Fast and Accurate Detection of Masked Faces Using CNNs and LBPs

1  Introduction

With the spread of COVID-19, millions of people have been infected worldwide, causing over 2 million deaths [1]. The infection can be transmitted through nasal secretions or saliva when an infected person sneezes or coughs since there is no particular treatment for this virus [2]. According to the research work of Eikenberry et al. [3], wearing a mask can limit the danger of coronavirus transmission. Therefore, World Health Organization mandates wearing masks in public places; however, not everyone complies with the rules. Moreover, it is difficult to track down all the offenders in crowded places.

One of the most important fields is computer vision focuses on training machines to understand digital images or video sequences [4]. Therefore, computer vision could be used as an automatic real-time tool for monitoring people who do not comply with the rules. To this end, this work introduces an efficient computer vision approach to classifying human faces as masked and unmasked. Three standard benchmark datasets are used to assert the efficiency of the proposed method: the Real-World Masked Face Dataset (RMFD), the Simulated Masked Faces Dataset (SMFD), and the Labelled Faces in the Wild (LFW). Specifically, the proposed method exploits deep learning (DL) and the local descriptor local binary pattern (LBP) for face mask detection. First, the Haar cascade algorithm was employed to detect faces as one of the most popular and efficient face detection algorithms. Haar cascade algorithm was first introduced by Viola et al. [5] and has been commonly used to identify faces in images and real-time videos, proving its efficiency compared to alternative modern DL techniques [6].

The rest of this work is structured as follows. Section 2 summarizes related work on masked face detection. Section 3 presents materials and methods. Experimental results and discussion are included in Section 4, while Section 5 concludes the paper.

2  Related Work

Nowadays, due to the COVID-19 pandemic, masked face detection is a topic of interest. It should be noted that a search on Scopus on the term “masked face detection” returned 257 documents from 2019, when the pandemic started (February 2023). Due to the extensive number of related research, it is impossible to reference all of them; however, some were selected to be indicatively mentioned, focusing on recent publications that used the same datasets as the current work.

Verulkar et al. used viola jones for driver fatigue detection and could achieve an accuracy of 99.9% [7]. Sunar et al. could improve the Viola-Jones algorithm and introduce an approach to detect ovarian and breast cancer. He used the Viola-Jones to segment the images to find the region of interest. He could achieve an accuracy of 97% [8]. Another breast cancer detection technique built on top of Viola-Jones was proposed by [9]. Lu et al. [10] presented an algorithm for detecting spliced images for images that have humans, and then they used SVM to detect whether these images were fake or not. Many other applications use the Viola-Jones algorithm [11–13]. Anitha et al. [14] proposed a driver drowsiness technique using the Viola-Jones face detection algorithm to monitor the movements of drivers’ eyes for a pre-determined amount of time. If the drivers’ eyes were closing continuously, an alarm buzzed. Fatima et al. [15] used Viola-Jones for driver fatigue detection and reported an accuracy of 99.9%. Hussein et al. [16] improved the Viola-Jones algorithm and introduced an approach to detect ovarian and breast cancer. The authors used the Viola-Jones to segment the images and find the region of interest. Experimental results reported an accuracy of 97%. Another breast cancer detection technique built on top of Viola-Jones was also proposed by Magnuska et al. [17]. Hamad et al. [18] also proposed a face recognition system that uses the Viola-Jones algorithm to detect facial areas and could achieve high accuracy. The Viola-Jones algorithm has also been used in multiple alternative applications [19–23].

This work uses the Viola-Jones algorithm for face mask detection, combined with local binary patterns (LBPs) and a deep learning model. More specifically, this work introduces an efficient algorithm that, for the first time to the authors’ knowledge, combines LBPs and a convolutional neural network (CNN) to detect masked faces. Moreover, three well-known faced masked datasets, RMFD, SMFD, and LFW, are comparatively used to evaluate the performance of the proposed method to the problem under study.

Ryumina et al. [24] implemented a method for automatically recognizing the presence of a mask on a face. They combined Convolutional Neural Networks (CNN) with image histograms and tested their algorithm on the RMFD dataset. In [25], the authors developed a masked face detection system based on deep CNN (DCNN) models in conjunction with Extreme Learning Machines (ELM) as a classification technique. The RMDF dataset was also used in [26], where the authors employed the You Only Look Once (YOLOv3) object detection algorithm. The SMDF dataset was used in [27], where a CNN model and VGG16 were employed to detect masked and unmasked faces. All three datasets were employed in [28]; a hybrid model was proposed based on ResNet50 combined with classical machine learning (ML) algorithms. Rahman et al. [29] proposed a model developed by fine-tuning the pre-trained MobileNet. The LFW dataset was used in [30]; the authors introduced an implementation of FaceNet for masked face detection in extreme environments. Bae et al. [31] also trained their encoder with the images of LFW. Singh et al. [32] employed MobileNetV2 to determine the mask’s presence. Sharma et al. [33] combined infrared temperature sensors with a CNN and used the RMFD dataset to detect masked faces. In [34], the authors used MobileNetV2 to determine mask presence. They used the SMFD dataset only during their experiment, with an accuracy of 99.016%. In [35], the authors presented a new masked face detection approach using hybrid machine learning techniques and achieved an accuracy of 97.32%.

Table 1 summarizes all information on the referenced research, including methods used, dataset, and performance results. It is obvious from the Table that the proposed method displays competitive performance compared to all previously referenced methods of the recent literature in all three datasets, as will be proved in the following sections.

3  Materials and Methods

3.1 Preliminaries

3.1.1 Viola-Jones Face Detector

The Viola-Jones detector is trained to recognize several object classes. It has also been applied to deal with the challenge of face detection and has proven its efficiency in real-time face detection. The algorithm uses a sliding window, which segments the image into sub-images. Then it searches these sub-images for Haar features, indicating the face characteristics, i.e., eyes, nose, mouth, etc. Fig. 1 shows the main Haar features.

Figure 1: The main Haar features

3.1.2 Local Binary Pattern (LBP) Descriptor

LBP is a local visual descriptor that describes the local texture features of the input image. It was first introduced in 1994 [36,37]. The most important advantage of the LBP descriptor is that it is invariant to rotation and robust to illumination and transition. The traditional LBP works by using a window of size 3 × 3. It compares the central pixel with each of the surrounding eight neighbors. If it is greater than or equal to the central pixel, it returns the value 1; otherwise, it returns the value 0. After that, the binary number is converted to its corresponding decimal value. Fig. 2 illustrates an example of the traditional LBP descriptor.

Figure 2: Uniform LBP operator

Therefore, LBPs work by thresholding the neighborhood of each pixel in the image and then converting them to binary numbers according to the following equations:

LBP(x,y)=∑n=07C(In−Ithrsh)2n(1)

where C(r)={1r≥00otherwise, and Ithrsh=I(x,y) is the central pixel. Then the LBP histogram (LBPH) is used and calculated as follows:

LBPH(t)=∑δ{t,LBP(x,y)}t={0,1,…,7}(2)

Ojala et al. [37] proposed an improved variation that could use neighbors of different sizes to take any radius and neighbors surrounding the central pixel, called a uniform local binary pattern. The uniform local pattern presents less feature vector length, and it is computed as follows:

LBPn,r(Ithrsh)={∑n=0n−1C(In−Ithrsh)2nif U(LBPn,r)>2n+1otherwise(3)

LBPHn=∑f{I(x,y)=n}n={0,1,…,n−1}

where r is the radius, and n is the number of neighbors. Fig. 3 shows the resulting codded image data.

Figure 3: Uniform LBP operator

3.1.3 Convolutional Neural Network (CNN) Model

The convolutional neural network is an artificial neural network widely used in image recognition. CNN models have proven their efficiency in computer vision, which motivated us to deploy them in this work. The motivation for using the CNN model for feature description is the deep neural network’s ability to extract high-level features that are the main points to distinguish images. Our proposed model includes the following: input layer, convolution layer, max pooling, average pooling layer, fully connected layer, and output layer.

3.2 The Proposed Methodology

The proposed method comprises two main steps: (1) the first step is face detection using the Viola-Jones detector; (2) the second step is the masked face classification using a new CNN model and LBP descriptor. The proposed method uses established techniques that have proven their efficiency and combined them for the first time towards masked face detection.

The Viola-Jones descriptor puts a bounding box around the face region [37]. Then, a cropping technique is applied to each box to identify the region of interest, as shown in Fig. 4. After that, the LBP algorithm is used to extract local features of the face area to get the LBP image which is used as an input image to the CNN model. The following section discusses the proposed CNN model in detail.

Figure 4: Generation of the region of interest

3.2.1 Features Extraction and Classification

The architecture of the CNN model proposed in this work is shown in Fig. 5. It comprises six convolution layers, six-max pooling layers, one average pooling layer, one flattened layer, finally dense layer, two fully connected layers, and one output layer. As illustrated in Table 2, the model is simple; the total number of parameters is 1,573,794, out of which 446,306 are non-trainable; hence, the number of trainable parameters utilized is 1,127,488, which is suitable for environments with limited resources such as Raspberry Pi. The number of parameters is small when compared to the state of art approaches, as shown in Table 3.

Figure 5: The proposed CNN model structure

The process is as follows:

•   The size of the input image is 224 × 224. At first, the local features are extracted from the input image using the LBP descriptor to generate the LBP-coded image of the same size. Then the LBP-coded image is passed as an input layer to the CNN model.

•   The convolution neural network model extracts the distinctive features using a convolution layer with kernel size 3 * 3. This work used six convolution layers, each followed by one max pooling layer to shrink the number of features; the filter size of the max pooling layers is 2 * 2. Then, one average pooling layer was added to minimize the total number of extracted features. Next, the flattened layer was used to get a 1D feature vector, and finally, the feature vector was passed to the fully connected layer for classification.

The steps of the proposed methodology can be summarized as follows:

-Application of Viola-Jones to detect face region.

-Cropping of the region of interest.

-Application of local binary pattern descriptor to get LBP-coded image.

-Applying a CNN model to an LBP-coded image to classify the faces as masked and unmasked.

3.2.2 System Architecture

In this work, the Raspberry Pi 4 computer model B with 1 GB RAM, which runs the Debian operating system, was used for all experiments; the tools used are summarized in Table 4. Raspberry Pi 4 was chosen because of its small size and high performance. When using these nano-devices, size is of great importance. The proposed model is suitable for raspberry pi’s abilities. The dataset is loaded and then fed to the CNN model for training.

4  Experimental Results and Discussion

The proposed technique was evaluated on three benchmark datasets: the Real-World Masked Face Dataset (RMFD), the Simulated Masked Faces Dataset (SMFD), and the Labelled Faces in the Wild (LFW). Experimental results and their analysis on each dataset occur in what follows.

4.1 Datasets

These datasets have been chosen because images in each dataset vary in background, head scale, translation, facial emotions, and image lighting.

Researchers from Wuhan University gathered and developed the RMFD dataset [38]. It contained 5,000 masked faces and 90,000 unmasked faces for 252 different identities. Fig. 6 shows some indicative samples of the RMFD dataset. Since this dataset was collected from the internet, it contains errors, blurry images, and duplicated images. Our experiment used 755 images for masked faces and 754 for unmasked faces. The same subgroup was introduced in [33].

Figure 6: Indicative images of the RMFD dataset [38]

The second dataset was SMFD [39,40]. It contained 1,570 images for different people, 758 images for people with simulated masked faces, and 758 images for unmasked faces. Samples of this dataset are shown in Fig. 7. Both RMFD and SMFD were used for training, validation, and testing.

Figure 7: Indicative images of the SMFD dataset [40]

The third dataset was LFW [41–43]. LFW dataset contained 13,000 simulated masked faces for different celebrities around the world. The LFW dataset was used only for the testing phase, and the proposed model was never trained on it. Fig. 8 shows some indicative samples of the LFW dataset.

Figure 8: Indicative images of the LFW dataset [42]

4.2 Preliminary CNN Model Results

In the first stage of the experiments, the proposed CNN model is tested on the data without using the LBP descriptor. In the second stage, the LBP-coded images were added to highlight their contribution and show that their combined use can further improve the results.

Therefore, at this first step, the extracted ROI (face area) from the input images using Viola-Jones [21] was applied to the CNN model for the RMFD and SMFD datasets. The training set was used to train the model. A ratio of 75:25 for training and testing (50% training, 25% validation, and 25% testing) was used. Fig. 9 shows the accuracy and loss rate results during the training and validation. As can be seen from the figure, the training accuracy rate reached 100% at epoch 16. In order to avoid overfitting, we used early stopping technique in which the training process is stopped when the performance on validation set stopes improving to prevent the model from continuing to learn the noise in training data.

Figure 9: Results of detection accuracy and loss rates during training and validation

Then, the testing set was used to evaluate the model’s performance. Accuracy and other performance metrics were calculated to evaluate the proposed innovative approach.

More specifically, the evaluation metrics that were considered were the following:

Accuracy=TP+FPTP+FP+TN+FN(4)

Recall=TPTP+FN(5)

Precision=TPTP+FP(6)

F1−score=2 ∗ Precision ∗ RecallPrecision+Recall(7)

The TP, TN, FP, and FN refer to the true positive, true negative, false positive, and false negative, respectively. Table 5 shows the evaluation of the proposed CNN model over the three datasets.

The performance was evaluated for the three datasets, and the confusion matrix and accuracy were calculated. The highest reported accuracies were 99.016% for the RMFD dataset, 99.48% for the SMFD dataset, and 99.65% for the LFW dataset.

Despite the high accuracy, further enhancement was pursued. Obtained results motivated the authors to think of LBPs as excellent descriptors able to extract local features from the image and detect the edges. LBPs have proved their efficiency in extracting distinctive features [44,45]. Therefore, the LBP descriptor was applied to the input images, and a second experimental trial took place to get LBP-coded images towards enhanced classification accuracy, as follows.

4.3 Proposed LBP + CNN Model Results

In the second step of the experiments, the LBP-coded images were used as an input layer to the CNN model. Results are included in Table 6.

The combination of LBPs with the CNN resulted in improved performance. The reported accuracy for the RMFD dataset was 99.86%, 99.98% for the SMFD dataset, and 100% for the LFW dataset.

A batch size of 35 and an Adam optimizer were utilized in the classifiers. The model was trained for 30 epochs. The initial learning rate was 0.04, the learning accuracy reached 100% at epoch 16, and the training was finished at the 30th epoch.

The obtained results justify the great utility of LBPs to uniquely describe an image’s texture in contrast to the common RGB images. In this context, the CNN that receives the LBP images as inputs can generate even more useful features using a smaller architecture depth. Obtained experimental results are also highly competitive compared to other state-of-the-art CNN models of the recent literature evaluated on the same dataset.

5  Conclusions

Since COVID-19 is one of the most difficulties humanity has ever encountered, wearing a face mask can effectively stop the spread of this deadly infection. This work proposes a new computer vision method for detecting masked faces, combining LBP with a CNN model. The motivation was to develop a simple technique to run on nano-devices like Raspberry Pi. The proposed CNN model was enhanced by using local feature descriptors such as LBP and, thus, achieved high performance because CNN algorithms use global features of the images. More specifically, the Viola-Jones algorithm was adopted to extract the face area from the input images. The method was experimentally evaluated on three different benchmark datasets. The evaluation results demonstrated that the proposed method outperformed recently proposed methods of the literature for the same problem under study and on the same datasets. The reported accuracy results of the proposed method were 99.86%, 99.98%, and 100% for the RMFD, SMFD, and LFW datasets, respectively.

Future work will focus on training the proposed CNN model to detect incorrectly worn masks. The proposed method can also be extended to regular occlusion objects like sunglasses and scarves. Environmental noise, such as lighting conditions, could also be investigated for detecting and recognizing masked faces under low lighting conditions. Moreover, current trends in computer vision, such as deep domain adaptation methods, could be investigated to see how they could boost the model’s performance.

Acknowledgement: This project is funded by Princess Nourah bint Abdulrahman University Researchers Supporting Project Number (PNURSP2023R442), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.

Funding Statement: Princess Nourah bint Abdulrahman University Researchers Supporting Project Number (PNURSP2023R442), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.

Author Contributions: Conceptualization, D.S.K, S.M.A, K.M.H, O.E, E.R.M, and A.Y.H; methodology, D.S.K, S.M.A, K.M.H, and A.Y.H; software, D.S.K, S.M.A, K.M.H, and A.Y.H; validation, D.S.K, S.M.A, K.M.H, O.E, E.R.M, and A.Y.H; formal analysis, D.S.K, S.M.A, K.M.H, O.E, E.R.M, and A.Y.H; investigation, D.S.K, S.M.A, K.M.H, O.E, E.R.M, and A.Y.H; resources, D.S.K, S.M.A, K.M.H, O.E, E.R.M, and A.Y.H; data curation, D.S.K, S.M.A, K.M.H; writing—original draft preparation, D.S.K, S.M.A, K.M.H, O.E, E.R.M, and A.Y.H; writing—review and editing, D.S.K, S.M.A, E.V and G.A.P; visualization, K.M.H, O.E, E.R.M, and A.Y.H; supervision, E.R.M; project administration, E.R.M; funding acquisition, G.A.P. All authors have read and agreed to the published version of the manuscript.

Availability of Data and Materials: The Real-World Masked Face Dataset (RMFD) is available online at https://github.com/X-zhangyang/Real-World-Masked-Face-Dataset (accessed on Nov 11, 2022) [36]. The Simulated Masked Faces Dataset (SMFD) is available online at https://github.com/prajnasb/observations (accessed on Nov 11, 2022) [38]. The Labelled Faces in the Wild (LFW) is available online at https://www.kaggle.com/muhammeddalkran/lfw-simulated-masked-face-dataset (accessed on Nov 11, 2022) [40].

Conflicts of Interest: The authors declare that they have no conflicts of interest to report regarding the present study.

References

1. Johns Hopkins University (JHU) & Medicine, “COVID-19 dashboard by the center for systems science and engineering (CSSE),” 2023. [Online]. Available: https://coronavirus.jhu.edu/map.html [Google Scholar]

2. Q. Cai, M. Yang, D. Liu, J. Chen, D. Shu et al., “Experimental treatment with favipiravir for COVID-19: An open-label control study,” Engineering, vol. 6, no. 10, pp. 1192–1198, 2020. [Google Scholar] [PubMed]

3. S. E. Eikenberry, M. Mancuso, E. Iboi, T. Phan, K. Eikenberry et al., “To mask or not to mask: Modeling the potential for face mask use by the general public to curtail the COVID-19 pandemic,” Infectious Disease Modelling, vol. 5, pp. 293–308, 2020. [Google Scholar] [PubMed]

4. N. Doulamis, A. Doulamis and E. Protopapadakis, “Deep learning for computer vision: A brief review,” Computational Intelligence and Neuroscience, vol. 2018, no. 7068349, pp. 1–13, 2018. [Google Scholar]

5. P. Viola and M. Jones, “Rapid object detection using a boosted cascade of simple features,” in Proc. of the 2001 IEEE Computer Society Conf. on Computer Vision and Pattern Recognition, CVPR 2001, Kauai HI, USA, pp. I, 2001. [Google Scholar]

6. J. M. Al-Tuwaijari and S. A. Shaker, “Face detection system based Viola-Jones algorithm,” in 2020 6th Int. Engineering Conf. “Sustainable Technology and Development” (IEC), Erbil, Iraq, pp. 211–215, 2020. [Google Scholar]

7. A. Verulkar and K. Bhurchandi, “Auto-luminance-based face image recognition system,” in Proc. of First Int. Conf. on Computational Electronics for Wireless Communications Springer, Singapore, vol. 329, pp. 553–565, 2022. [Google Scholar]

8. S. Sunar, S. K. Tripathi, U. Tiwari and H. Srivastava, “A comparative study on face recognition AI robot,” in Proc. Second Doctoral Symp. on Computational Intelligence, vol. 1374, pp. 211–221, 2022. [Google Scholar]

9. G. V. Sai Prasanna, K. Pavani and M. Kumar Singh, “Spliced images detection by using Viola-Jones algorithms method,” Materials Today: Proceedings, vol. 51, pp. 924–927, 2022. [Google Scholar]

10. C. T. Lu, C. W. Su, H. L. Jiang and Y. Y. Lu, “An interactive greeting system using convolutional neural networks for emotion recognition,” Entertainment Computing, vol. 40, pp. 100452, 2022. [Google Scholar]

11. S. Sharma, L. Raja, V. Bhatnagar, D. Sharma and S. N. Bhagirath, “Hybrid HOG-SVM encrypted face detection and recognition model,” Journal of Discrete Mathematical Sciences and Cryptography, vol. 25, pp. 205–218, 2022. [Google Scholar]

12. K. M. Hosny, M. Abd Elaziz and M. M. Darwish, “Color face recognition using novel fractional-order multi-channel exponent moments,” Neural Computing and Applications, vol. 33, pp. 5419–5435, 2021. [Google Scholar]

13. K. M. Hosny and M. A. Elaziz, “Face recognition using exact Gaussian-Hermit moments,” in Recent Advances in Computer Vision. Studies in Computational Intelligence, vol. 804. Cham: Springer, pp. 169–187, 2019. [Google Scholar]

14. J. Anitha, G. Mani and K. Venkata Rao, “Driver drowsiness detection using Viola Jones algorithm,” in Smart Intelligent Computing and Applications. Smart Innovation, Systems, and Technologies, vol. 159. Singapore: Springer, pp. 583–592, 2020. [Google Scholar]

15. B. Fatima, A. R. Shahid, S. Ziauddin, A. A. Safi and H. Ramzan, “Driver fatigue detection using Viola Jones and principal component analysis,” Applied Artificial Intelligence, vol. 34, pp. 456–483, 2020. [Google Scholar]

16. I. J. Hussein, M. A. Burhanuddin, M. A. Mohammed, M. Elhoseny, B. Garcia-Zapirain et al., “Fully automatic segmentation of gynecological abnormality using a new Viola-Jones model,” Computers, Materials & Continua, vol. 66, no. 3, pp. 3161–3182, 2021. [Google Scholar]

17. Z. A. Magnuska, B. Theek, M. Darguzyte, M. Palmowski, E. Stickeler et al., “Influence of the computer-aided decision support system design on ultrasound-based breast cancer classification,” Cancers, vol. 14, pp. 1–20, 2022. [Google Scholar]

18. A. Y. Hamad, K. M. Hosny, O. Elkomy and E. R. Mohamed, “Fast and accurate face recognition system using MORSCMS-LBP on embedded circuits,” PeerJ Computer Science, vol. 8, pp. 1–19, 2022. [Google Scholar]

19. K. D. Ismael and S. Irina, “Face recognition using Viola-Jones depending on python,” Indonesian Journal of Electrical Engineering and Computer Science, vol. 20, no. 3, pp. 1513–1521, 2020. [Google Scholar]

20. M. F. Hirzi, S. Efendi and R. W. Sembiring, “Literature study of face recognition using the Viola-Jones algorithm,” in Int. Conf. on Artificial Intelligence and Mechatronics Systems (AIMS), Bandung, Indonesia, pp. 1–6, 2021. [Google Scholar]

21. M. Ghosh, T. Sarkar, D. Chokhani and A. Dey, “Face detection and extraction using Viola-Jones algorithm,” in Computational Advancement in Communication, Circuits, and Systems, Singapore, Springer, vol. 786, pp. 93–107, 2022. [Google Scholar]

22. Y. Tang, Z. Pan, W. Pedrycz, F. Ren and X. Song, “Viewpoint-based kernel fuzzy clustering with weight information granules,” IEEE Transactions on Emerging Topics in Computational Intelligence, vol. 7, no. 2, pp. 342–356, 2023. [Google Scholar]

23. Y. M. Tang, L. Zhang, G. Q. Bao, F. J. Ren and W. Pedrycz, “Symmetric implicational algorithm derived from intuitionistic fuzzy entropy,” Iranian Journal of Fuzzy Systems, vol. 19, no. 4, pp. 27–44, 2020. [Google Scholar]

24. E. Ryumina, D. Ryumin, D. Ivanko and A. Karpov, “A novel method for protective face mask detection using convolutional neural networks and image histograms,” The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, vol. XLIV-2/W1, pp. 177–182, 2021. [Google Scholar]

25. C. Agarwal, P. Itondia and A. Mishra, “A novel DCNN-ELM hybrid framework for face mask detection,” Intelligent Systems with Applications, vol. 17, pp. 200175, 2023. [Google Scholar]

26. P. Kaviya, P. Chitra and B. Selvakumar, “A unified framework for monitoring social distancing and face mask wearing using deep learning,” Procedia Computer Science, vol. 218, pp. 1561–1570, 2023. [Google Scholar] [PubMed]

27. A. Negi, K. Kumar, P. Chauhan and R. S. Rajput, “Deep neural architecture for face mask detection on simulated masked face dataset against COVID-19 pandemic,” in Int. Conf. on Computing, Communication and Intelligent Systems (ICCCIS), Greater Noida, India, pp. 595–600, 2021. [Google Scholar]

28. M. Loey, G. Manogaran, M. H. N. Taha and N. E. M. Khalifa, “A hybrid deep transfer learning model with machine learning methods for face mask detection in the era of the COVID-19 pandemic,” Measurement, vol. 167, no. 5, pp. 108288, 2021. [Google Scholar] [PubMed]

29. Md H. Rahman, Mir K. Ara Jannat, Md S. Islam, G. Grossi, S. Bursic et al., “Real-time face mask position recognition system based on mobilenet model,” Smart Health, vol. 28, pp. 100382, 2023. [Google Scholar] [PubMed]

30. Y. Wang, Y. Li and H. Zou, “Masked face recognition system based on attention mechanism,” Information-an International Interdisciplinary Journal, vol. 14, no. 2, pp. 87, 2023. [Google Scholar]

31. G. Bae, M. de La Gorce, T. Baltrusaitis, Ch. Hewitt, D. Chen et al., “DigiFace-1M: 1 million digital face images for face recognition,” in Proc. of the CVF Winter Conf. on Applications of Computer Vision (WACV), Waikoloa, HI, USA, pp. 3526–3535, 2023. [Google Scholar]

32. J. Singh, L. Gupta, R. Kushwaha, T. Varshney and A. Chauhan, “Real-time face mask detection and analysis system,” in Proc. of Data Analytics and Management, Singapore: Springer, vol. 90, pp. 11–19, 2022. [Google Scholar]

33. A. Sharma, J. Dhingra and P. Dawar, “A check on who protocol implementation for COVID-19 using IoT,” in Internet of Things and its Applications. EAI/Springer Innovations in Communication and Computing, Cham: Springer, pp. 63–79, 2022. [Google Scholar]

34. S. K. Dey, A. Howlader and C. Deb, “MobileNet mask: A multi-phase face mask detection model to prevent person-to-person transmission of SARS-CoV-2,” in Proc. of Int. Conf. on Trends in Computational and Cognitive Engineering, Singapore, Springer, vol. 1309, pp. 603–613, 2021. [Google Scholar]

35. B. Wang, Y. Zhao and C. L. Philip Chen, “Hybrid transfer learning and broad learning system for wearing mask detection in the COVID-19 era,” IEEE Transactions on Instrumentation and Measurement, vol. 70, no. 5009612, pp. 1–12, 2021. [Google Scholar]

36. T. Ojala, M. Pietikainen and D. Harwood, “Performance evaluation of texture measures with classification based on kullback discrimination of distributions,” in Proc. of 12th Int. Conf. on Pattern Recognition, Jerusalem, Israel, vol. 1, pp. 582–585, 1994. [Google Scholar]

37. T. Ojala, M. Pietikäinen and D. Harwood, “A comparative study of texture measures with classification based on featured distributions,” Pattern Recognition, vol. 29, pp. 51–59, 1996. [Google Scholar]

38. Z. Wang, G. Wang, B. Huang, Z. Xiong, Q. Hong et al., “Real-world masked face dataset (RMFD),” 2020. [Online]. Available: https://github.com/X-zhangyang/Real-World-Masked-Face-Dataset [Google Scholar]

39. G. J. Chowdary, N. S. Punn, S. K. Sonbhadra and S. Agarwal, “Face mask detection using transfer learning of inceptionV3,” in Lecture Notes in Computer Science, Cham, Springer, vol. 2, pp. 81–90, 2020. [Google Scholar]

40. Prajnasb Prajna Bhandary, “Simulated masked faces dataset (SMFD),” 2020. [Online]. Available: https://github.com/prajnasb/observations [Google Scholar]

41. E. Learned-Miller, G. Huang, A. Roy Chowdhury, H. Li and G. Hua, “Labeled faces in the wild: A survey,” in Advances in Face Detection and Facial Image Analysis, Cham: Springer International Publishing, pp. 189–248, 2016. [Google Scholar]

42. M. Dalkiran, “LFW simulated masked face dataset,” 2020. [Online]. Available: https://www.kaggle.com/datasets/muhammeddalkran/lfw-simulated-masked-face-dataset [Google Scholar]

43. Z. Wang, G. Wang, B. Huang, Z. Xiong, Q. Hong et al., “Masked face recognition dataset and application,” IEEE Transactions on Biometrics, Behavior and Identity Science, vol. 5, no. 2, pp. 298–304, 2020. [Google Scholar]

44. S. Banerji, A. Verma and C. Liu, “LBP and color descriptors for image classification, cross disciplinary biometric systems,” Intelligent Systems Reference Library, vol. 37, pp. 205–225, 2012. [Google Scholar]

45. H. P. Truong, T. P. Nguyen and Y. G. Kim, “Statistical binary patterns for facial feature representation,” Applied Intelligence, vol. 52, pp. 1893–1912, 2022. [Google Scholar]