has been added;
— requirements on quality software have been added.
A list of all parts in the ISO/IEC 29794 series can be found on the ISO and IEC websites.
Any feedback or questions on this document should be directed to the user’s national standards
body. A complete listing of these bodies can be found at www.iso.org/members.html and
www.iec.ch/national-committees.

© ISO/IEC 2025 – All rights reserved
v
Introduction
Adoption of deep learning techniques has caused error rates associated with automated face recognition
tasks to be reduced. However, errors still occur and are often related to imaging, human factors, the level of
biometric capture subject cooperation, the comparison algorithm, and its associated threshold and decision
logic. Without significant modernisation of capture procedures, recognition errors will become more
prevalent as volumes increase. This document is aimed at reducing errors due to image quality, through
the use of quality assessment algorithms. Quality assessment algorithms have several roles (see Annex C),
primarily those related to sample capture. Drivers for improved capture are as follows.
— Need for improved usability — The general improvement of biometric systems has highlighted that
improved usability for both biometric capture subjects and human operators can reduce errors through
the improvement of capture. Without a careful consideration of both biometric capture subjects and
system operators, system designers risk seeing the limitations inherent in using technology alone.
— Increasing volumes — Vast numbers of face images are being collected in many commercial, civil identity
management and law enforcement applications. These photographs are used as reference enrolment
samples, or as recognition probes that, in turn, sometimes later serve as references.
— New programs — Future large-scale programs will employ face recognition: For example, in China the
railway transportation system uses face recognition for identity verification and to improve passenger
check-in efficiency. The European Union uses face recognition for biometric exit confirmation. The United
States currently uses face recognition for biometric exit confirmation and vessel boarding. In India, the
Aadhaar program allows face recognition for authentication.
— Face-blind cameras — Historically, many face images were collected using cameras that were not face-
aware. In contrast, in some situations concerning fingerprint and iris biometrics, capture devices run
in an auto-capture quality-assessment loop, with explicit awareness of the kind of image intended for
collection.
— Reliance on imaging design specifications — Faces collected for ID credentials and authoritative
databases are largely collected using cameras set up according to published documentary standards,
most recently ISO/IEC 39794-5, regulating geometry and photography. In the best case, face images
from such collections are then checked with image compliance tools. When photographs are collected
by a human photographer, this can be without any automated quality assessment, relying only on the
photographer to check conformance.
— Behaviour not intended by the relevant capture standard — Some recognition failures arise from
biometric capture subjects effecting differences in presentation in reference and probe images.
Standards define a canonical presentation to be centred and frontal with neutral expression, eyes-open
and without occlusions. Facial recognition systems are expected to operate accurately across a wide
range of individuals who vary in age, body size, ethnicity, language, culture, literacy and familiarity with
technology. Careful human factors design is vital to the acquisition of canonical images and improved
face image capture.
— Quality assessment is separated from the capture process — In many cases, a photograph is captured
and later submitted to a backend server while ensuring no image tampering occurs, where it is assessed
for quality. If poor quality is detected (by human or automated means), re-capture is initiated hours or
days later, when possible, with another encounter and attendant expense.
Regarding image quality, Table 1 lists characteristics of face image quality relating to the biometric
capture subject and characteristics relating to the capture process, demonstrating that issues due to mis-
presentation (often associated with human factors design) and issues related to imaging are in many cases
separable. For example, photographs can be systematically de-focused even when the biometric capture
subjects present perfectly.
© ISO/IEC 2025 – All rights reserved
vi
Table 1 — Characterization of face image quality
Biometric capture subject Capture process
characteristics
Static Biological characteristics; Capture process and capture device properties:
properties
— injuries and scars, — image resolution,
— dermatological conditions, — optical distortions,
— etc. — sub-optimal camera angle,
— field of view,
— etc.
Other static characteristics: Static properties of the background:
— thick or dark glasses, — (textured) wallpaper.
— permanent jewellery,
— makeup and cosmetics,
— etc.
Affordance:
— properties of a data capture subsystem that intuitively imply its
functionality and use to biometric capture subjects,
— human-centric system physical and process design.
Dynamic Behaviour: Scenery:
properties
— exaggerated expression, — background moving objects,
— hair across the eye, — variation in lightning.
Capture device variation:
— facial hair,
— de-focus,
— etc.
— camera vibration,
— sub-optimal camera angle,
— poor exposure,
— etc.
By defining image quality measurements, this document is intended to improve the accuracy of automated
face recognition systems. Quality can be tied to recognition accuracy (see Annex B). Improved quality can
also improve human review of images. The quality measures included in this document were selected because
guidance on how to control them has already been included in ISO/IEC 39794-5. The implementations of
[62]
some quality measures were evaluated for performance. The reference implementation defines quality
[58]
measures that use external algorithms with licence conditions.
[60]
This document recognizes the Open Face Image Quality (OFIQ) software as the reference implementation
[59]
of the requirements of the document. It is open-source. Other quality algorithm implementations can
conform to this document as described in Clause 5.
Some of the computations of this document can be effective on images captured with illumination at non-
visible wavelengths.
Encoding of quality data is defined in ISO/IEC 29794-1. The methodology for performance evaluation of
quality assessment algorithms is also defined in ISO/IEC 29794-1.
NOTE Use of this document can be subject to local regulations.

© ISO/IEC 2025 – All rights reserved
vii
International Standard ISO/IEC 29794-5:2025(en)
Information technology — Biometric sample quality —
Part 5:
Face image data
1 Scope
This document establishes requirements on implementations that quantify how a face image’s properties
conform with those of canonical face images, for example those specified in ISO/IEC 39794-5:2019,
Clause D.1, for three use-cases:
1) collection of reference samples for ID documents;
2) sample system enrolment; and
3) probes for instantaneous response.
This document also establishes terms and definitions for quantifying face image quality and specifies
methods for quantifying the quality of face images.
This document does not establish requirements on:
— assessing the quality of pairs or sequences of images;
NOTE This document establishes requirements for software that inspects exactly one image. This document
does not establish requirements for software that compares two or more images (such as biometric recognition).
However, the computations of this document can be applied separately to each image in a pair or sequence.
— assessing the quality of 3D captures;
— encodings of face image quality data;
— performance evaluation of face image quality assessment algorithms.
The use cases within scope of this document primarily address the assessment of images from data capture
subjects who consent to processing of their biometric data, or for whom biometric capture is operationally
authorized.
2 Normative references
The following documents are referred to in the text in such a way that some or all of their content constitutes
requirements of this document. For dated references, only the edition cited applies. For undated references,
the latest edition of the referenced document (including any amendments) applies.
ISO/IEC 29794-1, Information technology — Biometric sample quality — Part 1: Framework
ISO/IEC 39794-1:2019, Information technology — Extensible biometric data interchange formats — Part 1:
Framework
ISO/IEC 39794-5:2019, Information technology — Extensible biometric data interchange formats — Part 5:
Face image data
ISO/IEC 19794-1:2011, Information technology — Biometric data interchange formats — Part 1: Framework
ISO/IEC 2382-37, Information technology — Vocabulary — Part 37: Biometrics

© ISO/IEC 2025 – All rights reserved
IEC 61966-2-1:1999, Multimedia systems and equipment — Colour measurement and management — Part 2-1:
Colour management — Default RGB colour space — sRGB
IEC 61966-2-2:2003, Multimedia systems and equipment — Colour measurement and management — Part 2-2:
Colour management — Extended RGB colour space — scRGB
3 Terms and definitions
For the purposes of this document, the terms and definitions given in ISO/IEC 2382-37, ISO/IEC 29794-1,
ISO/IEC 39794-5 and the following apply.
ISO and IEC maintain terminology databases for use in standardization at the following addresses:
— ISO Online browsing platform: available at https:// www .iso .org/ obp
— IEC Electropedia: available at https:// www .electropedia .org/
3.1
canonical face image
face image conformant to an external standard or specification of a reference face image
Note 1 to entry: In many applications, the canonical face image is that given in ISO/IEC 39794-5:2019, Clause D.1, which
specifies a reference face image for a machine-readable travel document.
Note 2 to entry: Most of the computations of this document can be effective on images captured in automated border
control gates, visa images and other server-side images, which are in scope of this document.
3.2
de-focus
aberration in which an image or part of an image is out of focus
Note 1 to entry: De-focus tends to reduce the sharpness and contrast of the image.
3.3
face detection
process of determining whether and where faces are present in an image
3.4
face image
electronic image-based representation of the face of a capture subject
Note 1 to entry: Any image captured for use-cases 1 – 3 described in Clause 5 is considered as a face image.
Note 2 to entry: ISO/IEC 39794-5 includes a definition for face portrait as the visual representation of the capture
subject, which includes the full-frontal part of the head, including hair in most cases, as well as neck and possibly top
of shoulders. Face portraits appear in several places on and in a machine-readable travel document (MRTD).
Note 3 to entry: Given an image that has a roll angle of 90 ° or more (which is far from the presentation intended by
ISO/IEC 39794-5:2019), a QAA can assign low quality component values or fail to return a record
[SOURCE: ISO/IEC 39794-5:2019, 3.27, modified — Notes to entry have been added.]
3.5
face bounding box
rectangle containing the central region of interest of a face visible in the face image
Note 1 to entry: The face bounding box is used for the estimation of landmarks to restrict the face image to the region
of interest.
Note 2 to entry: The face bounding box is the result of the face detection process as defined in 6.4.

© ISO/IEC 2025 – All rights reserved
3.6
face landmarks
set of anthropometric points in the image marking the contour and different parts of the face
Note 1 to entry: The face landmarks are computed by the algorithm in 6.5.
3.7
landmarked region
minimal convex region of the face image containing all face landmarks
Note 1 to entry: The landmarked region of the face is defined by the landmark estimation algorithm, as described in
6.5. It encompasses the area from (and including) eyebrows to chin and from (but excluding) left ear to right ear.
Note 2 to entry: The inner region, as defined in ISO/IEC 39794-5:2019, 3.39 and D.2.2, is a coarse approximation of the
landmarked region.
3.8
face alignment
process of rotating, translating and scaling a face image so that the transformed image has certain
dimensions and the eyes, nose and mouth corners are approximately located at pre-defined locations in the
transformed image, and applying the same transformation to the face landmarks
3.9
face parsing
process of assigning semantic labels to regions contained in a face image specifying the part of the subject
depicted in the region
Note 1 to entry: The set of parts to which the pixels are assigned comprises various parts of the face, neck, hair,
clothing and various accessories. Pixels not belonging to the subject (background) are assigned to the value 0.
3.10
native quality measure
output of a quality assessment algorithm without constraints on data format and/or value range
[SOURCE: ISO/IEC 29794-1:2024, 3.10]
3.11
pose estimation
process of determining the 3-axis rotation of the head in an image
Note 1 to entry: Pose estimation requires a specified coordinate system for definition of the angles and their sense
(left hand vs. right hand). The default choice is given in ISO/IEC 39794-5:2019, 7.21.
3.12
sharpness
clarity of fine details in an image
Note 1 to entry: Sharpness will be improved with precise focus, good imaging resolution, and absence of biometric
capture subject motion relative to the camera.
3.13
ICC profile
International Color Consortium profile
set of data that characterizes colour input or output by defining the mapping between the data and a profile
defined colour space
Note 1 to entry: ICC profiles (e.g. sRGB and ROMM RGB) are intended to provide a standard approach to colour
management needs.
Note 2 to entry: The captured face portrait is a true-colour representation of the holder in a typical colour space such
as sRGB. Other true-colour representations, such as Adobe RGB (1998) or ProPhoto RGB (ROMM RGB), are used as
specified in ISO/IEC 39794-5: 2019, D.1.4.2.9.

© ISO/IEC 2025 – All rights reserved
3.14
quality assessment algorithm
quality algorithm
algorithm to calculate a quality measure
Note 1 to entry: The ISO/IEC 19785 series uses the term "quality algorithm".
[SOURCE: ISO/IEC 29794-1:2024, 3.13]
4 Abbreviated terms
For the purposes of this document, the acronyms and abbreviated terms given in Table 2 apply.
Table 2 — Acronyms and abbreviated terms
Acronym/ Definition
abbreviated term
BGR colour space with the order of colours (blue, green, red)
BU background uniformity
CNN convolutional neural network
DFT discrete Fourier transform
EVZ eye visibility zone
ICC International Color Consortium
IED inter-eye distance
IM illumination mean
IU illumination uniformity
NaN not a number
px pixels
QAA quality assessment algorithm
QC quality component
QCV quality component value
OFIQ Open Source Face Image Quality
QM quality measure
OpenCV OpenCV version 4.5.5
QS (unified) quality score
RGB colour space with the order of colours (red, green, blue)
UC use-case
5 Conformance
A claim of conformance to this document is made by asserting support for one or more use-case(s) (see
below) and a dated edition of this document.
To conform with this document, a face image quality assessment implementation shall:
— implement the computations marked as mandatory (M) in Table 3 for the claimed use case(s);
— conform to the requirements of Clauses 6 and 7;
— conform to the requirements of Clause 8 (face image quality block encoding); and
— meet the requirements specified in Annex A.

© ISO/IEC 2025 – All rights reserved
Conformance to the requirements of Clause 8 fulfils Level 1 and Level 2 conformance as specified in
ISO/IEC 39794-1:2019, Annex C. Conformance to the requirements of 7.2, 7.3 and 7.4 fulfils Level 3
conformance as specified in ISO/IEC 39794-1:2019, Annex C.
NOTE 1 Organizations deploying quality assessment algorithms in the various use cases can choose to compute the
provided quality measures and to relate their values to any use-case specific thresholds.
NOTE 2 For standardized interchange of quality values produced by implementations of this document,
ISO/IEC 29794-1 defines a standardized quality block and requires quality measures to be an integer between 0 and 100.
Three use-cases are considered:
— UC1:   Collection of reference samples for ID documents. The face image will be stored on a document,
used for example for a maximum of 10 years, and should support human examination.
— UC2:   System enrolment, current or later creation of a reference, delayed recognition. Acquisition of
face images where quality should be high enough to ensure later usage and interoperability.
— UC3:   Collection of probe samples for instantaneous recognition. Single use of a face image with
instantaneous response.
Table 3 — Conformance requirements by activity
# Face image quality measure Sub- UC1 UC2 UC3
clause
(ID (System (Probe for instantaneous
Documents) enrolment) recognition)
1. Quality score (unified) 7.2 M M M
2. Background uniformity 7.3.2 M O O
3. Illumination uniformity 7.3.3 M O O
4. Luminance mean 7.3.4.2 O O O
5. Luminance variance 7.3.4.3 M O O
6. Under-exposure prevention 7.3.5 O O O
7. Over-exposure prevention 7.3.6 O O O
8. Dynamic range 7.3.7 M O O
9. Sharpness 7.3.8 M O O
10. No compression artefacts 7.3.9 O O O
11. Natural colour 7.3.10 O O O
12. Single face present 7.4.2 M M O
13. Eyes open 7.4.3 M O O
14. Mouth closed 7.4.4 M M O
15. Eyes visible 7.4.5 M M O
16. Mouth occlusion prevention 7.4.6 M M O
17. Face occlusion prevention 7.4.7 M M O
18. Inter-eye distance 7.4.8 M M M
19. Head size 7.4.9 M M M
20. Leftward crop of face in image 7.4.10.1 M M M
21. Rightward crop of face in image 7.4.10.2 M M M
22. Margin above face in image 7.4.10.3 M M M
23. Margin below face in image 7.4.10.4 M M M
Key
M  mandatory for the QAA to implement this quality measure
O  optional
© ISO/IEC 2025 – All rights reserved
TTabablele 3 3 ((ccoonnttiinnueuedd))
# Face image quality measure Sub- UC1 UC2 UC3
clause
(ID (System (Probe for instantaneous
Documents) enrolment) recognition)
24. Pose angle yaw frontal alignment 7.4.11.2 M M O
25. Pose angle pitch frontal alignment 7.4.11.3 M M O
26. Pose angle roll frontal alignment 7.4.11.4 M M O
27. Expression neutrality 7.4.12 M O O
28. No head covering 7.4.13 M O O
29. Radial distortion D.2.1 O O O
30. Pixel aspect ratio D.2.2 O O O
31. Gaze D.3.1 O O O
32. Shoulder presentation D.3.2 O O O
33 Camera subject distance D.3.3 O O O
34 Motion blur prevention D.3.4 O O O
Key
M  mandatory for the QAA to implement this quality measure
O  optional
6 Common computations
6.1 Overview
Quality measures defined in this document can be applied to either 8 bit encoded greyscale images, or 24 bit
encoded colour images. Clause 6 supports the computations in Clause 7 by defining certain steps that appear
in more than one of its subclauses.
Several subclauses of Clauses 6 and 7 reference the usage of the pre-trained neural networks listed in
Table 4. For each, their implementation description, model, weights, and alternative implementations, are
provided in Reference [58].
Table 4 — List of neural networks used in computations of Clauses 6 and 7
# Clause Model Implementation Dataset
1 6.4 Face detec- See Reference [30] NA
tion
2 6.5 Face See Reference [53] Wider Facial Landmarks in the Wild (WFLW)
landmark
See Reference [31]
estimation
[69]
3 6.8 Face parsing See Reference [68] CelebMask-HQ
[69]
4 6.9 Face extrac- See Reference [32] CelebMask-HQ
tion
[70]
5 6.11 Head pose See Reference [50] 300W-LP
estimation
[54]
6 7.2 Unified qual- See Reference [54], [55] MS1MV2
ity scoring
[60]
7 7.3.9.2 No Com- See Reference [58] OFIQ Development Dataset
pression
artefacts
[9], [51], [52], [60]
8 7.4.12.2 Expression See References [9], [51], [52] OFIQ Development Dataset
neutrality
© ISO/IEC 2025 – All rights reserved
6.2 Conversion of 16 bits per channel images to 8 bits per channel images
When 16-bit scRGB images are given, conversion of 16-bit scRGB images shall be performed as defined in
IEC 61966 2-1, using conversions as defined in IEC 61966-2-2:2003, Clause A.2.
NOTE IEC 61966-2-2:2003, 3.1 and 4.1 show the conversion chain for 16-bit scRGB to 1931 CIEXYZ to 8-bit sRGB.
Images with other ICC profiles shall be converted according to their profile.
6.3 Conversion of high bit-depth images to 8 bit greyscale or 24 bit colour images
ISO/IEC 39794-5:2019 allows high-bit depth face captures with native camera formats that are encoded
as defined in ISO/IEC 39794-5:2019, D.1.4.2.4. ISO/IEC 39794-5:2019 allows also proprietary formats with
non-normative encoding specifications. Conversion, such as scale adaptation, should follow manufacturer
recommendation.
Conversion of high bit-depth images to 8 bits greyscale or 24 bits colour images shall be completed before
quality measures are computed. In the conversion, it is recommended to follow the methods for conversion
of high dynamic range content to standard dynamic range described in Reference [47].
6.4 Face detection
This subclause supports the computations of Clause 6 and Clause 7 that depend on detection of faces in a
face image. Face detection can be performed using a neural network pre-trained to produce face bounding
boxes using face image datasets annotated with face bounding boxes. The face detection algorithm should
accept any image and return a set of face bounding boxes for each face detected in the image.
The implementation should include the algorithm below. It may include an alternative detection algorithm if
the resulting Clause 7 quality outputs are reproduced according to Annex A. The algorithm takes as input a
face image I in the BGR colour channel order (8 bits per channel). It uses the model from Reference [30]. The
model is available from Reference [58].
1) Determine the height, h, and the width, w, of image I.
2) Pad image I from the left and right side by 0.2 w and from the top and bottom by 0.2 h, and update w and
h to the new width and height of I.
3) Resize image I to size (height, width) 300 px × 300 px using OpenCV with bilinear interpolation.
4) Normalize the values of I using mean µ = (104.0, 117.0, 123.0) and standard deviation σ = (1, 1, 1) using
6.19 to obtain the array A.
[0]
5) Encode A as input tensor T of dimensions (1, 3, 300, 300) for the neural network model.
[0] [1]
6) Run a forward pass through the model using T as input to obtain the output tensor T of dimensions
(1, 1, N, 7), where N is the number of detected faces prior to filtering.
[1]
7) Extract the last 2 dimensions of T as face bounding box table F of size (N, 7), where each row of F
contains the data of the detected face bounding boxes:
a) column 3 specifies the confidence;
b) columns 4 to 5 specify the relative (i.e. scaled to the interval [0,1]) x- and y-coordinates, respectively,
of the upper left corner;
c) columns 6 to 7 specify the relative (i.e. scaled to the interval [0,1]) x- and y-coordinates, respectively,
of the lower right corner.
8) Delete all rows in F, where one of the following conditions is fulfilled:
a) the confidence is lower than 0.4, i.e. F ≤ 0.4;
i,3
© ISO/IEC 2025 – All rights reserved
b) the width of the face bounding box is smaller than 1/20 of the image width, i.e. F - F < 0.05;
i,6 i,4
c) the face bounding box protrudes outside the image boundary, i.e. F ≤ 0 or F ≥ 1 or F ≤ 0 or F ≥ 1.
i,4 i,6 i,5 i,7
9) If no rows remain, the abstract value failureToAssess is returned.
10) For each row i , obtain the coordinates (a ,b ) and (c ,d ) of the upper left corner and the lower right
i i i i
corner of the face bounding box within the input image, respectively, as:
a = wF· −p +05.  ,   b = hF· −p +05.  ,
i  i,41  j  i,52 
c = wF· −p +05. ,   d = hF· −p +05. .
   
i i,61 j i,72
   
11) Determine the row j with the largest face bounding box area, i.e.:
j = argmax((c - a )∙(d - b )).
j j j j
If there are several face bounding boxes with largest area, choose the first one of these.
12. Return the face bounding box coordinates (a , b , c , d ).
j j j j
As an example, Figure 1 shows the typical output for a face image with significant roll angle. The image is
selected specifically to not be a canonical face image.
NOTE 1 The coordinates ab,,cd, computed in step 11 can lie outside the image region, i.e.
()
ii ii
ab≥≥00,,cw≤−11,dh≤− does not necessarily hold. The same holds for the coordinates of the largest face
ii ii
bounding box output by the algorithm.
NOTE 2 The face detection definitions of this subclause accommodate images where one or two eyes are visible, or
can be closed or occluded, and images where multiple faces are present.
Figure 1 — Example face bounding box
6.5 Face landmark estimation
This subclause supports the computations of Clause 6 and Clause 7 that either require specific landmarks or
the landmarked region (see Figure 2 for an illustration of location and naming of landmarks in an example
face image). Face landmark estimation can be performed using a neural network pre-trained to compute the
position of face landmarks using face image datasets annotated with face landmarks to be estimated.

© ISO/IEC 2025 – All rights reserved
Figure 2 — Location and indices of landmarks in an example face image
The landmark estimation algorithm should accept an image along with a face bounding box returned by face
detection. The estimation algorithm should return the 98 face landmarks (see the example in Figure 2).
An implementation should include the algorithm below. It may include an alternative landmark estimation
algorithm if the resulting Clause 7 quality outputs are reproduced according to Annex A.
[53]
The algorithm uses the ADNET CNN which has been trained on the Wider Facial Landmarks in the Wild
[31]
(WFLW) dataset. The model is ava
...