FaceCounter: Massive Attendance Taking in Educational Institutions

Through Facial Recognition

Adrian Moscol and Willy Ugarte

Universidad Peruana de Ciencias Aplicadas, Lima, Peru

Keywords:

Massive Attendance, Facial Recognition, Face-Recognition Library, Educational Institution, Benchmarking

Facial Algorithms.

Abstract:

Our purpose is to implement a facial recognition system that will improve efﬁciency when taking assistance in

educational institutes, as well as reducing the possible cases of identity theft. To achieve our objective, a facial

recognition system will be created that, upon receiving a photograph of the students present in the classroom,

will identify them and conﬁrm their attendance in the database. The investigation of pre-trained models using

the agile benchmarking technique will be important, the analyzed and compared models will serve as a basis

for the development of the facial recognition system. This program will be connected to an application that

will use a simple interface so that teachers can save class time or evaluation’s time by taking attendance or

conﬁrming the identity of the students present. Also, it will increase security by avoiding possible identity

theft with tools such as false ﬁngerprint molds (admission exams) or partial and/or ﬁnal exams (false ID).

1 INTRODUCTION

Attendance tracking is a common and important task

for educators, but it can be time-consuming and prone

to errors, which can cause delays and inaccuracies in

recording attendance

(Xu et al., 2017).

Particularly in a big classroom, the currently used

conventional method has shown to be unreliable, im-

precise, and time-consuming. On the basis of the tra-

ditional attendance marking approach, it is challeng-

ing to detect absentees and proxy participants (Nordin

and Fauzi, 2020).

In many industries, including business, biometric

technologies with facial recognition systems are in-

creasingly necessary. One such application is the at-

tendance marking system, which is a vital repeating

transaction requirement since it relates to employee

productivity. The removal of the need to make di-

rect contact with the scanning device is one of the nu-

merous advantages of attendance recording utilizing

a person’s face from an ethical standpoint (Wati et al.,

2021).

https://orcid.org/0000-0002-7510-618X

“Why is school attendance so important and what are

the risks of missing a day?” - The Education Hub (Gov UK)

- https://educationhub.blog.gov.uk/2023/05/18/school-att

endance-important-risks-missing-day/

Face detection, one of the preprocessing steps be-

fore face recognition, seeks to identify the presence of

a face image with an eye, a nose, a mouth, and other

facial features.

In Peruvian universities, attendance is mandatory.

During exams, it is especially important to main-

tain security and accuracy in identifying students and

recording attendance, as the teachers in charge of

supervising the classroom may not know personally

most of the students.

In April 2022, the prosecutor’s ofﬁce reported that

the ﬁrst place in the admission exam received help

from impersonators, who charged exorbitant amounts

to ensure admission

To tackle this problem, we propose an automated

process for attendance tracking the will solve possible

delays and human error by providing greater security

during classes and exams by ensuring accurate iden-

tiﬁcation of students and avoiding impersonation.

There are others works that aim to solve this prob-

lem such as (Yang and Han, 2020; Xu et al., 2017;

Nordin and Fauzi, 2020; Wati et al., 2021), but the

“Prosecutor’s Ofﬁce denounces that ﬁrst place in the

admission exam received help from impersonators” (in

spanish) - Peru21 - https://peru21.pe/lima/universidad-s

an-marcos-unmsm-primer-puesto-del-2021-habria-sido-b

eneficiada-por-mafia-encargada-de-suplantar-examenes-d

e-admision-rmmn-noticia

Moscol, A. and Ugarte, W.

FaceCounter : Massive Attendance Taking in Educational Institutions Through Facial Recognition.

DOI: 10.5220/0012181100003584

In Proceedings of the 19th International Conference on Web Information Systems and Technologies (WEBIST 2023), pages 84-94

ISBN: 978-989-758-672-9; ISSN: 2184-3252

main difference is that Yang and Han use real time

video for face recognition and Xu et al. does not

present an app for the simplicity of it use.

Nevertheless, our approach differs from existing

solutions ﬁrst of all by creating a mobile user-friendly

interface app that allows teachers to easily take and

upload pictures. Secondly, we have chosen to output

the attendance data as a simple txt ﬁle, for an eassier

implementation to school systems.

Finally, we have prioritized simplicity in our sys-

tem, without requiring specialized hardware or com-

plex setup procedures. These three aspects make our

system stand out in the market, providing a valuable

and innovative solution for schools and universities

looking to streamline their attendance tracking pro-

cesses.

The app, built with the framework React Native,

allows teachers to take a picture of the classroom,

which is then uploaded and processed by the system.

The system will recognize the faces of students and

add them to the database as either present or absent.

The primary aim of our work is to reduce attendance

time taken and increase security during exams identi-

ﬁcation of students so they don’t do identity theft or

miss out.

Some limitations of our system is that it relies on

the accuracy of face recognition technology, which

can be affected by factors such as lighting and occlu-

sion.

Our main contributions are described as follows:

• We develop an easy friendly interface app that al-

lows users to easily manage attendance tracking.

• We develop a system capable of analysing one or

multiple pictures with one or multiple faces for

their recognition.

• Simple system with no need of complex hardware,

common components and minimal dependencies,

reducing potential compatibility issues and facili-

tating deployment across a range of devices.

The paper details the following: First, related

works will be shown in Section 2. Second, we will

discuss our main contributions, talking about the the-

ory and method used for the solution in Section 3.

Also, we will talk about the experimentation process,

with details on the analysis and selection of the algo-

rithm library of face recognition, the design and de-

velopment and the results of this experiments in Sec-

tion 4. Finally, we will discuss the main conclusions

and possible future work that can be added in Sec-

tion 5.

2 RELATED WORKS

In the related works of attendance tracking systems

with face recognition, several papers have explored

different aspects of the technology. Some studies have

investigated cheating in exams and methods to pre-

vent it, while others have focused on the challenges

of recognizing faces with masks during the COVID-

19 pandemic. Real-time face recognition in video has

also been explored. Additionally, some studies have

aimed to reduce the complexity and expense of atten-

dance tracking systems. These papers provide valu-

able insights into the development and application of

facial recognition technology, as well as addressing

the problem attendance tracking and exam cheating.

In (Cerd

a-Navarro et al., 2022), the authors pro-

vide an article provides a detailed analysis of the

policies and strategies for academic integrity used by

different educational institutions in Spain to prevent

evaluative fraud. This is a good example to set context

about why this is a problem and how we can solve it,

as we can see the results show that only 27.5% of in-

stitutions use an identiﬁcation device for online eval-

uation. Private universities are the ones that use iden-

tity veriﬁcation software the most (75%) compared to

those that do not use it (9.5%).

There are other example of face recognition appli-

cation in other problems that are interesting to take

a look like in (Kocac¸inar et al., 2022), the authors

developed a system to distinguish between masked,

unmasked, and incorrectly masked individuals using

a mobile application called MadFaRe (Masked Face

Recognition application), this because of the COVID-

19 pandemic we had during this last years. This is a

good example to show how our technique can be used

and implemented, in this case to recognise with the

use of a mask. They developed a deep learning and

CNN-based facial recognition algorithm (Kocac¸inar

et al., 2022). The author achieved a validation as-

sertiveness of 80.88% for partial facial recognition.

For the results of facial recognition applying CNN,

an increase in validation was obtained from 78.41%

to 90.40%.

We now can take a look to papers that used the

face recognition as a way to solve the attendance

tracking. Starting in (Yang and Han, 2020), the au-

thors propose a complete assistance system that com-

bines multiple modules to reduce the complexity of

the program and make the code reusable. The system

consists of a video terminal module, a cable transmis-

sion module, data storage, a facial recognition mod-

ule, and a computer terminal module. Tests were con-

ducted in two universities, where 200 students who

must register with ID cards were selected. The facial

FaceCounter: Massive Attendance Taking in Educational Institutions Through Facial Recognition

recognition rate was high (around 82%). The system

aims to reduce absenteeism, and the results showed

that the rate of students skipping classes decreased by

13% compared to the control group.

Then we can analyze the creation of a facial recog-

nition system for attendance taking that aims to re-

duce time and maintenance costs. In (Xu et al., 2017),

the authors have measured the accuracy of their pro-

posal using two tables that focused on scenarios such

as distance, angles, and lighting. The results showed

high accuracy, with recognition rates of 97.1% to

98.8% for face positions in the range of -15

◦

to +15

◦

and an average accuracy of 96.47% under low-light

conditions. This system can recognize faces with an

accuracy of 99% to 98% when the face is 4 to 5 me-

ters away from the camera under normal lighting con-

ditions.

Finally, in (Khan et al., 2020) the authors con-

ducted 6 tests where the number of recognized indi-

viduals increased per test. In all tests, the program

was able to detect the number of people in the photo

and achieved a 0% false recognition rate. It achieved

100% accuracy in each of the tests, with the most no-

table being case 6, which involved 12 people.

We can then say our proposed multi-facial recog-

nition attendance tracking system for teachers offers

several advantages compared to previous approaches.

While (Yang and Han, 2020) and (Xu et al., 2017)

have demonstrated the potential of facial recognition

for attendance tracking, our system offers a more

user-friendly interface for teachers to take a picture

of the classroom, which is then uploaded and pro-

cessed to recognize all students in the image. Our

system also extends previous work by allowing for

recognition of multiple faces in a single image and

automatically marking students as present or absent

in a database. Additionally, our approach offers the

potential for improved accuracy and convenience over

traditional attendance taking methods.

3 MAIN CONTRIBUTION

We aim to present the contributions of a new atten-

dance tracking system. The system is designed to

provide an easy-to-use app for users to manage at-

tendance tracking, with a simple interface that elimi-

nates the need for complex hardware and reduces po-

tential compatibility issues. The attendance data is

stored in both txt and csv ﬁles, which can be eas-

ily imported into other software systems for further

processing or analysis. The system also includes a

face recognition model, created using the python li-

brary face-recognition, which is integrated into an

android app that teachers can use to take a picture

of the classroom. However, we have several restric-

tions, such as the requirement for a private connec-

tion to the database, a minimum camera resolution of

2 megapixels, and a minimum Android version of 10.

3.1 Context

3.1.1 Concepts

Henceforth, we list the main concepts used on our

work:

• Database: A database is an organized collection

of information that is stored and can be electron-

ically accessed from a computer. Data is typ-

ically structured into tables and ﬁelds to facili-

tate searching, sorting, and retrieving speciﬁc in-

formation. The database can be used for a vari-

ety of purposes, such as inventory management,

billing, customer tracking, project management,

and much more. An example of the entities can

be seeing in Fig. 1

Figure 1: Database example.

• Massive attendance: Massive attendance is the

process of taking attendance for a large group of

people, such as in a conference, lecture, or other

event. It often involves the use of technology to

collect and analyze data quickly and efﬁciently.

• Facial recognition: Facial recognition refers to the

technology of identifying or verifying the identity

of a person based on their facial features. It uses

algorithms to analyze and compare the unique

characteristics of a person’s face (Miller, 2023).

• Educational institution: An educational institu-

tion refers to an organization that provides for-

mal education and is recognized by the educa-

tional authorities of a particular country or region.

This can include schools, colleges, universities,

and other institutions that offer educational pro-

grams.

WEBIST 2023 - 19th International Conference on Web Information Systems and Technologies

3.1.2 Tools

Henceforth, we list the main tools used on our work:

• Firebase: Firebase is a cloud platform that en-

ables the development of mobile and web applica-

tions without the need to create and manage your

own server infrastructure. It provides solutions for

real-time data storage, user authentication, push

notiﬁcations, among other services. According to

Firebase, their platform integrates with multiple

languages and frameworks

. You can see an ex-

ample in Fig. 2.

Figure 2: Example of ﬁrestore database.

• React Native: React Native is a mobile applica-

tion development framework that uses JavaScript

and React to create native applications for iOS and

Android. It was developed by Facebook and its

community of developers

3.2 Method

We’ve used the face-recognition Python library

which uses other libraries for their facial recognition

such as:

• dlib for the detection and align of faces in images;

• OpenCV for the manipulation and upload of im-

ages, as well as to manipulate images; and

• NumPy to represent and manipulate pixel ma-

trixes from images.

This algorithm is pre-trained with data to recog-

nise and identify what a face is but we still need to use

our data. We then pre-trained the algorithm with pic-

tures of the users that were going to be tested, which

involved taking a video of twenty to thirty seconds

and using a small program that extracted every frame

of the video and saved it in a folder.

https://ﬁrebase.google.com/

https://reactnative.dev/

https://realpython.com/face-recognition-with-python/

The algorithm was then trained with these folders,

saving the list of images with the name or code of

the user. This training process allowed the algorithm

to accurately recognize the faces of the users. The

algorithm’s recognition capabilities were not limited

to single faces, as it was also able to identify multiple

faces in a single picture. This approach proved to be

highly effective in achieving our goals and could be

further optimized to enhance the precision and speed

of the recognition process.

To ensure that the recognition has been properly

registered, the team has developed a database that

simulates teachers, assigned classrooms, and their

students. Firebase has been chosen as the platform

for storing not only the videos but also the database

that will be connected to the app.

Firebase is a comprehensive platform that pro-

vides tools and services for mobile and web devel-

opment, including app building, authentication, real-

time databases, storage, and hosting. As shown in

Figure 3:

• The data is stored in Firebase and connected to the

app.

• You login into the app, go to classroom and take a

picture.

• After the picture is sent, the API is called and runs

the algorithm to recognise faces.

• The image analysis starts using its 68 facial land-

marks.

• The image recieved begins to be analysed.

• The image already analysed gets uploaded to the

database to maintain a record.

Figure 3: App Arquitecture.

FaceCounter: Massive Attendance Taking in Educational Institutions Through Facial Recognition

(a) Login Screen. (b) Home Screen. (c) Classroom Screen. (d) Menu.

Figure 4: Screen Mockups.

• A csv and txt ﬁle is uploaded with all the data of

students present or absent.

When a picture is taken and sent, the face recog-

nition system runs and recognizes the students from

the folder with images. Every time a student is iden-

tiﬁed, the program checks whether the student be-

longs to the classroom the picture was taken for. If

the student is present in the classroom, he or she is

marked as present in the database. Otherwise, the stu-

dent is marked as absent. Additionally, each picture

analyzed is saved in another folder in Firebase stor-

age to provide a record of the time and date. Finally,

the database is updated with the information of the

present and absent students, and a txt and csv ﬁle are

created for easy implementation in institutions.

To seamlessly integrate the process of face recog-

nition into the mobile app, we needs to develop an

API that connects the app and the recognition pro-

gram. The API will enable the user to simply take

a picture, which will then be automatically uploaded

to the system’s database. Subsequently, the recogni-

tion algorithm will be triggered and will determine

if the student is present or absent. One of the main

advantages of this approach is that it eliminates the

need for teachers to wait for the app to respond, as the

recognition process will be handled automatically in

the background. This will save time and reduce the

burden on teachers, allowing them to focus on other

important tasks. Additionally, this seamless integra-

tion will help ensure that attendance data is collected

accurately and in real-time, providing teachers with

up-to-date information on their students.

3.2.1 Database Building

After the creation of our python algorithm in its most

primitive state, we then decided to create a database

where all the data of students, teachers, curses and

list of assitance will be saved. For this we ﬁrst started

using postgreSQL. We created the entities but later

on we found some problems with compatibility, then

during the search of new database we found ﬁrebase.

Firebase is an online no SQL database, we used it

with our google account and we were able to create

a new app to use all the functions on it. Firebase has

a cost in long term use for big apps but in our case

with the ”pay as you go” selection we were able to

use all functions with no cost. Firebase helped us to

save the pictures, the trained model, the list of assis-

tance that we will talk later on and the videos of the

students plus our database and its own tool of authen-

tication for our app. We used ﬁrebase storage for pic-

tures, videos, and ﬁles and our model; we used ﬁre-

store database for our no SQL fatabase; and we used

ﬁrebase authentication for the login of the app.

3.2.2 Database Connection

After polishing the code it was time to connect it with

our ﬁrebase. The connection was easy using the key it

was given to us to get access to the database. We then

used the ﬁrebase library. The changes we made in

the code were basically to use the database to indetify

students and for changing the information in order to

see which students where recognised and if they were

from that speciﬁc classroom. Then with this informa-

tion we could create the assistance ﬁles (csv and txt),

WEBIST 2023 - 19th International Conference on Web Information Systems and Technologies

this where store in the ﬁrebase storage. We used ﬁre-

base storage to take from there our trained model and

also to access the folder with the images for recogni-

tion (the app will send here the pictures taken with the

camera).

3.2.3 App Development

For the app development, we utilized React Native

and the Visual Studio Code IDE. The app interface

followed our university’s color scheme and consisted

of a login screen (Fig. 4a), a main page displaying all

the subjects assigned to the teacher (Fig. 4b), and a

subject-speciﬁc screen with relevant data and the op-

tion to capture a picture (Fig. 4c).

The purpose was to associate the captured image

with a speciﬁc classroom, allowing the recognition

code to access the corresponding information. The

primary objective of the app was to provide teachers

with a convenient way to view their classrooms and

select the desired one for attendance tracking.

We also got a menu option for common app tools

suchs us: change my password, logout, home (return-

ing main page), proﬁle and help menu like shown in

Fig. 4d.

3.2.4 App Database Connection

For the database connection with our app we mainly

needed the access to all 3 functions of ﬁrebase. First

we used ﬁrebase authentication to create the authen-

tication method for our login screen. Then, we used

the ﬁrestore database to get access of the teacher who

is login in, the classrooms they had to get the ID of

the classroom selected and from which the image was

taken, this to send the ID like a parameter to the recog-

nition code. Finally, we needed access to the ﬁrebase

storage to upload the pictures taken in into a folder

from where the recognition code will download them

for their analysis.

3.2.5 Deployment

For the deployment of our recognition code in python

we ﬁrst needed to use ﬂask for creating an app web

with python, after this, we create an image with

docker of the ﬁle and ﬁnally uploaded it to heroku.

This deployment generated an url that we added to our

app to access the url everytime an picture was taken

and uploaded to the ﬁrebase store. In this way, every

time a picuted was uploaded the code run and our face

recognition program will do the recognition, the cre-

ation of assistance ﬁles, and the upload of the image

recognised to another folder.

4 EXPERIMENTS

In this section we will show all the results from the

decision of the algorithm chosen in the benchmark-

ing section, to the development of the app and ﬁ-

nally to its testing. We will ﬁrst show and explain a

small benchmarking summarising the things we took

into consideration for pickingour algorithm of python

face-recognition. Then, in develop of the app we

will show images of the app, explain the python algo-

rithm, explain the no SQL database we used, ﬁrebase

and ﬁnally the deployment of the python algorithm in

heroku for the connection with the app.

4.1 Benchmarking

For the benchmarking we analysed different pre

trained models that we can use for trainning and im-

plement it with our app, we tried 4 models (Yang and

Han, 2020; Khan et al., 2020)

. In Table 1, these

models were tested to see if they detect and recognise

faces or if they only have detection; which metrics

they used to test the algorithm; which datasets they

have use; ﬁnally, the results.

As we can see accuracy is really good in most of

this models. Payment and information of the model

were also taken into consideration at the moment of

selection. In the end we chose the library of python

face-recognition due to all the information and active

community they have. Also because of the easy use

for windows and libraries they included on it, also

with all the trainning this model had.

4.2 Image Testing

For our testing we devided into the categories of dis-

tance, lighting, twins and classroom escenario. for the

distance and lighting test we used only 3 people for

the pictures and a total of ten images per test. In this

pictures the members appearing on it changed their

positions so we don’t use the same pictures with dif-

ferent gestures.

For the twins test we only used one pair of twins

and 5 images of them with no distance or lighting

speciﬁcation.

Finally, for the classroom test we had 2 class-

rooms, a regular and a computer lab one. In each we

used 5 pictures that differ one from other in zoom-

ming done and angles. Also, some pictures (like in

the lab test) had the students on it cutoff because of

“Face recognition with OpenCV, Python, and deep

learning” - PyImageSearch - https://pyimagesearch.com/

2018/06/18/face-recognition-with-opencv-python-and-dee

p-learning

FaceCounter: Massive Attendance Taking in Educational Institutions Through Facial Recognition

Table 1: Benchmarking.

Results

Name Category Dataset Accuracy

Percentage of

Blurry

Images

Face-recognition

Face detection

and recognition

Around 3 million images

Labeled Faces in the Wild (LFW) dataset

99.38%

Microsoft Azure API

Face detection

and recognition

20 photos from 12 individuals

240 image dataset

99.57%

YoloV3 (Redmon and Farhadi, 2018)

Face detection

20 photos from 12 individuals

240 image dataset

99.60%

OpenCV (Bradski, 2000)

Face detection

and recognition

Photos from 200 students

from 2 universities

82.00% 15%

Table 2: Comparison of Pictures in Different Contexts.

(a) Pictures with Different Distances.

Distance

(in meters)

(b) Pictures with Different Lightnings.

Lightning

Backlighting

High lighting

Low lighting

zoomming. Students that did not want to participate

in the experiment where censored in the ﬁnal pictures

and not taken into consideration.

https://face-recognition.readthedocs.io/

https://learn.microsoft.com/en-us/azure/cognitive-ser

vices/computer-vision/overview-identity

4.2.1 Distance Images Testing

In Table 2a, the distance testing for pictures from 2,

5 and 10 meters with 3 people. For 2 meter distance,

there is no problem for the detection. For 5 meter

distance, there is no problem for the detection, even

if we inially thought that longer distance will carry

some problems in the recognition, there are no detec-

tion problems. For 10 meter distance, we start to see

WEBIST 2023 - 19th International Conference on Web Information Systems and Technologies

some issues, the distance seem to see really far and

the recognition does not always works correctly.

4.2.2 Lighting Images Testing

Lighting test will use 10 images per test as before

with same members starting with backlight, then good

lighting and ﬁnally bad lighting.

In Table 2b, the image is a little darker than it

should be with low lightning but still you can be able

to distinguish faces. For the good lighting pictures,

there is no problem at all with a high accuracy. Fi-

nally, we have the bad lighting images, in this test we

believe it will have some issues in the recognition be-

cause of the lack of lighting.

4.2.3 Twins Images Testing

In Fig. 5, a test with twins is depicted, showing that

the method can distinguish them accurately

4.2.4 Classroom Images Testing

In Fig. 6, a classroom controlled environment where

we took two examples. First, we got a computer lab

with separate desktops and equipment that will inter-

fere in the image like obstacles. Second, we got a

regular classroom with desks close to each other and

a more clearer view. As it was mentioned before, not

Figure 5: Twins recognised picture.

Figure 6: Lab Classroom recognised picture.

Figure 7: Regular Classroom recognised picture.

all students wanted to participate in the test and they

were not taken into consideration and censored from

the images test.

In Fig. 7, in the lab test we are able to observe how

it was able to recognise most of students but comput-

ers make it difﬁcult to get a clean view and searching

for some complicated angles to take the picture. We

will go deep into its analysis in the results section.

Here we got the image of the regular classroom

image. As we can obserb students appear clearer than

in the lab class, this could be because of the loca-

tion of desks, in rows and columns like it usually is

making students align and always looking at the front.

As we can see, it did not had problems to detect and

recognise all students in the picture, including the one

that was the farthest away. We will see the results of

this tests later on.

4.3 Results

Here we will talk about the results of our testing of

images. We are going to divide this into distance,

lighting, twins and classrooms testing. In this case

for the accuracy we use the equation (1):

a =

r−e

(1)

where a : accuracy

r : number of recognised faces

e : number of unrecognised faces

d : total number of faces

FaceCounter: Massive Attendance Taking in Educational Institutions Through Facial Recognition

Table 3: Accuracy Results for Different Distances.

(a) For 2 meter distance.

Facial

Detection

Facial

Recognition

Facial

Error

Accuracy

1 3 3 0 100%

2 3 3 0 100%

3 3 3 0 100%

4 3 3 0 100%

5 3 3 0 100%

6 3 3 0 100%

7 3 3 0 100%

8 3 3 0 100%

9 3 3 0 100%

10 3 3 0 100%

Total 30 30 0 100%

(b) For 5 meter distance.

Facial

Detection

Facial

Recognition

Facial

Error

Accuracy

1 3 3 0 100%

2 3 3 0 100%

3 3 3 0 100%

4 3 3 0 100%

5 3 3 0 100%

6 3 3 0 100%

7 3 3 0 100%

8 3 3 0 100%

9 3 3 0 100%

10 3 3 0 100%

Total 30 30 0 100%

Facial

Detection

Facial

Recognition

Facial

Error

Accuracy

1 3 0 0 0.0%

2 3 2 0 66.7%

3 3 1 0 33.3%

4 3 3 0 100.0%

5 3 3 0 100.0%

6 3 3 1 66.7%

7 3 1 0 33.3%

8 3 3 0 100.0%

9 3 2 0 66.7%

10 3 3 0 100.0%

Total 30 21 0 67.0%

The facial detection will be refered to how many

people is in the picture with their faces looking at

the camera for it to be recognised. Facial recogni-

tion will be refered to how many people were labeled

not taking in consideration a wrong label. Finally, fa-

cial error will be associated to how many people was

wrong labeled, which means they were given a label

of someone they are not.

During the experiments as you may see we have

reduced the images original size by 50% so it will be

faster and accurate. Originally the image was reduced

up to 15% of its original size, this because larger im-

ages take longer time to be analysed but are more ac-

curate than smaller images. In this test we have re-

duced most of them up to its 50% and in some cases

we leaved the original size where distance had a big

impact, like 10 meters distance and classroom test.

For the testings we’ve used 10 images in which we

had 3 people on it except in the twins and classroom

testing. During the tryouts, the 3 people in the pic-

tures have changed their positions to right or left with

the purpose of not having the same image ten times

with different facial expressions.

4.3.1 Distance Testing

In this case we have done three experiments of dis-

tance with the 2, 5 and 10 meters to know what would

be the limit of our face-recognition model. Our hy-

pothesis is that farther from 5 meters it will start to

have issues at the momento of recognition, this due to

the fact of image quality downgrade.

As we can see in Table 3a we managed to get

100% of accuracy in the 2 meter test. All faces were

recognised with no problems. This is a good start be-

cause 2 meters is close but what would be a minimun

distance from where the teacher will take a picture.

In Table 3b we also got a 100% accuracy. This dis-

tance was starting to be farther away from the camera

but still close, it is almost double the distance of the

ﬁrst test.

For Table 3c we started to have some errors, orig-

inally in the test of 50% of image size reduction the

program was not able to recognise nor detect any face

in the picture, this is probably because of the image

quality downgrade in the moment of modifying the

size of the image, reducing pixels. We decided to

leave the original size of the image which needed 11

minutes per images for analysis. In the end, after leav-

ing the original size we got the result of a 67

4.3.2 Lighting Testing

For the lighting experiments we will take in consid-

eration lighting factors that can affect the recognition

directly or indirectly. This would be the case of a good

lighting classroom that will make easier the recogni-

tion for the model, but also some other variaties like

bad lightint (with no light at all) and backlight (in case

the classroom got windows).

For the lighting test we start with the backlight, to

see if the model has problems to recognise faces with

a bad lighting. Table 4a shows us that it got 100%

of accuracy, independent from the distance, backlight

will affect the quality of the image and how clean the

faces will be for the recognitions, in this cases it had

no effect.

For the good lighting test in Table 4b we also got

a 100% accuracy as expected, with artiﬁcial lighting

of a classroom.

In the ﬁnal lighting test which was with bad light-

ing (no light at all) we got a 100% accuracy as shown

in Table 4c. Even though there was no good light-

ing and the image was darker the face-recognition

alogrithm was able to detect the people in the picture

with no problem.

WEBIST 2023 - 19th International Conference on Web Information Systems and Technologies

Table 4: Accuracy Results for Different Lightnings.

(a) Backlight.

Facial

Detection

Facial

Recognition

Facial

Error

Accuracy

1 3 3 0 100%

2 3 3 0 100%

3 3 3 0 100%

4 3 3 0 100%

5 3 3 0 100%

6 3 3 0 100%

7 3 3 0 100%

8 3 3 0 100%

9 3 3 0 100%

10 3 3 0 100%

Total 30 30 0 100%

(b) Low Lightning.

Facial

Detection

Facial

Recognition

Facial

Error

Accuracy

1 3 3 0 100%

2 3 3 0 100%

3 3 3 0 100%

4 3 3 0 100%

5 3 3 0 100%

6 3 3 0 100%

7 3 3 0 100%

8 3 3 0 100%

9 3 3 0 100%

10 3 3 0 100%

Total 30 30 0 100%

Facial

Detection

Facial

Recognition

Facial

Error

Accuracy

1 3 3 0 100%

2 3 3 0 100%

3 3 3 0 100%

4 3 3 0 100%

5 3 3 0 100%

6 3 3 0 100%

7 3 3 0 100%

8 3 3 0 100%

9 3 3 0 100%

10 3 3 0 100%

Total 30 30 0 100%

Table 5: Accuracy Results for Twins Tests.

Facial

Detection

Facial

Recognition

Facial

Error

Accuracy

1 2 1 0 50.0%

2 2 2 0 100.0%

3 2 2 0 100.0%

4 2 2 0 100.0%

5 2 2 0 100.0%

6 2 2 0 100.0%

Total 12 11 0 91.6%

4.3.3 Twins Testing

In this section of the experiments we had the oppor-

tunity to test twins in regular scenarios to see if it

was able to see the differences. Distance, lighting, or

background was not taken into consideration for this

experiment, the only goal was to see if it was able to

recognised each of them.

As Table 5 shows, the program was able to recog-

nised the twins with no problem and no error taken.

Only in image one it could not detect one of them but

the one recognised was correct.

4.3.4 Classroom Testing

In Table 6a, we took 5 pictures of a classroom with 12

students. All 5 pictures were different, in some cases

some students did not appear due to a zoom done and

in other cases students where not showing their faces.

We only considered for the column of facial detection

the faces that were clean looking at the camera and

not the faces covered by monitors or other students.

Even though our expectations were low in a labora-

tory classroom due to the separation of desktops and

Table 6: Accuracy Results for Different Rooms.

(a) Lab Class.

Facial

Detection

Facial

Recognition

Facial

Error

Accuracy

1 10 10 0 100%

2 10 7 4 30%

3 11 7 1 55%

4 11 9 2 64%

5 9 7 3 44%

Total 51 40 10 59%

(b) Regular Classroom.

Facial

Detection

Facial

Recognition

Facial

Error

Accuracy

1 9 9 0 100%

2 9 9 0 100%

3 9 9 0 100%

4 9 9 0 100%

Total 36 36 0 100%

students positions from one another, we still managed

to get in image 1 a 100% facial recognition of the stu-

dents in the picture. In some cases wwe had errors

in recognition, some because of the distance, others

probably because of the use of glasses, etc. We got in

the end an average of 59% of facial recognition.

In Table 6b, we took four pictures in a regular

classroom with 14 students and desks in rows and

columns. However, we did not managed to get all stu-

dents to participate and they where censored from the

testing. Then, there were nine students from which

we took the data and nine faces to detect from each

FaceCounter: Massive Attendance Taking in Educational Institutions Through Facial Recognition

image (in this one all were looking to the front mainly

because of desk position). For this opportunity we

had the hypothesis that the accuracy was going to be

good due to the fact students are closer and there are

less objects that obstruct their faces. The results ﬁ-

nally conﬁrms our thoughts with an accuracy of 100%

in all pictures and no errors of mistaken recognitions.

This results shows that our work is better in an envi-

ronment where people is closer, looking to the front

and with less objects that obstruct the picture taken.

5 CONCLUSIONS AND

PERSPECTIVES

We can conclude that we have successfully devel-

oped a user-friendly and straightforward application

for teachers to efﬁciently record classroom attendance

tracking. Through the algorithm testing, we observed

that it is highly effective, although image size reduc-

tion can impact image quality while accelerating the

recognition process. Furthermore, based on our ex-

perimentation with regular classrooms and laboratory

classrooms, we can infer that the algorithm performs

better in regular classrooms, beneﬁting from its opti-

mal viewing angle and fewer obstructions caused by

objects.

To achieve this accuracy the image needs to loose

the minimum of its original size and quality, that is

the main reason the image reduction is to its 50% and

not less. If the image is reduced lower it can affect its

quality downgrade and have more errors in the recog-

nition. The objective would be to not loose quality

or size at all but this takes a long quantity of time

(11 minutes per image) for the program to process.

This work is deployed in heroku because of its lower

prices for deployment, but this gives a low quantity of

ram memory and time for the program to run. When

the image reduction is set to its 50% it crashes due

to time. For this reason, we reduce the images to a

15% of their original size. It was demonstrated in the

section of results that the algorithm has good results

depending on the quality an size of the image.

For future works it would be centered mainly in

the upgrade of this run time issue, to be able of putting

images to its 50% size or even its original size, simi-

lar to (Rodriguez-Meza et al., 2022; Rodr

ıguez et al.,

2021). Also, it would analyse to see if theres a better

way to speed the process of analysis and not reducing

the image quality and size (Leon-Urbano and Ugarte,

2020). Also, the algorithm can be upgrade using other

tecnologies to be able of recognising people that is

farther away, this by using another program that that

polishes the image to have better quality.

REFERENCES

Bradski, G. (2000). The OpenCV Library. Dr. Dobb’s Jour-

nal of Software Tools.

Cerd

a-Navarro, A., Touza, C., Morey-L

opez, M., and

Curiel, E. (2022). Academic integrity policies against

assessment fraud in postgraduate studies: An analy-

sis of the situation in spanish universities. Heliyon,

8(3):e09170.

Khan, S., Akram, A., and Usman, N. (2020). Real time

automatic attendance system for face recognition us-

ing face API and opencv. Wirel. Pers. Commun.,

113(1):469–480.

Kocac¸inar, B., Tas, B., Akbulut, F. P., Catal, C., and Mishra,

D. (2022). A real-time cnn-based lightweight mo-

bile masked face recognition system. IEEE Access,

10:63496–63507.

Leon-Urbano, C. and Ugarte, W. (2020). End-to-end elec-

troencephalogram (EEG) motor imagery classiﬁcation

with long short-term. In SSCI, pages 2814–2820.

IEEE.

Miller, K. W. (2023). Facial recognition technology: Navi-

gating the ethical challenges. Computer, 56(1):76–81.

Nordin, N. and Fauzi, N. H. M. (2020). A web-based mobile

attendance system with facial recognition feature. Int.

J. Interact. Mob. Technol., 14(5):193–202.

Redmon, J. and Farhadi, A. (2018). Yolov3: An incremental

improvement. CoRR, abs/1804.02767.

Rodr

ıguez, M., Pastor, F., and Ugarte, W. (2021). Classi-

ﬁcation of fruit ripeness grades using a convolutional

neural network and data augmentation. In FRUCT,

pages 374–380. IEEE.

Rodriguez-Meza, B., Vargas-Lopez-Lavalle, R., and

Ugarte, W. (2022). Recurrent neural networks for de-

ception detection in videos. In ICAT, pages 397–411.

Springer.

Wati, V., Kusrini, K., Fatta, H. A., and Kapoor, N. (2021).

Security of facial biometric authentication for atten-

dance system. Multim. Tools Appl., 80(15):23625–

23646.

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

bum: Towards automatic photo management based on

person identity on mobile phones. In ICASSP, pages

3031–3035. IEEE.

Yang, H. and Han, X. (2020). Face recognition attendance

system based on real-time video processing. IEEE Ac-

cess, 8:159143–159150.

WEBIST 2023 - 19th International Conference on Web Information Systems and Technologies