PHYSIOGNOMY RECOGNITION

BASED CNN
1
KALEESWARI.M, 2MATHUMITHAA.V, 3JONNALAGADDA LAVANYA, 4SOWMIYASREE.C

5
DR. V. SAI SHANMUGARAJA.V, 6DR. G. GUNASEKARAN

1,2,3
IV Year B. Tech CSE Students, 4Assistant Professor, 5,6Professor

Dept of Computer Science and Engineering, DR MGR EDUCATIONAL AND RESEARCH INSTITUTE,
Maduravoyal,Chennai-95, Tamil Nadu, India

2 5
[email protected], [email protected], [email protected]

Abstract-- In a groundbreaking fusion of Keywords: Convolutional Neural

cutting-edge technology and law enforcement, Networks(CNNs), CCTV systems, Criminal
our project embarks on a paradigm-shifting Identification, Face detection, Criminal
journey by harnessing the power of database, Deep learning.
Convolutional Neural Networks (CNNs) to
transform CCTV systems in criminal 1. INTRODUCTION
identification and investigation. By automating
key processes like face detection and matching Facial recognition is a technology that has found
against criminal databases, our technique extensive applications in various fields,
revolutionizes the efficiency of suspect particularly in computer vision. This system is
identification, streamlining the investigative designed to identify and verify human faces from
workflow and allowing law enforcement to images or video footage. It serves critical security
deploy resources more effectively. Notably, the purposes, such as monitoring individuals entering
demonstrated effectiveness of CNNs in a location or searching for specific individuals
distinguishing individuals with criminal within a given area. While not as sophisticated as
records promises a high level of accuracy in biometric or iris scanning, facial recognition is
suspect identification, ultimately improving favoured for its ease of implementation.
public safety through faster responses to
criminal activities and more targeted The core of a face recognition system involves
investigations. This represents a significant the use of algorithms to compare the captured face
leap forward in applying deep learning to in a camera or video footage with a database of
address real-world challenges in law stored facial images. These algorithms can utilize
enforcement and criminal justice. grayscale representations of the stored images,
making it easier for efficient storage and
processing. Several algorithms are employed in
this process. One algorithm stores human images
in a grayscale format in a database, while another footage with a database of images. Inspired by
searches the database for matching faces, and a high-accuracy related works, the CNN model
third algorithm compares the shape, area, or size undergoes preprocessing and training, resulting in
of the face to identify matches within the database. a notable improvement in accuracy. A key
These algorithms work together to determine innovation lies in incorporating metadata for
whether the captured image corresponds to any mapping, enhancing applications in criminal
face stored in the database. tracking and security. This integration allows the
system to extract additional details from images,
Convolutional Neural Networks (CNN) have such as flashlights, location, device name, and
revolutionized the field of face recognition. These camera quality, offering a multidimensional
networks have simplified the development of approach beyond facial recognition. The system's
facial recognition systems, making it more versatility positions it as a robust tool for advanced
accessible and game- changers in the various biometric security applications.
computer vision related applications and provide
solutions for face recognition challenges. 3. RELATED WORK

The face recognition process typically involves The facial recognition system is a advanced
three fundamental steps: face detection, face computer vision designed to link a human face
recognition, and face feature extraction. The with digital images or frames extracted from
system captures the image and locates the position videos. The task of detecting and verifying faces
of the face within it. Face detection may also poses a notable challenge within the field of
include background removal to focus solely on the computer vision, given its real-time applications
facial region, excluding any other distractions. across various domains. Early research articles
However, it's worth noting that facial recognition delved into the intricacies of face recognition.
systems still have limitations, such as difficulty in Post 1990s, the research landscape evolved to
counting heads when faces overlap or when two emphasize the development of models capable of
individuals have very similar facial features. autonomously detecting human faces. Turk and
Pentland emerged as pioneers in this arena,
In this representation, the facial recognition presenting a groundbreaking research model.
system employs a trained CNN model to perform Their approach involved creating a system
the recognition task. Once a matching face is proficient in both detecting and recognizing
found in the input feed, the system utilizes human faces. This system operated in a semi-real-
metadata to track the location where the image time mode, tracking a person's head and
was captured. This information can be valuable for identifying them by scrutinizing facial features. At
various applications, including security the core of their innovation was the introduction
and surveillance. of the "eigenface" algorithm, This algorithm
stands out as the initial algorithm to produce
2. OVERVIEW praiseworthy and efficient outcomes in the realm
of face recognition in a semi-real-time setting.
The Face Recognition System employs This algorithm represented a notable progress in
Convolutional Neural Networks (CNN) for secure the functionalities of face recognition system,
identification by matching individuals in video
laying the foundation for subsequent 4. METHODOLOGY
advancements in the field.
The described methodology focuses on
In the realm of face detection research, significant developing a facial recognition system with image
strides have been made over the years. Early metadata for location tracking. Technological
efforts, exemplified by the seminal work of Viola- advancements have led to improvements in the
Jones, leveraged Haar features and AdaBoost to tracking system. The proposed face recognition
achieve commendable accuracy in real-time model follows a basic architecture, as depicted in
applications. However, subsequent research aimed the figure.
at enhancing performance led to the exploration of
sophisticated hand-crafted features and more The process begins by obtaining images from
potent classifiers, often optimizing individual social media or other sources, which undergo pre-
components in isolation, resulting in sub-optimal processing. During this stage, the images are
face detection pipelines. resized to meet the specifications required for
training on a CNN model. Subsequently, the
Recently, there has been a rise in the use of CNN model is trained.
based face detectors. Cascade CNN, for instance,
introduced a cascade architecture leveraging The training phase involves constructing the
CNNs to enhance discriminative capabilities and proposed CNN model, employing convolutional
performance. Other significant advancements and dense layers comprising artificial neurons.
include the joint training of Cascade CNN for end- Convolutional layers are effective in detecting
to-end optimization, the Face Ness approach spatial patterns and features within the images,
utilizing CNNs for facial attribute recognition to while dense layers contribute to learning complex
detect partially occluded faces, and Sun et al.'s relationships among these features. This training
methods for enhancing Faster R-CNN for face process enables the model to recognize and
detection tasks. differentiate faces.

In the context of our work, we also draw

inspiration from advancements in live location
fetching. Incorporating real-time location data
into the face detection process adds a dynamic
dimension to our methodology. This integration
aligns with the broader goal of creating more
context-aware and versatile face detection
systems. Inspired by the Region Proposal Network
(RPN) in Faster R-CNN and the multi-scale
mechanism in SSD, our work builds upon these
advancements, incorporating live location
fetching to create a state-of-the-art face detector
that not only excels in accuracy and speed but also
integrates seamlessly with real-time Fig 4.1 System Model
location information.
4.1 Open CV images, along with their associated class labels,
are bundled together and stored. This file-centric
OpenCV is like a toolbox for working with images strategy not only enhances data organization but
and videos. This open-source library encompasses also facilitates ease of access during subsequent
2500 optimized algorithms,including a stages of the workflow.
comprehensive set of machine learning
algorithms. Notably, OpenCV is compatible with As the workflow progresses, a file-reading
various operating systems, making it highly mechanism is implemented to retrieve the stored
adaptable. One of the key strengths of OpenCV image-class pairs. Subsequently, a critical step
lies in its language support, facilitating integration involves shuffling the data. This randomization
with Java, C++, Python, and MATLAB. Its process is imperative to prevent any inherent
adaptability makes it a perfect option for a wide biases that might be introduced during model
range of projects. training due to the order of the data. Shuffling
contributes to the robustness of the machine
OpenCV's primary functions include image learning model by ensuring that it learns patterns
manipulation tasks like reading and writing without being influenced by the sequential
images, as well as capturing and saving videos. arrangement of the input data.
Additionally, it excels in real-time video analysis,
enabling facial recognition and feature detection.
This wide range of capabilities makes OpenCV a
powerful and accessible tool for researchers and
developers working on computer vision
applications.

4.2 Image Processing

In the initial phase of this image processing, the

Fig 4.2 Proposed methodology
images are collected and organized within a
designated folder structure. To facilitate effective
4.3 Convolutional Neural Network
data management, a multidimensional array is
employed, wherein each row encapsulates a Convolutional Neural Networks (CNNs) have
numerical representation of an image, obtained emerged as a powerful tool in machine learning,
through the use of the OpenCV library, along with and their success can be attributed to a
its corresponding class label. OpenCV, a widely- combination of factors. First and foremost, the
used computer vision library, is instrumental in architecture of a CNN is carefully designed to
transforming the images into numerical arrays, manage intricate patterns and relationships present
enabling their utilization in subsequent machine in the data. Additionally, the availability of large
learning tasks. amounts of training data is crucial for fine-tuning
the model's parameters, allowing it to learn and
Further, to streamline the handling of large
generalize effectively. The choice of an
datasets, a systematic approach is adopted to save
appropriate loss function further refines the
the image arrays and classes into a file. Set of
network's ability to make accurate predictions. At
the heart of the CNN architecture is the outputs of size 25x25. After applying the max
convolutional layer, which plays a pivotal role in pooling operation, the dimensions are reduced to
feature extraction. In the initial convolutional three tensor outputs of size 6x6. The reduction in
layer, two 5x5 filters are applied to a 25x25 input dimensions simplifies line-by-line sampling and
image across its three channels, and an activation computation, leading to a more computationally
function is subsequently applied. The equation is: efficient model.

𝑍 =𝐴×𝑆
𝑃 = (ℎ − 𝑆 + 1)/𝑙 × (𝜔 − 𝑆 + 1)/𝑙 × 𝑐
The efficiency of the convolutional layer is In a fully connected layer, every neuron form links
quantified by parameters such as the number of
with the layers in neurons before it. The equation
input channels (ci), output channels (co), and
kernel dimensions (kw and k h). This is expressed
through the formula: 𝑍 = 𝑊𝑇 ⋅ 𝐴 + 𝑏
Computation = [(ci x kw x k h)+(ci x kw x
calculates the weighted sum of inputs and adds a
K h – 1) +1] x co x W x H
bias term.
The ReLU activation function enhances efficient
and quick training neural networks by turning off The parameter of the fully connected layer
determines the extent of computation involved in
the output for negative input values. This helps in
the equations provided below.
speeding up the learning process and has become
a popular choice in deep learning models due to its
simplicity and effective 𝑥1 = 𝑊1 1 × 𝑎1 + 𝑊1 2 × 𝑎2 + 𝑊1 3 × 𝑎3 + 𝑏1

ReLU: output=max (0, input) 𝑥2 = 𝑊2 1 × 𝑎1 + 𝑊2 2 × 𝑎2 + 𝑊2 3 × 𝑎3 + 𝑏2

𝑥3 = 𝑊3 1 × 𝑎1 + 𝑊3 2 × 𝑎2 + 𝑊3 3 × 𝑎3 + 𝑏3
This nonlinear transformation boosts the model's
capacity to grasp intricate patterns and
connections within the dataset. A1, A2, A3 Input for fully connection layer x1, x2,
x3 for export.
In a convolutional neural network's pooling
𝑝𝑎𝑟𝑎𝑚𝑠 = (𝑖 + 1) × 𝑜 = 𝑖 × 𝑜 + 𝑜
layer, the primary objective is to decrease the
spatial dimensions of the input volume. Max Each O/P connects with I/P
pooling, the most prevalent type of pooling,
involves sliding a filter across the input, retaining 𝐹𝐿𝑂𝑃𝑠 = [ 𝑖 + (𝑖 − 1) + 1] × 𝑜 = (2 × 𝑖) × 𝑜
only the highest value within the filter's receptive
field. A max pooling operation is applied with a This formula estimates the total number of
filter (also known as a kernel) of size 4x4. This arithmetic operations (multiplications and
additions) required to compute the neuron's
means that, for each region of the input feature
output. Each input connection requires one
map, only the maximum value within the 4x4 multiplication and one addition, and there's an
window is retained. You have three previous additional addition operation for the bias term.
Then this total is multiplied by the number of we aim to harness the potential for improved
output neurons since the computation for each convergence, reduced training instability, and
neuron in the layer needs to be done for each ultimately, a more effective and robust CNN
output neuron.
model.
The soft max function is utilized to transform raw
scores into probabilities, enabling the CNN algorithm employed in this project is
interpretation of the output as class probabilities in designed to extract features from images, saving
a multiclass classification scenario. these features as encodings in a dedicated file.
These encodings, represented as digital numbers,
consist of 128 values for each full-face image.
𝑒 𝑧𝑖
𝑠𝑜𝑓𝑡𝑚𝑎𝑥(𝑍)𝑖 = 𝐾 𝑧𝑗 During image testing, these 128-dimensional
Σ𝑗=1 𝑒 encodings are extracted once again and compared
in contrast to the many encodings. The
Softmax ensures that the output values are comparison involves identifying aspects in the
between 0 and 1 and sum to 1. This is useful when photos that are extremely similar, enabling the
the output is interpreted as probabilities across recognition of faces. Our faces possess distinctive
different classes. ResNet architecture is employed features, including eyes, mouth, nose, and more.
to train Convolutional Neural Network (CNN) To detect these features, we utilized the Dlib
models, contributing to enhanced both efficiency algorithm, which provides a facial recognition
and steadiness. Within ResNet, the residual output in the form of a 128-dimensional vector.
scaling factor is a parameter that can be manually This vector captures essential facial features and
set. In our approach, we posit that adjusting their spatial relationships. The Dlib algorithm
training parameters, particularly by utilizing a produces a map of points surrounding each facial
smaller value for the remaing scaling factor, holds feature, facilitating accurate and detailed feature
the potential to augment model stability. This extraction. In the context of Dlib facial
belief stems from the understanding that subtle recognition, the 128-dimensional output is
changes in these parameters can have a profound obtained by training the network using a triplet
impact on the training dynamics and overall dataset. In triplet-based training, the algorithm
robustness of the model. Moreover, we introduce learns to distinguish between anchor images
a subtle yet effective technique involving the (representing a specific individual) and positive
alteration of activation function values within the and negative examples. Positive examples share
Rectified Linear Unit (RELU) layer of the CNN the same identity as the anchor, while negative
model. This adjustment, although leading to a examples belong to different identities. This triplet
slight increase in training parameters, loss training strategy enables the model to
demonstrates a notable improvement in both efficiently learn a compact representation of facial
model performance and training stability. In features that is robust to variations in pose,
essence, our proposed method seeks to strike a lighting, and other factors.
balance between parameter adjustments and
resultant model enhancements. By strategically After completing the training phase of the
modifying these aspects of the training process, proposed model, it transitions to the testing
process where images can be sourced from various
locations and undergo preprocessing to identify 5.1.1 True Positives (TP):
accurate matches or classes. Subsequently, the These are instances where the facial
workflow involves node-to-node plotting, recognition system correctly identifies individuals
incorporating metadata associated with the who are criminals as criminals. Each true positive
images. represents a successful detection of a criminal,
contributing positively to the system's
Image metadata typically contains details about performance. It indicates the system's ability to
various aspects such as flash settings, location, accurately recognize and identify criminal
device name, camera quality, and more. While this individuals, which is a critical aspect of its
information is embedded or hidden within the effectiveness.
image and may not be visible to everyone, our
methodology involves extracting and utilizing this 5.1.2 True Negatives (TN):
metadata. Specifically, we focus on leveraging These are instances where the system correctly
location information, i.e., coordinates, extracted identifies individuals who are non-criminals as
from the metadata. This node-to-node plotting not non-criminals. True negatives indicate that the
only enhances the interpretability of the results but system can distinguish between non-criminals and
also provides valuable contextual information criminals accurately. While the primary focus
about the origin and distribution of the images in might be on identifying criminals, the ability to
the testing phase. correctly classify non-criminals is equally
important to prevent false accusations or
5. EVALUATION AND PERFORMANCE misidentifications.
ANALYSIS METHOD
5.1.3 False Positives (FP):
5.1 Confusion matrix table These are instances where the system
incorrectly identifies individuals who are non-
A confusion matrix can be used to evaluate the criminals as criminals. False positives can have
performance of a facial recognition system in serious implications, as they might lead to
correctly identifying individuals, particularly wrongful accusations or unnecessary
criminals, based on their facial features interventions against innocent individuals. It's
essential to minimize false positives to maintain
the system's credibility and avoid causing harm to
innocent people.

5.1.4 False Negatives (FN):

These are instances where the system
incorrectly identifies individuals who are
criminals as non-criminals. False negatives
represent missed opportunities to detect and
apprehend criminals, which can compromise
Fig 5.1 Confusion matrix
public safety and security. Minimizing false
negatives is crucial for ensuring that the system
effectively fulfills its intended purpose of
identifying and apprehending criminal 6. CONCLUSION
individuals.
This paper focuses on the development of a
5.2 Interpretation Convolutional Neural Networks (CNN) model
designed specifically for face recognition
5.2.1 Accuracy: purposes. The model utilizes matched face
The model's overall accuracy can be assessed by images, and the corresponding metadata from
computing, the ratio of correct predictions (TP + these images is employed for mapping the
TN) in comparison to the total predictions recognized faces onto a geographical map. The
provides a measure of overall accuracy applications of this model extend to criminal
tracking, suspect identification, and the gathering
𝑇𝑃 + 𝑇𝑁 of digital footprints for enhanced security
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁 measures. In practical scenarios, the
implementation of these CNN-based face
5.2.2 Precision: recognition models has considerable potential in
Precision is a metric for positive forecast accuracy. law enforcement and security contexts. By
It is calculated as employing matched face images, the model allows
for the plotting of identified individuals on a map,
𝑇𝑃 utilizing the metadata associated with each image.
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
𝑇𝑃 + 𝐹𝑁 This metadata contains valuable information, such
as flash settings, location, device name, and
indicating the proportion of correctly identified camera quality, providing additional context to the
face images among all images predicted as facial recognition results.
belonging to a specific individual. The implications of this technology are
particularly significant in criminal tracking, where
5.2.3 Recall : the ability to match and map faces could aid in
Recall assesses the model's capacity to correctly locating and apprehending individuals of interest.
recognise positive instances. It is calculated as Moreover, the system could play a crucial role in
𝑇𝑃
𝑅𝑒𝑐𝑎𝑙𝑙 = suspect tracking, contributing to investigations
𝑇𝑃 + 𝐹𝑃 and ensuring public safety. The gathered digital
footprints, enriched with metadata details, offer a
indicating the proportion of correctly identified
comprehensive overview of the circumstances
face images among all images that truly belong to
surrounding the recognition events. In essence, the
a specific individual.
implementation of the CNN model and its
integration with metadata-driven mapping not
5.2.4 F1 Score:
only enhances face recognition accuracy but also
Recall and precision are combined into one
opens avenues for advanced applications in law
statistic, the F1 score, which offers a balance
enforcement, security, and digital forensics,
between the two.
ultimately contributing to public safety and crime
prevention.
𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 × 𝑟𝑒𝑐𝑎𝑙𝑙
𝐹1 𝑆𝑐𝑜𝑟𝑒 = 2 ×
𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑟𝑒𝑐𝑎𝑙𝑙
References [8] S3FD: Single Shot Scale-Invariant Face
Detector, Shifeng Zhang, Xiangyu Zhu,
[1] FaceNet: A Unified Embedding for Face Zhen Lei, Hailin Shi, Xiaobo Wang, Stan Z.
Recognition and Clustering, Florian Li; Proceedings of the IEEE International
Schroff, Dmitry Kalenichenko, James Conference on Computer Vision (ICCV),
Philbin; Proceedings of the IEEE 2017, pp. 192-201
Conference on Computer Vision and Pattern [9] Face detection with end-to-end integration
Recognition (CVPR), 2015, pp. 815-823 of a convnet and a 3d model, Y Li, B Sun, T
[2] DeepFace: Closing the Gap to Human- Wu, Y WangComputer Vision–ECCV
Level Performance in Face Verification, 2016: 14th European Conference,
Yaniv Taigman, Ming Yang, Marc'Aurelio Amsterdam,... 2016•Springer
Ranzato, Lior Wolf; Proceedings of the [10] Finding Tiny Faces in the Wild
IEEE conference on Computer Vision and With Generative Adversarial Network,
Pattern Recognition (CVPR), 2014, pp. Yancheng Bai, Yongqiang Zhang, Mingli
1701-1708 Ding, Bernard Ghanem; Proceedings of the
[3] Joint face detection and alignment using IEEE Conference on Computer Vision and
multitask cascaded convolutional networks, Pattern Recognition (CVPR), 2018, pp. 21-
K Zhang, Z Zhang, Z Li, Y Qiao IEEE 30
signal processing letters, 2016 [11] DSFD: Dual Shot Face Detector,
[4] Accurate Face Detection for High Jian Li, Yabiao Wang, Changan Wang,
Performance, Faen Zhang, Xinyu Fan, Guo Ying Tai, Jianjun Qian, Jian Yang,
Ai, Jianfei Song, Yongqiang Qin, Jiahong Chengjie Wang, Jilin Li, Feiyue Huang;
Wu [Submitted on 5 May 2019 (v1), last Proceedings of the 30 IEEE/CVF
revised 24 May 2019 (this version, v3)] Conference on Computer Vision and
[5] Efficient Detection of Objects and Faces Pattern Recognition (CVPR), 2019, pp.
with Deep Learning, Najibi, Mahyar. 5060-5069
University of Maryland, College Park [12] A Convolutional Neural Network
ProQuest Dissertations Publishing, 2020. Cascade for Face Detection, Haoxiang Li,
27837540. Zhe Lin, Xiaohui Shen, Jonathan Brandt,
[6] You Only Look Once: Unified, Real-Time Gang Hua; Proceedings of the IEEE
Object Detection, Joseph Redmon, Santosh Conference on Computer Vision and
Divvala, Ross Girshick, Ali Farhadi; Pattern Recognition (CVPR), 2015, pp.
Proceedings of the IEEE Conference on 5325-5334
Computer Vision and Pattern Recognition [13] Face Detection & Recognition
(CVPR), 2016, pp. 779-788 from Images & Videos Based on CNN &
[7] K. Zhang, Z. Zhang, Z. Li and Y. Qiao, Raspberry Pi M Zamir, N Ali, A Naseem,
"Joint Face Detection and Alignment Using A Ahmed Frasteen, B Zafar, M Assam, M
Multitask Cascaded Convolutional Othman, EA Attia Computation, 2022.
Networks," in IEEE Signal Processing [14] Face recognition using machine
Letters, vol. 23, no. 10, pp. 1499-1503, learning, Ishan Ratn Pandey, Mayank Raj,
Oct.2016,doi:10.1109/LSP.2016.260334. Kundan Kumar Sah , Tojo Mathew, M. S.
Padmini International Research Journal of
 Engineering and Technology Volume: 06 Indonesia, 26–27 September 2018; Volume
Issue: 04 | Apr 2019 170, p. 032110.
[15] A real-time framework for human [21] Aydin, I.; Othman, N.A. A new
face detection and recognition in cctv IoT combined face detection of people by
images R Ullah, H Hayat, AA Siddiqui, using computer vision for security
UA Siddiqui… - Mathematical …, 2022 application. In Proceedings of the 2017
[16] E. Nasiri, M. Milanova and A. International Artificial Intelligence and Data
Nasiri, "Video Surveillance Framework Processing Symposium (IDAP), Malatya,
Based on Real-Time Face Mask Detection Turkey, 16–17 September 2017; pp. 1–6.
and Recognition," 2021 International [22] X. Sun, P. Wu, and S. C. Hoi. Face
Conference on INnovations in Intelligent detection using deep learning: An improved
SysTems and Applications (INISTA), faster rcnn approach.
Kocaeli, Turkey, 2021, pp.1-7, arXiv:1701.08289, 2017.
doi:10.1109/INISTA52262.2021.9548475. [23] H. Qin, J. Yan, X. Li, and X. Hu.
[17] Design of face detection and Joint training of cascaded cnn for face
recognition system to monitor students detection. In CVPR, pages 3456–
during online examinations using Machine 3465, 2016.
Learning algorithms," M. Geetha, R. S.
Latha, S. K. Nivetha, S. Hariprasath, S.
Gowtham and C. S. Deepak, International
Conference on Computer Communication
and Informatics (ICCCI),2021, pp. 1-4,
doi: 10.1109
[18] Huang, Y.; Wang, Y.; Tai, Y.;
Liu,X.; Shen, P.; Li, S.; Li, J.; Huang, F.
Curricularface: Adaptive curriculum
learning loss for deep face recognition. In
Proceedings of the IEEE/CVF Conference
on Computer Vision and Pattern
Recognition, Seattle, WA, USA, 13–19 June
2020; pp. 5901–5910.
[19] Yousafzai, B.K.; Khan, S.A.;
Rahman, T.; Khan, I.; Ullah, I.; Ur Rehman,
A.; Baz, M.; Hamam, H.; Cheikhrouhou, O.
Student-performulator: Student academic
performance using hybrid deep neural
network. Sustainability 2021, 13, 9775.
[20] Wang, J.; Li, Z. Research on face
recognition based on CNN. In Proceedings
of the IOP Conference Series: Earth and
Environmental Science, Banda Aceh,