Introduction

The outbreak of the novel coronavirus infection (COVID-19) has caused concern around the world as it causes illness, including death, and continues to spread from person to person in many countries [1] CoVs are a large family of viruses, including Middle East Respiratory Syndrome (MERS) CoV and Severe Acute Respiratory Syndrome (SARS) CoV [2]. On February 11, 2020, the World Health Organization (WHO) declared 'COVID-19' as a disease [3], Artificial intelligence (AI) is gradually changing practical medicine and recent advances in digitized data collection, making machine learning (ML), computing infrastructure and AI applications expand into areas previously reserved for human experts [4]. Therefore, the main objective of this research is to be able to analyze the use of recognition and detection systems that allow the identification of control measures established for the mitigation of SARS-CoV-2 contagion using masks and distancing as the main strategy of the study through the review of articles on image processing and computer vision, using algorithms for a clear understanding of the prediction and use of these computer tools in support of science and its application. The challenges and limitations encountered will allow us to propose opportunities for improvement and a clear focus on the main findings found in the information reviewed.

Methods

This study followed the literature search style recommended by the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) method [5], where uses 8 databases, namely Ebsco, Dynamed, IEEE, IOP, Sage, Scopus, Science Direct, Taylor and Francis. Trusted science and technology journals containing contemporary research papers in computer science, electronics, and interdisciplinary research. The results of this study can help researchers in the area of image processing and computer vision to know detailed information on technological advances with existing image detection and recognition systems.

Search Strategy

A bibliographic search in English was carried out in the eight academic repositories between the years 2019 and 2022, considering precise exclusion criteria. The selection of the research articles was due to the similar characteristics with our research, considering that the new methods identified require greater computational power. This research carried out a search strategy using several keywords related to the coronavirus and keywords related to the detection, diagnosis and classification of masks under the concept of AI and ML. We use these query methods to improve the search and investigation of distancing and mask applications for various AI and machine learning systems.

Inclusion criteria

• Articles are journals or conferences in English.

• The focus is on the development of various applications, systems, algorithms, methods and technologies in artificial intelligence and machine learning. • Development focused on the detection and classification of masks and social distancing.

Exclusion criteria

• Articles minor to the year 2019.

• Less with less than 15 bibliographical references.

Study Selection

The process begins with the elimination of duplicate articles, Unique articles were screened by title and abstract to verify their compliance with our inclusion and exclusion criteria. The relevant articles have been carefully read. The process of collecting, extracting research data, and developing a review document.

Data extraction and classification

Given the multidisciplinary topic of this systematic review, data extraction and classification were performed for selected studies, including CoV data using AI applications, especially ML techniques, to assess the effectiveness of viruses in detection, diagnosis, prevention and classification. Enhanced data factors were extracted from the academic literature, including author nationality, publication date, number of articles per year, and number of articles in the database. To provide a comprehensive understanding of CoV, this study The discussed CoV and the growing scale of the global pandemic in the context of artificial intelligence are analyzed using various ML and data mining algorithms, such as classification, regression, and prediction. For each study, the document distinguishes the name of the significant characteristics, the evaluation methods used and the exact status of each method. From the analyzed literature, brief motivations, challenges, limitations and recommendations have been extracted to address the serious health problems associated with CoV.

Results

The results of the search query performed in this study are shown in Figure 1. During data collection, four queries were performed to cover all databases and search mechanisms. The first result included 8578 articles from eight databases. The number of duplicate articles in all databases was 626. The first process was to select articles based on a relevant keyword selection filter, which resulted in 1141 articles. The second process was to select articles based on title and then map inclusion and exclusion criteria, resulting in 351 articles. The third process carried out was to select articles by abstract, reading each one for its selection, which resulted in 275 articles. The final process was to read all the articles in their entirety, with only 48 articles meeting the inclusion and exclusion criteria.

Statistical results

The results of metodology from Figure 1 has 48 papers that then realizes a histogram of tecniques and datasets in each query. Each query delimited a group of interested acording to proposals and datasets, 1st query (red color) is only papers that proposal and datasets to social distance, 2do query (green color) is face mask and 3rd query (blue color) are proposal that include social distance and face mask. In the Figure 2 shows to summary of algorithms and methods used in the literature review for 1st, 2nd and 3rd query, when has techniques like YOLO V3, Faster R-CNN, YOLO V2, MobileNet, DensNet, YOLO V4, YOLO V5 and YOLO-LITE, where consideres that techniques with best result using mAP metric. The analysis from histogram shows that YOLO V3 is that more used with 12, 12 and 3 papers to 1st, 2nd and 3rd query respectively.

Acording to state of art YOLO V5 has better in accuracy and inference time compared YOLO V3 but in Figure shows that YOLO V5 is only uses in 1, 0 and 1 paper to 1st, 2nd and 3rd query respectively. For 3rd query that include both a proposal to face mask and social distance have 3, 1, 1, 0, 1 papers to YOLO V3 [6,7,8], Faster R-CNN [9], MobileNet [10] and YOLO V5 [11]. Analysing Figure 3, it shows to summary of datasets used in the literature review for 1st, 2nd and 3rd query, when has techniques like COCO, Oxford Town Center Dataset, KITTI 3D dataset, CCTV, RMFD, Medical Masks Dataset, Own dataset, Youtube, FaceMask, SAI-YOLO, Face dataset Kaggle, Benchmark mask dataset and Imagenet-21k. The analysis from histogram show that COCO is that more used with 5 papers to 1st query, Own dataset with 4 and 2 papers to 2nd and 3rd query respectively. For 3rd query that include both a proposal to mask fask and social distance have 1, 2, 1 papers to Own dataset [10,8,11] and Benchmark mask dataset [9].

Discussion

Techniques

In Table 2 shows results of analyzes techniques from 45 papers according to 1st, 2nd 3rd query from Table 1 using methodology in section of methods. The first 17 papers from Table 2 shows most used algorithms for the recognition of social distancing are YOLO, Faster R-CNN and SSD with evaluations oriented on the average accuracy (mAP) and FPS to compare which algorithm gives better results in different environments. [12,13,14,15,16,17,18,10,19,20,21,22,23,24]. YOLO is the most used algorithm, with the percentage of use by versions being (YOLO V2 is with 5%, YOLO V3 with 75% and YOLO V4 with 10%), but those that use (F-RCNN and SDD) [25,26,24,27,21] go for the performance side and response time in FPS with higher results. Regarding the versions, it can be seen that there is a tendency to look for the latest updates.

Then 22 papers from Table 2 shows most used algorithms for face mask detection as YOLO, Mobile Net, ResNet, and topen-source computer vision library of python OpenCV. YOLO V3 [28], YOLO V4 [29] [30] and YOLO V5 [31] [11]; last one shows a high level of accuracy, major that 95%. After the analysis, it can notice the article [32] has the highest accuracy with 99.7%, the algorithm used was MobileNetV2 with the dataset Face-Mask-Detection. Last 6 papers from Table 2 shows about proposal that include face mask detection and social distancing where it is analyzed 4 of them work with YOLO V3 [9,6,7,8], with an average of 94.5% accuracy corresponding to the test evaluation scope; Faster R-CNN algorithm [9], with an average accuracy of 99%; MobileNet algorithm [10] with an average of 99%, YOLO V5 [11], with a precision/recall of 0.6/0.98 and finally the ResNet algorithm [10,7,11], with an average of 98.5%.

Datasets

In Table 4 includes a short description of the research purpose of Face mask dataset. In addition, it shows the principal classes used, images, labels, total weight, metrics and access type. Next in Table 3 shows a brief of the main features of social distancing dataset as: Duration, resolution, weight, metric and access type.

Challenges and limitations

The studies carried out to develop applications that use artificial intelligence techniques present many challenges and limitations in research repositories that must be addressed with urgency and interest. The current challenges are related to the transmission and contagion of COVID-19 due to the lack of

Ref. ML Classification algorithms

Evaluation Accuracy [12] YOLO V3, SSD and Faster RCNN accuracy by mAP and FPS 46.5% and 55.3% [13] YOLO V3 -- [33] Faster R-CNN, R-FCN and SSD screen refresh rate 2 FPS [14] YOLO V3 detection with various angles 90% [15] YOLO V3 -- [16] YOLOV3 mean average precision 57.9% [17] YOLO V3, Dist-YOLO V3 G and Dist-YOLO V3 W mean average precision 74.3% and 77.1% [18] YOLO V2, Fast R-CNN and R-CNN accuracy, precision and recall 86% and 96% [19] YOLO V4, YOLO V3 and Faster RCNN performance metrics acurracy 0.96, 0.84 and 0.6 [20] YOLO V3 and R-CNN performance metrics FPS 0.5 and 0.45 [21] YOLO V3, Faster RCNN and SSD evaluation acurrancy 0.5, 0.59 and 0.35 [27] YOLO V3, SSD and Faster RCNN acurrancy 0.80, 078 and 074 [22] YOLO V3 and YOLO V4 acurrancy 94.75% and 95% [23] YOLO-LITE, R-CNN and Darknet-53 -- [25] OpenCV HOG averaged error (cm) 6.72 [26] YOLO V3 and MobilNetSSD acurracy and Speedd comparison 30 / 26.66 [24] YOLOV3, FPN FRCNN and SSD300 performance metrics mAP and FPS 55.3 / 59.1 / 41.2 [34] MobileNet accuracy 94.5% [35] Openpose and FMRN accuracy daytime/nighttime 95.8% / 94.6% [36] MobileNet (HoG) accuracy 95.67% [37] YOLO Nano approach and Gaussian Mixture Model -- [38] YOLO V3 accuracy 98% [31] YOLO V5 accuracy 97.9% [29] YOLO V4 and SpineNet-190 accuracy 94.7% [30] YOLO v4, Tiny and Lightweight mAP/ AP 86% / 88% [39] Gabor Wavelet accuracy 97% [32] MobileNetV2 accuracy 99.7% [40] OpenCV and TensorFlow accuracy 97% [41] ResNet50 and ResNet101 accuracy 91% [42] OpenCV, Keras and TensorFlow accuracy 97.05% [11] YOLO V5 and DBSCAN accuracy precision and recall 85%, 97% and 80% [43] Computer vision accuracy/ precision 9% / 89% [44] TensorFlow and MobileNetV2 accuracy 99.64% [28] Haar Cascade Classifier connection speed (seconds) 0.001695977 [45] ResNet50 accuracy 98.2% [46] Quantum Transfer Learning and ResNet-18 accuracy 99.05% [47] mAP, MaskedFaceNet and MaskedFaceNet Light 0.9813 / 0.9812 [48] ResNet-50 accuracy 98.74% [49] Pynq-YOLO-Net accuracy 97% [9] Faster R-CNN and YOLOv3 Precision with mask/Without mask 0.99 / 0.88 [10] MobileNet and ResNet F1-score/sensitivity/ specificity/accuracy 0.99 / 0.99 / 0.99 / 1.0 [6] YOLO V3 -- [7] YOLO V3 and Resnet50 -- [8] YOLO V3 accuracy/F1 score 0.9120 / 0.9079 [11] YOLO V5 and Resnet50 Precision/recall 0.6 / 0.98 Another important point to take into consideration with the shadows, given that in some detections made by the angle of the camera, it is filtered as a person itself. These are cases that cause an impact within the percentage of final accuracy for applications.

Recommendations

The objective of this research is to help researchers learn about research on artificial intelligence issues and limitations and advances to mitigate the contagion of the pandemic generated by COVID-19 and thus be able to contribute to generating new research. Studies, such as the one by [32] propose a deep convolutional neural network (CNN) based on the MobileNetV2 architecture as a learning algorithm. The results show an accuracy of 99.7 % in mask detection with a run time of 1.54 s. Another study [46] uses ResNet-18; with 99.05% accuracy in the classification of protective masks. Finally [10] focuses on implementing a face mask detection and social distancing model as an integrated vision system. Pre-trained models such as MobileNet, ResNet Classifier, and VGG with an F1 score of 99%, a sensitivity of 99%, a specificity of 99%, and an accuracy of 100%.

Conclusions

The research carried out allows knowing the technological advances in the area of artificial intelligence aimed at solving problems generated by COVID-19. This research analyzed the different solution proposals based on mask detection algorithms and social distancing methods, considering the performance of machine learning models. Additionally, recommendations were established that will serve as a guide for the selection of research proposals in the area of artificial intelligence that try to solve problems of COVID-19. Finally, it is suggested to use specific terms in database queries to obtain optimal results considering the quotes and logical connectors.