AI-Powered Management of Identity Photos for Institutional Staff

Directories

Daniel Canedo

, Jos

e Vieira

, Ant

onio Gonc¸alves and Ant

onio J. R. Neves

IEETA, DETI, LASI, University of Aveiro, 3810-193 Aveiro, Portugal

Keywords:

Computer Vision, Face Veriﬁcation, Deep Learning, Identity Photos, Photo Management.

Abstract:

The recent developments in Deep Learning and Computer Vision algorithms allow the automation of several

tasks which up until that point required the allocation of considerable human resources. One task that is getting

behind the recent developments is the management of identity photos for institutional staff directories because

it deals with sensitive information, namely the association of a photo to a person. The main objective of this

work is to give a contribution to the automation of this process. This paper proposes several image processing

algorithms to validate the submission of a new personal photo to the system, such as face detection, face

recognition, face cropping, image quality assessment, head pose estimation, gaze estimation, blink detection,

and sunglasses detection. These algorithms allow the veriﬁcation of the submitted photo according to some

predeﬁned criteria. Generally, these criteria revolve around verifying if the face on the photo is of the person

that is updating their photo, forcing the face to be centered on the image, verifying if the photo has visually

good quality, among others. A use-case is presented based on the integration of the developed algorithms as

a web-service to be used by the image directory system of the University of Aveiro. The proposed service is

called every time a collaborator tries to update their personal photo and the result of the analysis determines

if the photo is valid and the personal proﬁle is updated. The system is already in production and the results

that are being obtained are very satisfactory, according to the feedback of the users. Regarding the individual

algorithms, the experimental results obtained range from 92% to 100% of accuracy, depending on the image

processing algorithm being tested.

1 INTRODUCTION

In this paper, we propose image processing algo-

rithms to be integrated into a photo management sys-

tem, presenting a use-case implemented at University

of Aveiro. With the help of the proposed solutions,

the collaborators can update their identity photo au-

tonomously, with the guaranty of a good image qual-

ity, following the institutional guidelines and validat-

ing the user identity.

Generally, photo management is done manually

with considerable costs associated with human re-

sources, especially in large institutions such as univer-

sities. Automating this process to some extent would

lower those costs. However, one needs to be wary

of automating certain tasks, such as face recognition,

since it deals with sensitive information that if forged

could undermine the purpose of such systems. Since

Deep Learning algorithms are not perfect nor trans-

https://orcid.org/0000-0002-5184-3265

https://orcid.org/0000-0002-4356-4522

https://orcid.org/0000-0001-5433-6667

parent in their decisions (Xu et al., 2019), the photo

management system must deal with failures by pro-

viding feedback to users and Human Resources.

The proposed photo management system takes a

submitted photo as an input, processes it through sev-

eral image processing algorithms to extract relevant

information, matches this data against the validity cri-

teria imposed by the institution in which the system is

operating and provides feedback regarding each sin-

gle criterion as the output. Based on the provided

feedback, the user can correct the submission, or con-

tact the Human Resources reporting the problem. The

whole pipeline of the photo management system, in-

cluding its image processing algorithms, are detailed

in Section 3. A use-case for this system at Univer-

sity of Aveiro is presented in Section 4. The proposed

photo management system allows the collaborators of

University of Aveiro to update their personal photo

autonomously under certain guidelines by uploading

a photo on its website. The image processing block

of this system will verify and transform the submitted

photo to follow those guidelines.

804

Canedo, D., Vieira, J., Gonçalves, A. and Neves, A.

AI-Powered Management of Identity Photos for Institutional Staff Directories.

DOI: 10.5220/0011649000003417

In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 4: VISAPP, pages

804-811

ISBN: 978-989-758-634-7; ISSN: 2184-4321

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

This document is structured as follows: Section

2 presents the related work; Section 3 presents the

methodology; Section 4 presents the use-case; Sec-

tion 5 presents the results; Section 6 presents the con-

clusion.

2 RELATED WORK

Most of the institutions nowadays have a digital staff

directory, public or private. There are clear advan-

tages to keep this information updated. However, de-

pending on the dimension of the institution, this task

can be complex. Creating procedures to help each

person to update their personal information are of ex-

treme importance. Taking into consideration the fo-

cus of this work, there are cases where the user or

collaborator have not updated their picture for a long

time, which leads to outdated personal biometric in-

formation. This can be a problem for a photo man-

agement system during the face veriﬁcation.

In the use-case of University of Aveiro that is

presented later in this paper, some identity photos

of students and collaborators come from their iden-

tity card and were taken in their adolescence. From

adolescence to adulthood, there are signiﬁcant facial

changes. The aging problem in face recognition is not

new (Singh and Prasad, 2018), however verifying the

identity of someone using the photo of the identity

card is a particular problem that the proposed photo

management system may face.

(Albiero et al., 2020) proposed AIM-CHIYA (Ar-

cFace Identity Matching on CHIYA) to tackle this

problem. Firstly, they collected a dataset called

Chilean Young Adult (CHIYA), that contains a pair

of images for each person. Each pair of images is one

image of the person’s national identity card issued at

an earlier age, and one current image acquired with

a contemporary mobile device. They also detected

and aligned the faces using RetinaFace (Deng et al.,

2020). For the training strategy, they chose a few-

shot learning with triple loss approach. Since they are

only interested in matching between photos and iden-

tity card images, all the triplets are selected as photos

to identity cards. That is, if the anchor is a photo face,

the positive can only be an identity card face, con-

straining the negative to also be an identity card face.

Then they performed transfer learning using a model

that was trained in a larger in the wild dataset. This

approach can be relevant for our work since depend-

ing on where the photo management system operates,

it may have to deal with identity card faces as the ref-

erence faces for the face veriﬁcation. This would have

to be done at least one time per user, since after suc-

cessfully updating their photo, the old identity card

one is no longer the reference for future updates.

Another work found in the literature that re-

inforces this approach is DocFace+ (Shi and Jain,

2019). The authors of this work proposed a pair of

sibling networks for learning domain-speciﬁc features

of identity card faces and photo faces with shared

high-level parameters. Afterwards, they trained their

model with a larger in the wild dataset and then ﬁne-

tuned it using an identity card dataset. To overcome

underﬁtting, the authors proposed an optimization

method called dynamic weight imprinting to update

the classiﬁer weights.

A similar photo management system found in the

literature is MediAssist (Cooray et al., 2006). Me-

diAssist is a web-based personal photo management

system that groups all the photos into meaningful

events based on time and location information which

are automatically extracted when uploading the pho-

tos to the system. The authors found that such events

are useful both in the search and indexing operations.

This is a semi-automatic system where the users can

annotate who appears in the photos. The system re-

ceives the uploaded photo and performs face detec-

tion and identiﬁcation based on body-patch similar-

ity matching. Then, during the annotation, this photo

management system can generate an automatic name

suggestion for someone appearing in a given photo.

Another photo management system found in the

literature is Face Album (Xu et al., 2017). This sys-

tem is basically a mobile application that organizes

photos by person identity. This system is also semi-

automatic since users can correct misidentiﬁed faces.

The authors use a light convolutional neural network

(CNN) for face recognition and proposed an algo-

rithm consisting of two pools: the certain pool which

consists of clusters of identiﬁed faces, and the uncer-

tain pool which consists of clusters of faces that are

yet to be identiﬁed. If some faces form a convincing

cluster within the uncertain pool, this cluster is moved

to the certain pool as a new identity. Besides this au-

tomatic management, the user can interact with the

system to correct misidentiﬁed faces and to identify

faces in the uncertain pool.

While the purpose of the photo management sys-

tems found in the literature is to index and organize

photos, mainly around identity, our system’s purpose

is to allow for an automatic picture update. This goal

not only requires face recognition, but also a set of im-

age processing algorithms to validate and transform

the submitted pictures into an adequate format.

AI-Powered Management of Identity Photos for Institutional Staff Directories

805

3 METHODOLOGY

This section addresses the pipeline of the proposed

photo management system. Basically, this system

is a web-service that implements several image pro-

cessing algorithms. They should be computationally

optimized to keep the processing time under control

for the system to be responsive. This is intended for

time constrained applications that may have numer-

ous users using the system at the same time.

The main goal of this system is to receive a pic-

ture, analyze it, make the necessary adjustments to

ﬁt within certain guidelines, and return the adjusted

picture with the analysis feedback. Therefore, the

following image processing algorithms were imple-

mented: color veriﬁcation, face detection, face recog-

nition, face alignment, cropping, head pose estima-

tion, sunglasses detection, blink detection, gaze esti-

mation, brightness estimation, and image quality as-

sessment. The following subsections brieﬂy explain

each algorithm.

3.1 Color Veriﬁcation

Institutions generally have colored images for the

identity photos of their collaborators (a full RGB im-

age). Therefore, the photo management system must

be able to identify what is the color space of the sub-

mitted photo. A straightforward way to do this is

to extract this information from the image metadata.

However, this metadata can be edited before the sub-

mission, that is why the proposed system should have

a way to complement the metadata information. The

proposed algorithm splits and compares the channels

of the image. If the channels are equal, it means the

image is grayscale, if the channels are different, it

means the image is colored. Thus, if the image meta-

data indicates that the image is RGB and the channels

of the image are different, this algorithm validates the

image as having the adequate color space.

3.2 Face Detection

The face detector on the proposed system uses Dlib

library (King, 2009). This face detector uses the His-

togram of Oriented Gradients (HOG) feature com-

bined with a linear classiﬁer, an image pyramid, and

sliding window detection scheme. The main reasons

behind this choice are the lower processing time when

compared to heavier Deep Learning options and the

fact that in addition to the bounding box, it also re-

turns a set of 68 facial landmarks. Some image pro-

cessing algorithms implemented in this photo man-

agement system take advantage of these landmarks.

To avoid false positives or multiple faces being de-

tected in a single picture, only the largest face is re-

turned, which generally is the face of the user.

3.3 Face Recognition

This algorithm is one of the most important in the pro-

posed system since it deals with sensitive information,

namely personal identity. Therefore, the Dlib ResNet

model was deployed as the basis of the face recogni-

tion algorithm, which is reported to attain 99.38% ac-

curacy on the standard LFW face recognition bench-

mark (Huang et al., 2008). This model maps a face to

a 128-dimensional vector space, therefore two vector

spaces will be needed for this algorithm: the candi-

date face and the reference face. These two vector

spaces are compared with each other by calculating

their Euclidean distance. The closer the distance, the

more likely it is that they are the same person.

3.4 Face Alignment and Cropping

Generally, institutions want a uniform and adequate

directory of pictures of their collaborators. To reach

that, face alignment is performed in every submitted

picture, using the facial landmarks returned by Dlib.

This algorithm considers two facial landmarks on the

eyes and if they form an angle of zero in the horizon-

tal axis, the face is aligned. If the angle is not zero,

a rotation transformation is applied to align those two

facial landmarks with the horizontal axis, until the an-

gle formed by them is zero. After the face is aligned,

the cropping is done.

We also consider as requisite that photos should

have a similar cropping as the identity card photos.

Thus, to achieve that, the cropping algorithm of the

photo management system takes the left, right, top,

and bottom extremities of the facial landmarks. With

these four points, the center point is found. From this

center point, the algorithm expands a bounding box

around it with a certain length, emulating the style of

identity card photos. Finally, the resulting cropped

image is resized to a ﬁxed resolution to keep unifor-

mity between all collaborators’ pictures.

3.5 Head Pose Estimation

Another issue taken into consideration is the rota-

tion of the face, considering that the photo should

be frontal towards the camera. Therefore, the photo

management system is equipped with an algorithm to

estimate the head pose of the face present in the sub-

mitted photo. This algorithm makes use of the facial

landmarks outputted by the Dlib face detector. By

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

806

calculating the rotation and translation of key facial

landmarks, it is possible to transform those points in

world coordinates to 3D points in camera coordinates

and project them onto the image plane. This results

in the Euler angles which represent the orientation of

the facial landmarks. For a face to be frontal, the pitch

and yaw need to be close to zero degrees, the roll is

disregarded because the face is previously aligned us-

ing this angle.

3.6 Sunglasses Detection

For an identity photo in a professional proﬁle, gen-

erally sunglasses are not adequate. For this reason,

the photo management system implements two algo-

rithms to detect sunglasses: a low-level algorithm and

a Deep Learning algorithm. The low-level algorithm

uses the Dlib facial landmarks to get regions of in-

terest right below the eyes and a region on the nose.

The idea is to make a color comparison between the

regions below the eyes and the nose. If there are sun-

glasses, the regions of interest below the eyes repre-

sent a part of the lenses instead of skin, resulting in

a color disparity between those regions and the re-

gion of the nose. This algorithm triggers a difference

for sunglasses, but not for glasses since the lenses are

transparent. This is obviously desirable since there

are people who need to use glasses in their daily life.

Figure 1 illustrates this algorithm.

Figure 1: Low-level algorithm for sunglasses detection. A

color comparison between the blue regions and the red re-

gion is made.

The photo management system implements an-

other algorithm for this task based on Deep Learning.

A hybrid dataset was created by combining a Kag-

gle dataset for glasses detection

and a Kaggle dataset

https://www.kaggle.com/code/jorgebuenoperez/computer-

vision-application-of-cnn

for sunglasses detection

. This was done to improve

the quality and robustness of the dataset and to tackle

class imbalance, since sunglasses photos were quite

outnumbered. Then, a model was trained using a

transfer learning approach on a VGG16 network that

was trained on the ImageNet dataset (Simonyan and

Zisserman, 2014). This model is the heavier and more

accurate solution of the photo management system for

the sunglasses detection problem.

3.7 Blink Detection and Gaze

Estimation

Since having collaborators with their eyes closed on

their identity photos is generally undesirable, the

photo management system implements an algorithm

that estimates the eye aspect ratio (EAR) (Soukupova

and Cech, 2016). This algorithm uses the facial land-

marks of Dlib, namely the eyes landmarks. Figure 2

illustrates which landmarks are used by Equation 1 to

estimate the EAR.

Figure 2: Eyes landmarks used by the blink detection algo-

rithm.

EAR =

||P2 − P6|| + ||P3 − P5||

2||P1 − P4||

(1)

The numerator of Equation 1 basically calculates

the vertical distance of the eye, while the denominator

calculates the horizontal distance. The EAR is mostly

constant when the eyes are opened, but it is approxi-

mately zero when the eyes are closed. Therefore, with

this algorithm, it is possible to verify if the eyes are

opened as desired.

Not only it is desired that there are no pictures

with eyes closed, but it is also desired that the gaze

is directed to the camera. For this reason, the photo

management system has a gaze estimation algorithm

following the ideas presented in (Canedo et al., 2018).

This algorithm also uses the Dlib facial landmarks,

and Figure 3 shows which ones are used.

As it is possible to observe in Figure 3, an average

between points P1 and P2 is made to obtain the top

left corner and an average between points P4 and P5

is made to obtain the right bottom corner. With these

https://www.kaggle.com/datasets/amol07/sunglasses-

no-sunglasses

AI-Powered Management of Identity Photos for Institutional Staff Directories

807

Figure 3: Eyes landmarks used by gaze estimation algo-

rithm in green and the resulting region of interest.

corners, a region of interest is extracted. These av-

erages are made such that the region of interest only

catches the eye and not unnecessary features like eye-

lashes or skin which could deteriorate the algorithm

performance. After obtaining the region of interest,

a series of low-level image processing operations are

conducted to detect the pupils. Firstly, this region

is converted to grayscale and a bilateral ﬁlter is per-

formed to smooth the image. Then a morphological

operation is conducted to erode the image to remove

unnecessary features and noise. Finally, it is applied

an inverted binary threshold followed by Otsu thresh-

olding (Xu et al., 2011). With this, the pupil is prop-

erly segmented. Then, the contours of the pupil are

found, and the centroids of the eyes are calculated us-

ing the moments of the contours. By averaging the

distance of both centroids to the eye’s extremities, it

is possible to estimate where the user is looking at.

3.8 Brightness Estimation and Image

Quality Assessment

It is desirable that the submitted pictures are not

overexposed nor underexposed. Therefore, the photo

management system has an algorithm to estimate the

brightness of the picture. This is simply done by con-

verting the image to the HSV (Hue, Saturation, Value)

color space, splitting the channels, and averaging the

Value channel, since this channel corresponds to the

brightness. The resulting average is then compared to

a certain threshold to indicate if the image has ade-

quate brightness.

It is also desirable that the submitted pictures

have good quality. Therefore, the photo manage-

ment system is equipped a model provided by the

OpenCV library (OpenCV, 2022) called BRISQUE

(Blind/Referenceless Image Spatial Quality Evalua-

tor) (Mittal et al., 2012). This model uses scene statis-

tics of locally normalized luminance coefﬁcients to

quantify possible losses in the genuineness of the im-

age due to distortions, which leads to the measure of

quality. The features used derive from the distribu-

tion of locally normalized luminance and products of

locally normalized luminance under a spatial natural

scene statistic model. This algorithm not only has

a very low computational complexity, but it also is

highly competitive within the image quality assess-

ment ﬁeld, which makes it an ideal choice for an al-

ready feature-heavy photo management system.

4 IMPLEMENTATION OF A

REAL USE-CASE

In order to validate the proposed photo manage-

ment system, a prototype called FotoFaces was im-

plemented at University of Aveiro, which serves as a

reliable use-case. Figure 4 illustrates the process fol-

lowed by this prototype.

As it can be observed in Figure 4, the user can

submit a photo in the University of Aveiro’s website

and this photo goes through the FotoFaces Analysis

block, which basically runs all the image process-

ing algorithms described in Section 3 except for the

face veriﬁcation. If the picture meets all the require-

ments, then it goes through the face veriﬁcation al-

gorithm, and if it is successfully veriﬁed, the photo

is updated. Otherwise, the human resources depart-

ment SGRH will re-analyze the photo manually. This

is still a closed prototype that is going through testing

with the staff before opening it to the students. Figure

5 shows the website of the photo management system

implemented in University of Aveiro.

To test this prototype, a popular and challenging

dataset was chosen: LFW (Huang et al., 2008). This

dataset contains challenging images that were taken in

the wild, meaning they were not taken in a controlled

environment. A new dataset

with 50 pairs of images

was created using images from LFW. Each pair corre-

sponds to a reference picture and a candidate picture.

The reference picture represents a picture that is al-

ready in the system and that is used to verify the iden-

tity of a submitted photo. The candidate picture repre-

sents a submitted photo. LFW is already challenging

by nature, but the pictures chosen for this dataset con-

sider speciﬁc challenges that the image processing al-

gorithms need to tackle. Therefore, it was chosen im-

ages with glasses, sunglasses, different ethnicities and

genders, different lighting conditions, different head

poses and gaze directions, different age and looks be-

tween the reference and candidate pictures, and so on.

Finally, the candidate pictures were manually anno-

tated with 1 or 0 for eight different categories sequen-

tially (1 = True, 0 = False): frontal face, colored im-

age, same person, no sunglasses, opened eyes, frontal

gaze, adequate brightness, adequate image quality.

For instance, if a picture meets all these require-

Available in tinyurl.com/y4vmyx27

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

808

Figure 4: Photo management system prototype of University of Aveiro.

Figure 5: Webpage available to the staff of University of

Aveiro for personal photo update.

ments the label is 11111111, if a picture does not

meet the third and fourth requirements the label is

11001111, and so on. Figure 6 illustrates a pair of

images considered for the dataset.

Figure 6: A pair from the dataset. Reference picture on the

left, candidate picture on the right.

Looking at the candidate picture in Figure 6, it

meets all the requirements except for the no sun-

glasses category. Since he is wearing sunglasses, the

opened eyes and frontal gaze categories were also an-

notated as 0 (False) because the eyes are obstructed

by the dark lenses. Therefore, the ﬁnal label is

11100011. Table 1 shows what the dataset is con-

sisted of.

AI-Powered Management of Identity Photos for Institutional Staff Directories

809

Table 1: Dataset based on the LFW dataset, with 50 pairs of

images.

Algorithm True False

Frontal 30 20

RGB 50 0

Veriﬁcation 44 6

No Sunglasses 46 4

Opened Eyes 41 9

Gaze 16 34

Brightness 47 3

Quality 49 1

5 RESULTS

Two experiments were conducted. In the ﬁrst one,

each pair of images of the dataset was consecutively

fed to the prototype shown in Figure 4 and the out-

puts of each image processing algorithm were directly

compared with the labels for the pair in question. The

outputs of each image processing algorithm do not in-

ﬂuence each other. Table 2 shows the results of this

experiment.

Table 2: First experiment: algorithms operate indepen-

dently.

Algorithm Accuracy

Frontal 1.00

RGB 1.00

Veriﬁcation 0.96

No Sunglasses 0.96

Opened Eyes 0.88

Gaze 0.80

Brightness 0.92

Quality 1.00

As it can be observed in Table 2, the accuracy

of each algorithm composing the photo management

system is high. The poorest result is the gaze direc-

tion, with 80% accuracy.

For the second experiment, an improved approach

was taken. It is clear that the image processing al-

gorithms of the photo management system synergize

and are dependent on each other. Most algorithms use

the facial landmarks of Dlib (Figure 1) and are based

on a geometric approach. For this reason, the ideal

condition for most algorithms to function properly is

when the face is frontal to the camera. Not only that,

but it is possible to go further: whenever sunglasses

are present, it is not desirable to run the algorithms

that check if the eyes are opened or if the gaze is

frontal to the camera, since the eyes are obstructed by

the dark lenses. Additionally, if the eyes are closed,

it is not desirable to run the algorithm that checks the

gaze direction since it will not detect the pupils. Fig-

ure 7 shows how these changes impacted the previ-

ously independent pipeline and Table 3 shows the re-

sults.

Figure 7: Pipeline for the second experiment.

Table 3: Second experiment: key algorithms operate depen-

dently.

Algorithm Accuracy

Frontal 1.00

RGB 1.00

Veriﬁcation 1.00

No Sunglasses 0.97

Opened Eyes 0.96

Gaze 0.96

Brightness 0.92

Quality 1.00

As it can be observed in Table 3, the results were

signiﬁcantly improved from Table 2. This was possi-

ble by understanding the dependency and synergy be-

tween some image processing algorithms of the photo

management system. For instance, constraining the

sunglasses’ detection algorithm and face veriﬁcation

algorithm only to frontal faces boosted the accuracy

from 96% to 97% and 100%, respectively. Constrain-

ing the blink detection algorithm that checks if the

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

810

eyes are opened to faces that are not wearing sun-

glasses boosted the accuracy from 88% to 96%. Fi-

nally, constraining the gaze estimation algorithm to

faces that are not wearing sunglasses and have their

eyes opened boosted the accuracy from 80% to 96%.

These experiments validate the potential of this

prototype deployed at University of Aveiro. It is

worth mentioning that the sunglasses detection algo-

rithm used in this prototype was the low-level one

and not the Deep Learning one (Subsection 3.6). This

choice was done to keep the system responsive when

in high demand.

6 CONCLUSION

A photo management system was presented in this

document. This system analyzes submitted photos

with image processing algorithms, enabling users to

update their identity photo automatically. A use-case

was tested, and it is currently in use at University of

Aveiro. In this use-case, a prototype was implemented

to allow the staff to update their identity photo. To

validate this prototype, a dataset was built and anno-

tated based on images from the LFW dataset. Two

experiments were conducted to test the several im-

age processing algorithms of the prototype. It was

concluded that it is possible to take advantage of the

dependency of certain image processing algorithms

to boost their accuracy. The results obtained were

quite satisfactory, considering the challenging nature

of the LFW images. The results ranged from 92% to

100% accuracy, depending on the image processing

algorithm being tested. The most crucial algorithm

of such systems, which is face veriﬁcation, attained

100% accuracy.

As for future work, this prototype can be improved

in several aspects. The face veriﬁcation algorithm

needs to employ strategies to deal with aging, as dis-

cussed in Section 2, since there are institutions that

have not updated the identity photos of their collab-

orators for a long time. This prototype also needs to

be equipped with an emotion recognition algorithm,

since generally identity photos require a neutral ex-

pression. Furthermore, it also needs an algorithm to

detect hats and other accessories that are not usually

welcome in a professional setting.

ACKNOWLEDGEMENTS

This work was supported in part by Institute of

Electronics and Informatics Engineering of Aveiro

(IEETA), University of Aveiro.

REFERENCES

Albiero, V., Srinivas, N., Villalobos, E., Perez-Facuse, J.,

Rosenthal, R., Mery, D., Ricanek, K., and Bowyer,

K. W. (2020). Identity document to selﬁe face match-

ing across adolescence. In 2020 IEEE Int. Joint Conf.

on Biometrics (IJCB), pages 1–9. IEEE.

Canedo, D., Trifan, A., and Neves, A. J. (2018). Monitor-

ing students’ attention in a classroom through com-

puter vision. In Int. Conf. on Practical Applications

of Agents and Multi-Agent Systems, pages 371–378.

Springer.

Cooray, S., O’Connor, N. E., Gurrin, C., Jones, G. J.,

O’Hare, N., and Smeaton, A. F. (2006). Identifying

person re-occurrences for personal photo management

applications.

Deng, J., Guo, J., Ververas, E., Kotsia, I., and Zafeiriou, S.

(2020). Retinaface: Single-shot multi-level face local-

isation in the wild. In Proceedings of the IEEE/CVF

Conf. on computer vision and pattern recognition,

pages 5203–5212.

Huang, G. B., Mattar, M., Berg, T., and Learned-Miller,

E. (2008). Labeled faces in the wild: A database

forstudying face recognition in unconstrained envi-

ronments. In Workshop on faces in’Real-Life’Images:

detection, alignment, and recognition.

King, D. E. (2009). Dlib-ml: A machine learning toolkit.

The Journal of Machine Learning Research, 10:1755–

1758.

Mittal, A., Moorthy, A. K., and Bovik, A. C. (2012).

No-reference image quality assessment in the spatial

domain. IEEE Transactions on image processing,

21(12):4695–4708.

OpenCV (2022). Opencv. https://opencv.org/. Accessed:

2022-08-26.

Shi, Y. and Jain, A. K. (2019). Docface+: Id document

to selﬁe matching. IEEE Transactions on Biometrics,

Behavior, and Identity Science, 1(1):56–67.

Simonyan, K. and Zisserman, A. (2014). Very deep con-

volutional networks for large-scale image recognition.

arXiv preprint arXiv:1409.1556.

Singh, S. and Prasad, S. (2018). Techniques and challenges

of face recognition: A critical review. Procedia com-

puter science, 143:536–543.

Soukupova, T. and Cech, J. (2016). Eye blink detection

using facial landmarks. In 21st computer vision winter

workshop, Rimske Toplice, Slovenia.

Xu, F., Uszkoreit, H., Du, Y., Fan, W., Zhao, D., and Zhu,

J. (2019). Explainable ai: A brief survey on history,

research areas, approaches and challenges. In CCF

Int. Conf. on natural language processing and Chi-

nese computing, pages 563–574. Springer.

Xu, X., Xu, S., Jin, L., and Song, E. (2011). Characteristic

analysis of otsu threshold and its applications. Pattern

recognition letters, 32(7):956–961.

Xu, Y., Peng, F., Yuan, Y., and Wang, Y. (2017). Face

album: towards automatic photo management based

on person identity on mobile phones. In 2017 IEEE

Int. Conf. on acoustics, speech and signal processing

(ICASSP), pages 3031–3035. IEEE.

AI-Powered Management of Identity Photos for Institutional Staff Directories

811