Wasit Journal for Pure Sciences Vol. (2) No.

(3)

Face Recognition approach via Deep and Machine

Learning
Ola N. Kadhim
Department of Computer Science, Faculty of Computer Science & Mathematics, University of
Kufa, Najaf, Iraq
Technical Institute of Al-Mussaib, Al-Furat Al-Awsat Technical University, Najaf, Iraq
ola.najah@atu.edu.iq

Mohammed Hasan Abdulameer

Department of Computer Science, Faculty of Education for Girls, University of Kufa, Najaf,
Iraq
mohammed.almayali@uokufa.edu.iq

Abstract— Face recognition is a biometric technology that involves identi-

fying and verifying individuals based on their facial features. It finds applications
in security, surveillance, and user authentication systems. The extraction of facial
image features and classifier selection are more challenging to identify with con-
ventional facial recognition technologies, and the recognition rate is lower. The
paper present proposed model combined between deep wavelet scattering trans-
form network regarding the extraction of features and machine learning for clas-
sification purposes. The proposed model consists four stage: obtaining images,
performing pre-processing, extracting features, and then applying classification
techniques. using both SoftMax classifier (part of deep learning model) and Sup-
port Vector Machine classifier (SVM). We used property collected dataset called
MULB dataset. The experimental result shows that SVM classifier provide better
results than SoftMax classifier. The results from the experiments conducted on
the MULB face database showcased the efficacy of the suggested face recogni-
tion approach. The proposed method achieved an outstanding recognition accu-
racy of 98.29% with SVM classifier and 97.87% with SoftMax classifier.

Keywords—Wavelet Scattering Network, Face recognition, Biometric, deep

learning.

1 Introduction

Biometric recognition technology relies on physiological or behavioral attributes to

identify individuals, has found extensive application across diverse sectors of society.
Face recognition technology has emerged as a groundbreaking advancement in the field
of biometric identification and surveillance [1]. With its ability to accurately identify
individuals through their distinct facial characteristics, this technology has garnered
considerable interest and extensive implementation in diverse fields. From enhancing
security measures at airports and organizations to improving user authentication sys-
tems on smartphones, face recognition has revolutionized the way we interact with

152
Wasit Journal for Pure Sciences Vol. (2) No. (3)

technology and ensure safety in different environments [2]. Face recognition, utilizing
traditional techniques, has long been a fundamental approach in the field of computer
vision and biometric identification. Before the advent of deep learning and advanced
algorithms, traditional face recognition methods relied on the extraction of handcrafted
features and the application of statistical classifiers. These techniques involved analyz-
ing facial characteristics such as shape, texture, and spatial relationships to establish
identity. Although traditional face recognition methods may not possess the same level
of accuracy and robustness as deep learning approaches, they have laid the foundation
for the development of modern face recognition systems [3]. With deep learning, face
recognition has revolutionized the field of computer vision and biometric identification.
With its ability to automatically extract and analyze intricate facial features, deep learn-
ing has meaningfully improved the accuracy and efficiency of face recognition systems.
Through the utilization of deep neural networks, these systems these systems have the
capacity to acquire knowledge and understand complex patterns, allowing for robust
identification and verification of individuals [4]. The deep wavelet scattering transform,
known for its ability to capture multi-scale and invariant representations, has emerged
as a promising approach for extracting robust features from facial images. By decom-
posing facial data into different frequency bands and orientations, the deep wavelet
scattering transform effectively captures both local and global information, enabling
enhanced face recognition accuracy [5]. The upcoming sections of this paper follow
this organization: Section 2 examines related literature, Section 3 outlines the research
background, Section 4 presents the methodology and the proposed approach in detail,
Section 5 analyzes the experimental results, and finally, Section 6 concludes the paper
along with discussing future work.

2 Related Literature

In 2018 Neamah H. Alskeini, et al. [6] To improve face recognition performance,

two algorithms were proposed. The initial algorithm utilizes Sparse Representation-
based Classification (SRC), Training Image Modification (TIM), Histograms of Ori-
ented Gradients (HOG) descriptors, while also selecting the Maximum Number of Im-
ages (MNI) from sub-databases. In order to tackle the dimensionality problem, the en-
tire database is partitioned into multiple sub-databases. The second algorithm involves
the use of Convolutional Neural Networks (CNNs). In 2019 Soad Almabdy, et al. [7]
employed various approaches using pre-trained Convolutional Neural Network (CNN)
architectures. In the first approach, the researchers utilized pre-trained CNN models for
example AlexNet and ResNet-50 for the extraction of features, followed by classifica-
tion using SVM. In the second strategy, they utilized transfer learning based on the
AlexNet model. to extract features and perform classification. The accuracy level
ranges from 94% to 100%. %. In 2020 Arpita Gupta, et al. [8] proposed architecture for
facial emotion recognition involves incorporating attention blocks, residual connec-
tions, and convolution networks into a deep self-attention network. Each block within
the network serves a specific purpose - residual connections address the vanishing gra-
dients issue, the convolution network extracts features, and attention significantly en-

153
Wasit Journal for Pure Sciences Vol. (2) No. (3)

hances the network's visual clarity. The suggested model demonstrated exceptional per-
formance when compared to other CNN-based networks, achieving an impressive train-
ing accuracy of 85.76% and a validation accuracy of 64.40%. In 2020 Fahima Tabas-
sum, et al. [9] The combination of the Discrete Wavelet Transform (DWT) coherence
with four different algorithms, namely principal component analysis (PCA) error vec-
tor, PCA eigen vector, Linear Discriminant Analysis (LDA) eigen vector, and Convo-
lutional Neural Network (CNN), was utilized. The results of these four approaches were
fused using detection probability entropy and a Fuzzy system, resulting in an impres-
sive recognition rate of 89.56% even in the worst-case scenario. and an impressive
93.34% for the best-case scenario. In 2022 Showkat Ahmad Dar, et al. [10] developed
a novel Real Time Face Recognition (RTFR) system, comprising three primary stages.
In the initial phase, the system captures real-time video footage and collects 1056 face
images from 24 individuals using a camera with a resolution of 112*92. The subsequent
stage employs an efficient RTFR algorithm to recognize faces by comparing them
against a pre-existing database. This algorithm leverages two distinct deep learning ap-
proaches, namely CNN and VGG-16 with Transfer Learning, to improve the perfor-
mance of the RTFR system. The proposed algorithm demonstrates impressive accuracy,
yielding a result of 99.37.

3 Research Background

In this section, a brief explanation of the various tools (algorithms and techniques)
employed in this proposal will be provided.

3.1 Multi-task Cascaded Convolutional Neural Networks (MTCNN)

MTCNN is a deep convolution neural network-based approach Regarding the task

of detecting faces. and alignment that is capable of simultaneously doing both tasks.
When compared to the traditional method, MTCNN demonstrates superior perfor-
mance, locates the face precisely, moves at a faster rate, and has the ability to detect in
real time. It consists of three cascaded neural networks, each designed to perform spe-
cific tasks in face detection. These networks work together to accurately detect and
localize faces in an image. Before employing these networks, the original image should
be scaled to various scales to create an image pyramid in order to achieve face recog-
nition on a unified scale [11]. The overall structure of MTCNN shown in Figure 1.
P-Net (Proposal Network): The first stage of MTCNN is P-Net, which is responsible
for generating candidate face regions (proposals) in the input image. It utilizes a fully
convolutional network to propose potential face bounding boxes along with confidence
scores for face detection. R-Net (Refine Network): The second stage, R-Net, further
refines the candidate face regions proposed by P-Net. It filters out false positives and
improves the accuracy of face detection by employing a more sophisticated network
and fine-tuning the bounding box predictions. O-Net (Output Network): The final stage,
O-Net, is responsible for accurately regressing landmarks on the face within the refined

154
Wasit Journal for Pure Sciences Vol. (2) No. (3)

face regions obtained from R-Net. It also performs additional face classification to de-
termine if the detected regions indeed contain faces.

Fig. 1. Overall structure of MTCNN

3.2 Wavelet Scattering Transform Network

Wavelet techniques prove to be powerful instruments for creating meaningful data

representations and extracting features, making them compatible with various classifi-
cation algorithms. The wavelet scattering transform, specifically, enables the genera-
tion of dependable features that exhibit local stability against minor deformations.
These features can be effectively combined with a deep neural network, enhancing the
network's capabilities for tasks like pattern recognition and classification. The Scatter-
ing Wavelet Network (ScatNet) was first developed by Mallat [12] Using a series of
wavelet transforms in conjunction with a modulus operator. A deep wavelet consists of
many layers, where the input of one layer is used as the output of the following layer.
Each layer consists of three operations [13], as shown in Figure 2.

Convolution Nonlinearity Averaging

(wavelet) (modulus) (scaling fun)
𝑰∗ 𝜓 |𝑰 ∗ 𝜓| |𝑰 ∗ 𝝍| ∗ 𝝋

Fig. 2. Operations of Wavelet Scatter where 𝐼 represents the input data, 𝜓 denotes a wavelet
function and 𝜑 represents an averaging low-pass filter.
Suppose we have an image (𝐼). The first scattering coefficient is the average of the
image and can be obtained by convolving the image with scaling filter (low pass filter)
𝜑𝑗 , as Equation 1 [14]:
𝑆0,𝑗 (𝐼) = 𝐼 ∗ 𝜑𝐽 (1)
Where 𝑗 a certain scale.
First-layer scattering coefficients are obtained by applying convolution to the input im-
age (𝐼) using a wavelet filter (𝜓𝜆1 ) at a certain scale (𝑗) and taking the modulus of the

155
Wasit Journal for Pure Sciences Vol. (2) No. (3)

resulting coefficients, followed by low-pass filtering with a scaling filter (𝜑𝑗 ). The co-
efficient is determined by Equation 2 [14]:
𝑆1,𝑗 (𝜆1 , 𝐼) = |𝐼 ∗ 𝜓𝜆1 | ∗ 𝜑𝑗 (𝐼) (2)
Second-layer scattering coefficients are obtained by convolving the image (I) by a
wavelet filter (𝜓𝜆1 ) at a certain scale (𝑗), taking the modulus of the resulting coeffi-
cients, followed by convolving again with another wavelet filter (𝜓𝜆2 ) at a different
scale and taking the modulus, and finally low-pass filtering with a scaling filter (𝜑𝑗 ).
The coefficient is determined by Equation 3 [14]:
𝑆2,𝑗 ((𝜆1 , 𝜆2 ), 𝐼) = ||𝐼 ∗ 𝜓𝜆1 | ∗ 𝜓𝜆2 | ∗ 𝜑𝑗 (𝐼) (3)
Note that the length of local translation invariance is determined by the parameter (𝑗),
which is the breadth of the low-pass filter. In addition to the amount of scales the trans-
form can produce, 𝜑 signifies a low-pass filter, 𝜓 denoted wavelet, 𝜆 is the rotation op-
erations, since 𝑆1 , 𝑗 and 𝑆2 , 𝑗 are the outputs of low-pass filters, they can be down-sam-
pled according to the filter width 2 power 𝑗. The last layer of the Wavelet Scattering
network is the average pooling layer, which computes the average of each scattering
coefficient over the spatial domain. The network's final output is a vector of all the
calculated averages, which captures the low-level and high-level texture features of the
input signal or image. Equation 4 determines the final layer's output [14]

𝑆𝑘,𝑗 (((𝜆1 , 𝜆2 , … , 𝜆𝑘 ), 𝐼) = |||𝐼 ∗ 𝜓𝜆1 | ∗ … |𝐼 ∗ 𝜓𝜆𝑘 | ∗ 𝜑𝑗 (𝐼) (4)

where 𝑆 represents the final output vector, and 𝑘 represents the network's layer count.
The values of 𝜆1 , 𝜆2 , … , 𝜆𝑘 represent the possible values of the wavelet filter orientation
angles at each layer. The values of 𝑗 for each layer determine a measure of how many
scales the transform can produce. The final output vector 𝑆 contains information about
the local spatial arrangement of texture features, which can be used for various image
analysis and classification tasks [14] [15]. Figure 3 show wavelet scattering transform
network.

156
Wasit Journal for Pure Sciences Vol. (2) No. (3)

Fig. 3. Wavelet Scattering Network [16]

3.3 SoftMax Classifier

SoftMax classifier is a type of classification algorithm commonly used in machine

learning for multi-class classification tasks. It evaluates the probabilities of the input
belonging to various classes and allocates it to the class exhibiting the highest proba-
bility. Mathematically, given an input vector z = [z₁, z₂, ..., zₖ] of k real numbers, the
SoftMax function calculates the probability distribution as Equation 5:
𝑃(𝑦 = 𝑖|𝑧) = exp(𝑧𝑖 ) / ∑𝑁 𝑗 (exp ( 𝑧𝑗 )) 𝑓𝑜𝑟 𝑖 = 1,2,3, … , 𝑘 (5)
Here, 𝑃(𝑦 = 𝑖 | 𝑧) represents the likelihood of the input belonging to class 𝑖 .
𝑧𝑖 stands for the 𝑖-th element of the input vector 𝑧. 𝛴 denotes the summation over all
elements 𝑗 in the vector 𝑧. 𝑁 indicates the total number of classes. The SoftMax classi-
fier is often used as the output layer in deep neural networks for multi-class classifica-
tion tasks, and it plays a crucial role in training and optimizing the network's parameters
to minimize the cross-entropy loss between the predicted probabilities and the actual
target labels [17].

3.4 Support Vector Machine (SVM) Classifier

SVM is a widely used supervised machine learning algorithm utilized for classifica-
tion and regression purposes. It is particularly renowned for binary classification tasks,
though it can be adapted for multi-class scenarios too. SVM is especially valued for its
proficiency in dealing with high-dimensional data and its capability to handle complex
datasets with well-defined margin separation. It operates by locating the optimal hyper-
plane in a high-dimensional space that divides the data into various classes [18].

4 Methodology

Through the use of a wavelet scattering transform network, we are investigating the
performance of face recognition in this study. This section explains in detail the pro-
posed model, which consists of four stages: 1) Image Acquisition,2) Pre-processing
stage, 3) Features Extraction stage, 4) Classification stage. Figure 4 shows the steps of
a proposed model.

4.1 Image Acquisition Stage

In the implementation we start with image Acquisition from face sub-dataset in
MULB dataset, then spilt the dataset to 70% for training, 10% for validation, 20% for
testing.

157
Wasit Journal for Pure Sciences Vol. (2) No. (3)

4.2 Pre-processing Stage

In this stage, first we detected face in image by using Multi-task Cascaded Convo-
lutional Neural Networks (MTCNN) that presented in section 4.1. Second, we cropping
the face images region, third we resize all images to the suitable dimensions. (200×200)
for the next stage, finally we transform all input images into grayscale representations.

Load Dataset
Image Acquisition

Splitting Dataset

Training & Validation Testing

Input image Detect face Cropping Resize & RB to gray

Pre-Processing
Feature Extraction

Nonlinearity

Nonlinearity
Convolution

Convolution
Wavelet

Wavelet
Averaging

Averaging

FC

Feature vector
Classification

SoftMax Classifier
&
SVM Classifier

Classified image with class label

158
Wasit Journal for Pure Sciences Vol. (2) No. (3)

Fig. 4. General Structure of The Proposed Model

4.3 Features Extraction Stage

During this phase, we applied the Wavelet Scattering Transform Network to extract
features from the face image, as discussed in section 4.2. The images underwent pre-
processing before being passed through a network of wavelet filters, which comprised
3 levels and 10 nodes for each level. At each layer, the image was processed through a
group of filters, resembling convolution filters used in a convolutional neural network.
After that we use fully connected layer with activation function RELU, this layer ena-
bles the network to learn about nonlinear feature combinations by linking Every neuron
in the preceding layer to each neuron in the subsequent layer. The outcome at this stage
is a set of feature vectors that proves valuable for the classification process.

4.4 Classification Stage

Once the feature vectors are ready, the authentication process utilizes a classifier.
During this phase, the test data is classified using the feature vectors stored in the fea-
ture database, leading to the identity recognition procedure. This study employs both
the SoftMax and SVM classifiers for classification purposes. The architecture of the
proposed wavelet combined with SoftMax classifier for Classification is shown in table
1.

Table 1. A concise overview of the proposed wavelet Scattering Model

Layer (type) Shape of Output Parameter’s number

Input Layer (200,200) 0
Scattering 2D (331,25,25) 0
Flatten (206875) 0
Dense (512) 105920512
Dense (176) 90288
Total params: 106,010,800
Trainable params: 106,010,800
Non-trainable parameters: 0

5 Experimental Results and Analysis

Within this specific section, the proposed model’s efficiency is evaluated using face
sub-dataset in MULB dataset showed in section 3.
3520 frontal-face images with a bit depth of 24 are included in the MLUB and are
divided among 176 participants. Images in (.jpg) format type were taken under various
bright lighting situations using a range of expressions, positions, and accessories. We

159
Wasit Journal for Pure Sciences Vol. (2) No. (3)

divided it into 70% for training (2464 images), 10% for validation (352 images), and
20% for testing (704).

5.1 MULB dataset

The MULB is a property multimodal biometric dataset, which contains homologous
biometric traits. It comprises 20 images of each person's face in various poses, facial
expressions, and accessories, 20 images of their right hand from various angles, and 20
images of their right iris from various lighting positions. The database contains real
multimodal biometrics from 176 people, and all biometrics were accurately collected
using the micro camera of the iPhone 14 Pro Max. The MULB dataset was put together
at Al-Furat Al-Awsat Technical University in Kufa, Iraq, during the winter of 2023. A
group of 176 individuals, consisting of 118 males and 58 females, with ages ranging
from 17 to 54, took part in the data gathering procedure. Each participant had their face,
hand, and iris biometric features gathered, resulting in the creation of three sub-dataset
in MULB. This dataset will soon be available online and can be accessed without any
cost for research and academic purposes. In this paper we use face sub-dataset for face
recognition technique. It comprises of 3520 color images of 176 people’s faces.

5.2 Evaluation Metrics

To evaluate the model, this study uses a variety of metrics, including accuracy, loss
function, F1 Score, precision, and recall.
Accuracy: It is the predominant metric utilized for assessing classification models.
It represents the proportion of accurate predictions out of all the model's predictions.
To calculate accuracy, the number of correct predictions is divided by the total number
of predictions made by the model. The accuracy can be determined as Equation 6 [7]:
(𝑇𝑝𝑜𝑠 +𝑇𝑛𝑒𝑔 )
𝐴𝐶𝐶 = (6)
(𝑇𝑝𝑜𝑠 +𝑇𝑛𝑒𝑔 +𝐹𝑝𝑜𝑠 +𝐹𝑛𝑒𝑔 )
Loss Function: It is a mathematical tool employed to measure the alignment between
the model's predictions and the actual target values during the training process. It rep-
resents the discrepancy or error between the predicted values and the true values. The
primary objective of the model during training is to minimize the loss function, aiming
for predictions that closely match the actual target values.

Precision: It is an evaluation metric that gauges the proportion of true positive pre-
dictions (accurately predicted positive samples) out of all positive predictions made by
the model. It serves as an indicator of the model's capability to minimize false positives,
indicating how well it identifies positive instances correctly. It can be determined as
Equation 7 [7]:
𝑇𝑝𝑜𝑠
𝑃𝑟𝑒 = (7)
(𝑇𝑝𝑜𝑠 +𝐹𝑝𝑜𝑠 )
Recall: It quantifies the fraction of true positive predictions among all actual positive
samples present in the dataset. It evaluates the ability of model to avoid false negatives.
The recall can be determined as Equation 8 [7]:
𝑇𝑝𝑜𝑠
𝑅𝑒𝑐 = (8)
(𝑇𝑝𝑜𝑠 +𝐹𝑛𝑒𝑔 )

160
Wasit Journal for Pure Sciences Vol. (2) No. (3)

F1-Score: It is determined by taking the harmonic mean of precision and recall. A

high F1 Score indicates a balanced trade-off between precision and recall. The F1 score
can be determined as Equation 9 [7]:
(𝑃𝑟𝑒∗𝑅𝑒𝑐)
𝐹1𝑠𝑐𝑜𝑟𝑒 = 2 ∗ (9)
(𝑃𝑟𝑒+𝑅𝑒𝑐)
Where, True Positive (𝑇𝑝𝑜𝑠 ): The number of samples that fall under the positive cat-
egory and that the model correctly identifies as positive. True Negative (𝑇𝑛𝑒𝑔 ): The
proportion of samples that the model properly predicts are negative and fall into the
negative category. False Positive (𝐹𝑝𝑜𝑠 ): The number of samples that should be classi-
fied as negative but are instead the model mistakenly predicts as positive. False Nega-
tive (𝐹𝑛𝑒𝑔 ): The proportion of samples that fall into the positive category but that the
model mistakenly predicts as negative.

5.3 The Evaluation of the Proposed Model Using Various Evaluation Metrics

First, we evaluated the performance when extracting the image features from deep
wavelet scattering transform network followed by SoftMax as a classifier. Second, Im-
age features were extracted from the same network, and subsequently, SVM was em-
ployed as the classifier. After a number of training rounds, the best accuracy rate for
the testing model was found to be 97.87%, and the loss function was 0.0948 with Soft-
Max classifier. During classification with the SVM classifier, the highest accuracy rate
for the testing model was achieved, reaching 98.29%. Figure 6 and Figure 7 show the
accuracy and loss function respectively for proposed model with SoftMax classifier.
Table 2 show the evaluation metrics for proposed model. The chart in figure 8 show the
accuracy of the model with both classifiers. From Table 2 and Figure 8 we notice the
SVM classifier give better result for face recognition model.

Table 2. The Evaluation Metrics for Proposed Model

The Model Accuracy Precision Recall F1-score

Wavelet+SoftMax 0.9787 0.9899 0.9759 0.9785
Wavelet+SVM 0.9829 0.9865 0.9829 0.9827

161
Wasit Journal for Pure Sciences Vol. (2) No. (3)

Fig. 6. The Proposed Model Accuracy Fig. 7. The Proposed Model Loss func-
tion
98.4
98.2
Accuracy

98
97.8
97.6
Type of classifier

SotMax classifier SVM classifier

Fig. 8. The Relationship Between the Classifier and the Accuracy of the Model

6 Conclusion and Future Work

In this paper, deep wavelet scattering transform network were applied for feature
extraction from face image and machine learning technique (SoftMax & SVM) for clas-
sification. According to the experimental results, the SVM classifier has demonstrated
superior performance compared to the SoftMax classifier. The proposed face recogni-
tion method has been thoroughly evaluated on the MULB face database, showing its
effectiveness. Specifically, the SVM classifier achieved an impressive recognition ac-
curacy of 98.29%, while the SoftMax classifier achieved a slightly lower accuracy of
97.87%. In our future research, we plan to incorporate face recognition into multimodal
biometric techniques, along with other types of biometric data that have demonstrated
distinct and discriminative features in face images.

162
Wasit Journal for Pure Sciences Vol. (2) No. (3)

7 Acknowledgment

Our research is not receiving support from any funding agency.

8 References

[1] M. Hasan Abdulameer and R. Adnan Kareem, "Face Identification Approach Using Legendre
Moment and Singular Value Decomposition," International Journal Of Computing Digital
System, vol. 13, no. 1, pp. 389-398, 2021.
[2] Y. Kortli, M. Jridi, A. Al Falou, and M. Atri, "Face recognition systems: A survey," Sensors, vol.
20, no. 2, pp. 3-36, 2020.
[3] Z. Xie, J. Li, and H. Shi, "A Face Recognition Method Based on CNN," in Journal of Physics:
Conference Series, vol. 1395, no. 1, pp. 1-8: IOP Publisher, 2019.
[4] P. S. Prasad, R. Pathak, V. K. Gunjan, and H. Ramana Rao, "Deep learning based representation
for face recognition," in Proceedings of the 2nd International Conference on Communications and
Cyber Physical Engineering, pp. 419-424: Springer, 2020.
[5] A. Rehman, M. Harouni, M. Omidiravesh, S. M. Fati, and S. A. Bahaj, "Finger Vein
Authentication Based on Wavelet Scattering Networks," Computers, Materials Continua, vol. 72,
no. 2, pp. 3369-3383, 2022.
[6] N. H. Alskeini, K. N. Thanh, V. Chandran, and W. Boles, "Face recognition: Sparse representation
vs. Deep learning," in Proceedings of the 2nd International Conference on Graphics and Signal
Processing, pp. 31-37, 2018.
[7] S. Almabdy and L. Elrefaei, "Deep convolutional neural network-based approaches for face
recognition," Applied Sciences, vol. 9, no. 20, pp. 1-21, 2019.
[8] A. Gupta, S. Arunachalam, and R. Balakrishnan, "Deep self-attention network for facial emotion
recognition," Procedia Computer Science, vol. 171, pp. 1527-1534, 2020.
[9] F. Tabassum, M. I. Islam, R. T. Khan, and M. R. Amin, "Human face recognition with
combination of DWT and machine learning," Journal of King Saud University-Computer
Information Sciences, vol. 34, no. 3, pp. 546-556, 2022.
[10] S. A. Dar and S. Palanivel, "Performance Evaluation of Convolutional Neural Networks (CNNs)
And VGG on Real Time Face Recognition System," Advances in Science, Technology and
Engineering Systems Journal, vol. 6, no. 2, pp. 956-964, 2021.
[11] K. Zhang, Z. Zhang, Z. Li, and Y. Qiao, "Joint face detection and alignment using multitask
cascaded convolutional networks," IEEE signal processing letters, vol. 23, no. 10, pp. 1499-1503,
2016.
[12] S. Mallat, "Group invariant scattering," Communications on Pure Applied Mathematics, vol. 65,
no. 10, pp. 1331-1398, 2012.
[13] B. Soro and C. Lee, "A wavelet scattering feature extraction approach for deep neural network
based indoor fingerprinting localization," Sensors, vol. 19, no. 8, pp. 1-12, 2019.
[14] J. Andén and S. Mallat, "Deep scattering spectrum," IEEE Transactions on Signal Processing, vol.
62, no. 16, pp. 4114-4128, 2014.
[15] Z. Liu, G. Yao, Q. Zhang, J. Zhang, and X. Zeng, "Wavelet scattering transform for ECG beat
classification," Computational mathematical methods in medicine, pp. 1-11, 2020.

163
Wasit Journal for Pure Sciences Vol. (2) No. (3)

[16] P. Pandey, R. Singh, and M. Vatsa, "Face recognition using scattering wavelet under Illicit Drug
Abuse variations," in International Conference on Biometrics (ICB), pp. 1-6: IEEE, 2016.
[17] I. Kouretas and V. Paliouras, "Simplified hardware implementation of the softmax activation
function," in 8th international conference on modern circuits and systems technologies
(MOCAST), pp. 1-4: IEEE, 2019.
[18] C.-C. Chang and C.-J. Lin, "LIBSVM: A Library for Support Vector Machines," ACM
Transactions on Intelligent Systems and Technology, vol. 2, no. 3, pp. 1-27, 2011.

Article submitted 13 July 2023. Accepted at 21 August 2023. Published at 30 Septem-

ber 2023.

164