1. Introduction
recognition system. The knowledge and learned
representations from the pre-training on the ImageNet dataset
Advancements in technology have led to the widespread [
8
] were transferred and fine-tuned on our specific face
adoption of automatic face recognition in several appli- recognition task. The proposed system is designed to
cations, which have become increasingly popular for tackle the challenges associated with recognizing faces
personal identification purposes and have gained pop- in uncontrolled environments, where factors like lighting,
ularity due to their non-intrusive and user-friendly na- pose variations, and occlusion can significantly impact
ture. Handcraft face recognition methods, like LBP [
1
], the accuracy of face recognition systems. Moreover, the
PCA [
2
], and LDA [
3
], have shown commendable per- system aims to address the limitations posed by small
formance in constrained environments. However, their datasets, which often hinder the training of deep learning
performance significantly degraded in unconstrained sce- models. The primary objective of this study is to
thornarios involving occlusion and variations in illumination, oughly evaluate the performance of our proposed face
pose, and expression[
4
]. To tackle this issue, Deep Learn- recognition system and compare it against existing
stateing (DL) approaches, particularly Convolutional Neural of-the-art systems. The remaining sections of this paper
Networks (CNNs), have obtained high accuracy rates are arranged as follows: Section 2 presents an overview
in image recognition, including biometric applications of the related works. The proposed system is introduced
[
5
], [
6
]. This paper introduces a novel and efective face in Section 3. Detailed experimental results, along with
recognition system relying on a new Deep CNN architec- a comparison to previous approaches, are provided in
ture inspired by the Xception model [
7
]. Transfer learn- Section 4. Finally, Section 5 concluded the paper.
ing was employed in our study to address the challenge
of limited data that can negatively impact the training
performance of our proposed deep CNN model. By lever- 2. Related Work
aging the power of transfer learning, we used a part of the
pre-trained Xception model as a foundation for our face
Over the past few years, there has been a proliferation
of studies focused on face recognition. This section
pro6th International Hybrid Conference On Informatics And Applied Math- vides an overview of some of these research endeavors.
ematics, December 6-7, 2023 Guelma, Algeria Wang et al. [
9
] presented a hybrid system designed
* Corresponding author. for facial recognition. The authors initially employed
† These authors contributed equally. Local Binary Patterns (LBP) to generate LBP images,
$ a.hattab@univ-batna2.dz (A. Hattab); a.behloul@univ-batna2.dz which were then utilized as input for a Deep
Convo(A.0B0e0h0l-o0u00l)3-1796-9267 (A. Hattab); 0000-0002-8664-7301 lutional Neural Network (CNN). The proposed approach
(A. Behloul) demonstrated a commendable accuracy rate in the task
© 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License of face recognition. In their study, Ouyang et al. [
10
]
Attribution 4.0 International (CC BY 4.0).
employed the Improved Kernel Linear Discriminant Anal- aims to enhance sparsity in the representation. However,
ysis (IKLDA) technique along with Probabilistic Neural despite its advantages, the image reconstruction methods
Networks (PNNs) for face recognition. The proposed based on this approach sufer from several well-known
approach showed promising results, particularly when drawbacks. These drawbacks include the requirement
dealing with datasets that contained only a small number of an overcomplete dictionary, a significant increase in
of samples. Using a widely-used multi-sample dataset, the number of gallery images, which introduces
comMin et al. [
11
] trained a Deep Convolutional Neural Net- plexity issues, and limitations in terms of generalization
work (CNN) model. Subsequently, the authors utilized capability. These limitations pose challenges in practical
this pre-trained model for face recognition on various implementation and afect the overall efectiveness and
datasets. To enhance the accuracy rate, they employed scalability of the proposed method. Liao and Gu [21]
the K Class Feature Transfer (KCFT) technique, which en- introduced a novel method for face recognition called
riched the intra-class variation. In their research, Kamen- Dictionary Learning and Subspace Learning (DLSL) . This
cay et al. [
12
] introduced a novel face recognition system approach efectively addresses the challenges posed by
that utilized Convolutional Neural Networks (CNNs) for corrupted data, such as noise, occlusion, and significant
feature extraction. The proposed system was evaluated pose variations. DLSL incorporates a subspace learning
using the ORL database and demonstrated an impressive algorithm that utilized both sparse and low-rank
conaccuracy rate of 98.3 Ȯusliman et al. [
13
] introduced a straints. Through extensive experiments conducted on
set of Rotation-Invariant features based on directional several face databases, the results demonstrate that DLSL
coding for texture classification. These features have outperforms many state-of-the-art algorithms. This
highbeen used in the face recognition field and exhibited ex- lights the superior performance of DLSL in achieving
ceptional accuracy when tested on diverse face image improved accuracy and robustness in face recognition
databases. Sapijaszko et al. [
14
] conducted a study where tasks. In their study, Zeghina et al. [22] utilized the
Harthey built a face recognition system that employed a ris Detector to identify significant regions within facial
preprocessing algorithm to enhance the quality of fa- images. These regions were then extracted and fed into
cial images. The system extracted features from the im- a proposed Convolutional Neural Network (CNN) for
proved images based on Discrete Cosine Transform and face recognition. The system’s performance was
evaluDiscrete Wavelet Transform algorithms. In the classifi- ated using multiple datasets, and it achieved satisfactory
cation stage, they employed a multilayer sigmoid neural results, particularly in handling occluded face images.
network. In their research, Hattab and Behloul [
15
] em- Ayyavoo and Suseela [23] introduced a robust method
ployed the SIFT technique to identify important regions that combined the use of 2D Discrete Wavelet
Transwithin facial images. Subsequently, the authors utilized form (DWT) and Contrast Limited Adaptive Histogram
Adaptive Local Ternary Pattern (ALTP) to extract fea- Equalization (CLAHE). The proposed approach involved
tures from the identified regions. For the classification several steps. Firstly, the image was decomposed into
stage, they employed the K-Nearest Neighbors (KNN) high-frequency and low-frequency components using
algorithm. The proposed method demonstrated a no- 2D DWT. Next, CLAHE was applied specifically to the
tably high accuracy rate compared to traditional hand- low-frequency components, enhancing their contrast.
Ficrafted methods. In another study, Hattab and Behloul nally, the features were extracted using Gabor Magnitude
[16] extract facial features from face images based on (GM) analysis. This method aimed to mitigate the
imthe pre-trained models AlexNet-v2 [17] and VGG16 [18]. pact of illumination efects and improve the accuracy
They then employed Linear Support Vector Classifier and robustness of face recognition systems. Yuan et al.
(LinearSVC) for classification purposes. The proposed [24] introduced a collaborative representation technique
method achieved a high accuracy rate when evaluated that captures the relationships between data points. By
on the ORL database. However, it should be noted that integrating Fisher Discriminant Analysis (FDA) and LDA,
the models used in this approach contained a large num- the authors were able to efectively represent the data
ber of parameters, reaching into the tens of millions. in a reduced-dimensional space. This novel
represenMehdipour Ghazi and Kemal Ekenel [19] conducted a tation yielded promising outcomes, particularly when
comprehensive analysis of the Visual Geometry Group applied to a limited training dataset. Notably, it achieved
(VGG) Face Network, a Convolutional Neural Network a satisfactory level of accuracy in face recognition on
(CNN) architecture widely used for face recognition. The the AR database, which includes variations in
illuminaVGG Face Network utilized a massive dataset consisting tion. Due to its sensitivity to noise, the traditional Linear
of 2.6 million facial images for training. In [20], a sparse Discriminant Analysis (LDA) did not perform well in
regularized Non-negative Matrix Reconstruction (NMR) classifying images captured under unconstrained
envimethod is presented, wherein the traditional L2-norm ronments. In response to this limitation, To address the
constraint on the NMR framework’s representation is challenges posed by illumination variations in face
recogreplaced by an L1-norm constraint. This modification nition, Wen et al. [25] introduced a novel approach for
feature extraction called Robust Sparse Linear
Discriminant Analysis (RSLDA). This method was specifically
designed to address the challenges posed by noise. The
researchers evaluated RSLDA on the AR database, which
includes illumination variations, as well as five other
databases. The experiments conducted demonstrated the
efectiveness of the proposed RSLDA method. Wen et
al. [
26
] introduced a highly efective linear regression
method known as Adaptive Locality Preserving
Regression (ALPR). This method proposed a single-view
dimension reduction technique that obtained an excellent
classification accuracy rate. The eficacy of this linear
regression method has been thoroughly demonstrated across
various databases. Dahmouni et al. [
27
] introduced a
recognition method tailored for educational applications.
Their proposed technique was specifically designed to
address the unique requirements and challenges in
educational contexts. The authors evaluated their method
using the AR database, which comprises images captured
under varying illumination conditions. The results of
the evaluation demonstrated the efectiveness of the
proposed method, achieving an acceptable level of
performance. In their research, Hu et al. [
28
] presented a novel
approach for extracting latent low-dimensional features.
They introduced a strategy that utilized the non-squared
L2 norm to enhance the local intraclass relationships
within the data. To validate the efectiveness of their
proposed low-dimensional representation, the authors
conducted experiments using multiple databases. Hattab
and Behloul [
29
] presented a face recognition method
that is robust to illumination variations. The proposed
method utilized four blocks of the pre-trained VGG16
architecture, followed by an Average Polling layer, to
extract features from the facial images. Subsequently,
Principal Component Analysis (PCA) was applied to
reduce the dimensionality of the extracted features. In the
classification task, Linear Support Vector Classifier
(LinearSVC) was employed. The efectiveness of the proposed
approach was evaluated using the Extended Yale B and
AR databases, both of which contain images captured
under varying illumination conditions.
the ImageNet Large Scale Visual Recognition Challenge
(ILSVRC) [
7
]. The Xception model draws inspiration
from the inception-v3 [
30
] model, incorporating
depthwise separable convolution layers instead of inception
modules. It has been built with 14 residual blocks and
more than 22 million parameters. The architecture of the
Xception model is illustrated in Figure 1 (a).
One notable limitation of the Xception model is its high
demand for a substantial volume of training data to
efectively learn all its parameters. This characteristic poses a
challenge when applying the model to tasks with small
datasets. In order to address this challenge, researchers
have employed Transfer Learning. They trained the
Xception model on a large dataset, then used the pre-trained
3. The Proposed System model to classify the images of a new smaller dataset
[
33
]. The Xception model, pre-trained on the ImageNet
In recent years, Deep Learning has achieved exceptional dataset ofers promising potential for face recognition
performance in image recognition tasks, particularly tasks due to the characteristics of the ImageNet database.
through the utilization of deep Convolutional Neural With a vast collection of 1.4 million images,
approxiNetworks (CNNs). In the past decade, numerous deep mately 17% of which contain at least one face [
34
], the
CNN models have been developed, such as Xception [
7
], ImageNet dataset provides a rich source of diverse facial
inception-v3 [
30
], ResNet [
31
], AlexNet [
32
] and VGG16 images. This extensive coverage of facial data enhances
[18]. The Xception model has showcased its robustness the efectiveness of the Xception model pre-trained on
and efectiveness in various image recognition tasks. The the ImagNet database when applied to face recognition
performance of Xception has been exemplary, surpassing tasks.
other renowned deep CNN models, as demonstrated in Through a series of experiments, we have successfully
built an efective face recognition system based on the
Xception model pre-trained on the Imagnet dataset. Our
experiments findings suggest that a CNN model, inspired
by the Xception architecture, and incorporating only five
residual blocks, proves to be an exceptionally efective
model for achieving accurate face recognition results,
particularly on limited face datasets captured under
unconstrained conditions. The paper proposes a novel CNN
model that aims to achieve a balance between computa- Figure 2: Samples from the ORL database.
tional eficiency and high-performance face recognition.
The proposed model comprises only five residual blocks,
which are followed by Global Average Pooling (GAP) and Google Colab and leveraging the functionalities ofered
Softmax layers. The model architecture includes 5 Con- by the open-source Keras library. We employed the Adam
volutional layers, 9 Separable Convolutional layers, and optimizer with a learning rate of 0.001 to train our model
4 MaxPooling layers. A GAP layer is used to minimize for a maximum of 100 epochs, using a batch size of 32.
the feature dimensions from 361*728 to 728. Overall, the Additionally, we incorporated early stopping measures
model consists of less than three million parameters. to address the risk of overfitting by halting training in
To achieve high performance while conserving compu- the absence of improvement in validation accuracy over
tational resources, we adopt a strategic approach during 10 consecutive epochs.
the training stage. Specifically, we freeze the initial four
pre-trained residual blocks, which are responsible for
extracting low-level face features. This allows us to con- 4.1. Dataset:
centrate our eforts on fine-tining the fifth residual block, We conducted evaluation experiments on two diferent
which plays a crucial role in capturing high-level features face datasets:
capable of extracting representative facial features. At
the same time, we train the Softmax classifier, responsi- 4.1.1. The ORL database:
ble for outputting the person’s identity. The architecture
of our proposed model, inspired by the Xception model, The ORL database1 consists of a collection of 400 face
is depicted in Figure 1 (b). images captured from 40 individuals. Each person is
represented by ten images, all of which have a resolution
of 112×92 pixels. The images in this database were
cap4. Experiments tured under diverse conditions, encompassing variations
in illumination, facial expressions, time intervals, and the
presence of eyeglasses. Samples from the ORL database
are depicted in Figure 2.
.
In this work, we proposed a face recognition system
relying on the Xception model pre-trained on the ImageNet
dataset. To demonstrate the efectiveness of our proposed
system, we conducted comprehensive evaluations using
the cross validation protocol. Our performance
evaluations were conducted on the ORL and AR databases, 4.1.2. The AR Face Database:
which served as the testbeds for assessing the system’s The AR database [
35
] consists of more than 4,000 face
recognition capabilities. Two experiments were con- images capturing 126 individuals (56 women and 70 men).
ducted using the AR database to assess the performance The images in this dataset were captured under
conof our system. In the first experiment, an AR dataset con- strained conditions, including various facial expressions,
taining persons with glasses and scarves was employed to lighting conditions, and occlusions such as sunglasses
evaluate the robustness of our method against occlusions. and scarves. In this work, a subset of 100 subjects was
In the second experiment, a dataset containing images selected, including 50 men and 50 women. Each subject
captured under illumination variation was utilized to is represented by 26 images. In the first experiment, we
demonstrate the system’s capability to robustly recog- used the selected dataset to evaluate the robustness of our
nize faces despite changes in illumination conditions. To system against occlusions; only the images containing
establish a comparative analysis, we benchmarked our fully visible faces were included in the training subset,
system’s recognition rate against other recent works in resulting in 14 images per subject (as shown in Figure 3
the field. This allowed us to gauge the performance of our (a)). The remaining images, which portray individuals
proposed system relative to existing methodologies and
highlight its potential advancements. The experiments
for this study were performed utilizing the resources of
1“AT&T laboratories of cambridge university ,the database of faces”,
Available online: http://cam-orl.co.uk/facedatabase.html, Accessed
on 07/06/2023
than three million parameters. This highlights the
efiwearing glasses or scarves, were reserved for a testing ciency and efectiveness of our approach in achieving
subset (see Figure 3 (b)). In the second experiment, we comparable or superior accuracy rates with significantly
applied a two-fold cross validation protocol to evaluate fewer parameters.
the robustness of our system against illumination vari- In order to validate the robustness of our proposed
sysation, where we used a dataset containing fully visible tem against occlusions, we conducted an additional
experfaces captured under diferent illumination variations (as iment using the AR database. The training set consisted
shown in Figure 3 (a)). only the fully visible faces, while the testing set
com. prised faces intentionally occluded by glasses or scarves.
This experimental setup allowed us to assess the
system’s performance under challenging conditions where
4.2. Experimental results: occlusions are present. The accuracy and loss function
To validate the robustness of our proposed system under values of our system on the AR database are visualized
uncontrolled conditions, we performed experiments us- in Figure 4. In Figure 4 (a), it can be observed that our
ing the ORL database. Our system exhibited exceptional model achieved a commendable accuracy of 95.17%. This
robustness, achieving a remarkable accuracy rate in the demonstrated the efectiveness of our system in
accuexperiments. By employing the five-fold and two-fold rately recognizing faces even in the presence of
occluCross-Validation protocol (refer to Table 1), we achieved sions. Furthermore, Figure 4 (b) provides evidence of
a perfect accuracy of 100%. the rapid convergence of the loss function. This further
To demonstrate the robustness of our proposed sys- supports the robustness of our system. Table 3 proves
tem, we conducted a comparative analysis of our accu- our system’s efectiveness against occlusion compared
racy rates against state-of-the-art techniques commonly to state-of-the-art methods on the AR database.
employed in face recognition research. The outcomes, .
as displayed in Table 2, clearly indicate that our system By applying the two-fold Cross Validation protocol,
surpassed the accuracy rates achieved by recent works we ensured a comprehensive evaluation of our system’s
on the ORL database. For instance, the system developed performance. The outstanding performance of our
proby Hattab and Behloul [16] achieved a notable recogni- posed system in handling illumination variations is
evition accuracy of 100% on the ORL dataset. However, it is dent from the results presented in Table 4. Our system
important to note that their system incorporated tens of achieved an impressive accuracy rate of 99.93% on the AR
millions of parameters, while our system utilizes fewer dataset, surpassing the performance of the
state-of-theart methods evaluated on the same dataset that contain
fully visible faces captured under illumination variation.
Table 4 provides a comparative analysis of the accuracy
rates obtained by state-of-the-art methods, clearly
indicating that our proposed system outperforms other
approaches in terms of recognizing faces under varying
illumination conditions.