[Link].

com/scientificreports

OPEN An improved YOLOv8 safety

helmet wearing detection network
Xudong Song , Tiankai Zhang  & Weiguo Yi *

In the field of industrial safety, wearing helmets plays a vital role in ensuring workers’ health. Aiming
at addressing the complex background in the industrial environment, caused by differences in
distance, the helmet small target wearing detection methods for misdetection and omission detection
problems are needed. An improved YOLOv8 safety helmet wearing detection network is proposed
to enhance the capture of details, improve multiscale feature processing and improve the accuracy
of small target detection by introducing Dilation-wise residual attention module, atrous spatial
pyramid pooling and normalized Wasserstein distance loss function. Experiments were conducted
on the SHWD dataset, and the results showed that the mAP of the improved network improved to
92.0%, which exceeded that of the traditional target detection network in terms of accuracy, recall,
and other key metrics. These findings further improved the detection of helmet wearing in complex
environments and greatly enhanced the accuracy of detection.

Keywords YOLOv8, Attention mechanism, Pooled pyramid, Loss function, Safety helmet wearing detection

With the acceleration of industrialization, the safety supervision of construction sites has gradually received
increased amounts of attention. Production safety accidents in housing and municipal engineering usually
occur because site construction workers do not wear safety helmets. Although it has been clear that employees
must correctly use safety equipment at work, there are still cases in which safety regulations are ignored during
actual operation; therefore, it is especially important to detect the wearing of safety helmets.1However, traditional
methods for detecting safety helmet, such as manual inspection and video surveillance, have problems such
as incomplete coverage and slow response, making it difficult to meet the real-time detecting requirements of
modern construction sites. This background has given rise to the study of safety helmet wearing detection using
target detection technology.
Traditional safety detecting mainly relies on manual supervision by on-site safety personnel, a method that
faces many challenges and drawbacks.2 Firstly, construction sites usually have a vast footprint, complex and
changing work scenarios, and a high degree of worker mobility, so it is difficult for manual supervision to achieve
comprehensive coverage of the helmet-wearing status of all workers. This limitation makes it difficult to effectively
carry out safety detecting and capture every worker’s violation in real time. Secondly, in the traditional safety
management system, a large number of safety supervisors are required to carry out prolonged supervision work,
which not only consumes a large number of human resources, but also makes the supervisors fatigue easily by
concentrating for a long period of time, thus affecting the accuracy and continuity of supervision. This problem
is further exacerbated by the decentralized nature of workers’ operations, the difficulty of safety management, and
the lack of clarity of safety responsibilities. Finally, the manual supervision process inevitably involves subjective
judgment, which may lead to inconsistency and inaccuracy in the supervision results. At the same time, due to
the inefficiency of supervision, even if violations are detected, they may not be effective in preventing accidents
due to the lack of timeliness.
With the progress of technology, the popularization of intelligent high-definition video surveillance system
based on target detection ­technology3 provides production and operation units with an intelligent safety detecting
method that avoids human operation and has high real-time performance. Such systems are not only able to
detect the working environment in real time, but also send timely warning messages to effectively prevent safety
accidents. Although the application of intelligent detecting equipment has brought significant safety management
improvements, detecting equipment still faces many challenges in helmet detection in the actual construction
work environment. On the one hand, the detection effect of safety helmets is often affected by factors such as light
intensity, changes in the background environment, equipment blockage, and changes in viewing angle and scale,
and these environmental factors may lead to misdetection and omission of the detecting system, thus reducing
the accuracy and reliability of detecting. On the other hand, in the complex construction environment, the
safety helmet, as a relatively small target, is more difficult to detect. This not only requires the detecting system

Computer and Communication Engineering Institute, Dalian Jiaotong University, 794 Huanghe Road, Shahekou
District, Dalian, Liaoning, China. *email: jiekexun98@[Link]

Scientific Reports | (2024) 14:17550 | [Link] 1

Vol.:(0123456789)
 [Link]/scientificreports/

to accurately recognize the helmet, but also to accurately determine its wearing status to ensure effective safety
management.
In recent years, with the rapid development of deep learning t­ echnology4,5, the target detection algorithm has
also been significantly improved and perfected, especially in terms of detection accuracy and detection speed.
Therefore, deep learning technology is utilized to improve the accuracy of detecting of helmet as well as small-
target helmet wearing in complex backgrounds, which can not only effectively enhance the safety management
of the construction site and safeguard the lives of the workers, but also reduce the potential safety accidents, and
provide a strong guarantee for the safety of production in the construction industry.
In this experiment, we chose to use YOLOv8 network, and based on this, we introduced ­DWR6 attention
module in the backbone layer, which enhanced the feature extraction capability by multi-scale cavity convolution.
Additionally, by introducing the A ­ SPP7,8 pooling pyramid, the information of different scales is captured, which
more effectively handles the helmet detection task for helmets at different distances and background conditions.
­ WD9 loss function improves the regression accuracy and compensates the deficiency
Finally, the introduction of N
of CIoU in small target detection. The improved model proposed in this paper can effectively capture the details
and improve the helmet detection accuracy in different scenarios, especially in the detection of small targets, thus
greatly reducing the cases of misdetection and omission, and providing further technical support for ensuring
the production safety in complex construction environments.

Related work
In recent years, a series of target detection networks based on deep learning have been proposed. The YOLO series
­algorithm10–16 has gradually become one of the mainstream target detection algorithms due to its accuracy. For
example, Liu et al.17 proposed an improved YOLOv7 network (YOLOv7-AC) for underwater target detection,
which improves the speed of feature extraction and network inference. Wang et al.18 proposed a new network
lightweight model, HV-YOLOv8, which effectively improves the detection accuracy of small targets and improves
the adaptability to different small targets.
For helmet detection, Deng et al.19 addressed the complexity and high resource demands of the YOLOv3
object detection algorithm by designing a new, lightweight version, ML-YOLOv3. The approach integrated
the cross stage partial network (CSPNet) and GhostNet to form a more efficient residual network, CSP-Ghost-
ResNet, and combines CSPNet with Darknet53 to create the new backbone network, ML-Darknet. Additionally,
a lightweight multiscale feature extraction network, PAN-CSP-Network, was introduced. The resulting
ML-YOLOv3 significantly reduced the floating point operations (FLOPs). Zhang et al.20 presented a practical
algorithm aimed at improving helmet detection, utilizing an enhanced version of the YOLOv5s algorithm.
Firstly, the K-means method was utilized to recalibrate the size of the anchor boxes based on the dataset’s label
characteristics, which aims to boost the model’s feature extraction accuracy. Secondly, an additional layer was
integrated into the algorithm to bolster the model’s capability in recognizing small-sized targets. Finally the
attention mechanism was incorporated and the CIOU_Loss function was replaced with the EIOU_Loss function
within the YOLOv5 framework to refine the model’s precision. Li et al.21 introduced YOLO-PL, an innovative
and lightweight helmet detection algorithm derived from YOLOv4, which enhanced detection accuracy and
efficiency. The development began with YOLO-P algorithms, which refined the network’s ability to detect small
objects and improved anchor assignment. The introduction of the Enhanced PAN (E-PAN) structure allowed for
effective merging of information across different layers, improving detection accuracy. The study progressed by
lightening the design through the Dilated Convolution Cross Stage Partial with X res units (DCSPX), optimizing
the structure for enhanced lightness while maintaining performance, and replacing the conventional spatial
pyramid pooling (SPP) module. Xia et al.22 added a new feature output to the YOLOv5 network to detect small
target helmets and used clustering methods to obtain a more appropriate prior anchor frame. Yi et al.23 used the
YOLOv5 algorithm to detect the helmet wearing situation of operators in complex scenes, which can accurately
detect operators in motion, and also has better detection effect for the obscured helmets. Dai et al.24 improved
the sensitivity of the network for small target detection based on SSD using multilayer fusion to consider both
shallow and deep semantic information. Tan et al.25 improved YOLOv5 by introducing DloU-NMS to increase
the accuracy of suppressing the predicted bounding box. Fang et al.26 established a large-scale data set and used
the method of deep learning to detect the helmet. The optimization approaches for the YOLO network generally
include the introduction of the attention ­mechanism27–30, the improvement of the loss ­function31,32, and the
optimization of the pyramid pooling l­ ayer7.
In summary, advancements have been made in the detection of safety helmets. Yet, there is a gap in targeted
research within construction and engineering fields, where challenges such as identifying small-scale objects
in complex backgrounds frequently result in detection failures or inaccuracies. Addressing the existing gaps in
research, this paper explores a improved model for the detection of safety helmets, focusing on enhancing the
accuracy of detection to better protect the safety of workers in the construction industry.

Methods
In the field of helmet detection, opting for an anchor-free YOLOv8 model circumvents the issues associated with
fixed sizes and ratios inherent in traditional anchoring methods, which is particularly crucial for identifying safety
helmets of various sizes and shapes. The anchor-free design streamlines the detection process, enhancing model
training and generalization capabilities. Moreover, the method of directly regressing bounding boxes increases
the detection accuracy for small objects and overlapping targets, a key advantage in complex construction
scenarios. Additionally, YOLOv8 maintains the high-speed detection characteristics of the YOLO series and
further optimizes speed and accuracy by eliminating anchors, making it more suitable for real-time safety

Scientific Reports | (2024) 14:17550 | [Link] 2

Vol:.(1234567890)
[Link]/scientificreports/

monitoring. Therefore, based on these considerations, choosing to improve based on the anchor-free YOLOv8
model becomes a rational choice.

YOLOv8 network
YOLOv8 is a leading end-to-end target detection network model that continues and builds on the core ideas
of the YOLO series. Its structure is divided into four main parts: the input, the backbone layer, the neck hybrid
feature network layer and the detect layer. On the input, YOLOv8 employs a variety of data enhancement
techniques, including mosaic data enhancement and adaptive image scaling, to effectively enrich the training
dataset. The backbone layer consists of an attention mechanism module, a cross-stage local network, and a
spatial pyramid pooling structure, which work together to efficiently extract image features. The neck hybrid
feature network layer utilizes the path aggregation network and feature pyramid network structure for multi-
scale feature fusion, which enables the model to efficiently handle images of different scales. Finally, in the detect
layer network, YOLOv8 employs decoupled detection headers optimized for classification and localization tasks,
respectively, to further improve the accuracy and efficiency of detection.

Improved safety helmet wearing detection network

Despite YOLOv8’s advancements in object detection, its performance in safety helmet detection can be
compromised by distance variations and complex backgrounds typical in industrial settings, leading to potential
misdetections and omissions. To address these issues, the network has been enhanced for more precise helmet
detection.
As illustrated in Figure 3, the enhanced YOLOv8s model for helmet detection incorporates significant
modifications: the DWR attention module is integrated into the backbone layer, and the SPPF module is
substituted with the ASPP module in the network’s backbone. This alteration not only broadens the receptive
field of the feature map but also preserves the resolution of the original image. The model gains the ability to
handle and interpret spatial information across various scales through the inclusion of receptive fields of differing
sizes. Additionally, the adoption of NWD within the regression loss framework markedly diminishes localization
errors while boosting accuracy. These modifications enable the model to precisely detect helmet usage across
varying distances and in complex backgrounds, enhancing the detection of smaller targets.

DWR attention mechanism

The DWR module is also called Dilation-wise residual attention module, and its structure is realized in residual
mode. As shown in Fig. 1, the multiscale information is efficiently extracted by the "RR–SR" two-step method
(where RR is Region Residualization, and SR is Semantic Residualization), and then the generated feature maps
with multiscale receptive fields are fused.
Specifically, the first step is RR, which is implemented by a conventional 3 × 3 convolution combined with
a bulk normalization (BN) layer and a ReLU layer and is used to generate a series of concise feature maps of
different sizes. These feature maps will be used as materials for the second step of morphological filtering. The
second step is SR, which employs multirate null depth convolution to morphologically filter the features of
regions of different sizes separately. As a result of this step, only one desired receptive field is applied to each
channel feature to exclude possible redundant receptive fields.
With the above approach, the role of multirate null depth convolution changes from trying to obtain as much
complex information as possible to simply morphologically filtering each succinctly expressed feature map to
obtain multiscale information more efficiently.
After extracting multiscale information, multiple outputs are aggregated. All the feature maps were
concatenated and then batch normalized. Then, the features are merged using point-by-point convolution to form
the final residuals. Finally, the final residuals are added to the input feature maps to construct a more robust and
comprehensive feature representation DWR that enhances the feature extraction capability through multiscale
null convolution, which enables the network to effectively capture details and contextual information, thus
improving the detection accuracy in complex scenes. In addition, the improved ability of DWR to handle targets
at different scales enhances the network’s generalization capability and makes it more adaptable to environmental
changes, which is crucial for stability and reliability in practical applications.

ASPP pooled pyramid

In the original YOLOv8, when utilizing spatial pyramid pooling fast (SPPF) for feature extraction, since the
main focus is on global feature extraction, it is not effective enough to capture fine-grained local features; this
approach may become a bottleneck when accurately locating helmets occupying smaller portions of the frame
and not providing sufficient contextual information.
In this regard, this paper introduces ASPP, whose structure is shown in Fig. 2. ASPP consists of two parts
in parallel, the first of which performs dimensionality reduction via one 1 × 1 convolution and constructs the
pooling pyramid. The corresponding 3 × 3 null-space convolutional layers are stacked to extract the features
at different scales. The second part first applies global average pooling, which reduces the feature maps to
a numerical value that captures the global contextual information of the image. Then, a 1 × 1 convolutional
layer with a convolutional kernel of 256 is connected to output a single mean feature, which is subsequently
upsampled back to the size of the original feature map by bilinear interpolation. Finally, the outputs of the two-
part null convolutional and global pooling layers are connected, and a 1 × 1 convolutional layer is used to fuse
the compressed multiscale features to generate the final output feature map.
ASPP is able to capture information at different scales through its null convolution with different sampling
rates, and through global context fusion, it is able to handle the task of safety helmet detection at different

Scientific Reports | (2024) 14:17550 | [Link] 3

Vol.:(0123456789)
 [Link]/scientificreports/

Figure 1.  Improved YOLOv8 safety helmet wearing detection network(CBS modules are used to extract
the initial features. The C2f. module is a residual feature learning module that enriches the gradient flow of
the model through cross-layer connections, resulting in a neural network module with a stronger feature
representation capability. The DWR is add to enhance the model’s focus on relevant features, improving feature
representation and overall detection accuracy. The ASPP module replaces the SPPF module, which uses a
combination of serial and parallel maximum pooling operations to amplify different receptive fields and output
feature maps with adaptive sizes. The loss function consists of Complete Intersection over Union and NWD.)

distances and under different background conditions more efficiently, especially when accurately locating small
targets. This improvement enables the network to achieve greater accuracy and robustness in handling helmet
detection in diverse and complex site environments.

NWD loss function

The original YOLOv8 used the CIoU as the default loss letter number, which is defined as follows:
ρ 2 b, bgt
 
RCIoU = + αv (1)
c2
where b and bgt represent the centers of the two bounding boxes of the prediction box and the real box,
respectively; ρ is the Euclidean distance between the centers of the two bounding boxes; c is the diagonal length
of the smallest closed region containing the two bounding boxes; and α is a weight parameter for balancing the
effects of overlapping regions and aspect ratios, defined as follows:
v
α=
(1 − IoU) + v (2)

IoU denotes the intersection and concurrency ratio between the predicted bounding box and the true
bounding box, and v measures the consistency of the aspect ratio, defined as follows:

Scientific Reports | (2024) 14:17550 | [Link] 4

Vol:.(1234567890)
[Link]/scientificreports/

Figure 2.  Flowchart of the DWR.

2
w gt

4 w
v= arctan − arctan (3)
π2 hgt h
w,h and w gt,hgt are the width and height of the predicted and real bounding boxes, respectively. Thus, the
complete CIoU is defined as follows:
ρ 2 b, bgt
 
LCIoU = 1 − IoU + + αv (4)
c2
The CIoU has several limitations in small target detection. The CIoU is more sensitive to slight positional
deviations in small targets, especially when the helmet occupies a small area in the image, which may lead
to performance degradation. To address the above limitations, this paper introduces NWD as part of the
regression loss in the YOLOv8 network. The network is first modeled by representing the bounding box as a two-
dimensional Gaussian distribution. Specifically, for the horizontal border R = (cx, cy, w, h), where (cx, cy), w and
h denote the center coordinates, width and height, respectively. Its interior elliptic equation can be expressed as:
2
(x − µx )2

y − µy
+ =1 (5)
σx2 σy2

µx , µy is the center coordinates of the ellipse and σx , σy are the lengths of the semi-axes along the x and y
 
axes. Therefore, µx = cx,µy = cy,σx = w/2,σy = h/2.
The probability density function of a two-dimensional Gaussian distribution is:

Scientific Reports | (2024) 14:17550 | [Link] 5

Vol.:(0123456789)
 [Link]/scientificreports/

Figure 3.  Schematic diagram of ASPP.

exp − 12 (x − µ)T � −1 (x − µ)
 
f (x|µ, �) = 1 (6)
2π|�| 2
where x,µ, are the covariance matrices of coordinate (x, y), mean vector and Gaussian distribution. When
(x − µ)T � −1 (x − µ) = 1, The ellipse in Eq. 5 will be the density profile of a two-dimensional Gaussian
distribution. Thus the horizontal bounding box R = (cx, cy, w, h) can be modeled as a 2D Gaussian distribution
N(µ, �), where:
 2 
w
 
cx 4 0
µ= ,� = 2
cy 0 h4

Then, the similarity between the predicted and real targets is calculated by comparing their corresponding
Gaussian distributions. The Wasserstein distance from Optimal Transport theory is used to calculate the
distribution distance. For two Gaussian distributions µ1 = N(m1 , �2 ) and µ2 = N(m2 , �2 ), the Wasserstein
distance between µ1 and µ2 is:
  
1/2 1/2

1/2
W22 (µ1 , µ2 ) = �m1 − m2 �22 + Tr �1 + �2 − 2 �2 �1 �2 (7)

The above equation can be simplified as:

1/2 2
 
 1/2
W22 (µ1 , µ2 ) = �m1 − m2 �22 + �1 − �2  (8)
F

 || · ||F is Frobenius
where  norm. And for Gaussian distributions Na and Nb modeled by A = cxa , cya , wa , ha and
 
B = cxb , cyb , wb , hb , the above equation can be further simplified as:
  2
  wa ha T wb hb T 
 
2
W2 (Na , Nb ) =  cxa , cya , , , cxb , cyb , · (9)
 
2 2 2 2

 
2

But W22 (Na , Nb ) is a distance metric and cannot be used directly as a similarity metric (i.e., values between 0
and 1 as IoU). Therefore it is normalized using its exponential form to obtain a new metric, called Normalized
Wasserstein distance (NWD). It is defined as follows:
 � 
W22 (Na , Nb )
NWD(Na , Nb ) = exp −  (10)
C

C is a constant closely related to the dataset. Na and Nb represent the predicted bounding box and the true
bounding box, respectively, which are modeled as two-dimensional Gaussian distributions.
The NWD is independent of the target scale and is suitable for measuring the similarity between small targets.
In helmet detection, the prediction accuracy is effectively measured even when the size of the helmet in the image
is low. So the NWD indicator is chosen and designed as a loss function:

Scientific Reports | (2024) 14:17550 | [Link] 6

Vol:.(1234567890)
[Link]/scientificreports/

(11)
 
LNWD = 1 − NWD Np , Ng

Np is the Gaussian distribution model for the prediction frame p and Ng is the Gaussian distribution model
for Ground Truth box g.
With the above method, the introduction of NWD can effectively measure the similarity between the
predicted frame and the real frame, reduce positioning errors, compensate for the shortcomings of the CIoU in
small target detection, and improve the accuracy and robustness of detection.

Experiments and analysis of results

Datasets
In this paper, we use the publicly available SHWD safety helmet wearing detection dataset, which contains 7581
images covering 9044 helmet-wearing heads (positive samples) and 111,514 ordinary heads without helmets
(negative samples). The dataset also contains images of workers taken at different distances in different scenes,
which can effectively demonstrate the effects of multiple scales. To prevent overfitting due to insufficient data
volume, the original dataset was augmented with data in this paper. The augmented dataset contains 10,581
images, and a Python script was used to randomly assign the training set (train) and test set (val) with 8:2
weighting (Fig. 3).

Experimental environment
In this study, the experimental environment utilizes the Windows10 operating system, with programming carried
out in Python. Model training, and result testing are all conducted within the PyTorch , leveraging the CUDA
(compute unified device architecture). The configuration details are outlined in Table 1.

Network training
During the training process of the YOLOv8 model, in order to optimize the model performance, this study
defines specific hyper-parameters during the training process, as shown in Table 2.

Experimental evaluation indicators

In this paper, the precision (P), recall (R), average precision (AP) and mean average precision (mAP) are used
as the main evaluation metrics of performance. The commonly used mAP0.5 and mAP0.5:0.95 are chosen as
evaluation indices in terms of accuracy. Its specific formula is as follows:
TP
P= (12)
FP + TP

TP
R= (13)
TP + FN

Experimental configuration Specification

Operating system Windows10
CPU Intel Xeon W-2265
GPU NVIDIA GeForce RTX 3090
IDE PyCharm
Development language Python3.8.15
Deep learning frame Pytorch1.13.1
CUDA CUDA11.6

Table 1.  Experimental configuration.

Name of the training parameter Parameter value

Epochs 100
Batch 16
Initial learning rate 0.01
Momentum 0.937
Weight_decay 0.0005

Table 2.  Network training hyperparameters.

Scientific Reports | (2024) 14:17550 | [Link] 7

Vol.:(0123456789)
 [Link]/scientificreports/

1
AP = (R)dR (14)
P

1
mAP = AP (15)
n
TP is the number of correctly identified positive samples,FP is the number of incorrectly identified positive
samples,FN is the number of positive samples not correctly identified, and n is the number of categories.

Analysis of the experimental results of the improved network

Performance measures
The developed network’s efficacy is shown in graphs, which show different metrics of the performance of training
and validation sets. Two separate types of loss are depicted in Fig. 4 where box_loss indicates the error between
the prediction box and the calibration box and cls_boss indicates the accuracy of calculating the anchor box
and the corresponding calibration classification. By observing the changes in each loss, it can be seen that the
loss value of the improved network decreases and gradually converges, which proves that the model training
has achieved good results.
Figure 5. shows the change in each evaluation index. Observing the changes in the four evaluation indices,
it can be seen that the detection accuracy increases and stabilizes at the highest value, which shows that the
improved network performs very well in terms of both the convergence effect and learning effect.
We better represent the performance improvement of the model proposed in this study by comparing the loss
curve as well as the precision-recall curve with other models. As shown in Fig. 6, the commonly used evaluation
metric for precision-recall curve is the area under the PR curve, i.e., PR AUC (Area Under the PR Curve), which
ranges from 0 to 1, and the larger the value, the better the performance. The maximum large area under the PR
curve of the model proposed in this study indicates the best performance of the classification model with the
best trade-off between precision and recall. As shown in Fig. 7, compared with other networks, our network
maintains a high level of convergence speed while the loss value is lower than other networks. This shows that
the improved network proposed in this paper reflects a great advantage in performance.
The detection results of the improved network are shown in Fig. 8. Due to the default use of mosaic data
augmentation, each picture was stitched together by a random number and size of pictures, in which 0 represents
the person wearing a helmet and 1 represents the person not wearing a helmet. The figure shows that even a
target that occupies a small part of the picture can effectively detect whether the helmet is worn, which shows that
the improved network in this paper greatly improves the quality of helmet detection, which is highly effective.

Figure 4.  Loss profile of the improved network on the helmet dataset.

Scientific Reports | (2024) 14:17550 | [Link] 8

Vol:.(1234567890)
[Link]/scientificreports/

Figure 5.  Evaluation metrics of the improved network on the helmet dataset.

Figure 6.  The precision-recall curve of other networks versus our network.

Ablation experiments
DWR module ablation experiments
Due to the complex background of helmet detection and different distances from the detection equipment, these
factors can lead to misdiagnosis and omission of detection results, so the DWR module is introduced to enhance

Scientific Reports | (2024) 14:17550 | [Link] 9

Vol.:(0123456789)
 [Link]/scientificreports/

Figure 7.  The loss curve of other networks versus our network.

Figure 8.  Safety helmet wearing detection results after network improvement.

the feature extraction ability of the network. The DWR module is added on the basis of YOLOv8s network and
compared with other different mainstream attention for experiments to verify its effectiveness.
Table 3 illustrates the comparative analysis of the network’s performance with the introduction of the
DWR module. It shows an enhancement of 2.58% in Precision, 6.54% in Recall, and 4.97% in mAP over the
conventional network. Conversely, the addition of the ECA attention module marginally increases the mAP by
0.35% but does not achieve improvements in other performance metrics. Although the CBMA attention module
slightly surpasses DWR in Precision by 0.11%, it demonstrates lower performance in Recall and mAP by 2.36%
and 1.32%, respectively. Therefore, the overall effectiveness in detection is superior with the DWR module
compared to the other attention module.
From the analysis in Table 3, it can be seen that the introduction of the DWR module enhances the feature
extraction capability, the multi-scale processing effect, and the model generalization performance of YOLOv8
in helmet target detection.

Scientific Reports | (2024) 14:17550 | [Link] 10

Vol:.(1234567890)
[Link]/scientificreports/

mAP@ mAP@
Network P/% R/% 0.5/% 0.5:0.95
v8_s 89.2 79.5 86.5 0.542
v8_s_CBAM 91.6 82.7 89.6 0.582
v8_s_ECA 86.9 79.4 86.8 0.505
v8_s_DWR 91.5 84.7 90.8 0.604

Table 3.  Attention module ablation experiment results.

ASPP module ablation experiments

To capture multi-scale helmet details, the original network’s SPPF module was replaced with the ASPP module.
This change was compared against SimSPPF and RFB modules. As shown in Table 4, the SimSPPF led to a modest
increase in Precision by 0.22% and Recall by 0.51%, albeit with a slight reduction in Precision by 0.12%, compared
to the original setup. The RFB improved all performance metrics, yet it fell short of the ASPP’s enhancements,
with Precision, Recall, and mAP lower by 0.44%, 2.31%, and 1.34% respectively. In stark contrast, the integration
of the ASPP module significantly boosted the model’s performance: precision increased by 1.12%, Recall by
3.40%, and mAP by 2.77%, all compared to the original network configuration.
From this analysis, it can be seen that ASPP is able to capture details and contextual information more
effectively through its multi-scale null convolution, especially when dealing with different scale targets, which
improves the accuracy of model detection.

NWD loss function ablation experiments

Table 5 presents a comparative analysis of the effects of various loss functions. Incorporating NWD into CIoU
significantly enhances the model’s performance, yielding a 3.12% higher mAP and reducing the Box loss by
29.6% compared to the original network. In contrast, the WIoU loss function does not match the effectiveness
of the original network configuration. The introduction of the SIoU loss function improves mAP and decreases
Box loss, however, when compared to the NWD, there’s a deficit of 2.01% in mAP and a 40.03% increase in Box
loss. Similarly, the performance improvements with DIoU do not rival those achieved by NWD. This evidence
strongly supports that the integration of NWD with the original network not only boosts the detection accuracy
of small targets but also significantly diminishes localization errors.

Integral network ablation experiments

To validate the improvement effects of the YOLOv8 network, a series of ablation experiments were designed to
explore how these improvements affect the network’s performance on the helmet detection task by introducing
different combinations of modules, such as ASPP, NWD, and DWR, into YOLOv8. As shown in Table 6, the
performances of the various combinations are evaluated by key metrics to reveal the specific contributions of
the different modules to the detection effect.
According to the data presented in the Table 6, the addition of DWR, ASPP, and NWD individually
contributes to notable improvements in the original network’s metrics. However, the collective integration of
these modules optimizes the model, resulting in significant enhancements: precision is augmented by 2.91%,

mAP@ mAP@
Network P/% R/% 0.5/% 0.5:0.95
v8_s 89.2 79.5 86.5 0.542
v8_s_SimSPPF 89.4 79.9 86.4 0.505
v8_s_RFB 89.8 80.3 87.7 0.554
v8_s_ASPP 90.2 82.2 88.9 0.556

Table 4.  Pooled pyramid module ablation experiment results.

network [email protected]/% [email protected]:0.95 Box_loss

v8_s(CIoU) 86.5 0.542 1.395
v8_s_WIoU 86.5 0.533 1.411
v8_s_SIoU 87.4 0.545 1.385
v8_s_DIoU 88.6 0.522 1.366
v8_s_NWD 89.2 0.561 0.982

Table 5.  Loss function ablation experiment results.

Scientific Reports | (2024) 14:17550 | [Link] 11

Vol.:(0123456789)
 [Link]/scientificreports/

mAP@ mAP@
Network P/% R/% 0.5/% 0.5:0.95
v8_s 89.2 79.5 86.5 0.542
v8_s_DWR 91.5 84.7 90.8 0.604
v8_s_ASPP 90.2 82.2 88.9 0.556
v8_s_NWD 91.2 82.1 89.2 0.561
v8_s_DWR_ASPP 91.5 84.9 90.9 0.606
v8_s_DWR_NWD 92.5 85.6 91.6 0.617
v8_s_ASPP_NWD 90.7 84.2 90.5 0.601
Ours 91.8 86.6 92.0 0.622

Table 6.  Ablation experiment results.

mAP@ mAP@
Network P/% R/% 0.5/% 0.5:0.95
Faster R-CNN 89.5 80.5 87.4 0.541
RFBNet 82.6 73.4 75.6 0.416
YOLOv5 88.4 78.6 85.7 0.554
YOLOX 91.0 81.7 88.3 0.549
YOLOv8 89.2 79.5 86.5 0.542
SSD24 88.3 76.7 84.2 0.516
DAAM-YOLOv525 87.7 78.6 85.3 0.526
Ours 91.8 86.6 92.0 0.622

Table 7.  Comparative experiment results.

Recall by 8.93%, and mAP by 6.36%. These results highlight the complementary nature of the modules: DWR
improves detail recognition, ASPP enhances multi-scale feature processing, and NWD reduces localization
errors while ensuring the accuracy of small target detection. Their unified application not only boosts the model’s
overall accuracy but also its robustness and adaptability across varied scenarios, culminating in optimal helmet
detection performance.

Comparative experiments
To comprehensively evaluate the performance of the improved YOLOv8 network on the helmet detection task,
comparative experiments were conducted with several classical target detection networks, including traditional
networks such as Faster R-CNN, RFBNet, YOLOv5 and YOLOX, and networks in the literature referenced in
papers 24 and 25. As shown in Table 7, the network proposed in this paper outperforms Faster R-CNN by 5.26%
in mAP, RFBNet by 21.69%, YOLOv5 by 7.35%, and YOLOX by 4.19%, and also outperforms other methods in
all other metrics. It can be seen that the method proposed in this thesis reflects obvious advantages in terms of
detection accuracy compared with other methods.

Conclusion
The improved YOLOv8 network proposed in this paper provides significant improvement in safety helmet
wearing detection. In response to the problem of a large number of small-target helmets being detected due to
the small frame occupied by helmets caused by distance and other reasons, the detection capability for small
targets has been significantly improved by the introduction of the DWR, ASPP and NWD modules. The DWR
module enhances the feature extraction capability of the network, the ASPP module improves the network’s
multiscale processing effect, and the NWD module optimizes the network’s detection accuracy for small targets.
The experimental results show that these improvements and enhancements cause the network to outperform the
traditional target detection network in key metrics such as accuracy, recall and mAP. Therefore, the improved
method proposed in this paper effectively improves the accuracy of safety helmet wearing detection, especially
in small target helmet detection.

Data availability
The data used in this paper is public and has been deposited on GitHub at [Link]
Safety-​Helmet-​Weari​ng-​Datas​et. The data is from a third party and the ownership of the data is Njvisionpower.
The above website provides a usage License. Permission is hereby granted, free of charge, to any person obtaining
a copy of this software and associated documentation files.

Received: 3 February 2024; Accepted: 23 July 2024

Scientific Reports | (2024) 14:17550 | [Link] 12

Vol:.(1234567890)
[Link]/scientificreports/

References
1. Zhang, J., Qu, P., Sun, C. & Luo, M. Safety helmet wearing detection method based on improved YOLOv5. J. Comput. Appl. 42,
1292–1300 (2022).
2. Geng, J. & Ren, B. N. Application of fuzzy comprehensive evaluation in the bid evaluation of municipal engineering construction
projects. Appl. Mech. Mater. 584, 2159–2164 (2014).
3. Li, W., Feng, X. S., Zha, K., Li, S. & Zhu, H. S. In Journal of Physics: Conference Series. 012003 (IOP Publishing).
4. Qi, S. et al. Two-dimensional electromagnetic solver based on deep learning technique. IEEE J. Multiscale Multiphys. Compu. Tech.
5, 83–88 (2020).
5. Sadad, T. et al. Brain tumor detection and multi-classification using advanced deep learning techniques. Microsc. Res. Tech. 84,
1296–1308 (2021).
6. Wei, H. et al. DWRSeg: Dilation-wise residual network for real-time semantic segmentation. arXiv preprint arXiv:​2212.​01173
(2022).
7. Lian, X., Pang, Y., Han, J. & Pan, J. Cascaded hierarchical atrous spatial pyramid pooling module for semantic segmentation. Pattern
Recognit. 110, 107622 (2021).
8. He, H., Yang, D., Wang, S., Wang, S. & Li, Y. Road extraction by using atrous spatial pyramid pooling integrated encoder-decoder
network and structural similarity loss. Remote Sens. 11, 1015 (2019).
9. Yu, Z. et al. Yolo-facev2: A scale and occlusion aware face detector. arXiv preprint arXiv:​2208.​02019 (2022).
10. Chen, W., Huang, H., Peng, S., Zhou, C. & Zhang, C. YOLO-face: a real-time face detector. Visual Comput. 37, 805–813 (2021).
11. Adibhatla, V. A. et al. Applying deep learning to defect detection in printed circuit boards via a newest model of you-only-look-
once (2021).
12. Jocher, G. et al. ultralytics/yolov5: v6. 0-YOLOv5n’Nano’models, Roboflow integration, TensorFlow export, OpenCV DNN support.
Zenodo (2021).
13. Guo, Z., Wang, C., Yang, G., Huang, Z. & Li, G. Msft-yolo: Improved yolov5 based on transformer for detecting defects of steel
surface. Sensors 22, 3467 (2022).
14. Kim, J.-H., Kim, N., Park, Y. W. & Won, C. S. Object detection and classification based on YOLO-V5 with improved maritime
dataset. J. Mar. Sci. Eng. 10, 377 (2022).
15. Wang, G. et al. UAV-YOLOv8: a small-object-detection model based on improved YOLOv8 for UAV aerial photography scenarios.
Sensors 23, 7190 (2023).
16. Zhang, Y. et al. Complete and accurate holly fruits counting using YOLOX object detection. Comput. Electron. Agric. 198, 107062
(2022).
17. Liu, K. et al. Underwater target detection based on improved YOLOv7. J. Mar. Sci. Eng. 11, 677 (2023).
18. Wang, W., Meng, Y., Li, S. & Zhang, C. Hv-Yolov8 by Hdpconv: Better Lightweight Detectors for Small Object Detection. Available
at SSRN 4632283
19. Deng, L., Li, H., Liu, H. & Gu, J. A lightweight YOLOv3 algorithm used for safety helmet detection. Sci. Rep. 12, 10981 (2022).
20. Zhang, Y.-J., Xiao, F.-S. & Lu, Z.-M. Helmet wearing state detection based on improved YOLOv5s. Sensors 22, 9843 (2022).
21. Li, H., Wu, D., Zhang, W. & Xiao, C. YOLO-PL: Helmet wearing detection algorithm based on improved YOLOv4. Digit. Signal
Process., 104283 (2023).
22. Xia, Z. & Xiao, H. A study of campus environment security cap detection system based on YOLO v4. Network Security Technology
and Applications, 40–41 (2021).
23. Yi, Z., Wu, G., Pan, X. & Tao, J. in 2021 33rd Chinese Control and Decision Conference (CCDC). 769–773 (IEEE).
24. Dai, B., Nie, Y., Cui, W., Liu, R. & Zheng, Z. In Proceedings of the 2nd International Conference on Artificial Intelligence and Advanced
Manufacture. 95–99.
25. Tan, S., Lu, G., Jiang, Z. & Huang, L. In 2021 IEEE International Conference on Intelligence and Safety for Robotics (ISR). 330–333
(IEEE).
26. Fang, Q. et al. Detecting non-hardhat-use by a deep learning method from far-field surveillance videos. Autom. Constr. 85, 1–9
(2018).
27. Huang, H., Liang, Q., Luo, D. & Lee, D. H. Attention-enhanced one-stage algorithm for traffic sign detection and recognition. J.
Sens. 2022 (2022).
28. Guo, M.-H., Liu, Z.-N., Mu, T.-J. & Hu, S.-M. Beyond self-attention: External attention using two linear layers for visual tasks.
IEEE Trans. Pattern Anal. Mach. Intell. 45, 5436–5447 (2022).
29. Huang, H., Chen, Z., Zou, Y., Lu, M. & Chen, C. Channel prior convolutional attention for medical image segmentation. arXiv
preprint arXiv:​2306.​05196 (2023).
30. Yu, Y., Zhang, Y., Cheng, Z., Song, Z. & Tang, C. MCA: Multidimensional collaborative attention in deep convolutional neural
networks for image recognition. Eng. Appl. Artif. Intell. 126, 107079 (2023).
31. Gevorgyan, Z. SIoU loss: More powerful learning for bounding box regression. arXiv preprint arXiv:​2205.​12740 (2022).
32. Zhang, S. et al. Diag-IoU Loss for Object Detection. IEEE Transactions on Circuits and Systems for Video Technology (2023).

Acknowledgements
This work was supported by the National Natural Science Foundation of China [No. 52175379] and the Liaoning
Provincial Science and Technology Department [No. 2022JH2/101300268].

Author contributions
X.D.S. and T.K.Z. designed the concept and the experimental approach. T.K.Z. developed the model and
performed the experiments. T.K.Z. wrote the first draft of the manuscript. X.D.S. and W.G.Y. reviewed the
manuscript and corrected the manuscript.

Competing interests
The authors declare no competing interests.

Additional information
Correspondence and requests for materials should be addressed to W.Y.
Reprints and permissions information is available at [Link]/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Scientific Reports | (2024) 14:17550 | [Link] 13

Vol.:(0123456789)
 [Link]/scientificreports/

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-

NoDerivatives 4.0 International License, which permits any non-commercial use, sharing,
distribution and reproduction in any medium or format, as long as you give appropriate credit to the original
author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the
licensed material. You do not have permission under this licence to share adapted material derived from this
article or parts of [Link] images or other third party material in this article are included in the article’s Creative
Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the
article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the
permitted use, you will need to obtain permission directly from the copyright [Link] view a copy of this
licence, visit [Link]

© The Author(s) 2024

Scientific Reports | (2024) 14:17550 | [Link] 14

Vol:.(1234567890)