Volume 6, Issue 7, July – 2021 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Face Mask Detection

Aman Goyal, Abhinav Ararwal, Shunham Kumar Pandey, Ravi Shanker Pathak, Kunal Yadav
Galgotias College of Engnieering & Technology

Abstract:- The World Health Organization (WHO) PyTorch for implementation. This method first teaches deep
stated that the COVID-19 virus can be spread in two learning, which must be MobileNetV2, and then uses it to
ways: droplet infection and physical contact, so we want detect image/video channel masks.
to take some precautions to prevent the spread of the
virus. In addition to wearing masks, they are also II. LITERATURE REVIEW
socially estranged. Wearing a mask is considered an
important aspect of the spread of the COVID-19 virus, The facial recognition system is effective in a variety
because droplets may fall on any surface between these of areas. One of several specific areas is Presentation Attack
two preventive measures. People who wear masks or not Detection for recognizing this face, which was featured by
are more important. Here, we provide a mask with a the authors in this report. They used silicon, which is 3D
detection system that can detect any type of mask and masks for real people. You will have used the database, it is
masks of various shapes through the film channel in certainly 8 new and custom 3D silicone masks with the real
accordance with the guidelines of the federal presentation. Here in this report, the authors have proposed a
government. It uses deep learning algorithms, and the mask detection system for continuous health workers in the
Python PyTorch library can be used to identify masks operating room. Wellness care requires a mask to be used
based on image/video sequences. The recommended privately in the operating room, and the recommended
system can detect mask-wearing people wearing masks. system warns almost anyone who is not wearing the mask.
There are two detection systems for the face and the mask,
I. INTRODUCTION which is health. His system got practically 90% recovery
and less than 5% false rate, which is good. You have
The COVID19 pandemic started in December 2019, generally worked for the mask, which is certainly healthy
and this virus will definitely start in China. Then the virus for images taken by digital cameras from a distance of 5 m
began to spread all over the world, actually in this world [8]. The writers have worked on this report since the film's
country Atlanta [1]. Staying at home, relaxing alone, masked face recognition. Anyone who is masked will be
avoiding traffic, and not polluting the city, these measures recognized in this presented method and there are mainly 4
can prevent the spread of Covid 19 disease [2]. The WHO steps to recognition, namely estimating the length between
usually cited the two main reasons for the spread of this the camera and the person, recognizing the attention area,
virus, because the inhalation and contact of droplets must be recognizing part of the face and recognizing attention .
physical [3]. When a person is infected or coughs, human various video surveillance systems and achieved good
respiratory droplets can reach many other contacts (within 1 reliability [9]. It presents a design based on the paper for
m), these droplets can spread into the atmosphere, and even masked face detection when it comes to protection. Three
bring many surfaces closer together. This virus is still layers of CNN for masked face detection. In addition, the
present in almost all areas, which may also lead to contact authors worked on the data set FACE ENMÁSCARA,
transmission. The preventive measures formulated by the which they developed themselves, since the existing face is
national regulatory agencies in each country are usually certainly masked, it is not enough to evaluate the algorithms
equivalent to social distancing, which is why masks are used much more precisely. it worked well and masked face
[4]. People usually use many types of masks, such as detection is also performed accurately [10]. Another
medical masks. , Procedural masks, and masks of various operation in this industry is in relation to masked face
shapes (glass shapes) of some fashion designers [5]. These detection and then detection in conjunction with the initial
masks are used to prevent the spread of respiratory droplets face. The proposed work will be carried out in two phases,
from infected persons. This mask can also provide sufficient the first is to identify the masks that address a larger part of
air permeability, liquid permeability and a high degree of the face than necessary and the second to obtain the real face
purification. Social distance is a trend in which the distance that is not present in the teaching record. They received a
between objects and people is at least 2 meters.It should be used GAN-based algorithm and a used data set, which is a
noted that the role of the Internet of Things, synthetic CelebA training and has achieved higher precision [11].
intelligence, blockchain, drones, etc. in the detection of Here, the authors continue to work continuously on the
Covid19 disease is likely to be extremely important [6]. subject of masks. We worked on face recognition through
Therefore, in our recommended strategy, we will create a faces that can be covered with the help of masks. The
mask. This recognition method can of course recognize authors found out that the photos are covered with masks
many different masks and different designs. Our strategy from the film, then the original photos of this surely masked
allows you to use the computer view to identify the mask of face are recognized. That is why they suggested CNN
the person used for this deep learning algorithm, possibly multitasking, this is certainly cascading, as well as in the
not through image and video sources. Use the Python library category The recommended system has the capacity. to

IJISRT21JUL1046 www.ijisrt.com 1402

Volume 6, Issue 7, July – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
recognize faces masked from the original faces of people pre-processed image can be provided as the input of neural
[12]. The authors here have proposed a facial recognition points.
system that is certainly based on a complete system, this is 1.Resizing the feedback image (256 x 256)
certainly a convolution. The authors worked on generating 2. Applying the color filtering (RGB) throughout the
the face segmentation from any image measurements. FCN channels (Our model MobileNetV2 aids 2D 3 channel
acts here as an instruction form for the feature extraction picture)
and the descent, this is certainly a gradient with which one 3. Scaling / Normalizing images using the mean that is
can work with binomial cross entropy reduction work. With standard of build in weights
the proposed approach, you have achieved an average pixel 4. Center cropping the picture with the pixel price of
accuracy of 93.884% [13]. This is exactly where the authors 224x224x3
worked on facial area detection, which is very fast and 5. Finally changing all of them into tensors (comparable to
reliable b and viola, this algorithm uses the skin mask for array that is numPy
face detection and leads to a result 4 times faster. Here we
train to model the system with 2 eyes or 1 attention and 1 IV. PROPOSED APPROACH
nose in order to lower the detection rate that is wrong. They
received 2.4 ls negative instead of 10% [14]. The recommended mask recognition method and the
implementation involved in the method are listed here. The
III. IMPLEMENTATION following is a flowchart describing the general flow of the
method. The model starts with the entire set load. Mask
It shows how to use the recommended methods, recognition data In addition, the preprocessing of the
various techniques, and the data used to implement it. In information is done by Torch Vision PyTorch. Then, after
addition, this area also specifies the selection of collections, the assembly is completed, the MobileNetV2 model training
structures, algorithms, and data sets. is completed. Then perform the serialization associated with
the actual mask classifier. The model described in the
Data at Source above statement and the training process of the
Here, we load the natural image in this article on MobileNetV2 classifier are completed using the deep
PyImageSearch for our experiment, and the task of learning framework PyTorch. Now that we have the mask
enhancing this image is done by OpenCV. "Mask" and "No serialization, we load the mask classifier, and then load the
Mask" are the tags that pictures initially liked. The size and face from the available image and video stream. This is the
resolution of the photos you take will vary because they are step after PyTorch and Python OPENCV transformation
taken by various devices with different settings. preprocessing. Only in the end, the mask detector can
identify the person wearing a mask or not wearing a mask.
Data Preprocessing End in Figure 1.
The pre-processing techniques to make the image free
and visible for recognition have been outlined here, and the

Figure 1. Flowchart for the Proposed Mask Detection

IJISRT21JUL1046 www.ijisrt.com 1403

Volume 6, Issue 7, July – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Deep Learning Frameworks
Depending on the availability of implementing such
deep learning networks, the following options can be seen.
1. TensorFlow
2. Keras
3. PyTorch
4. Caffee
5. MxNet
6. Microsoft toolKit this is certainly intellectual

We have now used PyTorch with ordinary Python

knowledge, and it will use it to create our own deep learning
projects. In addition, you will see several advantages that
can be quickly implemented from the Python system,
namely:
1.Data Parallelism
Figure 2. MobileNetV2 Architecture
2.Framework like design
Face Mask Detector:-
Algorithm development using PyTorch is completed
In mask training, we use a smaller data set, that is,
through the following modules
extended data, as we have already mentioned here. Using
• PyTorch DataLoader: Load information from the image
PyTorch transformation, PIL and dataset, the original image
folder.
is preprocessed first. The training file mask is shown in
• PyTorch DataSets ImageFolder-Hosts image sources and
Figure 3.
provides ready-made modules for marking target variables.
• Pytorch conversion-integrated help for image
preprocessing steps provided when reading from the source
folder.
• PyTorch device: used to identify related device capacity,
such as processor and GPU performance to train this model.
It can be used to change the way the functional system is
used.
• Pytorch TorchVision-With it, you can load libraries that
have already been created. For example, pre-trained models,
image sources, etc. In PyTorch, it is provided as
corePyTorch nn-it is considered the main module. This is
mainly used to build neural models of deep
systems.Libraries are provided because they may need to
build models. As a linear layer, 1D convolution plane,
conv2d, conv3d, consistency, cross entropy loss (target
loss), Softmax, ReLu, etc.
• PyTorch Optim-This defines the layout of the optimizer.
This module ensures correct identification of data such as
ADAM and SDg. Figure 3: Example for Training Set
• Pytorch PIL: Load images through resources in this way.
In this component. Just like the signal line. The gradient Model Overwritten:-
calculation can be performed with the inverse function We use the MobileNetV2 development model to
().This is considered very useful when doing feedforward in ensure that it can be deployed on any mobile device. These
DNSA. The image category algorithm used by 4 fully connected layers are built on MobileNetV2, as 4
PyTorchCNN provides different architectures, such as layers, they are consistent throughout the development
ResNet, Inception, AlexNet, MobileNet, etc. We focus on process of the model shown in Figure 4. The layer is
the MobileNetV2 architecture because it is lightweight. It's 1. typical Pooling layer with 7×7 weights
really effective. 2. Linear layer with ReLu activation function
3. Dropout Layer
MobileNetV2:- 4. Linear layer with Softmax activation function using the
The model builds on the previous version of total result of 2 values.
MobileNetV1 and uses the building block split level.
MobileNetV2 provides us with some additional functions, This level, the last Softmax level, provides the final
such as: a) linear bottlenecks between layers, b) between result according to two categories: "mask" or "no mask".
bottlenecks, providing connections as shortcuts. Figure 2
below shows the structure of MobileNetV2.

IJISRT21JUL1046 www.ijisrt.com 1404

Volume 6, Issue 7, July – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Figure 4. Final NetworkModel Architecture/Flow

V. RESULTS

The accuracy is within the range of the proposed mask

recognition method. The data set is divided into two data
sets: training and validation. The following figure shows the
accuracy of the image classification used for training and the
verification rate. The training set here contains 4000 masked
images and 3000 unmasked images, as shown in Figure 5.

Figure 6. Mask recognition by image

Fig 5:- Training vs Validation Accuracy

Next, we show some embodiments related to the

image mask detection in Figure 6, expressing the result of
the video channel in Figure 7.

Figure 7. Find live video masks

IJISRT21JUL1046 www.ijisrt.com 1405

Volume 6, Issue 7, July – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
VI. CONCLUSION Intell. Lect. Notes Bioinformatics), vol. 9117, pp.
138–145, 2015, doi: 10.1007/978-3-319-19390-8.
The COVID19 pandemic has brought various [10]. G. Deore, R. Bodhula, V. Udpikar, and V. More,
challenges to the global world, and the spread of this virus “Study of masked face detection approach in video
needs to be controlled because the virus has been infected analytics,” Conf. Adv. Signal Process. CASP 2016, pp.
and there are still millions of people in the world. The 196–200, 2016, doi: 10.1109/CASP.2016.7746164.
infected person should stop inhaling droplets when coughing [11]. W. Bu, J. Xiao, C. Zhou, M. Yang, and C. Peng, “A
or sneezing, and healthy people should wear a mask. So here cascade framework for masked face detection,”
we are going to introduce a way to use deep learning and 2017 IEEE Int. Conf. Cybern. Intell. Syst. CIS 2017
use the MObileNetV2 framework to run it. Combining IEEE Conf. Robot. Autom. Mechatronics, RAM 2017
Python PyTorch and OPENCV, the results show that the - Proc., vol. 2018-January, pp. 458–462, 2017, doi:
proposed model can detect masked or unmasked people 10.1109/ICCIS.2017.8274819.
based on the images in each movie stream, and the accuracy [12]. N. Ud Din, K. Javed, S. Bae, and J. Yi, “A Novel
of comparison and detection verification and training is GAN-Based Network for Unmasking of Masked
80.28%. Face,” IEEE Access, vol. 8, pp. 44276–44287, 2020,
doi: 10.1109/ACCESS.2020.2977386.
REFERENCES [13]. M. S. Ejaz and M. R. Islam, “Masked face recognition
using convolutional neural network,” 2019 Int. Conf.
[1]. A. Waheed, M. Goyal, D. Gupta, A. Khanna, F. Al- Sustain. Technol. Ind. 4.0, STI 2019, vol. 0, pp. 1–6,
Turjman, and P. R. Pinheiro, “CovidGAN: Data 2019, doi: 10.1109/STI47673.2019.9068044.
Augmentation Using Auxiliary Classifier GAN for [14]. T. Meenpal, A. Balakrishnan, and A. Verma, “Facial
Improved Covid-19 Detection,” IEEE Access, vol. 8, Mask Detection using Semantic Segmentation,” 2019
pp. 91916–91923, 2020, doi: 4th Int. Conf. Comput. Commun. Secur. ICCCS 2019,
10.1109/ACCESS.2020.2994762. pp. 1–5, 2019, doi: 10.1109/CCCS.2019.8888092.
[2]. P. He, "Study on Epidemic Prevention and Control [15]. Y. Heydarzadeh, A. T. Haghighat, and N. Fazeli,
Strategy of COVID -19 Based on Personnel Flow “Utilizing skin mask and face organs detection for
Prediction," 2020 International Conference on Urban improving the Viola face detection method,” Proc. -
Engineering and Management Science (ICUEMS), UKSim 4th Eur. Model. Symp. Comput. Model.
Zhuhai, China, 2020, pp. 688-691, doi: Simulation, EMS2010, pp. 174–178, 2010, doi:
10.1109/ICUEMS50872.2020.00150.. 10.1109/EMS.2010.38.
[3]. A. A. R. Alsaeedy and E. Chong, “Detecting Regions
At Risk for Spreading COVID-19 Using Existing
Cellular Wireless Network Functionalities,” IEEE
Open J. Eng. Med. Biol., pp. 1–1, 2020, doi:
10.1109/ojemb.2020.3002447.
[4]. R. F. Sear et al., “Quantifying COVID-19 Content in
the Online Health Opinion War Using Machine
Learning,” IEEE Access, vol. 8, pp. 91886–91893,
2020, doi: 10.1109/ACCESS.2020.2993967.
[5]. J. Barabas, R. Zalman, and M. Kochlan, “Automated
evaluation of COVID-19 risk factors coupled with
real-time , indoor , personal localization data for
potential disease identification , prevention and smart
quarantining,” pp. 645–648, 2020.
[6]. V. Chamola, V. Hassija, V. Gupta and M. Guizani, "A
Comprehensive Review of the COVID-19 Pandemic
and the Role of IoT, Drones, AI, Blockchain, and 5G
in Managing its Impact," in IEEE Access, vol. 8, pp.
90225-90265, 2020, doi:
10.1109/ACCESS.2020.2992341.
[7]. R. Ramachandra et al., “Custom silicone face masks:
Vulnerability of commercial face recognition systems
& presentation attack detection,” 2019 7th Int. Work.
Biometrics Forensics, IWBF 2019, pp. 1–6, 2019,
doi: 10.1109/IWBF.2019.8739236.
[8]. R. Paredes, J. S. Cardoso, and X. M. Pardo, “Pattern
recognition and image analysis: 7th Iberian
[9]. Conference, IbPRIA 2015 Santiago de Compostela,
Spain, june 17–19, 2015 proceedings,” Lect. Notes
Comput. Sci. (including Subser. Lect. Notes Artif.

IJISRT21JUL1046 www.ijisrt.com 1406