Face Mask Detection System Using Mobilenetv2: Mayank Arora, Sarthak Garg, Srivani A
Face Mask Detection System Using Mobilenetv2: Mayank Arora, Sarthak Garg, Srivani A
Face Mask Detection System Using Mobilenetv2: Mayank Arora, Sarthak Garg, Srivani A
Published By:
Retrieval Number: 100.1/ijeat.D24040410421 Blue Eyes Intelligence Engineering
DOI:10.35940/ijeat.D2404.0410421 127 & Sciences Publication
© Copyright: All Rights Reserved
Face Mask Detection System using MobileNetV2
ones by introducing random jitters and perturbations (but at Due to the limited availability of data, the model showed a
the same time ensuring that the class labels of the data are not high accuracy of 99% on the training data set but a fairly low
changed). When we use data augmentation, we want to accuracy on the test data set, hence it was overfitted, and
improve the model's generalizability. hence not suitable [5]. It is also very important to know that
how face recognition techniques have evolved and improved
D. Evaluation Metrics
from the past. Various researchers and professionals have
Accuracy - The foremost intuitive indicator of success is recently concentrated on gray-scale face images.
accuracy, and it's merely a quantitative relation of properly Although some were based purely on pattern recognition
foreseen observations to total observations. One can assume models, with no prior knowledge of the face model, others
that our model is best if we've high exactitude. Therefore, to
relied on AdaBoost, an excellent classifier for training.
see the potency of your model, you have got to appear at
Real-time face detection became possible thanks to the
alternative parameters. Accuracy = TP+TN /
TP+FP+FN+TN Viola-Jones Detector, which offered a breakthrough in face
Recall (Sensitivity) - Recall is that the quantitative relation detection technology [6]. It had several issues, including the
of positive observations expected properly to all or any orientation and brightness of the face, which made it difficult
observations within the actual class-yes. Recall = TP / to intercept. So, in a nutshell, it didn't fit in dark or dim light.
TP+FN As a result, researchers began looking for a new alternative
Precision – Precision is that the quantitative relation of the model capable of detecting faces as well as masks on the face.
positive observations properly foreseen to the positive Many datasets for face recognition have been developed in
observations foreseen in total. the past to develop a perception of face mask detection
F1 Score - F1 Score is that the Precision and Recall weighted models. WiderFace, IJB-A, MALF, and CelebA are
average. This score conjointly takes into thought each false examples of recent datasets that were generated by taking
positives and false negatives. it's not as simple to grasp online photos. In these datasets, annotations are given for
intuitively as accuracy, however, F1 is usually a lot of helpful current faces, as opposed to earlier ones. Large datasets are
than accuracy, particularly if you have got associate degree far more needed for improved training and testing data, as
uneven distribution of categories. well as for performing practical applications in a much more
Accuracy functions well if there are unit equal prices for false straightforward manner.
positives and false negatives. If the expense of false positives This necessitates a variety of deep learning algorithms that
and false negatives area unit somewhat completely different, can read faces and masks directly from the user's data. Face
each Precision and Recall area unit easier to appear at. F1 Mask detection templates come in a variety of shapes and
Score = 2*(Recall * Accuracy) / (Recall + Accuracy)
sizes. These can be classified into many groups. The
Viola-Jones face detector, which was mentioned earlier in
II. LITERATURE SURVEY
this section, was used to accept boosted cascades with easy
This section provides an overview of the main human face haar features in boosting-based recognition. The Viola-Jones
recognition techniques that apply primarily to frontal faces, detector model was then used to build a Multiview face mask
as well as the benefits and drawbacks of each approach. detector. In addition, decision tree algorithms were used to
Eigenfaces (Eigen features), neural networks, dynamic build a face mask detector model.
relation architecture, hidden Markov model, geometrical This category's face mask detectors were good at detecting
feature matching, and prototype matching are among the masks [7]. Another study employs deep learning techniques
methods considered. to discern facial recognition and determine whether or not an
The methods are evaluated based on the facial individual is wearing a facemask. The dataset collected
representations they employ [4]. The principal components includes 25,000 images with a resolution of 224x224 pixels,
of the distribution of faces, or the eigenvectors of the and the qualified model performed with a 96 percent
covariance matrix of the set of face images, are known as accuracy rate [8]. However, the test accuracy was also lower
eigenfaces in mathematics. The eigenvectors are arranged to in this case. Although it was put into practice in real life, it
reflect varying amounts of difference among the faces. A could be improved.
linear combination of the eigenfaces will accurately represent Many datasets for face detection have been created in the
each face. Only the “best” eigenvectors with the largest past to shape an understanding of face mask detection
eigenvalues can be used to approximate it. The appeal of models. Earlier datasets were made up of images taken in
using neural networks may stem from the network's supervised environments, while more recent datasets are
nonlinearity. made up of images taken online [9]. The MobileNetV2
As a result, the feature extraction stage could be faster than architecture classifier was an upgrade over the MobileNetV1
the linear Karhunen-Loève methods. A single-layer adaptive architecture classifier since it used a three-layer
network called WISARD, which includes a separate network convolutional layer as the initial layer, followed by 13 times
for each stored person, was one of the first artificial neural the previous building blocks. MobileNetV2 architecture, on
network (ANN) techniques used for face recognition. For the other hand, consists of 17 3X3 convolutional layers in a
efficient identification, the method for constructing a neural row, as well as a 1 × 1 convolution, an average layer of max
network structure is critical. It is largely determined by the pooling, and a classification layer [10].
intended application. Some of the other approaches use the
pre-trained state-of-the-art deep learning model,
InceptionV3. In this approach, the dataset used is Simulated
Masked Face Dataset (SMFD).
Published By:
Retrieval Number: 100.1/ijeat.D24040410421 Blue Eyes Intelligence Engineering
DOI:10.35940/ijeat.D2404.0410421 & Sciences Publication
128
© Copyright: All Rights Reserved
International Journal of Engineering and Advanced Technology (IJEAT)
ISSN: 2249-8958, Volume-10 Issue-4, April 2021
III. EXPERIMENTAL RESULTS was gathered from a variety of other sources and the images
used in the dataset were manually cleaned to improve the
Before conducting the experiment, an image data generator
accuracy of the results, the issue of multiple incorrect
was used to augment the dataset. IDG ensures that multiple
predictions would be successfully eliminated from the model.
variations of any given image are taken into account while
Real-world implementations will be a much more difficult
training the model. Post that, the entire input was passed
problem in the future. The new proposed model should,
through MobileNetV2 architecture which makes the entire
ideally, assist the concerned authorities in dealing with this
model unique and novel. After that, the regular CNN process
massive pandemic that has spread across most of the globe.
continued, ranging from feature enhancement, flattening,
pooling, and finally displaying whether the person standing
REFERENCES
in front of the camera is wearing a mask or not. The above
module was able to perform operations and run diagnostics at 1. Nagrath, P., Jain, R., Madan, A., Arora, R., Kataria, P., & Hemanth, J.
a faster rate as compared to CNN. The model used less (2020). SSDMNV2: A real time DNN-based face mask detection
system using single shot multibox detector and MobileNetV2.
number of parameters without compromising the accuracy of Sustainable cities and society, 102692.
the model. “Relu” activation function was used in the hidden 2. Chen, Y., Hu, M., Hua, C., Zhai, G., Zhang, J., Li, Q., & Yang, S. X.
layers of CNN because of the presence of non-linear use (2020). Face Mask Assistant: Detection of Face Mask Service Stage
cases (images). Soft-max activation was used in the output Based on Mobile Phone. arXiv preprint arXiv:2010.06421.
3. Jiang, M., & Fan, X. (2020). RetinaMask: a face mask detector. arXiv
layer to predict whether or not a person is wearing the mask preprint arXiv:2005.03950.
with the expected probability. 4. Militante, S. V., & Dionisio, N. V. (2020, August). Real-Time
Precision Recall F1-score Support Facemask Recognition with Alarm System using Deep Learning. In
2020 11th IEEE Control and System Graduate Research Colloquium
With_mask 1.00 0.83 0.90 383 (ICSGRC) (pp. 106-110). IEEE.
Without_mask 0.85 1.00 0.92 384 5. Chowdary, G. J., Punn, N. S., Sonbhadra, S. K., & Agarwal, S. (2020,
The below-mentioned table depicts the statistics December). Face mask detection using transfer learning of inceptionv3.
post-experiment. In International Conference on Big Data Analytics (pp. 81-90).
Springer, Cham.
The above model showcased an accuracy of 96 percent. 6. Loey, M., Manogaran, G., Taha, M. H. N., & Khalifa, N. E. M. (2020).
Precision, Recall, f1-score, support for with_mask attribute A hybrid deep transfer learning model with machine learning methods
are 1.00, 0.83, 0.90, 383 respectively. The values obtained for for face mask detection in the era of the COVID-19 pandemic.
Measurement, 167, 108288.
without_mask attribute for the above-mentioned parameters 7. Guo, J. M., Lin, C. C., Wu, M. F., Chang, C. H., & Lee, H. (2011).
are 0.85, 1.00, 0.92, and 384 respectively. Complexity reduced face detection using probability-based face mask
The following images showcase the working of our prefiltering and pixel-based hierarchical-feature adaboosting. IEEE
Signal Processing Letters, 18(8), 447-450.
proposed model. The software captures the faces via a 8. Sachdeva, R. (2020). Face Mask Detection System. Available at SSRN
live-stream and tells the user whether a person is wearing a 3755508.
mask or not. 9. Pooja, S., & Preeti, S. (2021). Face Mask Detection Using AI. In
Predictive and Preventive Measures for Covid-19 Pandemic (pp.
293-305). Springer, Singapore.
10. Loey, M., Manogaran, G., Taha, M. H. N., & Khalifa, N. E. M. (2020).
Fighting against COVID-19: A novel deep learning model based on
YOLO-v2 with ResNet-50 for medical face mask detection.
Sustainable Cities and Society, 102600.
AUTHORS PROFILE
Mayank Arora, UG Student, School of Computer
Science and Engineering, Vellore Institute of Technology
Figure 1: With Mask Demo (VIT), Vellore. Core Interests: Machine Learning, Deep
Learning, Natural Language Processing, Neural
Networks Mail ID- mayank.arora2017@vitstudent.ac.in.
Published By:
Retrieval Number: 100.1/ijeat.D24040410421 Blue Eyes Intelligence Engineering
DOI:10.35940/ijeat.D2404.0410421 129 & Sciences Publication
© Copyright: All Rights Reserved