16.1 Introduction

The COVID-19 outbreak has been major healthcare and social problem in the world since November 2019. As per the WHO (World Health Organization), the pandemic is causing a global health emergency, making this the most recent human health virus outbreak throughout the last century so wearing a mask is required. Before this pandemic, people used to carry the mask only to protect themselves from air pollution. This pandemic is circulated through the respiratory system which is spreading very fast. Given the fact that many states mandate people to wear masks in public areas, a lot of people forget or fail to wear masks or wear masks incorrectly [1]. Such facts will increase the speed of disease and bring a greater burden on the public healthcare system. Despite the effectiveness of many vaccinations, wearing a mask is also one of the most successful and cost-effective ways to avoid 80 % of respiratory infections. As a result, many monitoring systems have been established to provide efficient supervision in airports, hospitals, public transit systems, sporting events, and retail locations to detect the mask. This pandemic affects many areas like the institute, Organization.

Medical masks are surgical or procedure masks that are flat or pleated (some appear like cups) that are attached to the head with belts. Wearing a surgical mask is among the ways of preventing the transmission of such respiratory infections, such as COVID-19, in afflicted regions. Masks, on the other hand, should be worn in compliance with good practice guidelines to be safe. Single-use options include FFP2 face masks, surgical face masks, and N95 face masks. Even so, wearing a mask alone would be inadequate and provide proper security, and additional, similarly important steps should be taken. To avoid the spread of COVID-19, masks should be used in conjunction with hand washing as well as other IPC interventions.

Presently, there is no single drug or vaccine available to prevent. Over five million cases across 188 countries have been infected with COVID-19 in less than 6 months and because of close contact. So, the only choice left is just to take the best care and keep away from the disease. For example, keep social distancing, wear a mask, and regularly wash hand. Many countries have regulations that require people to wear masks in public. This legislation has been started in many countries as an approach to the rapid rise in cases and deaths. Researchers have discovered that wearing masks will minimize the risk of COVID-19.

The method of screening requires the identification of someone who does not wear a face mask. Here, the dataset is a collection without mask and with mask images and is using real-time mask detection through a webcam. Mask detection by using the popular deep learning technique is useful for figuring out who wears the mask and who does not and that can be used on any common device.

Deep transfer learning was used to extract features, and it was paired with three traditional machine learning algorithms. The greatest thing regarding deep learning is that there is deep architecture in all the models. Deep architecture has many layers, becoming the biggest opponents of deep architecture that has a few hidden layers. The CNN is used to function extraction from images, and then, several hidden layers learn these features. Also, SSDMNV2 technology and SVM technology are used to detect the mask detection.

Here, a comparison among them to identify the most appropriate algorithm that reached the best accuracy while taking the least amount of time during the training and detection processes is performed.

16.2 Literature Review

Mostly, studies focused on deciding whether or not such a person wears a mask. For that, many algorithms are there to determine this like LLECNN, SVM, SSDMNV2, and K-NN classifier.

Z. Wang et al. [1] introduce an article and mentioned the challenges faced by the United States in this Epidemic, such as how the epidemic developed day by day due to respiratory droplets created by breathing, talking, coughing, and sneezing [3, 4]. To avoid this, mask is mandatory to keep safe from this epidemic, and also mask can reduce respiratory infection up to 80%, and for this, the mask detection system which can monitor and provide supervision for many places like hospitals, transport, venues, airports, and other location is made [1, 2]. In this article, the author proposed browser-based edge computing which can detect the face mask. This edge computing-based solution is serverless and can be handled on any devices like tablet, mobile, and desktops through internet connections with the help of the browser. Edge computing solution reduces the hardware costs and this is integrating with YOLO (deep learning model), NCNN, and virtual machine. It is design with minimal limitation and risk and has the low computation, low network bandwidth, and high speed for detection. This solution assists in determining whether or not the individual is wearing a mask, as well as whether or not the mask is appropriate.

This pandemic is generating a health issue. M. Loey et al. [2] introduced the paper using the hybrid model like machine learning and deep learning to detect the mask. Here, models have two parts, the first part is having extraction of feature by ResNet50 and the other second part is for the process of classification for face mask by using SVM, decision tree, and another ensembled base algorithm. Here, the authors choose three datasets for determining the masked face. The datasets are the Simulated Masked Face Dataset (SMFD), the Real-World Masked Face Dataset (RMFD), and the Labeled Faces in the Wild (LFW). Herein, the result shows 99.64% accuracy in the dataset of RMFD by using SVM classifier, 99.49% accuracy in the SMFD dataset, and 100% accuracy in the LFW dataset.

Several techniques and algorithm are used in mask detection with various models. P. Nagrath et al. [3] introduced masked detection by using computer vision and image processing. Here, the approaches used in this paper to detect mask are deep learning, Keras, OpenCV, and Tensorflow, which can be helpful in the purpose of safety. The authors introduced one term that is SSDMNV2 approach which is the combination of MobileNetV2 and single-shot detector. These are the models which can be used for training and development of dataset and having very lightweight and can be used for mask detection using webcam in devices like Rasberry Pi and NVIDIA Jetson Nano. Dataset has been collected from different sources. Other researchers can be used this dataset for an advanced model like a facial landmark, part of face detection, and recognition of face process. By this technique and algorithm, the accuracy is 92.64% and the F1 score is 93%.

Almost every sector which belongs to development has collapsed because of COVID-19. Md. M. Rahman et al. [4] introduced a paper in which it talks about the healthcare system which gets affected in COVID-19 and one of the major precautionary measures is to use the mask [2]. The major part of this paper is to find out those people who are not carrying the mask in a smart city by using CCTV cameras which can reduce the growth of this pandemic. When an individual without a mask is identified, the city network notifies the appropriate authority. The main motive of this work is to find out who does not wear a mask. Dataset is collected from various sources which consist of with mask and without mask label and the architecture of deep learning is trained on these datasets. The dataset includes 1539 samples, with 80% of them being used for the training process and 20% for the testing. There are 1,231 and 308 images throughout the training and testing datasets, respectively. Here, by using this deep learning architecture, the accuracy is 98.7%.

To spread awareness about the mask in this epidemic, Y. Chen et al. [5] proposed a detection method based on the cell phone to fix the cause of the challenge in this epidemic. The gray level co-occurrence matrices (GLCMs) of the mask are first used to retrieve four elements and the K-Nearest Neighbor (KNN) algorithm is used to build a three-result detection technique. By using this technique, it shows an accuracy of 82.87%. The type I precision and type III recall is reached up to 92.00% and 92.59%, respectively. The authors also focused on future work to add more mask types to the detection objects. This research demonstrated that the developed mobile microscope device can be used as a helper while wearing a face mask.

Furthermore, many private and public service providers require customers to utilize the service only if they correctly wear masks. S. K. Dey et al. [6] proposed image processing and artificial intelligence to detect mask detection. This paper proposed a deep learning-based MobileNet Mask for multi-phase mask detection. Two datasets are trained and tested that have more than 5,200 images to detect the mask from the webcam and images. MobileNet Mask obtained an accuracy of 93% with 770 test data and nearly 100% with 276 test data. Finally, the proposed MobileNet Mask model can be implemented on light-weight computing technologies such as smartphones or embedded devices. Furthermore, this developed model incorporates cutting-edge technology to aid public and government health recommendations in the implementation of compulsory face mask legislation around the world.

Because of this pandemic, the academy faces lots of problems. S. V. Militante and N. V. Dionisio [7] proposed a certain situation that can affect the institute because still there is no vaccine available and preventive measures are only to keep safe. One of the preventive measures is to wear a face mask and use a 70% alcohol base sensitizer. Here, for detecting that a person is wearing a mask or not, deep learning is effective in detecting and classifying through the image process. The dataset comprises 25,000 images with a resolution of 224 x 224 pixels, and the model’s accuracy after training is 96%. If the person being identified is not wearing a mask, the computer produces a Raspberry Pi-based real-time face mask recognition system that alerts and captures facial features. This paper provided a study on real-time mask detection with an alarm system using deep learning techniques and CNN. This can get more accurate and faster using this tool. By using the VGG-16 CNN model, the qualified model was able to complete its task with an output accuracy of 96%.

Detecting mask manually is a very difficult task in a crowded place. G. Jignesh Chowdary, Narinder Singh Punn et al. [8] used a transfer learning method to simplify the process of recognizing people who do not wear masks [3, 4]. The model is built by fine-tuning InceptionV3, a deep learning model that has already been trained. The proposed model is trained and tested using the SMFD. It achieved 99.9% accuracy during preparation and 100% accuracy during testing.

COVID-19 creates a major impact in every area like academic and organization, and manual checking of every person for the mask is a very difficult task. A. Nusrat Zereen et al. [9] used a technology that can automatically detect the face mask. So, two-stage mask detection is introduced. Clustering and extraction is the first phase and the second is to analyze nose reason. So, here, the total accuracy is 97.13%.

C. Li et al. stated in an article [10] that, to identify masked faces, the information can be completely extracted and transformed. Studies on observational and functional datasets demonstrate the feasibility of the proposed solution.

Hariri Walid introduced an article [11]. Here, author is concerned about the security during COVID-19 pandemic and gives some efficient solution based on deep learning [1]. In this paper, a method like CNN to get the best feature is applied and the MLP problem is used for convolution neural network [2].

Locally Linear Embedding-CNN (LLE-CNN) is also working for detecting face mask. LS. Ge et al. introduced an article [12] that used LLE-CNN to find the masked faces [2]. The first one is a large dataset of mask face that is absent and the other one is the absence of facial signals from the masked areas. To handle these issue authors introduced a dataset which is known as MAFA which has 35,806 masked faces and 30,811 images. Based on that, LLE-CNN comes into the picture and experimental findings on the MAFA dataset showed that the suggested technique substantially performs at least 15.6% of six state-of-the-art approaches.

A. Anwar and A. Raychowdhury introduced a paper [13]. This paper looks at how to improve existing facial datasets by using devices that enable masked faces to be identified with lower detection costs and higher overall classification without having to recreate the user dataset by taking new recognition photos. It is an open-source platform to effectively mask faces to create a massive dataset of masked faces [3]. The dataset developed with this tool is then used for the training of an efficient facial recognition system for masked faces with target accuracy. For the FaceNet scheme, an increase of about 38% in the positive result is reported. On a specific real-world MFR2 dataset, the performance of the retrained model is also improved and comparable accuracy is reported. In this article, the authors explored the problem of identifying masked faces with accurate accuracy across existing face detection.

16.3 Implementation Approach

A short description of the methods shown in this paper is given in this section of the paper. The approach developed here involves four major phases. The first phase is data collection, the second phase is model creation, the third phase is model training, and the final phase is evaluation of the model produced.

They have used the existing facial datasets by improving them with tools that allow the identification with a low false-positive rate of masked faces and higher overall accuracy despite allowing the user dataset to be reproduced for identification by introducing better photographs.

16.3.8 Data Collection

Here, for facial recognition, two kinds of datasets, namely, with_mask and without_mask are taken. From with_mask datasets, persons who wear masks are identified, and from without_mask dataset, persons who do not wear masks are identified.

Here, this dataset is consists of 3,846 images with two classes: without mask have 1,930 images and with mask have 1,916 images as shown in Figures 16.1 and 16.2 [2].

Here, the training dataset is taken, and a facial recognition system deep learning-based is designed. A high number of masked faces and unmasked faces are needed for this.

Here, it did not sustain the training images directly. For all of that, the images are eventually needed to be marked. It is among the significant processes related to the processing of data. The researcher has used a technique known as “Labeling” in this project. This tool allows developers to access labels for both the training process mostly on images and stores the details.

16.4 Observation and Analysis

We got the result after studying papers by using different technologies like Tensorflow, YOLO, Keras, SSDMNV2, SVM, CNN model.

16.4.1 CNN Algorithm

Since the further result of training in overfitting on the training data, the built architecture is trained for 100 epochs. When an algorithm performs the undesirable patterns of the training samples, this is known as overfitting. As a result, training accuracy improves while test accuracy declines. Here, the trained model showed accuracy of 98.7% [4].

For about 100 epochs, the training and testing accuracy curve is shown and shows that the accuracy of both is the same. This indicates that the model is capable of generalizing to previously unknown data while avoiding overfitting.

As the number of epochs increases, the training loss decreases. The study loss is smaller than the training loss for about 30 epochs, but after that, it begins to increase, implying that now the prediction confidence begins to decline. The research loss varies within an appropriate range, reaching a peak around the 98th epoch.

The proposed framework’s receiver operating characteristic (ROC) curve. The graph of ROC shows the classifier’s ability to predict at various thresholds. The true positive rate and the false positive rate abbreviated as TPR and FPR, respectively, which are determined using (16.1) and (16.2), are shown on the ROC curve. For different threshold, TPR and FPR are calculated. For every possible threshold binary classifier’s performance get measures in the ROC curve (AUC), here, the range of AUC is 0 to 1. When 100% correctly predicted by model, then AUC is 1, and for 100% wrong prediction, the value of AUC is 0. Here, AUC is 0.985, which means that it is a strong classifier [4].

(16.1)

(16.2)

Here, a deep learning mechanism is used to obtain the result, which has a 98% accuracy rate, as shown in Table 16.1.

By using CNN model, the accuracy achieved is 99%. The binary classification problem is solved using a deep learning model. Keras is a high-level artificial neural network API that can be used to construct a classification model.

	Precision	Recall	F1 score	Support
With mask	0.99	1.00	0.99	384
Without mask	1.00	0.99	0.99	386
Accuracy			0.99	770
Macro avg	0.99	0.99	0.99	770
Weighted avg	0.99	0.99	0.99	770

16.4.3 SVM

By using the SVM classifier on three different datasets, i.e., RMFD, SMFD, and LFW here, datasets split into three-part (testing phase, training phase, validation phase). The testing phase has 20%, the training phase has 70%, and the validation phase has 10% of data. Performance matrices will be studied in this research to test the performance of the various classifiers.

(16.3)

(16.4)

(16.5)

(16.6)

Datasets	Decision tree classifier	SVM classifier	Ensemble classifier
RMFD	~92% to 94%	98%	97%
SMFD	96%	100%	94%
Combination of RMFD and SMFD	98%	99%	100%

This accuracy, recall, precision and F1-score [2, 3] are commonly used. In the confusion matrix, true positive, true negative, false positive, and false negative sample count are all abbreviated as TP, TN, FP, and FN. The experimental findings will be summarised in three subsections, with the first describing the results of the decision trees classifier and the second presenting the results of the SVM classifier and the last shows the result of the ensemble classifier [2] (Table 16.5).

Higher accuracy achieved in the SVM classifier is 100% for the SMFD dataset, while the highest accuracy of the decision tree classifier achieved 98% and the ensemble classifier achieved 100%.

16.5 Conclusion

COVID-19 is causing havoc on the global health system. Governments around the world are fighting to keep this virus away. A hybrid model like SSDMNV2, SVM, deep learning, and machine learning for face mask detection is presented in this paper. To train the CNN model, labeled image data were used, with the images being facial images with masks and without masks. With a 98.7% accuracy, the proposed device detects a face mask.

The SSDMNV2 model was compared to other pre-existing models by training them on the same dataset. For this reason, LeNet-5, AlexNet, VGG-16, and ResNet-50 were chosen. The methodology used in this paper yields a 0.9264 accuracy score and a 0.93 F1 score. The MobilenetV2 image classifier was used to accurately identify images, which is one of the proposed approach’s unique features. The SVM classifier was able to reach the highest level of accuracy while consuming the least amount of time during the training period. In RMFD, the SVM classifier had a testing accuracy of 99.64%. It achieved 99.49% testing accuracy in SMFD and 100% testing accuracy in LFW. With related works, a comparative result is obtained. In terms of research precision, the proposed model outperformed the related works.