SANSKRITI

UNIVERSITY
REVIEW PAPER

ON

“FACE RECOGNITION WITH MASK”

SUBMITTED TO

DEPARTMENT OF COMPUTER SCIENCE

Submitted in the partial fulfilment for the award of degree of

“MASTER OF COMPUTER APPLICATIONS”

(BATCH – 2021-2023)

Under the Guidance of

(PROJECT GUIDE)

PROF. ANUPAM SHUKLA

PROF. EESHA GUPTA

Submitted by

Archit barua Enroll:(2102209002)

Shivani Enroll:(2102209028)

Shivani Singh Enroll:(2102209029)

 Abstract

Face recognition with masks has become a critical area of research due to the
COVID-19 pandemic. The widespread use of masks has made traditional facial
recognition techniques ineffective. Machine learning algorithms have been
developed to identify faces with masks. These algorithms use deep learning
techniques, such as convolutional neural networks, to detect features on the upper
half of the face that are still visible despite the mask. Transfer learning is also used
to improve accuracy in the identification of faces with masks. Several datasets have
been created to train and test these algorithms. This abstract provides an overview
of the state-of-the-art techniques used in face recognition with masks in machine
learning.

Face recognition technology has become a vital tool in various applications, from
security to user authentication. However, the ongoing COVID-19 pandemic has
made face recognition with masks a crucial challenge to tackle. Traditional face
recognition algorithms rely on analyzing the entire face, including the nose and
mouth regions, which are covered by masks, making recognition accuracy difficult
to achieve. This paper presents a review of the recent advances in face recognition
with masks, including the development of innovative approaches that combine
machine learning algorithms with deep neural networks to accurately recognize
individuals wearing masks. These approaches involve capturing and analyzing the
unique facial features that are still visible when a mask is worn, such as the eyes and
forehead, and using them to identify individuals. The paper also discusses the
potential applications of face recognition with masks in various fields, including
public safety, security, and user authentication, highlighting the importance of this
technology in the current pandemic situation.
 Introduction

Face recognition is a rapidly advancing field of research, with many practical

applications in security, healthcare, and social interactions. However, the emergence
of the COVID-19 pandemic and the widespread use of masks have posed significant
challenges for face recognition systems. Masks can obscure important facial features
that are used by face recognition systems to identify individuals, leading to reduced
accuracy and reliability of these systems. As a result, researchers have been
investigating various techniques to improve the performance of face recognition
systems when dealing with masked faces.

The goal of this review paper is to provide an overview of recent research on face
recognition with masks in machine learning. This paper will discuss the impact of
masks on face recognition systems and describe various techniques used to improve
their performance. These techniques include traditional feature-based methods, deep
learning-based methods, and hybrid models. The review will also highlight the
importance of using diverse datasets to train face recognition systems to recognize
faces with masks. Furthermore, the paper will discuss the challenges and limitations
of existing approaches and propose future research directions to address these
challenges.

Face recognition technology has become a vital tool in various applications, from
security to user authentication. However, the ongoing COVID-19 pandemic has
made face recognition with masks a crucial challenge to tackle. Traditional face
recognition algorithms rely on analyzing the entire face, including the nose and
mouth regions, which are covered by masks, making recognition accuracy difficult
to achieve. This paper presents a review of the recent advances in face recognition
with masks, including the development of innovative approaches that combine
machine learning algorithms with deep neural networks to accurately recognize
individuals wearing masks. These approaches involve capturing and analyzing the
unique facial features that are still visible when a mask is worn, such as the eyes and
forehead, and using them to identify individuals. The paper also discusses the
potential applications of face recognition with masks in various fields, including
public safety, security, and user authentication, highlighting the importance of this
technology in the current pandemic situation.
 Related Work

How many organizations rely on image processing and object recognition to identify
security issues and image classification, notably in health image processing .The
prime objective for this research is to create a face recognition algorithm that
combines real-time computer vision (CV2) with CNN (an advanced learning
algorithm) to recognition a person who is wearing a mask .Figure 1.1 shows that the
classification results are being shown independent of whether the individual is
wearing a mask or not.

Figure 1.1: Image Classification Steps

The proposed model’s main principle is that in the first phase, a random image
dataset that is easy to access online is collected and processed even-before being
uploaded to the server. Following the generation of the dataset, the next step in this
field is to include image pre-processing to recover elements and improve the model’s
accuracy. Subsequent to preprocessing, the model must be trained to predict the final
output, which can take some time dependent on the hardware setup. The second step
is to recognize real-time computer vision data using an CV2-built library and,
finally, to present the results to determine whether or not a person is wearing a mask,
as shown in figure 1.2.

Image processing and data augmentation Pre-processing data assists in the cleaning
of data, the extraction of important characteristics from a dataset, and the
improvement of model performance in classification With an augmentation
algorithm, image pre-processing helps in transmitting information, improving
images, and making new data with smaller image datasets. The size, contrast,
position, and orientation of a classified data obtained may vary. Image pre-
processing is therefore the systematic, algorithm-based step done to ensure that
images are correctly edited and ready for training and inference. The image dataset
is pre-processed using augmentation methods. It may randomly crop, color filter,
position, flip, and add noise to the training data using Image Generator and CV2
built-in functions to minimize model over fitting. This study uses other pre-
processing methods and augmentation to create a verified dataset of various types.
As shown in Figure 1.2, the system will send preprocessed images to any advanced
learning algorithms to extract features. Several data augmentation methods are used
to ensure that the dataset increases in size and quality. This assisted in reducing over
fitting issues and improving the model’s overview ability during training. Table 4.1
shows the image augmentation settings that were employed. So, first and primarily,
the data is resized by a value of 1/ 255. This aids in the normalization process.
Following that, a zoom range of 20% is used to allow for some random cropping of
the image. The rotation range is set to 40 degrees, and the horizontal flip is enabled.
 Literature Survey

2.1 Overview

The world is facing a huge health crisis due to the rapid transmission of the
coronavirus (COVID-19). According to WHO, the most effective preventive measure
against COVID-19 is wearing a mask in public places. It is very difficult to manually
monitor people in these areas. So in this paper, we propose a deep learning model
that detects people who are not wearing a mask, and a transfer learning model is
proposed to automate the process of identifying people who are not wearing a mask.
The proposed model is built by fine-tuning the pre-trained, state-of-the-art deep
learning model, InceptionV3. The proposed model is trained and tested on the Simulated
Masked Face Dataset (SMFD). An initiative was started by the French government to
identify passengers who are not wearing masks in public places with the help of artificial
intelligence (AI) techniques like machine learning (ML) and deep learning (DL), the
Internet of Things (IOT) and big data. These are being used for the faster diagnosis of
COVID-19.[2.]

Bosheng Qin and Dongxiao Li.[3.] designed a face mask identification method using
the SRCNet classification network and achieved an accuracy of 98.7 percent in
classifying the images into three categories, namely "correct face mask wearing,"
"incorrect face mask wearing," and "no face mask wearing."

In 2020, the COVID-19 pandemic became a major global issue. It is currently

impossible to determine when the COVID-19 pandemic will be fully contained.
Vaccines are being tested, but their efficacy cannot be assured. Following health
protocols, such as wearing face masks, is still the most effective strategy to avoid
transmission. In public places, people are required to wear a mask. However, many
people still continue to ignore health guidelines. Unmasked people can be detected and
alerted using an ai algorithm. In many cases, the k-Nearest-Neighbours (KNN) method
is a non-parametric classification method that is simple but successful. A residential
area is created when a data record’s k nearest neighbours are extracted in order to
classify it. The classification is normally decided by simple majority among the data
records in the neighbourhood, with or without use of distance-based weighting.
However, in order to use KNN, we should choose a proper value for k, and the
classification success is highly dependent on this number. In some ways, the KNN
approach is influenced by k. There are other methods for determining the k value,
but one of the most simple is to run the algorithm many times with different k values
and select the one that performed the best.[4.]

SVM is a powerful classification technique. Its goal is to establish a decision

boundary between two classes that allow for label prediction using one or more relevant
features. The hyperplane, or decision boundary, is oriented so that it is as far away from
the nearest data points from each of the classes as possible. Support vectors are the spots
that are the closest together. The SVM algorithm was first proposed as a way to
implement a linear classifier. The kernel method, which allows us to model higher-
dimensional, non-linear models, is an- other use for SVM. A kernel function could be
used to add additional dimensions to raw data in a non-linear problem, making it a
linear problem in the resulting higher-dimensional space. In a summary, a kernel
function can help in the speed up of certain calculations that would otherwise require
computations in high-dimensional space.[5.]

K-nearest neighbours (KNN), support vector machine (SVM), and deep learning
are examples of classification methods. A popular DL algorithm for classification
tasks is the convolutional neural network (CNN). Because it is created of neurons that
can learn, CNN is equal to a convolutional network. Each neuron will receive data and
perform tasks, and they will all be connected in a network. Convolutional layers,
pooling layers, and fully connected layers are also the three stages of CNN’s operation.
CNN is designed to work with information that has a two-dimensional shape. Each
layer of the network is made up of various two-dimensional planes, each of which
includes multiple neurons. Neurons in two adjacent layers will connect with others.
CNN is the expert in image recognition.[6]
Face recognition using KNN, SVM, and CNN has been the subject of much research.
On a low-power processor, a reference used KNN for face recognition and achieved 91.5
%. For face recognition, the reference used SVM and achieved accuracy of 90.82 %. In
the YALE and ORL datasets, Reference used SVM for face recognition and achieved 96
% and 96.5 % accuracy, respectively. Reference employed CNN to obtain 99.5 % on the
AR database and 85.13 % accuracy on the FERET database for face recognition with
pose and lighting variation. SVM, KNN, and CNN are effective algorithms for
classifying face photos, according to previous research. As a result, we examine KNN,
SVM, and CNN in the same dataset for face mask detection. We expect similar
results.[7.]

The goal of this study is to examine and compare the performance of algorithms such
as SVM and KNN with a deep learning algorithm such as CNN. Pixel value is a feature
that is used to classify face masks. Accuracy, precision, recall, F1 score, and execution
time are all factors to consider. Researchers should be able to use the data to choose the
best and most suitable algorithm to identify face masks. Md. Sabbir Ejaz et al.[8.]
implemented the Principal Component Analysis (PCA) algorithm for masked and
unmasked facial recognition. The main aim of this work is to identify people who are
not wearing masks in public places to prevent the further transmission of COVID-19.
As per the research, wearing a mask can prevent and control the transmission of
COVID-19. Therefore, the governments of various countries have made it mandatory
to wear masks in public places and crowded areas. The same work can further be
improved by employing large volumes of data and can also be extended to classify the
type of mask and implement a facial recognition system, deployed at various
workplaces to support person identification while wearing the mask.

Roshan et. a l [9.] His paper proposes a python-based GUI tool for face mask
detection. Convolution Neural Network (CNN), PCA, and HAAR Cascade
algorithms are used in the development of the tool. Keras is the library used to create
the CNN architecture. To extract the image contents, PCA features are used. The
processing is first, the image is given to the model as input. Now the image has gone
through the preprocessing stage. At this stage, the algorithm detects the mask and eyes.
Face recognition will be done with the help of the HAAR algorithm in the next stage.
From this, the features are extracted and they move to the classification stage. In this
stage, principal component analysis (PCA) classifies the input images with various
factors. PCA has been trained with images. The next stage is the training stage. At
this stage, with the help of CNN, the qualified model from the feature extraction will
be saved to the prediction section. In the final stage, that is the prediction stage, the
prediction is By using CNN, the accuracy of the model is high. For the implementation
of the CNN algorithm, the Python Keras library is used. This paper also describes
the modules and the system architecture of the proposed model.

The COVID-19 virus spreads globally in 2020, and it can be spread through contact.
So, Shuang et. [ 10.] as mentioned in this paper, identity authentication systems based
on passwords or fingerprints are unsafe. Using facial biometrics seems a safe choice,
but most people are wearing masks. That’s why the performance of many algorithms
related to faces is challenged. This paper is to study the head pose classification problem,
which can help improve the performance of many algorithms related to faces. And this
research also has great value for many applications, such as human attention monitoring,
driver fatigue monitoring, and interaction-based face liveness detection. So, to solve
this issue, we use a method called HGL to deal with head pose classification by adopting
colour texture analysis of images and line portraits. This method uses the color
texture analysis of images and line portraits. This experiment was carried out on the
MAFA dataset because most face datasets with masks do not have head pose
information. The training data provided by the MAFA dataset is not sufficient, so
it is important for our experiment to expand the data. First, we use the information
about facial location provided by the MAFA data set to crop the image. In this
process, we increase the size of the facial frame to 1.2, 1.4, and 1.6 times the original
size for cropping, and normalize these facial images to 8080. Then, we extract the H-
channel in the HSV color space of the image and normalize the pixel values of the
image in the range of 0-255. by extracting and processing the pixel information of
the H channel in the HSV color space to distinguish the face and mask in the image.
The results show that the method proposed in this paper has better performance than
other algorithms in both front accuracy and side accuracy. For the front accuracy, this
algorithm is 0.8% better than the best accuracy of other algorithms. And this
algorithm is 2.28% higher than the best accuracy of other algorithms for side accuracy.
The reason for this phenomenon is that the images processed through the H-channel can
provide focused information to the net. The evaluation of the MAFA data set shows
that our method is superior to the other methods listed. The solution of this problem
also provides help for the study of multi-angle problems. In practical applications, we
can use a face recognition algorithm that distinguishes between wearing a mask and not
wearing a mask. If the facial image has masks, we can use the method proposed in the
paper for classification, and if it is a normal facial image, we can use FSA-Net. The line
portrait algorithms can also be used.

The recent COVID-19 epidemic has increased the amount of hygiene and contactless
identity confirmation. However, this epidemic has led to the widespread use of face
masks, which are important to maintaining this epidemic under control. The effect
of wearing a mask on facial recognition in the interaction area is a sensitive issue at
the moment but unread. Naser et.al.[11.] In his paper, he says that facial recognition
has become as important in our daily lives as it is simple and intangible how to ensure
accurate identity. Procedures such as authentication in the default border control of
gates or secure access to electrical equipment are increasingly dependent on such
technology. Face occlusion has been repeatedly addressed in the scope of face recognition
solutions. Moreover, developing occlusion-invariant face recognition solutions has been
a growing research challenge. How- ever, most of these works address general occlusion
that commonly appears in in-the-wild capture conditions, such as sunglasses and partial
captures. Our work aims at studying this effect to enable the future development of
solutions addressing accurate face recognition in such scenarios. To provide a wide view
of the effect of wearing a mask on face recognition performance, we analyzed the
performance of three face recognition algorithms. Two of these algorithms are among
the top performing academic approaches, namely the Arc Face and Sphere Fac. The
third algorithm is a COTS algorithm from the vendor Neurotechnology. In the
following, this section provides more details on these algorithms. I chose Sphere Face
as it achieved competitive verification accuracy on Labeled Face in the Wild (LFW)
of 99.42% and YouTube Faces (YTF) of 95.0% using 64-CNN layers trained on the
CASIA-Web Face dataset . However, Arc Face achieved state-of-the-art performance
on several face recognition benchmarks, such as LFW 99.83% and YTF 98.02%. In the
recent NIST report addressing the performance of vendor-face verification products,
Neurotechnogy achieved one of the best performances in the recent NIST report.
Matching the faces by COTS produces almost similar scores. For the Arc Face and
Sphere Face, the Multi-task Cascaded Convolutional Networks (MTCNN)solution is
used, as recommended in, to detect(crop) and align (affine transformation) the face.
Both networks process the input aligned and cropped image and produce a feature
vector of the size 512. To compare two faces, a distance is calculated between their
respective feature vectors. This is calculated as the Euclidean distance for arc-face
features, as recommended in, and as the cosine distance for sphere-face features, as
recommended in. Our analyses pointed out the significant effect of wearing a mask
on comparison scores and separability between genuine and imposter comparisons in
all the investigated systems. Moreover, we point out a large drop in the verification
performance of the academic face recognition solutions.

Lingxue et. [12.] It proposes two models, PDSN and FDM, based on the CNN
algorithm for face mask recognition. With this, a three-stage model is proposed. In
the initial stage, PDSN is used to capture the correspondence between the feature
elements and facial blocks. On the second stage, they create a dictionary which includes
the mask details that they learned from the generator. And in the final stage, using
FDM, perform the recognition. With the help of various mathematical formulas, train
the CNN algorithm to perform the learning process for the mask generators. In the next
stage, establishing the mask dictionary And finally, recognition. In the evaluation of the
proposed model, providing a decent output It is able to provide a maximum of 98.26%
accuracy in its algorithms. The analysis is conducted with various other data sets and
it is able to perform with the same accuracy all over the datasets. As a result, it can be
used as a good proposal.

We all are aware about the outbreak of Coronavirus Disease (COVID-19),Chavda et.
al.[13.] states As the world recovers from the pandemic and plans to return to a state of
normalcy, there is a wave of anxiety among all individuals, especially those who intend
to resume in- person activity. Studies have proved that wearing a face mask significantly
reduces the risk of viral transmission as well as provides a sense of protection. But it is
not possible to track the implementation of this policy manually. We introduce a
Deep Learning-based system that can detect instances where face masks are not used
properly. Our system consists of a dual-stage Convolutional Neural Network (CNN)
architecture capable of detecting masked and unmasked faces and can be integrated with
pre-installed CCTV cameras. We all know that technologies such as Machine Learning
and Artificial Intelligence have made our lives easier and provided solutions to a
variety of complex problems. Modern computer vision algorithms are approaching
human-level performance in visual perception tasks. In this paper, we propose a two-
stage CNN architecture where the first stage detects human faces, while the second stage
uses a lightweight image classifier to classify the faces detected in the first stage as either
‘Mask’or ‘No Mask’faces and draws bounding boxes around them along with the detected
class name. We are working with CNN in two categories: Multi-Stage Detectors and
Single-Stage Detectors. The face mask detection methods are further divided into three
methods: correct face mask wearing, incorrect face mask wearing, and no face mask
wearing. The system takes an image, detects and crops faces, and then uses SRCNe to
perform image super-resolution and classify them. Our architecture consists of two major
stages. The first stage of our architecture includes a Face Detector, and the second stage
of our architecture is a CNN based Face Mask Classifier. The results from the second
stage are decoded, and the final output is the image with all the faces shown. We also
use Dlib Deep Learning Face Detector, MTCNN-It uses a cascade architecture with
three stages, and Retina Face-It is a staggering design with pixel-wise localization that
uses a multi-task learning strategy to simultaneously predict face box. The second stage
of our system is a face mask classifier. This stage takes the processed ROI from the
Intermediate Processing Block and classifies it as either Mask or No Mask.

P. Gupta et. al.[14.] presented research on facial recognition that provides a novel
approach to using a deep neural network (another type of deep network). Rather than
sup- plying raw pixel values as input, this proposed approach simply provides the
derived facial features. The Haar cascade is used to extract facial features, which are
then fed instead of raw pixel data. The complexity of the neural network-based
recognition framework has been reduced as the number of redundant input features has
been reduced. By employing DNN instead of a convolutional network, the procedure
is also lighter and faster. The accuracy of the framework is not guaranteed by the given
method, as the average accuracy obtained is 97.05%.
For the face recognition system, K. J. Bhojane et. al.[15] used the embedded face
recognition and face tracking system methods available in MATLAB and the
Raspberry Pi B. It uses the Haarlike function, which was utilised to recognise and
authenticate the authenticated user’s face, to establish a safe environment for ignition
and access to the car. In a secure setting, the face of a person is the most important
part of the car’s ignition. Future work will be informed by the findings of the gesture
identification and control studies.

M. Rahman et. [16.] released a paper. aiming at establishing a system for recognising
whether or not a person wears a mask and It informs the appropriate authorities in the
smart city network. It captures facial photos by filming them in real time at numerous
public spots throughout the city. Masked faces are being identified using facial pictures
de- rived from this movie. After extracting features from images Using the
convolutional neural network (CNN) learning technique, the features are learned
through several hidden layers. When the architecture recognises people who aren’t
wearing a mask, the information is sent to the proper authorities via the municipal
network so that they can take the right action. Based on data gathered from a variety
of sources, the proposed method assessed promising findings. They also lay out a
framework for enforcing basic health guidelines in this pandemic situation in these
materials.

The goal of this review paper is to summarise and assess numerous well-known
techniques in the multiple steps of a pattern recognition system, as well as to
highlight the analytical and application techniques that are at the forefront of this
exciting and complicated topic. Closer machine learning methodologies are used in the
literature to approach pattern recognition frameworks. Pattern recognition techniques
are required for applications such as data processing, retrieval of multimedia
information, internet searching, face recognition, and cursive handwriting recognition.
Information gathering, segmentation, and feature ex- traction are just a few of the
obstacles that come with fully realising this. [17]
Vinitha et. al.[18.] released an article in which they suggested a method that em-
ploys a deep learning algorithm and computer vision to recognise a person with a
masked face in an image or video stream. Tensor Flow, Open CV, Keras, and PyTorch
are some of the libraries that have been used. In two stages, the project will be
implemented. The first phase involves training a deep learning model, which is followed
by the application of a mask detector to a live image or video stream in the second phase.
Open CV is the framework that was utilised to detect faces in real time from a live
webcam feed. A COVID-19 face mask detector has been built using Python and
computer vision using a data set.

The suggested system seeks to recognise the masked and faces are depicted using the
advanced YOLOv3 architecture, according to R. Bhuiyan et. al.[19] The learning
algorithm Convolution Neural Network is used by YOLO (You Only Look Once)
(CNN).YOLO connects to CNN via hidden layers, research, and quick algorithm
retrieval, and can recognise and locate any image. The process starts with the input
of 30 distinct photos from the dataset into the model, which is then combined to
produce action-level predictions. It produces outstanding imaging and detecting
performance. This model is tested on a live video to see if the model’s fps rate and
detection performance with masked and unmasked two layers are acceptable. With an
average frame rate of 17, our model produces amazing video inside. This system is more
efficient and quicker than methods that use their own data sources.

T. Meenpal et. al.[20.] proposed a semantic segmentation that categorises each pixel in
an image into face and non-face. It constructs a binary classifier quickly and then
recognises the fragments as chunks. From RGB photos, including localised objects, you
can generate accurate face masks for human objects. The findings of the Multi Human
Parsing were shown by the author. At the pixel level, this dataset has good average
accuracy. Further more, the issue of incorrect predictions is addressed, and the correct
bounding box is constructed around the segmented area. The suggested network is capable
of detecting both non-frontal and numerous faces in a single image. Complex tasks, such
as detecting parts of the face, can be accomplished using this technology.
Face recognition is performed by the paper. The face recognition network detects the
face and then estimates the pose. To process these regions for face identification and
posture estimation, a region proposal network is used to discover probable regions that
may include a face, and a single deep multitask CNN is used to process them. Because
the network is not capable of achieving real-time performance, it must be improved.
Face recognition is tested using the FDDB benchmark. Pose estimation is based on the
AFW benchmark. The head position can be used to determine a person’s motive, intent,
and gaze. According to recent research, convolutional neural networks (CNNs) have
made a series of advancements in the tasks of face recognition and position estimation.
Two CNN frameworks exist for performing these two integrated objectives at the same
time. One method is to first recognize faces using a face recognition network, and then
estimate each face’s pose using a pose estimation network. The other option is to utilise
a single deep multi-task CNN to analyse these regions for simultaneous face recognition
and posture estimation after using a region proposal technique to produce several
candidate regions that may contain faces. Because the two networks are different, the
performance of posture estimation is influenced by the performance of the face
recognition network. The latter generates a high number of candidate regions, which
takes CNN a great deal of time and resources to analyze.[21]

In an unlimited environment, partial facial recognition (PFR) is a critical task,

especially in settings like video surveillance and mobile devices, when partial facial
photos are likely to be caught due to occlusions, out-of-view, and a wide viewing angle.
However, PFR has received little attention to date, leaving the problem of recognising
an arbitrary portion of a face image mostly unsolved. This research provides a dynamic
feature matching (DFM) strategy for partial face recognition. It uses a collection
of fully convolutional networks and sparse representation classification (SRC) to solve
the partial face recognition problem regardless of face size. The position of partial faces
versus a holistic face is not required for DFM.[22.]

Researchers in psychology, computer science, linguistics, neurology, and allied

disciplines are increasingly interested in developing a human-computer interface system
for automatic face recognition or facial expression recognition. A Facial Expression
Recognition System (AFERS) is proposed in this research. Face recognition , feature
extraction, and facial expression identification are the three steps in the proposed
method. The initial phase of face recognition entails detecting skin colour using the
YCbCr colour model, lighting adjustment for uniformity on the face, and
morphological operations to maintain the required face area. Using the AAM (Active
Appearance Model) approach, the first phase’s output is utilised to extract facial
features such as eyes, nose, and mouth. The third stage uses the simple Euclidean
distance approach to recognise facial expressions automatically. The Euclidean distance
between feature points in the training and query images is compared in this method.
The output picture expression is determined by the minimum Euclidean distance.
This approach has a true recognition rate of about 90% to 95%. An Artificial Neuro-
Fuzzy Inference System is used to further improve this approach (ANFIS). In
comparison to other approaches, this non-linear recognition system achieves a
recognition rate of roughly 100%.[23]

Facial expression recognition is presented in this study. Happy, surprised, fear, disgust,
sad- ness, anger, and a neutral state are the seven primary expressions that are identified
and classified. There are three major aspects to this system. To extract the face-centered
region, the initial step is to recognise the face and facial features. The next step is to
normalise and extract the edges of this interest zone. We have a face edge image at this
point, which we utilise to calculate the Euclidean distance between all pixels that form
edges. The SVM approach is used in the third stage to classify various emotional states.
Face Alignment for a Single Image is addressed in this study. Researchers show how an
ensemble of regression trees can be used to directly estimate the face’s landmark positions
from a sparse subset of pixel intensities, leading to super- real time performance and
high-quality predictions.[24.]

Using deep convolutional neural networks, this article presents a method for
automatically detecting faces, locating landmarks, predicting poses, and identifying
gender (CNN). Hyper is a suggested approach that uses a separate CNN to fuse the
intermediate layers of a deep network, then uses a multi-task learning algorithm to
operate on the fused features. It takes advantage of task synergy to improve individual
results.[25.]

Using a single deep convolutional neural network, we present a multi-purpose

algorithm for face recognition , facing the challenge, pose estimation, gender
recognition, smile detection, age estimation, and face recognition (CNN). The proposed
method uses a multitask learning framework to normalise shared CNN parameters and
facilitate collaboration across domains and tasks.[26.]

In this paper, researchers proposed a dynamic neural network for human face
recognition. That outperforms other algorithms. A self-organizing map (SOM) neural
network, A convolutional neural network is used in the system. The SOM analysis shows
image samples into a topological space for which inputs that are near one another in the
original space are also near one another in the output space, which actually results in
dimensionality. reduction and invariance to minor changes in the image sample, although
the convolutional neural net- work provides partial invariance to translation, rotation,
scale, and deformation.[27.]

The standard face recognition pipeline consists of four stages: detect = > align
=>represent = > classify in modern face recognition To use explicit 3D face models to
apply a piece-wise affine transformation and extract a feature representation from a
nine-layer deep neural network, the paper revisits both the alignment and the
representation phases.[28.]

Due to various poses, illumination variations, and partial occlusion, face recognition
and matching in an unconstrained environment is a challenge. In this paper, we describe
a deep- casted multitask framework that increases productivity by using the intrinsic
association between detection and alignment. To identify the position of faces and
objects in a coarse- to-fine way, the framework uses a cascading approach with three
different levels of specially designed deep convolutional networks.[29.]

In tasks such as face reconstruction, recognition, and verification, key-point detection

is one of the most important preprocessing phases. We propose an adaptive approach
for key-point estimation and pose analysis of defined faces by learning efficient H-
CNN linear regression (KEPLER) to solve the face alignment problem in this research.
The paper introduces H-CNN (Heatmap-CNN), a unique architecture that captures
structured global and regional features and allows for reliable keypoint recognition.
The visibility, fiducials, and 3D-pose of the face are now all given to H-CNN in a
group situation.[30.]

The detection of pedestrians is a subject with a lot of practical implications. We present

state-of-the-art and competitive findings on all major pedestrian datasets with a
convolutional network model, adding to the list of successful applications of deep
learning methods to vision. Multi-stage features, connections that skip layers to
integrate global shape in- formation with local distinctive motif information, and an
unsupervised method based on convolutional sparse coding to pre-train the filters at
each stage are just a few of the new features included in the model.[31.]

There has been a lot of discussion about sparse representation. The study of picture super-
resolution (SR) yielded significant results. In order to improve SR results, deep-
learning based SR approaches have also been proposed in the literature. For rapid and
high-quality image SR, the study employs a number of decision tree techniques. Our
proposed SR using the decision tree (SRDT) method employs a divide-and-conquer
strategy, in which an input low-resolution (LR) patch is classified into one of the leaf
nodes using a few simple binary tests, and then this LR patch is directly multiplied
with the regression model at that leaf node for regression. The classification and
regression processes both require very little time to complete.[32.]
 RESULTS

Adam Optimizer with a learning rate of 0.15 is employed in the proposed model.
The model is trained over 15 epochs using all of the training images. Some data in
each epoch was lost during model training, namely accuracy, validation loss, and
validation accuracy. As a result, the decrease and accuracy are indicators of training
progress. It made a prediction about the classification of the training data and then
matched it to the new label. While accuracy is specified as the percentage of right
predictions, validation accuracy is defined as the measurement using data that has
not been used in training. Following training, graphs are displayed to show the
training accuracy, one for a loss and the other for accuracy. The overall training
accuracy is 1.01, while the validation accuracy is 0.47, with such a training loss of
0.73 as well as a validation loss of 0.463. The observed classification accuracy is
0.90, while the test loss is 0.15. Figures 5.1-5.3 show the training effectiveness.
 CONCLUSION

To conclude, as a result of the global pandemic caused by COVID-19, many

governments are using a variety of approaches and innovations to fight the virus.
This category contains defensive systems such as a face mask. The primary objective
of this study is to determine (classify) whether or not a person is wearing a face mask
with human identity while working on the corona virus face mask detection and
classification algorithm CV2. This study used the Convolutional Neural Network
deep learning categorization algorithm CV2.
 References

[1.] Md Rafiuzzaman Bhuiyan, Sharun Akter Khushbu, and Md Sanzidul Islam. A deep

learning based assistive system to classify covid-19 face mask for human safety with

yolov3. In 2020 11th International Conference on Computing, Communication

and Networking Technologies (ICCCNT) , pages 1–5. IEEE, 2020.

[2.] G Jignesh Chowdary, Narinder Singh Punn, Sanjay Kumar Sonbhadra, and

Sonali Agarwal. Face mask detection using transfer learning of inceptionv3 In

International Conference on Big Data Analytics , pages 81–90. Springer,2020

[3.] Bosheng Qin and Dongxiao Li. Identifying facemask-wearing condition using

image super-resolution with classification network to prevent covid-19.

Sensors,20(18):5236, 2020.

[4.] Mohammad Farid Naufal, Selvia Ferdiana Kusuma, Zefanya Ardya

Prayuska, Ang Alexander Yoshua, Yohanes Albert Lauwoto, Nicky

Setyawan Dinata, and David Sugiarto. Comparative analysis of image

classification algorithms for face mask detection. Journal of Information

Systems Engineering and Business Intelligence, 7(1):56–66, 2021.

[5.] Abdellah Oumina, Noureddine El Makhfi, and Mustapha Hamdi. Control

 the covid-19 pandemic: Face mask detection using transfer learning. In

2020 IEEE 2nd International Conference on Electronics, Control,

Optimization and Computer Science (ICECOCS), pages 1–5. IEEE, 2020.

[6.] Anushree More, Harita Tilloo, Pradnya Kamble, Barkha Jagtap, and SN

Bhadane. Face mask detection using knn algorithm.

[7.] Gargi Kale, Prashant Bhaware, Rohit Ingle, Sayali Sulbhewar, Yash

Gugaliya, Mayur Kaware, and Parag Thakare. Real time face mask

detection-a survey. International Journal of Information Technology &

Computer Engineering (IJITC) ISSN: 2455-5290, 2(01):1–4, 2022.

[8.] G Jignesh Chowdary, Narinder Singh Punn, Sanjay Kumar Sonbhadra, and

Sonali Agarwal. Face mask detection using transfer learning of inceptionv3. In

International Conference on Big Data Analytics , pages 81–90. Springer, 2020.

[9.] Gargi Kale, Prashant Bhaware, Rohit Ingle, Sayali Sulbhewar, Yash

Gugaliya, Mayur Kaware, and Parag Thakare. Real time face mask

detection-a survey. International Journal of Information Technology &

Computer Engineering (IJITC) ISSN: 2455-5290, 2(01):1–4, 2022.

[10.] Shuang Li, Xin Ning, Lina Yu, Liping Zhang, Xiaoli Dong, Yuan Shi, and

Wei He. Multi-angle head pose classification when wearing the mask for face
 recognition under the covid-19 coronavirus epidemic. In 2020 international

conference on high performance big data and intelligent systems (HPBD&IS),

pages 1–5. IEEE,2020.

[11.] Naser Damer, Jonas Henry Grebe, Cong Chen, Fadi Boutros, Florian Kirch-

buchner, and Arjan Kuijper. The effect of wearing a mask on face recognition

performance: an exploratory study. In 2020 International Conference of the

Biometrics Special Interest Group (BIOSIG) , pages 1–6. IEEE, 2020.

[12.] Lingxue Song, Dihong Gong, Zhifeng Li, Changsong Liu, and Wei Liu.

Occlusion robust face recognition based on mask learning with pairwise

differential siamese network. In Proceedings of the IEEE/CVF International

Conference on Computer Vision, pages 773–782, 2019.

[13.] Amit Chavda, Jason Dsouza, Sumeet Badgujar, and Ankit Damani. Multi-

stage cnn architecture for face mask detection. In 2021 6th International Con-
ference

for Convergence in Technology (I2CT) , pages 1–8. IEEE, 2021.

[14.] Priya Gupta, Nidhi Saxena, Meetika Sharma, and Jagriti Tripathi. Deep neural

network for human face recognition. International Journal of Engineering and

Manufacturing (IJEM), 8(1):63–71, 2018.

[15.] Ketan J Bhojane and SS Thorat. Face recognition based car ignition and
 security system. International Research Journal of Engineering and Technology

(IRJET), 5(05):3565–3668, 2018.

[16.] Mohammad Marufur Rahman, Md Motaleb Hossen Manik, Md Milon Islam,

Saifuddin Mahmud, and Jong-Hoon Kim. An automated system to limit

covid19 using facial mask detection in smart city network. In 2020 IEEE

International IOT, Electronics and Mechatronics Conference

(IEMTRONICS), pages 1–5. IEEE, 2020.

[17.] Devyani Bhamare and Poonam Suryawanshi. Review on reliable pattern recog-
nition

with machine learning techniques. Fuzzy Information and Engineering, 10(3):362

377,

2018.

[18.] Pierre Sermanet, Koray Kavukcuoglu, Soumith Chintala, and Yann LeCun.

Pedestrian detection with unsupervised multi-stage feature learning. In

Proceedings of the IEEE conference on computer vision and pattern

recognition, pages 3626–3633, 2013..

[19.] Safa Teboulbi, Seifeddine Messaoud, Mohamed Ali Hajjaji, and

Abdellatif Mtibaa. Real-time implementation of ai-based face mask

detection and social distancing measuring system for covid-19 prevention.

 Scientific Programming, 2021.

[20.] Toshanlal Meenpal, Ashutosh Balakrishnan, and Amit Verma. Facial mask

detection using semantic segmentation. In 2019 4th International Conference

on Computing, Communications and Security (ICCCS), pages 1–5. IEEE,

2019.

[21.] Hao Wu, Ke Zhang, and Guohui Tian. Simultaneous face detection and

pose estimation using convolutional neural network cascade. IEEE Access,

6:49563– 49575, 2018.

[22.] Lingxiao He, Haiqing Li, Qi Zhang, and Zhenan Sun. Dynamic feature match-

ing for partial face recognition. IEEE Transactions on Image Processing , 2

8(2):791–802, 2018.

[23.] Anagha S Dhavalikar and RK Kulkarni. Face detection and facial expres-
sion

recognition system. In 2014 International Conference on Electronics and

Communication Systems (ICECS), pages 1–7. IEEE, 2014.

[24.] Sonia Kherchaoui and Amrane Houacine. Facial expression identification

using gradient local phase. Multimedia Tools and Applications,

78(12):1684316859, 2019.
. [25.] Steve Lawrence, C Lee Giles, Ah Chung Tsoi, and Andrew D Back. Face

recognition: A convolutional neural-network approach. IEEE transactions on

neural networks, 8(1):98–113, 1997.

[26.] Yaniv Taigman, Ming Yang, Marc’Aurelio Ranzato, and Lior Wolf.

Deepface: Closing the gap to human-level performance in face verification. In

Proceedings of the IEEE conference on computer vision and pattern

recognition, pages 1701– 1708, 2014.

[27] Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, and Yu Qiao. Joint face

detection and alignment using multitask cascaded convolutional networks.

IEEE signal processing letters, 23(10):1499–1503, 2016.

[28.] Amit Kumar, Azadeh Alavi, and Rama Chellappa. Kepler: Keypoint and pose

estimation of unconstrained faces by learning efficient h-cnn regressors. In 2017

12th

ieee international conference on automatic face & gesture recognition

[29.] R Suganthalakshmi, A Hafeeza, P Abinaya, and A Ganga Devi. Covid-19

facemask detection with deep learning and computer vision. Int. J. Eng. Res.

Tech.(IJERT) ICRADL, 2021.

[30.] Hao Wu, Ke Zhang, and Guohui Tian. Simultaneous face detection and pose

estimation using convolutional neural network cascade. IEEE Access,

6:49563– 49575, 2018.

[31.] Rajeev Ranjan, Vishal M Patel, and Rama Chellappa. Hyperface: A deep

multi-task learning framework for face detection, landmark localization, pose

estimation, and gender recognition. IEEE transactions on pattern analysis and

machine intelligence, 41(1):121–135, 2017.

[32.] Rajeev Ranjan, Swami Sankaranarayanan, Carlos D Castillo, and Rama

Chellappa. An all-in-one convolutional neural network for face analysis. In

2017 12th IEEE international conference on automatic face & gesture

recognition (FG 2017), pages 17–24. IEEE, 2017.