Novel Face Mask Detection Technique using Machine Learning to control COVID'19 pandemic.

The COVID-19 pandemic has been scattering speedily around the world since 2019. Due to this pandemic, human life is becoming increasingly involutes and complex. Many people have died because of this virus. The lack of antiviral drugs is one of the reasons for the spreading of COVID-19 virus. This disease is spreading continuously and easily due to some common mistakes by people, like breathing, coughing and sneezing by infected persons. The main symptom is the normal flu. Therefore, in the present condition, the best precaution for this disease is the face mask, which covers both areas of mouth & nose. According to the government and the World Health Organization, everyone should wear a face mask in busy places like hospitals and marketplaces. In today's environment, it's difficult to tell if someone is wearing a mask or not, and physical inspection is impractical since it adds to labour costs. In this research, we present a mask detector that uses a machine learning facial categorization system to determine whether a person is wearing a mask or not, so that it may be connected to a CCTV system to verify that only persons wearing masks are allowed in.

Introduction

As of 2020, respiratory infectious disease is spreading over the world, affecting the entire population. This virus causes respiratory tract infections in humans, which can range from mild to fatal. This sickness has spread throughout China and is now afflicting the entire population. Every country is working hard to develop a vaccine that will either prevent or cure this disease. This disease is dangerous because it is easily transmitted through simple means, such as visiting public areas, touching infected objects, coming into close contact with an infected person, or their nasal and cough droplets. This is how anyone can become sick and subsequently infect others. As a result, the authorities advised wearing masks whenever going out in public places and maintaining social distance, both now and in the future, as a precautionary measure to protect yourself from this sickness. It is now required, but some are breaching the law by not wearing a mask as they walk out, necessitating the presence of a person at the entrance to verify whether or not a person is wearing a mask. Security personnel, on the other hand, will be at risk, as it is impossible to track every individual, and any inaccuracy may accelerate virus propagation. In our investigation, we used a detector for the face with a mask; typical detectors for auto-focusing [1], [2] and human–computer interaction [3], [4] are available. For example, in [5], C2D-CNN is used to distinguish a person with a mask or a covered face using a mask detection technique. In essence, we're using a face detection algorithm to distinguish human faces in digital data, such as a picture or a video clip. It can also be considered an example of object-class detection. Face recognition has long been a difficult challenge for computer vision, but as technology progresses, it is becoming easier [6], [7]. There was semi-automated equipment for face identification, just like in the 1960s. The first semi-automatic facial recognition system was deployed in 1988, and the automated system was displayed at the Super Bowl event in 2001, capturing surveillance photographs and attracting the attention of a large number of people. Nowadays, everyone uses this technology for various reasons and current trends, and it is extremely important in the IT sector and the field of computer vision. It's employed in practically every task, from simple camera detection to a wide range of scientific applications. Face detection techniques include feature-based, appearance-based, knowledge-based, and template-matching methods, among others. Deep learning approaches such as face detection of masked and non-masked people, as well as image categorization of both datasets, are used in our research. This may be a notification system that works with any CCTV or sensor, allowing only those wearing masks on their faces to enter. We employed two datasets in this study, one with people wearing masks and the other without masks, and we used numerous stages, such as data pre-processing, face detection, and image classification utilizing convolution neural networks and the Alex Net architecture. This study can be used in a variety of settings, including supermarkets, schools, hospitals, and other public areas, to detect those who are breaking the rules and to make it easier for security officials to follow each and every individual passing through.Table 1, Table 2

Table 1

Programming input & output data.

________________________________________________________________Layer (type) Output Shape Param #=================================================================input_1 (InputLayer) (None, 250, 250, 3) 0_________________________________________________________________conv2d_1 (Conv2D) (None, 63, 63, 96) 34944_________________________________________________________________max_pooling2d_1 (MaxPooling2 (None, 32, 32, 96) 0_________________________________________________________________conv2d_2 (Conv2D) (None, 32, 32, 384) 332160_________________________________________________________________conv2d_3 (Conv2D) (None, 32, 32, 384) 1327488_________________________________________________________________conv2d_4 (Conv2D) (None, 32, 32, 256) 884992_________________________________________________________________max_pooling2d_2 (MaxPooling2 (None, 16, 16, 256) 0_________________________________________________________________flatten_1 (Flatten) (None, 65536) 0_________________________________________________________________dense_1 (Dense) (None, 4096) 268439552_________________________________________________________________dropout_1 (Dropout) (None, 4096) 0_________________________________________________________________dense_2 (Dense) (None, 4096) 16781312_________________________________________________________________dropout_2 (Dropout) (None, 4096) 0_________________________________________________________________dense_3 (Dense) (None, 1000) 4097000_________________________________________________________________dropout_3 (Dropout) (None, 1000) 0_________________________________________________________________dense_4 (Dense) (None, 1) 1001=================================================================Total params: 291898449Trainable params: 291898449Non-trainable params: 0_________________________________________________________________

Table 2

Metrics on testing data.

Accuracy: 0.9839666154184055

Precision: 0.9904761904761905

Recall: 0.839677047289504

Database used

The Real-World Masked Face Dataset (RMFD) [8], which is accessible on a GitHub repository and is the world's biggest masked face dataset, is the first dataset we use for masked persons. Celeb Faces Attributes (CelebA) [9] is the other dataset we used. As illustrated in Fig. 1, Fig. 2 , the Kaggle dataset comprises over 200 k photographs of celebrities with 40 binary feature annotations.

Fig. 1

Images of people without Mask (Celebrity face dataset).

Fig. 2

Images of people with mask (RMFD Dataset).

Face-detection

For face identification in this article, we employed C2D CNN [9], [10], [16], [17] or 2D colour PCA-Principal component Analysis [11] – convolutional neural networks [5], [13], [14], [15]. In this procedure, we estimate a mixture of features learnt from original pixels and the representation learnt by CNN, and then estimate the outcome or make a decision. A colour 2-dimensional principal component analysis-convolutional neural network is known as a C2D CNN.

Convolution neural network

Convolution is a process to mix two signals and create a new one. It is one of the most important techniques in signal processing, as image is a 2-dimensional signal so convolution is applied to image to create a new image with modified properties according to need. In order to understand convolution, consider an image of shape (10x10) and a kernel/filter of shape (3x3) having pixel intensities from I1 to I64 and F1 to F9 respectively as per followings. Convolution will be performed by moving the filter over the picture from top left to bottom right. Equation explains it mathematically (1),where, i = 1 to (M – m + 1) and j = 1 to (N – n + 1)

As seen in Fig. 3 , the convolution process is now repeated across layers to produce a network-like structure known as a convolutional neural network. This diagram describes the colour 2-dimensional principal component analysis-convolutional neural network topology of this network (C2D-CNN).

Fig. 3

The construction of convolutional neural network algorithm.

This network takes colored face training images into input and passes through 11 layers to generate SoftMax output. To train this network backpropagation algorithm is used and gradients for that are calculated by using the following equations,

In these equations, L is the loss function and φ and ψ are the parameters that the model will learn. Now, the same image that was given as input to above CNN is passed through PCA to obtain projection vectors.

Principal component examination

The Principal Component Analysis (PCA) is used to remove linearly dependent vector from a data/matrix. It removes the features by projecting the whole data to a new rotated vector space where the data acts as ideal data. The whole process revolved around making the covariance matrix of data to a sparse matrix. This makes the inter feature covariance negligible or very less which is used to make features less linearly dependent.

Fig. 4 explains the rotated vector space. The vectors with b1& b2 as basis vectors are original vectors and vectors with μ1& μ2 are rotated vectors. This is how whole data is projected to a new vector space and then vectors with maximum individual variance and rest of the features are removed to reduce the dimensions.

Fig. 4

Different vector space rotation.

Mask detection

In this paper we did the mask recognition by a special convolutional architecture. It is basically a classification algorithm that is best suited for the classification of RGB images. The architecture consists of eight layers, the first five of which are convolutional layers, followed by some max-pooling layers, and the last three of which are fully linked layers. It primarily employs the ReLU activation function, which outperforms tanh and sigmoid.

Topology with results

Data preprocessing

Data preprocessing is a mining technique that includes transforming any primitive data into an easily understandable format. Real data is sometimes inconsistent or lacking in some trends, and can contain many errors. We are taking an image data set, so it contains lots of unnecessary images, or shape issues, dimension issues so we are pre-processing our data to make it suitable for our problem. We have taken two data sets in our research i.e., first is RMFD which is a masked image dataset and second is a celebrity dataset, so we are plotting or reshaping them to process them as shown in Fig. 5 and Fig. 6.Fig. 6A Fig. 6B .

Fig. 5

Mean data of all the masked images.

Fig. 6A

Mean data of all the unmasked images.

Fig. 6B

Detecting masks on faces.

Detected faces out of the image

For face recognition, we employed a deep C2D CNN (colour 2-dimensional principal component analysis – convolutional neural network). We mix the attributes of original pixels with the picture representation learnt by CNN in this procedure, and then perform decision fusion to enhance our face recognition model. There are several steps in this network, first, the introduction of the normalization layer in the network to shorten the time for training. Second, the introduction of a layered activation function which will prevent gradient diffusion. Lastly, the application of probabilistic max-pooling for the preservation of feature information to the maximum level. We are using this algorithm because in HOG, the algorithm is based on shapes and people with a mask cannot be detected using that, so we are using this C2D-CNN as it is a pixel-oriented algorithm and favors our process.

AlexNet

In this research we are doing classification by AlexNet architecture. It is a type of convolutional neural network. It is basically a classification algorithm that is best suited for the classification of RGB images. Architecture includes eight layers, among them, the first 5 were convolutional layers, followed with some max-pooling layers which are used in down sampling and the last three layers were fully connected layers. It uses a rectified linear unit activation function. The basic description of Alex Net architecture is shown in Fig. 7 .

Fig. 7

Alex Net Architecture [12]

3.4 Results

.

Conclusion

COVID-19, as we all know, is a worldwide epidemic that has harmed all of the world's main powers. As a result, it becomes everyone's national responsibility to stop it from spreading and to teach people to follow the norms. Mask recognition may be used in public surveillance cameras to look for persons who aren't wearing a mask, and then analyzing the data may reveal a lot about the disease's progress. This corona-virus will become ingrained in our society; we will have to be more careful when interacting with crowds, and wearing masks will become the norm. Based on the findings, we can conclude that the ML based topology provides better results with higher accuracy and is more effective in controlling the COVID-19 pandemic.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

I1	I2	I3	I4	I5	I6	I7	I8
I9	I10	I11	I12	I13	I14	I15	I16
I17	I18	I19	I20	I21	I22	I23	I24
I25	I26	I27	I28	I29	I30	I31	I32
I33	I34	I35	I36	I37	I38	I39	I40
I41	I42	I43	I44	I45	I46	I47	I48
I49	I50	I51	I52	I53	I54	I55	I56
I57	I58	I59	I60	I61	I62	I63	I64

F1	F2	F3
F4	F5	F6
F7	F8	F9