







































47 | Vol.3 No.2, July 2022  

P-ISSN : 2715-2448 | E-ISSSN : 2715-7199 

Vol.3 No.2 July 2022 

Buana Information Tchnology and Computer Sciences (BIT and CS) 

 

Automatic Face Mask Detection on Gates to Combat 

The Spread of Covid-19 

 
Musa Dima Genemo 

Study Program Computing (Software Engineering) 

Gumushane University musa.ju2002@gmail.com 

 

 
 

‹β› 
 

 

Abstract— The COVID-19 pandemic has spread across the 

globe, hitting almost every country. To stop the spread of the 

COVID-19 pandemic, this article introduces face mask 

detection on a gate to assure the safety of Instructors and 

students in both class and public places. This work aims to 

distinguish between faces with masks and without masks. A 

deep learning algorithm You Only Look Once (YOLO) V5 is 

used for face mask detection and classification. This algorithm 

detects the faces with and without masks using the video frames 

from the surveillance camera. The model trained on over 800 

video frames. The sequence of a video frame for face mask 

detection is fed to the model for feature acquisition. Then the 

model classifies the frames as faces with a mask and without a 

mask. We used loss functions like Generalize Intersection of 

Union for abjectness and classification accuracy. The datasets 

used to train the model are divided as 80% and 20% for training 

and testing, respectively. The model has provided a promising 

result. The result found shows accuracy and precision of 95% 

and 96%, respectively. Results show that the model 

performance is a good classifier. The successful findings indicate 

the suggested work's soundness. 

 

Keywords: - Accuracy, Computer vision, Classification, Face 

recognition, Surveillance 

 

 

1. INTRODUCTION 

The COVID-19 pandemic has spread across the globe, hitting 

almost every country. The epidemic was initially discovered 

in Wuhan, China, in December 2019. The CODI-19 epidemic 

has allowed us to pave the road for digital learning to be 

introduced. To stop the spread of the COVID-19 Pandemic, 

this article introduces face mask detection at university gates. 

Face recognition applications employ various deep learning 

algorithms to look for human faces within larger pictures that 

often include non-facial items such as landscapes, buildings, 

and other human body parts such as feet or hands. 

The search for human eyes, which is one of the simplest 

characteristics to identify, is often where face identification 

algorithms get started. After that, the algorithm could make 

an effort to recognize the iris, the mouth, the nose, the 

nostrils, and the eyebrows. After the algorithm has concluded 

that it has located a facial area, it next conducts further tests 

to ensure that it has identified a face in the image [1]. The 

algorithms need to be trained on big data sets that include 

hundreds of thousands of examples of both face masks and 

without face mask pictures. This will assist guarantee that the 

results are accurate. The training helps the algorithms get 

better at determining whether a picture contains a face mask. 

Where those faces are located within the image. Knowledge- 

based, feature-based, template-matching or appearance- 

based approaches are some of the strategies that may be used 

in face identification [2]. The deep learning algorithm YOLO 

is introduced that detects the objects efficiently. The authors 

do research and improve its version to V5 as shown in figure 

1. Each one has positive and negative aspects to consider. The 

problem occurs that we need an algorithm that accurately 

recognized the faces. In this paper, we have highlighted the 

issue that the algorithm accurately recognized the faces with 

and without masks. 
 

To build a model that detects face masks we used a new deep 

learning algorithm YOLO V5. This algorithm efficiently 

resolves the problem and accurately recognizes the faces. 

 

Figure 1: YOLO detection process [4] 

 

The following is the order in which the manuscript is written. 

The introductory section and the literature review overview 

section are included in Sections 1 and 2, respectively. The 

proposed approach is covered in Section 3. Section 4 shows 

the findings and explanation of the performance evaluation 



48 | Vol.3 No.2, July 2022  

experiments. The paper's conclusions are presented in 

 Section 5.  

 
2. LITERATURE REVIEW 

Knowledge-based approaches, also known as rule-based 

methods, characterize a face by adhering to certain 

guidelines. The difficulty of formulating rules that are clearly 

stated is one of the drawbacks of using this method. Noise 

and light may have a detrimental effect on facial recognition 

techniques known as feature invariant methods [3]. These 

techniques employ distinguishing characteristics of a person's 

face, such as their eyes or nose, to identify a person's face. 

The detection of faces using template-matching approaches 

involves comparing pictures with conventional face patterns 

or traits that have been recorded in the past and connecting 

the two to establish a relationship between the three. The 

problem with these approaches is that they do not account for 

differences in position, size, or form. Finding the important 

attributes of face photos requires statistical analysis and 

machine learning, both of which are used by appearance- 

based approaches [4]. 

This approach, which is also used in the process of feature 

extraction for face recognition, is broken down into many 

sub-methods. 

The following is a list of some of the more specialized 

methods that are used in face detection: Taking off the 

backdrop image. For instance, if a picture has a backdrop that 

is a single color, or if it has a background that is pre-defined 

and unchanging, then removing the background from the 

image may assist show the facial borders. Sometimes the 

color of the subject's skin may help identify faces in color 

photographs; however, this is not guaranteed to work with 

every complexion. 

The use of motion to identify faces is yet another possibility. 

Because a face is generally always moving during the real- 

time video, users of this approach are required to determine 

the region of the face that is moving. This approach has a few 

drawbacks, one of which is the potential for 

misunderstanding with other moving objects in the 

background. 

A complete technique for detecting faces may be created by 

combining some of the tactics described in the previous 

paragraph [5]. Face recognition in photographs can be 

challenging because of the many variables that can affect the 

process, differences in camera gain, lighting conditions, and 

image quality, as well as differences in attitude, emotion, 

location, orientation, skin color, and pixel values, the 

existence of spectacles or facial hair, and the photograph's 

orientation are all factors to consider. 
 

Deep learning, which has been used to make advancements 

in face identification in recent years, has the benefit of greatly 

outperforming standard computer vision approaches. These 

advancements have been made in recent years. Artificial 

intelligence detects objects. Artificial intelligence required a 

large amount of data. Then as a solution authors use YOLO, 

however, a single YOLO version did not detect a large 

number of objects accurately [6]. In [7] authors, highlight the 

issue of car number plate detection. Authors use the YOLO 

algorithm that does not accurately detect all number plates. 

The authors in [8], and [9] highlight the issue of face 

recognition. Many other authors use other ML techniques; 

however, that techniques fail to capture the facial 

expressions. The accuracy of the algorithms is not good. The 

authors resolved the problem of detecting apple plants’ 

pictures. Many authors use deep learning algorithms that fail 

to recognize that YOLO v4 is used to detect pictures whose 

accuracy is also low [10-11]. 
 

Paul Viola and Michael Jones, computer vision researchers, 

developed a system in 2001 that could accurately recognize 

faces in real-time. These developments brought about 

significant advancements in the face detection approach [1, 

12]. The Viola-Jones framework relies on the concept of 

teaching a model to recognize what constitutes a face and 

what does not constitute a face. 

Once the model has been trained, it will extract certain 

features, which will then be saved in a file. This will allow 

the features extracted from fresh photos to be compared with 

the features that were previously recorded at different phases 

[13-15]. If the picture being analyzed is successful in passing 

through each step of the comparison of its features, then a 

face has been identified and the processes may continue. 

Even though it is still widely used, the Viola-Jones 

framework has certain shortcomings when it comes to the 

recognition of faces in real-time applications. For instance, 

the framework may not function correctly if a face is 

obscured by something like a mask or a scarf. Likewise, if a 

face is not orientated correctly, the algorithm might not be 

able to locate it [16]. 

Other algorithms have been created to help improve 

processes and eliminate the disadvantages of the Viola-Jones 

framework, such as the region-based convolutional neural 

network (R-CNN) and Single Shot Detector (SSD). This has 

helped to improve the overall quality of the process [3]. 

In the field of image identification and processing, an 

artificial neural network known as a convolutional neural 

network (CNN) is a form of neural network that was 

developed for the sole purpose of handling pixel input. To 

localize and categorize the objects seen in pictures, R-CNN 

will provide region recommendations based on a CNN 

framework. 

SSD requires only one shot to recognize several objects 

within an image, compared to two shots required by region 

proposal network-based techniques like R-CNN [27,28]. The 

first shot is used to generate region proposals, and the second 

shot is used to detect the object associated with each proposal 

[2]. As a result, SSD is much quicker than R-CNN. 

The COVID-19 outbreak has quickly wreaked havoc on our 

day-to-day lives, impeding the flow of commerce and travel 

across international borders. Protecting one's face by using a 

face mask has emerged as the new standard practice. In [17, 

18] the not too distant future, many suppliers of public 

services will need their customers to wear masks that are 

suited for the environment to get their services. 

As a result, the identification of face masks has developed 

into an essential obligation to support international culture. In 

[19], [20] using several essential Machine Learning 

technologies such as TensorFlow, Keras, OpenCV, and 

Scikit-learn, the method presented in this research offers a 

straightforward approach to accomplishing this goal. The 



49 | Vol.3 No.2, July 2022  

method that is proposed can correctly recognize the face in 

the picture or video, and it then decides whether or not the 

subject is wearing a mask [4]. In addition, it can recognize a 

face even when it is covered by a mask, both while it is 

moving and while it is being seen on video [21, 22,30], and 

[23,29]. The approach attains high precision. To properly 

determine the presence of masks and avoid producing 

overfitting, we study the ideal parameter values for the 
 Convolutional Neural Network model (CNN) [33-34].  

 
3. MATERIALS AND METHOD 

Because LSVM is very fast, we utilized linear SVM with 

non-linear X2-kernel to train the model. To achieve a balance, 

we use the homogeneous kernel map, which estimates 

explicit feature mappings to approximate the homogeneous 

additive kernels, to compute a linear approximation to X2- 

kernel [32]. 

The proposed model efficiently resolves the problem and 

accurately recognizes the faces. In the detection, the process 

algorithm collects all features of video frames. In the next 

step, the layering process started where all features are 

combined and sent for the prediction. In the last step, the 

   prediction step is taken where all features are gathered from 

In this section, we discuss materials and implementation used 

for facemask detection on gates. Our proposed method 

consists of four major steps that are: (i) Object Detection (ii) 

Object Localization (iii) Objectness + box (iv) classification 

and confidence score. We first use a detector to extract ROI 

and divide inputs into two parts: the object region, which is 

crucial, and the context region, which is secondary. The ROI 

score is used to restore the classification outcome. Finally, we 

do classification with a confidence score. The detailed 

proposed flow is shown in figure 2. 
 

 

Figure 2: Proposed model for Facemask Detection and 

classification. 

 

To recognize the faces, we employed the Viola-Jones 

algorithm. The three strategies employed by [14] are as 

follows: Cascade, Integral Image (representation), and 

AdaBoost (Classifier). Several pixel-wise operations are used 

to compute the integral frame. We used equation 1 to 

calculate the sum of the left and top pixels of the impacted 

pixel, and the integral image is calculated. 

𝖯(𝑥′, 𝑦′) = ∑ 𝐼(𝑥, 𝑦) 𝑤ℎ𝑒𝑟𝑒𝑥′ ≥ 𝑥𝑎𝑛𝑑𝑦′ ≥ 𝑦 
(1) 

 

The interest RIO locations are found using Hessian Matrix 

Approximation. RIO is determined using equation 3. 

𝛽(I, 𝛼) = [
Cxx(I, 𝛼) Cxy(I, 𝛼)] 
Cxy(I, 𝛼)   Cyy(I, 𝛼) 

(2) 
Where Cxx(I, 𝛼) , Cxy(I, 𝛼) , Cyy(I, 𝛼) are Gaussian 
convolution of second order derivative. 

By forming a rectangle region around the interest locations, 

the descriptor is extracted. Before computing the Haar 

wavelet, the region is partitioned into smaller subregions. 

We increase the size of the object bounding box by a factor 

of 1:1 while maintaining the bounding box's center the same. 

The obtained result illustrates that increasing the bounding 

box size can increase classification performance. 

the image and performed the prediction as shown in figure 3. 

 
 

Figure 3: Model Structure of YOLO V5[10] 

 

The model is exploited with fine-tuned parameters. YOLO 

V5 is used to classify face recognition into with and without 

masks, as shown in figures 4 [24 - 26]. The dataset is split 

into 80% training and 20% testing sets and classes are with 

masks and without a mask. Moreover, the 300 images are 

from Class 1 which is with a mask, and 450 without a mask 

after that images are augmented using the technique of 

rotation, translation, and scaling to get more images and 

overall images become 1200 for With mask and 1350 without 

a mask. Furthermore, we used Yolo Fine Tuned algorithm 

with change weights as well as activation function. 
 

Figure 4: Facemask detection 

 

4. RESULTS AND DISCUSSION: 

In this section, we discuss the steps and parameters that were 

employed during the computation of results. In the facemask 

detection process, we tested the proposed model. During 

testing, the proposed model accuracy and error rate were both 

taken into account while detecting fasemask. The result 

found are shown in figure 5 and 



50 | Vol.3 No.2, July 2022  

 

 

Figure 5: Confusion Matrix 

Table 1 Confusion matrix 

 

 

 

 

 

 

 

Figure 6: Fasemask detection results 
 

We used the Generalized Intersection of Union (GIoU) loss 

function for the YOLO V5. The GıoU maximizes the overlap 

area of actual and predicted boundary boxes. As shown in 

figure 6, you is initially high at the point of prediction then 

slowly overlaps with the actual value. This shows that you 

gives good results that achieve good precision which is 97%. 

The value you are used for training purpose, this graph shows 

that this loss function maximizes the overlap area of ground 

truth and predicted boxes as shown in figure 4. These results 

show that YOLO V5 accurately classifies face recognition 

with and without masks as shown in table 1 and figure 5. In 

the ground reality, there are 367 faces with masks and 476 

without. 

Moreover, YOLO V5 is used to detect the face accurately. In 

figure 6, the objectness plot shows how accurately YOLO V5 

classifies the faces with and without masks. The objectness 

loss is due to wrong face recognition in GIoU prediction. The 

classification plot shows that the classification loss function 

deviates from predicting 2 for actual classes and 0 for other 

classes. This shows that YOLO V5 classifies the classes more 

accurately as shown in the results. 

Furthermore, precision and recall plots show that YOLO V5 

detects the objects accurately. The high precision shows that 

YOLO V5 accurately predicts the objects. High precision 

shows the correctness of prediction. The Recall plot shows 

how accurately find all the positives by using the YOLO V5 

algorithm. The recall is high means our results are good and 

YOLO V5 detects the faces efficiently. 

The Map@0.5 plot shows the mean of average precision at 

0.5 you. The Map is gradually increasing which shows how 

accurate the results are and how efficiently the YOLO V5 

algorithm detects the faces with and without masks. The F1 

is a harmonic mean of precision and recall. As shown in the 

results, precision and recall are high which means our F1 

score should also be high which shows the better performance 

of our used algorithm YOLO V5 that classifies the faces with 

and without masks accurately. Our results depict that YOLO 

V5 detects the faces with and with masks more accurately and 

works efficiently in classification. We also used the public 

dataset to compare our results to the current state of the art. 

The outcome is shown in table 2. 
 

Table 2: State-of-the-art performance evaluation 

  Dataset Accuracy  

Thermal Cheetah Dataset 0.942 

Anki Vector Robot Dataset Dataset 0.985 

Drone Gesture Control Dataset 0.971 

EgoHands Dataset 0.982 

   Pascal VOC 2012 Dataset 0.980  

 
5. CONCLUSION 

In this paper, the deep learning algorithm YOLO V5 is 

exploited to accurately classify faces with and without masks. 

YOLO V5 used the loss functions GIoU, objectness, and 

classification. That shows how accurately YOLO V5 detects 

the objects. The dataset is split into 80% training and 20% 

testing sets and classes are with masks and without a mask. 

Moreover, the 300 images are from Class 1 which is with a 

mask, and 450 without a mask after that images are 

augmented using the technique of rotation, translation, and 

scaling to get more images and overall images become 1200 

for With mask and 1350 without a mask. The used algorithm 

is evaluated through simulations in which the used algorithm 

YOLO V5 precision is 96% and accuracy is 95%. The results 

show that YOLO V5 detects objects more accurately. 
 

REFERENCES 

 

[1] A. D. Miller, T. B. Murdock, and M. M. 

Grotewiel, "Addressing Academic Dishonesty 

Among the Highest Achievers," Theory Into 

Practice, vol. 56, pp. 121-128, 2017. 
[2] Li, E. “Research on Face   Detection 

Methods”. 4th International Conference on 

Signal Processing and Machine Learning, 2021. 

[3] Genemo, M.D. “Suspicious activity recognition 

for monitoring cheating in exams”. Proc.Indian 

Natl. Sci. Acad. 88, 1–10 (2022). 

[4] Genemo, M. D. (2022). Suspicious activity 

recognition for monitoring   cheating   in 

exams. Proceedings of the Indian National 

Science Academy, 1-10. 

[5] Reddy, S., Goel, S., & Nijhawan, R. “ Real-time 

Face Mask Detection Using Machine Learning/ 

Deep Feature-Based Classifiers For Face Mask 

Class n(Truth) n(classified) Accuracy Precision Recall F1 

Score 

With 

Mask 

367 370 95.61% 0.95 0.95 0.95 

Without 

Mask 

476 473 95.61% 0.96 0.96 0.96 

 

mailto:Map@0.5


51 | Vol.3 No.2, July 2022  

Recognition”. IEEE Bombay Section Signature 

Conference (IBSSC), 2021. 

[6] Luh, G. “Face detection using a combination of 

skin color pixel detection and Viola-Jones face 

detector” International Conference on Machine 

Learning and Cybernetics, 2014. 

[7] Vidal, A., Jha, S., Hassler, S., Price, T., & 

Busso, C. “Face detection and grimace scale 

prediction of white-furred mice. Machine 

Learning with Applications”, 8, 100312.2022. 

[8] Alakkari, S., & Collins, J. J. “Eigenfaces for 

Face Detection” A Novel Study. 12th 

International Conference on Machine Learning 

and Applications, 2014. 

[9]  Jeong, H. J., Park, K. S., & Ha, Y. G. (2018, 

January). Image preprocessing for efficient 

training of YOLO deep learning   networks. 

In 2018 IEEE International Conference on Big 

Data and Smart Computing (BigComp) (pp. 

635-637). 

[10] Garg, D., Goel, P., Pandya, S., Ganatra, A., & 

Kotecha, K. (2018, November). A deep learning 

approach for face   detection   using   YOLO. 

In 2018 IEEE Punecon (pp. 1-4). IEEE. 

[11] Al-Masni, M. A., Al-Antari, M. A., Park, J. M., 

Gi, G., Kim, T. Y., Rivera, P., ... & Kim, T. S. 

(2018). Simultaneous detection and 

classification of breast masses in digital 

mammograms via a deep learning YOLO-based 

CAD system. Computer methods and programs 

in biomedicine, 157, 85-94. 

[12]      Wu, D., Lv, S., Jiang, M., & Song, H. (2020). 

Using channel pruning-based YOLO v4 deep 

learning algorithm for the real-time and accurate 

detection of apple flowers in natural 

environments. Computers and Electronics in 

Agriculture, 178, 105742. 

[13] George, J., Skaria, S., & Varun, V. V. (2018, 

February). Using a YOLO-based deep learning 

network for real-time detection and localization 

of lung nodules from   low dose   CT scans. 

In Medical Imaging 2018: Computer-Aided 

Diagnosis (Vol. 10575, p. 105751I). 

International Society for Optics and Photonics. 

[14] Magalhães, S. A., Castro, L., Moreira, G., Dos 

Santos, F. N., Cunha, M., Dias, J., & Moreira, A. 

P. (2021). Evaluating the single-shot multibox 

detector and YOLO deep learning models for the 

detection     of     tomatoes     in      a 

greenhouse. Sensors, 21(10), 3569. 

[15] Loey, M., Manogaran, G., Taha, M. H. N., & 

Khalifa, N. E. M. (2021). 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, 65, 102600. 

[16] Zhuang, Zhemin, Guobao Liu, Wanli Ding, Alex 

Noel Joseph Raj, Shunmin Qiu, Jingfeng Guo, 

and Ye Yuan. "Cardiac VFM visualization and 

analysis based on YOLO deep learning model 

and modified 2D continuity 

equation." Computerized Medical Imaging and 

Graphics 82 (2020): 101732. 

[17] Ardhianto, P., Subiakto, R. B. R., Lin, C. Y., Jan, 

Y. K., Liau, B. Y., Tsai, J. Y., ... & Lung, C. W. 

(2022). A Deep Learning Method for Foot 

Progression Angle Detection in Plantar Pressure 

Images. Sensors, 22(7), 2786. 

[18] Pouyanfar, S., Sadiq, S., Yan, Y., Tian, H., Tao, 

Y., Reyes, M. P., ... & Iyengar, S. S. (2018). A 

survey on deep learning: Algorithms, 

techniques, and applications. ACM Computing 

Surveys (CSUR), 51(5), 1-36. 

[19] Guo, Yanming, Yu Liu, Ard Oerlemans, 

Songyang Lao, Song Wu, and Michael S. Lew. 

"Deep learning for visual understanding: A 

review." Neurocomputing 187 (2016): 27-48. 

[12]     Zhou, Xinyi, Wei Gong, WenLong Fu, and 

Fengtong Du. "Application of deep learning in 

object detection." In 2017 IEEE/ACIS 16th 

International Conference on Computer and 

Information Science (ICIS), pp. 631-634. IEEE, 

2017. 

[21] Xiao, Youzi, Zhiqiang Tian, Jiachen Yu, Yinshu 

Zhang, Shuai Liu, Shaoyi Du, and Xuguang Lan. 

"A review of object detection based on deep 

learning." Multimedia Tools and 

Applications 79, no. 33 (2020): 23729-23791. 

[22] Mittal, P., Singh, R., & Sharma, A. (2020). Deep 

learning-based object detection in low-altitude 

UAV datasets: A survey. Image and Vision 

computing, 104, 104046. 

[23] Wu, X., Sahoo, D., & Hoi, S. C. (2020). Recent 

advances in deep learning for object 

detection. Neurocomputing, 396, 39-64. 

[24] Srivastava, S., Narayan, S., & Mittal, S. (2021). 

A survey of deep learning techniques for vehicle 

detection from UAV images. Journal of Systems 

Architecture, 117, 102152. 

[25] Garcia-Garcia, A., Orts-Escolano, S., Oprea, S., 

Villena-Martinez, V., Martinez-Gonzalez, P., & 

Garcia-Rodriguez, J. (2018). A survey on deep 

learning techniques for image and video 

semantic segmentation. Applied Soft 

Computing, 70, 41-65. 

[26] Zujovic, J., Gandy, L., Friedman, S., Pardo, B., 

& Pappas, T. N. (2009, October). Classifying 

paintings by artistic genre: An analysis of 

features & classifiers. In 2009 IEEE 

International Workshop on Multimedia Signal 

Processing (pp. 1-5). IEEE. 

[27] Kobylin, O. A., Gorokhovatskyi, V. О., 

Tvoroshenko, I. S., & Peredrii, O. О. (2020). The 

application of non-parametric statistics methods 

in image classifiers based on structural 

description components. Telecommunications 

and Radio Engineering, 79(10). 
[28] Dredze, M., Gevaryahu, R., & Elias-Bachrach, 

A. (2007, August). Learning fast classifiers for 

image spam. In CEAS (pp. 2007-487). 

[29] Kobylin, O. A., Gorokhovatskyi, V. О., 

Tvoroshenko, I. S., & Peredrii, O. О. (2020). The 

application of non-parametric statistics methods 



52 | Vol.3 No.2, July 2022  

in image classifiers based on structural 

description components. Telecommunications 

and Radio Engineering, 79(10). 

[30] Wenger, J., Kjellström, H., & Triebel, R. (2020, 

June). Non-parametric calibration for 

classification. In International Conference on 

Artificial Intelligence and Statistics (pp. 178- 

190). PMLR. 

[31] P. F. Felzenszwalb, R. B. Girshick, D. 
McAllester, and D. Ramanan, “Object 
detection with discriminatively trained part 
based models," IEEE TPAMI , 2009. 

[32] A. Vedaldi and A. Zisserman, “Efficient 
additive kernels via explicit feature maps," in 
CVPR, 2010 

[33]  Wu, Z., Xiong, Y., Yu, S., & Lin, D. (2018). 

Unsupervised feature learning via non- 

parametric instance-level discrimination. arXiv 

preprint arXiv:1805.01978. 

[34] Gorokhovatskyi, V. О., Tvoroshenko, I. S., & 

Vlasenko, N. V. (2020). Using fuzzy clustering 

in structural   methods   of   image 

classification. Telecommunications and Radio 

Engineering, 79(9). 

[35] Chen, R. C. (2019). Automatic License Plate 

Recognition via sliding-window darknet-YOLO 

deep learning. Image and Vision Computing, 87, 

47-56. 


