presentation of results and outlining future research
plans.
2 Related Work
The field of face recognition encompasses the
processes of face identification and face
authentication/verification. Face identification refers
to the procedure of discerning an individual's identity
by utilizing an image of their facial features.
Face recognition systems are the subject of many
papers, like Duan, et al. [4], which suggested a local
binary feature learning method. Xu, et al. [5] compile
comprehensive reviews on sparse coding and
dictionary learning algorithms in facial recognition
software. When developing systems for mobile
devices. Hassan and Elgazzar [6] and Oravec, et al.
[7] concentrate on finding a solution while working
with limited resources. Alternative methods
gradually replace sparse coding, with CNN emerging
as the most well-known. Utilizing CNN for face
recognition [8] and [9]. To effectively train this
method, CNNs have a notable limitation that
demands many different images for each class.
Melekhov, et al. [10] and Khalil-Hani and Sung [11]
proposed using CNN with a Siamese architecture to
address this issue. The amalgamation of two parallel
networks is typically achieved by utilizing a cost
function, which primarily functions to classify
features derived from said networks. The Borghi, et
al. [12] presented an advanced Siamese architecture
known as JanusNet, which integrates depth, RGB,
and hybrid Siamese networks through fusion. Fan
and Guan [13] devised two CNN architectures that
were explicitly designed to cater to different
scenarios. The models underwent rigorous training
on a comprehensive dataset of facial images and were
subsequently enhanced through the application of
embedding triplet techniques.
Ameur, et al. [14] have utilized the Weighted
PCA-EFMNet deep learning feature extraction
method to address issues about expression, position,
illumination, and occlusion changes. Majumdar, et al.
[15] introduce a new method (Auto-Encoder) for face
verification called class sparsity supervised encoding
(CSSE). This method uses supervised training data to
teach feature representation. Xiong, et al. [16] have
proposed a part-based learning method for face
verification in which a convolutional fusion network
(CFN) is used to extract feature representation.
Chong, et al. [17] suggest a double layer block
(DLB)-based metric learning method for better
resolution of a pair of face images and a faster general
process in face verification.
Zhang, et al. [18] presented a method for
developing a new CNN and implementing it in
Siamese architecture to achieve 94.8% accuracy in
face recognition by training their model on a small-
sample dataset LFW. To achieve facial recognition.
Heidari and Fouladi-Ghaleh [19] use transfer
learning in a Siamese network structure comprising
two identical CNNs. The findings imply that the
suggested model performs on par with sophisticated
models that have been trained on substantial datasets.
Furthermore, compared to techniques that are trained
using datasets with a limited number of samples, it
increases the accuracy of facial recognition.
According to evaluations using the Labeled Faces in
the Wild (LFW) dataset, the precision is 95.62%.
Wu, et al. [20] present a novel convolutional
Siamese network architecture for the purpose of face
recognition. Like many conventional face
recognition systems, face detection is employed to
determine the precise location of the face within an
image. Deep learning techniques are employed to
leverage facial characteristics. The process of
comparing the detected faces with those stored in the
database is completed. Lai and Lam [21] present a
novel approach utilizing a deep Siamese network to
effectively tackle the challenge of low-resolution
face recognition (LRFR). The approach employed in
our study involves utilizing a Siamese network to
extract profound characteristics from facial images at
varying resolutions. Additionally, a shared classifier
is employed to facilitate the comparison of deep
features from different resolutions with identical
class center vectors.
3 Methodology
A facial recognition system compares a human face
captured in a digital image or video frame with a pre-
existing database of faces. This system is commonly
utilized for the purpose of verifying user identities
through ID verification services. It involves
identifying and measuring specific facial
characteristics extracted from a provided image.
Transfer learning is a widely adopted methodology in
the field of machine learning, particularly in the
domain of deep learning. It involves leveraging pre-
existing models to address various computer vision
tasks. This approach is primarily employed in
situations where there is a limited quantity of data
accessible for the purpose of modeling a novel
problem. Hence, it is possible to employ deep
learning models that have been pre-trained on large
datasets and possess established principles to
construct a transfer learning model that leverages the
WSEAS TRANSACTIONS on SIGNAL PROCESSING
DOI: 10.37394/232014.2023.19.7
Adil Hussain, Asad Ullah, Ayesha Aslam, Amna Khatoon