Occlusion Detection for Face Image Quality Assessment

Jacob Carnap

1 a

, Alexander Kurz

2 b

, Olaf Henniger

2 c

and Arjan Kuijper

1, 2 d

Technical University of Darmstadt, Department of Computer Science, Darmstadt, Germany

Fraunhofer Institute for Computer Graphics Research IGD, Darmstadt, Germany

Keywords:

Biometrics, Face Recognition, Face Image Quality Assessment.

Abstract:

The accuracy of 2D-image-based face recognition systems depends on the quality of the compared face im-

ages. One factor that affects the recognition accuracy is the occlusion of face regions, e.g., by opaque sun-

glasses or medical face masks. Being able to assess the quality of captured face images can be useful in

various scenarios, e.g., in a border entry/exit system. This paper discusses a method for detecting face oc-

clusions and for measuring the percentage of occlusion of a face using face segmentation and face landmark

estimation techniques. The method is applicable to arbitrary face images, not only to frontal or nearly frontal

face images. The method was evaluated by applying it to publicly available face image data sets and analyzing

the results obtained. The evaluation shows that the proposed method enables the effective detection of face

occlusions.

1 INTRODUCTION

The accuracy of a face recognition system depends

on the quality of the available images, where quality

means utility or usefulness for automatic face recog-

nition (ISO/IEC 29794-1, 2024). Assessing the qual-

ity of face images can be useful in various applica-

tions. Low-quality images can be discarded to reduce

the false non-match rate. This is particularly impor-

tant in large-scale systems such as entry/exit systems

for monitoring border crossings. Face image quality

assessment can also be used when verifying the com-

pliance of passport photographs.

Various factors inﬂuence whether a person can be

recognized in an image. These include the image res-

olution, underexposure and overexposure, the head

pose, and the degree of occlusion of the face. Occlu-

sions include hair, opaque sunglasses and any objects

in front of the face (such as medical face masks, hat

brims or hands). Beards, moustaches and eyebrows

are not considered as occlusions (ICAO, 2018). In

case of occlusions, the utility of the face image for

recognition decreases. Therefore, assessing the de-

gree of face occlusion can contribute to the overall

assessment of the quality of the face image and gives

https://orcid.org/0009-0007-8973-5749

https://orcid.org/0000-0001-6175-9203

https://orcid.org/0000-0003-2499-8613

https://orcid.org/0000-0002-6413-0061

actionable feedback to the user.

The contributions of this paper are:

1. To clarify the method to assess the percentage of

face occlusion such that it works on any face im-

age, not only on frontal and nearly frontal face

images (Section 3),

2. To evaluate the results of the proposed method on

customized and publicly available data sets (Sec-

tions 4 and 5),

3. To outline how to choose the maximum permis-

sible face occlusion for a biometric system (Sec-

tion 5.5).

2 RELATED WORK

A standard for how to calculate face image qual-

ity measures is under development (ISO/IEC FDIS

29794-5, 2024). Several state-of-the-art face im-

age quality assessment methods are available (Merkle

et al., 2022). To better understand face image quality

assessment, algorithms can currently be submitted to

NIST for comparative evaluation (Yang et al., 2024).

Just as the NFIQ (NIST Fingerprint Image Qual-

ity) software (Tabassi et al., 2021) has been es-

tablished as reference implementation for the stan-

dard for assessing the quality of ﬁngerprint images

(ISO/IEC 29794-4, 2024), the OFIQ (Open Source

Carnap, J., Kurz, A., Henniger, O. and Kuijper, A.

Occlusion Detection for Face Image Quality Assessment.

DOI: 10.5220/0013168500003905

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 657-664

ISBN: 978-989-758-730-6; ISSN: 2184-4313

657

Face Image Quality) software (Merkle et al., 2024)

has been developed to serve as reference implementa-

tion for the emerging standard for assessing the qual-

ity of face images (ISO/IEC FDIS 29794-5, 2024). In

addition to an overall quality score, OFIQ returns a

vector of quality measures. The percentage of face

occlusion is included in the OFIQ output vector. We

use it as benchmark in our experiments.

3 PERCENTAGE OF FACE

OCCLUSION

3.1 Overview

For assessing the degree of face occlusion, this pa-

per adopts and advances the approach of the emerg-

ing standard (ISO/IEC FDIS 29794-5, 2024) and of

NIST’s face analysis technology evaluation (Yang

et al., 2024), calculating the percentage of occlusion

in the area between the eyebrows and the chin. The

forehead is excluded because it contains little identi-

fying information.

To measure the percentage of face occlusion, two

regions need to be determined: The face region of in-

terest and the part of it that is occluded. The face

region of interest is referred to as landmarked region

because it is determined using facial landmarks (see

Section 3.2). The part of the landmarked region that is

occluded is determined using a binary pixel-wise seg-

mentation map that indicates for each pixel whether it

belongs to the face or not (see Section 3.3).

3.2 Landmarked Region

Several state-of-the-art facial landmark estimators are

available (Merkle et al., 2022). We use the landmark

estimator skps (Lz, 2023) to estimate the positions of

facial landmarks. It takes a face bounding box as in-

put and estimates the positions of 98 landmarks that

outline the face, eyebrows, eyes, nose, and mouth. To

detect faces and to draw face bounding boxes, many

face detectors are available (Merkle et al., 2022). We

use a modiﬁed version of the YOLO network (Red-

mon et al., 2016), trained to recognize faces (hpc203,

2024).

The facial landmarks form the basis for determin-

ing the landmarked region. The emerging standard

(ISO/IEC FDIS 29794-5, 2024) suggests computing

the convex hull of the landmarks. In mathematical

terms, the boundary of the convex hull of a set of

landmarks is the simple closed curve with minimum

perimeter containing all landmarks. This works well

for frontal and nearly frontal face images. For non-

frontal face images, however, the convex hull of the

landmarks can include portions of the background,

which is not intended. To avoid this, we draw a con-

cave polygon connecting the landmarks on the face

contour and on the upper boundaries of the eyebrows

and consider all pixels inside and on this concave

polygon as landmarked region.

Figure 1 shows examples of landmarked regions

determined using the proposed method, including

both very good and faulty results. The landmarked

region in Fig. 1c on a heavily occluded face is faulty.

(a) (b)

Figure 1: Landmarked regions determined using skps,

marked in blue on the original face image.

3.3 Face Segmentation Map

Several state-of-the-art face segmentation methods

are available (Merkle et al., 2022). We adopted

the face segmentation approach face3d0725 (Yin and

Chen, 2022), which is recommended in the emerg-

ing standard (ISO/IEC FDIS 29794-5, 2024) and also

used in OFIQ (Merkle et al., 2024). The method

utilizes an alignment technique based on facial land-

marks, resizing the image to a resolution of 256 ×256

pixels. The model generates a pixel-wise binary seg-

mentation map of the face image. As desired, hair

hanging over the face is counted as occlusion, while

facial features such as beards, mustaches, and eye-

brows are not. Side whiskers, however, are counted

as occlusions. Also, glasses are counted as occlusion

even if the frame is neither extremely thick nor oc-

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

658

cluding the eyes. This deviates from the requirements

(ICAO, 2018; ISO/IEC 19794-5, 2011). The baseline

model is structured on a ResNet-18 backbone, follow-

ing the U-Net architecture, and trained using a binary

cross-entropy loss combined with online hard exam-

ple mining.

Figure 2 shows examples of segmentation maps.

While the segmentation maps are very accurate in

general, image Fig. 2c highlights an imperfection,

where the segmentation model struggles to accurately

capture the left side of the face.

(a) (b)

Figure 2: Face segmentation maps marked in blue on the

original face image.

3.4 Occluded Region

Let F be the landmarked region, S the segmentation

map containing the face pixels that are not occluded,

and O = F − F ∩ S the area within the face that is

occluded. Then the percentage of occlusion α is the

ratio of |O| to |F| (ISO/IEC FDIS 29794-5, 2024):

α =

|O|

|F|

|F − F ∩ S|

|F|

(1)

The output is a value between 0 and 1. |Y| represents

the number of pixels in an image region Y , i.e., the

area of that region.

Figure 3 shows the relevant face regions. The per-

centage of face occlusion in the example is calculated

to be 28.9 %.

(a) Face image. (b) Landmarked region F.

Figure 3: Examples of the regions used in Eq. (1), marked

in blue on the original face image.

3.5 Sources of Errors

There are various reasons that can cause a deviation of

the determined value from the expected value. Such

deviations in the context of face occlusion can occur

due to a faulty landmarked region (see Section 3.2)

or due to a faulty face segmentation map (see Sec-

tion 3.3).

An issue occurs at the boundaries of the regions.

Even if there is no occlusion and both the landmarked

region and the segmentation map contain no signiﬁ-

cant errors, a small percentage of occlusion can still

be measured. This happens when the face contour of

the segmentation map is placed more centrally than

the landmarked region. An example is the left side

of the face image in Fig. 4. Although no occlu-

sion is evident on the left side, a minor occlusion is

mistakenly detected due to the deviation between the

landmarked region and the segmentation map. Ap-

proaches to solve this problem include rounding to

fewer decimal places, introducing a threshold above

which occlusion is counted, or automatic downsizing

of the landmarked region.

Another issue that deserves attention is the po-

tential for varying interpretations of the underlying

deﬁnitions. For instance, opaque lenses of glasses

are supposed to count as occlusion while transparent

lenses do not. However, the criteria for distinguishing

between opaque and transparent lenses are not explic-

itly deﬁned. This makes it challenging to differentiate

between both and leaves room for interpretations.

Occlusion Detection for Face Image Quality Assessment

659

(a) Landmarked

region F.

(b) Segmentation

map S.

region O.

Figure 4: Example of the issue occurring at the boundaries

of the regions.

4 EXPERIMENTAL SETUP

4.1 Face Image Data Sets

For evaluating the performance of the presented

method, we use the following face image data sets,

which are publicly available for research purposes:

• Caltech Occluded Faces in the Wild (COFW) data

set (Burgos-Artizzu et al., 2022): We chose this

data set because baseline segmentation maps (Yin

and Chen, 2022) and 68 baseline landmarks (Ghi-

asi and Fowlkes, 2015) are given for 507 images

from this data set. From these baselines, we cal-

culated baseline values for the percentage of face

occlusion, using Eq. (1).

• ONOT data set (Di Domenico et al., 2024): This

data set is a collection of synthetic high-quality,

well-controlled images with frontal pose, neu-

tral expression, uniform background as well as

no occlusion as required for enrolment (ISO/IEC

19794-5, 2011). We chose this data set because it

contains face images known to be without occlu-

sion.

• EURECOM’s Kinect face data set (Min et al.,

2014): We chose this data set because it contains

both, ICAO-compliant reference face images and

face images that are occluded but otherwise of

high quality. It includes face images of 52 sub-

jects (14 females, 38 males) obtained using a mul-

timodal Kinect device. Data was captured in two

sessions approximately half a month apart. Each

session included face images of each subject in

nine states: neutral facial expression, smile, open

mouth, left proﬁle, right proﬁle, eyes occluded by

sunglasses, mouth occluded by hand, half of the

face occluded by paper, and lights on.

4.2 Face Comparisons

To evaluate the effect of face occlusions on face com-

parison scores, we compared 2D face images from

EURECOM’s Kinect face data set (Min et al., 2014).

As reference images, we used 52 face images from the

ﬁrst session that show a neutral facial expression. As

probe images, we used 208 mated face images from

the second session: unoccluded face images as well as

face images partially occluded by sunglasses, a hand

or a sheet of paper, all showing a neutral facial ex-

pression. All images were taken under the same, very

good capture conditions, only the occlusion varied.

For feature extraction, we used ArcFace (Deng

et al., 2019), an open-source deep-learning-based face

feature extraction algorithm implemented in Python.

This algorithm aims at obtaining highly discrimina-

tive face embeddings for automatic face recognition

by incorporating margins in the loss function. For

comparing two face feature vectors, we calculated

their cosine similarity as similarity score in the range

from -1 to 1 (the more similar, the higher). Cosine

similarity measures the similarity of the orientation

of two feature vectors regardless of their magnitude.

ArcFace is designed to give best results with cosine

similarity.

5 EXPERIMENTAL RESULTS

5.1 Images with Known Occlusion

The evaluation metrics for face images with known

occlusion are aligned with those used by NIST in

their evaluation of special image defect detection al-

gorithms (Yang et al., 2024). However, the test data

set is different because NIST uses a sequestered data

set.

The COFW baseline values for the percentage of

face occlusion are used to evaluate the validity of the

presented method for calculating the percentage of

occlusion. In Fig. 5, for each of the 507 images a

point is plotted. The x coordinate corresponds to the

baseline value considered to be the ground truth of

the percentage of face occlusion. The y coordinate

corresponds to the value estimated by the presented

method. The line y = x represents the optimal agree-

ment. The closer the estimated value is to the baseline

value, the closer the plotted point is to the line y = x.

While our method achieved a mean absolute error

of 3.91 percentage points, OFIQ achieved a mean ab-

solute error of 3.86 percentage points in the same data

set. For images with known occlusion, both OFIQ and

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

660

Mean absolute error

3.9066 percentage points

Figure 5: Estimated vs. baseline percentage of occlusion.

The blue line represents perfect agreement.

our method slightly underestimate the presence of oc-

clusion.

An earlier version of the proposed method using

another occlusion segmentation model (Nirkin et al.,

2018) has been submitted to NIST’s special image de-

fect detection evaluation, in the category of face oc-

clusion detection when the forehead is excluded. It

achieved the best results so far, along with the OFIQ

software, which achieved the same mean absolute er-

ror value. Switching to the occlusion segmentation

model face3d0725 (Yin and Chen, 2022) improved

the mean absolute error of our method in the COFW

data set by more than 3 percentage points.

5.2 Face Images Without Occlusion

To test how well the proposed method recognizes the

absence of occlusions, we selected face images com-

pliant to the requirements for passport photographs

(ICAO, 2018). This goes beyond NIST’s evaluation

(Yang et al., 2024) and helps setting a face occlu-

sion threshold for unoccluded face images (see Sec-

tion 5.5).

We selected 353 such images from the ONOT data

set. The selected images show the face frontally, well-

lit, and without any occlusion. Therefore, we ex-

pected the percentage of occlusion to be 0. However,

as discussed in Section 3.5, this is not always the case.

Figure 6 shows the histogram of the calculated values

for the 353 face images without occlusion. It can be

seen that in about 50 % of the images, an occlusion of

less than 1 % is calculated.

The mean absolute error of our method is just

above 2.85 percentage points. OFIQ achieved an

mean absolute error of 2.83 percentage points in the

same data set. The main source of error was the pres-

ence of eyeglasses, resulting in up to 14.6 % face oc-

clusion. The spike of error caused by the presence of

glasses is also observable in Fig. 6. Additionally, a

slight deviation from no occlusion is caused by hair

covering the eyebrows or the side of the face. This

is observable in Fig. 4, which shows an image of the

ONOT data set.

5.3 Face Images with Eyeglasses

Figure 7 is a box and whisker plot showing the

distributions of the percentages of face occlusion

in the ONOT data set for 43 face images with

and 310 without transparent eyeglasses. A box is

drawn between the ﬁrst and third quartiles with a

line in between marking the second quartile (median

value); crosses represent mean values. Whiskers are

drawn at the greatest/smallest percentage of occlusion

smaller/greater than 1,5 times the inter-quartile range

(between the ﬁrst and third quartiles) above/below the

third/ﬁrst quartile. Scores beyond the whiskers are

outliers. Figure 7 shows that the measured percentage

of face occlusion is considerably higher when eye-

glasses are worn.

At a 95 % conﬁdence level, a two-sample t-test

shows that the difference between the mean values in

Fig. 7 is statistically signiﬁcant. There is, however, no

statistically signiﬁcant difference between the mean

values of the overall OFIQ quality scores for wearers

of eyeglasses and for people who do not wear glasses.

5.4 Comparison with OFIQ

When using face images with known occlusion and

when using face images without occlusion, a simi-

lar behaviour of OFIQ and the presented method was

observed. We had expected differently. Both meth-

ods utilize the same occlusion segmentation model

Figure 6: Histogram of percentages of occlusion measured

in unoccluded face images.

Occlusion Detection for Face Image Quality Assessment

661

face3d0725 (Yin and Chen, 2022), but OFIQ, as ref-

erence implementation of the draft standard (ISO/IEC

FDIS 29794-5, 2024), uses the OpenCV function con-

vexHull when calculating the landmarked region. The

similar behaviour of OFIQ and the presented method

leads us to suspect that the OpenCV function convex-

Hull does not return the convex hull in mathematical

terms but the contour around the landmarks. OFIQ

behaves as if the landmarked region is based on the

possibly concave polygon along the face contour.

5.5 Face Occlusion Thresholds

How much face occlusion is admissible depends on

the use case. The emerging standard (ISO/IEC FDIS

29794-5, 2024) considers three use cases:

1. Acquisition of reference face images for machine

readable travel documents (MRTDs),

2. Acquisition of reference face images for other

systems (e.g., entry/exit system),

3. Acquisition of probe face images for instanta-

neous recognition.

For use case 1 concerning passport photographs,

ICAO prohibits face occlusions except in speciﬁc

exceptional cases (ICAO, 2018). For use case 2

concerning system enrolment, weaker requirements

(ISO/IEC 19794-5, 2011) apply (European Commis-

sion, 2019), but occlusions are not permitted either.

Both documents (ICAO, 2018; ISO/IEC 19794-5,

2011) allow subjects to wear eyeglasses with trans-

parent lenses not occluding the eyes. That is why we

count face images with transparent eyeglasses in the

Figure 7: Percentage of occlusion for unoccluded face im-

ages with and without transparent eyeglasses.

Figure 8: Mated similarity score vs. percentage of occlusion

scatterplot.

unoccluded category. Given the fact that some occlu-

sion may be measured even in unoccluded face im-

ages (see Fig. 6) and to avoid bias against wearers of

eyeglasses, a discard threshold of about 15 % mea-

sured face occlusion should be chosen.

For use case 3 concerning probe images for in-

stantaneous recognition, the question arises as to how

much occlusion would be permissible without ex-

ceeding a given FNMR target. Figure 8 shows the re-

lationship between the percentage of occlusion of the

probe images from the Kinect face data set and their

mated similarity scores. Percentage of occlusion and

mated similarity score appear to be negatively cor-

related. The orange data points in Fig. 8 also show

that occlusions caused by transparent eyeglasses have

no noticeable negative effect on the mated similarity

scores of face images that are otherwise of high qual-

ity. In the given data set, in general, the sunglasses

cause less occlusion than a hand or a sheet of paper

partially occluding the face.

The maximum permissible percentage of face

occlusion for the probe images depends on the

application-speciﬁc decision threshold. For instance,

let the target FNMR be 5 % or less

and the decision

threshold be a similarity score value of 0.8 in Fig. 8,

then frontal probe images with a percentage of face

occlusion of 23.85 % or more should be discarded and

recaptured to be able to meet the target FNMR. If the

decision threshold is more relaxed, e.g., at a similar-

ity score value of 0.7, then face images would need

to be discarded and recaptured only if the percentage

of occlusion exceeds 40.9 %. The lower the required

similarity, the higher the admissible percentage of oc-

clusion.

Be aware that the error rates of a speciﬁc biomet-

Best practice guideline for automated border control is

an FRR of at most 5 % at an FAR of 0.1 % (Frontex, 2015).

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

662

ric system depend on its input data and that the choice

of thresholds depends on the costs of false accepts and

false rejects. To set operational thresholds, the exper-

iment should be run with commercial face compari-

son algorithms and more, operational face image data.

Real-world probe images differ not only in terms of

face occlusion but may exhibit other irregularities at

the same time.

6 CONCLUSIONS

Drawing on state-of-the-art face landmark estimation

and face segmentation methods, a C++ implementa-

tion to determine the percentage of face occlusion was

developed. When forming the landmarked region, a

concave polygon along the face contour landmarks is

used instead of the convex hull proposed in (ISO/IEC

FDIS 29794-5, 2024). For faces showing other than

frontal or near-frontal poses, a convex hull of the land-

marks incorrectly includes parts of the background.

The proposed method is applicable to any image on

which a face can be detected, not only to frontal or

near-frontal face images. The experiments in Sec-

tion 5 show that the presented method and OFIQ

achieve very similar results. This suggests that OFIQ

does not use the convex hull in mathematical terms

either but the possibly concave polygon bounded by

the face contour.

Both, the presented method and OFIQ, use the

same face segmentation model, which counts the

frame of transparent eyeglasses as occlusion. To

avoid bias against the demographic group of wearers

of glasses, a discard threshold of almost 20 % mea-

sured face occlusion should be chosen. As this allows

unwanted face occlusions to be ignored, it may be bet-

ter to retrain the occlusion segmentation model not to

count transparent eyeglasses as occlusions.

Some issues in the underlying face landmark es-

timation and face segmentation software have been

identiﬁed in Section 3.5. Because of the observable

continuous improvement of such algorithms (Merkle

et al., 2022), it can be expected that these issues will

be alleviated over time.

The proposed approach can be extended to other

face image quality components deﬁned in the emerg-

ing standard (ISO/IEC FDIS 29794-5, 2024): The

measurement of overexposure and underexposure can

be restricted to the unoccluded landmarked region,

and the face segmentation map can be used to check

for the visibility of the eyes and the presence of sun-

glasses and to determine the percentage of mouth oc-

clusion. Satisfactory results were also achieved in

NIST’s evaluation for these quality components.

ACKNOWLEDGEMENTS

This research work has been funded by the German

Federal Ministry of Education and Research and the

Hessian Ministry of Higher Education, Research, Sci-

ence and the Arts within their joint support of the

National Research Center for Applied Cybersecurity

ATHENE.

REFERENCES

Burgos-Artizzu, X., Perona, P., and Dollar, P. (2022). Cal-

tech occluded faces in the wild (COFW). https://doi.

org/10.22002/D1.20099.

Deng, J., Guo, J., Xue, N., and Zafeiriou, S. (2019). Ar-

cFace: Additive angular margin loss for deep face

recognition. In IEEE Conf. on Computer Vision and

Pattern Recognition (CVPR).

Di Domenico, N., Borghi, G., Franco, A., Maltoni, D., et al.

(2024). ONOT: a high-quality ICAO-compliant syn-

thetic mugshot dataset. In 18

IEEE International

Conference on Automatic Face and Gesture Recogni-

tion (FG).

European Commission (2019). Commission Implementing

Decision (EU) 2019/329 of 25 February 2019 laying

down the speciﬁcations for the quality, resolution and

use of ﬁngerprints and facial image for biometric ver-

iﬁcation and identiﬁcation in the Entry/Exit System

(EES).

Frontex (2015). Best practice technical guidelines for auto-

mated border control (ABC) systems. Frontex Tech-

nical Report.

Ghiasi, G. and Fowlkes, C. (2015). Occlusion coherence:

Detecting and localizing occluded faces. https://doi.

org/10.48550/arXiv.1506.08347.

hpc203 (2024). hpc203/yolov8-face-landmarks-

opencv-dnn. https://github.com/hpc203/

yolov8-face-landmarks-opencv-dnn.

ICAO (2018). Portrait quality (reference facial im-

ages for MRTD). ICAO Technical Report.

https://www.icao.int/Security/FAL/TRIP/Documents/

TR-PortraitQualityv1.0.pdf.

ISO/IEC 19794-5 (2011). Information technology – Bio-

metric data interchange formats – Part 5: Face image

data. International Standard ISO/IEC 19794-5.

ISO/IEC 29794-1 (2024). Information technology – Bio-

metric sample quality – Part 1: Framework. Interna-

tional Standard ISO/IEC 29794-1.

ISO/IEC 29794-4 (2024). Information technology – Bio-

metric sample quality – Part 4: Finger image data. In-

ternational Standard ISO/IEC 29794-4.

ISO/IEC FDIS 29794-5 (2024). Information technology

– Biometric sample quality – Part 5: Face image

data. Final Draft International Standard ISO/IEC

FDIS 29794-5.

Lz (2023). 610265158/peppa pig face landmark: new base

line. https://doi.org/10.5281/zenodo.7894492.

Occlusion Detection for Face Image Quality Assessment

663

Merkle, J., Rathgeb, C., Herdeanu, B., Tams, B., Lou, D.-

P., D

orsch, A., Schaubert, M., Dehen, J., Chen, L.,

Yin, X., Huang, D., Stratmann, A., Ginzler, M., Grim-

mer, M., and Busch, C. (2024). Open Source Facial

Image Quality (OFIQ) – Implementation and eval-

uation of algorithms. https://github.com/BSI-OFIQ/

OFIQ-Project/tree/main/doc/reports.

Merkle, J., Rathgeb, C., Tams, B., Lou, D.-P., D

orsch, A.,

and Drozdowski, P. (2022). State of the art of quality

assessment of facial images. https://doi.org/10.48550/

arXiv.2211.08030.

Min, R., Kose, N., and Dugelay, J.-L. (2014). Kinect-

FaceDB: A kinect database for face recognition. IEEE

Trans. on Systems, Man, and Cybernetics: Systems,

44(11):1534–1548.

Nirkin, Y., Masi, I., Tuan, A., Hassner, T., and Medioni,

G. (2018). On face segmentation, face swapping, and

face perception. In 13

IEEE Int. Conf. on Automatic

Face & Gesture Recognition (FG), pages 98–105.

Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.

(2016). You Only Look Once: Uniﬁed, Real-Time

Object Detection. In Proceedings of the IEEE Con-

ference on Computer Vision and Pattern Recognition,

pages 779–788.

Tabassi, E., Olsen, M., Bausinger, O., Busch, C., Figlarz,

A., Fiumara, G., Henniger, O., Merkle, J., Ruhland,

T., Schiel, C., and Schwaiger, M. (2021). NFIQ 2.0 –

NIST ﬁngerprint image quality. NIST IR 8382, NIST.

Yang, J., Grother, P., Ngan, M., Hanaoka, K., and Hom, A.

(2024). Face analysis technology evaluation (FATE) –

Part 11: Face image quality vector assessment: Spe-

ciﬁc image defect detection. NIST IR 8485, NIST.

Yin, X. and Chen, L. (2022). FaceOcc: A diverse, high-

quality face occlusion dataset for human face extrac-

tion. In Treatment and Analysis of the Information

Methods and Applications (TAIMA).

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

664