<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE article PUBLIC "-//NLM//DTD Journal Publishing with OASIS Tables v3.0 20080202//EN" "journalpub-oasis3.dtd">
<article xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:oasis="http://docs.oasis-open.org/ns/oasis-exchange/table" xml:lang="en" dtd-version="3.0" article-type="research-article">
  <front>
    <journal-meta><journal-id journal-id-type="publisher">MS</journal-id><journal-title-group>
    <journal-title>Mechanical Sciences</journal-title>
    <abbrev-journal-title abbrev-type="publisher">MS</abbrev-journal-title><abbrev-journal-title abbrev-type="nlm-ta">Mech. Sci.</abbrev-journal-title>
  </journal-title-group><issn pub-type="epub">2191-916X</issn><publisher>
    <publisher-name>Copernicus Publications</publisher-name>
    <publisher-loc>Göttingen, Germany</publisher-loc>
  </publisher></journal-meta>
    <article-meta>
      <article-id pub-id-type="doi">10.5194/ms-13-979-2022</article-id><title-group><article-title>A real-time and accurate detection approach for bucket teeth falling off
based on improved YOLOX</article-title><alt-title>A real-time and accurate detection approach for bucket teeth falling off</alt-title>
      </title-group><?xmltex \runningtitle{A real-time and accurate detection approach for bucket teeth falling off}?><?xmltex \runningauthor{J. Lu and Y. Liu}?>
      <contrib-group>
        <contrib contrib-type="author" corresp="no">
          <name><surname>Lu</surname><given-names>Jinnan</given-names></name>
          
        </contrib>
        <contrib contrib-type="author" corresp="yes">
          <name><surname>Liu</surname><given-names>Yang</given-names></name>
          <email>121185350@qq.com</email>
        </contrib>
        <aff id="aff1"><institution>School of Mechanical Engineering, Liaoning Technical University, Fuxin
123000, China</institution>
        </aff>
      </contrib-group>
      <author-notes><corresp id="corr1">Yang Liu (121185350@qq.com)</corresp></author-notes><pub-date><day>25</day><month>November</month><year>2022</year></pub-date>
      
      <volume>13</volume>
      <issue>2</issue>
      <fpage>979</fpage><lpage>990</lpage>
      <history>
        <date date-type="received"><day>16</day><month>July</month><year>2022</year></date>
           <date date-type="rev-recd"><day>12</day><month>October</month><year>2022</year></date>
           <date date-type="accepted"><day>9</day><month>November</month><year>2022</year></date>
      </history>
      <permissions>
        <copyright-statement>Copyright: © 2022 Jinnan Lu</copyright-statement>
        <copyright-year>2022</copyright-year>
      <license license-type="open-access"><license-p>This work is licensed under the Creative Commons Attribution 4.0 International License. To view a copy of this licence, visit <ext-link ext-link-type="uri" xlink:href="https://creativecommons.org/licenses/by/4.0/">https://creativecommons.org/licenses/by/4.0/</ext-link></license-p></license></permissions><self-uri xlink:href="https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022.html">This article is available from https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022.html</self-uri><self-uri xlink:href="https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022.pdf">The full text article is available as a PDF file from https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022.pdf</self-uri>
      <abstract><title>Abstract</title>

      <p id="d1e84">An electric shovel is a bucket-equipped mining excavator widely used in open-pit mining today. The prolonged direct impact between the
bucket teeth and the ore during the mining process will cause the teeth to
loosen prematurely or even break, resulting in unplanned downtime and
productivity losses. To solve this problem, we propose a real-time and
accurate detection algorithm of bucket teeth falling off based on improved
YOLOX. Firstly, to solve the problem of poor detection effect caused by uneven illumination, the dilated convolution attention mechanism is added to
enhance the feature expression ability of the target in complex backgrounds
so as to improve the detection accuracy of the target. Secondly, considering
the high computing cost and large delay of the embedded device, the deep
separable convolution is used to replace the traditional convolution in the
feature pyramid network, and the model compression strategy is used to prune
the redundant channels in the network, reduce the model volume, and improve
the detection speed. The performance test is carried out on the
self-constructed dataset of WK-10 electric shovel. The experimental results show that, compared with the YOLOX model, the mean average precision of the
algorithm in this paper reaches 95.26 %, only 0.33 % lower, while the
detection speed is 50.8 fps, 11.9 fps higher, and the model volume is 28.42 MB,
which is reduced to 29.46 % of the original. Compared with many other
existing methods, the target detection algorithm proposed in this paper has
the advantages of higher precision, smaller model volume, and faster speed.
It can meet the requirements of real-time and accurate detection of the
bucket teeth falling off.</p>
  </abstract>
    </article-meta>
  </front>
<body>
      

<sec id="Ch1.S1" sec-type="intro">
  <label>1</label><title>Introduction</title>
      <p id="d1e96">With the continuous progress of science and technology, the mining of
mineral resources is developing in a more complex, automated, and intelligent
direction (Wang et al., 2018; Chen et al., 2017). Electric shovels are the
main excavation equipment in mine production. The prolonged direct impact of
the bucket teeth on ore during excavation can cause the bucket teeth to be
subject to strong impact, friction, and bending force, which will cause the
bucket teeth to loosen or even break unexpectedly from the teeth base,
resulting in serious economic losses and waste of human and material
resources. Firstly, the failure of the bucket teeth will increase the
frictional resistance of the electric shovel during the excavation
operation, making it difficult for the bucket to excavate, thereby affecting
the efficiency of the excavation operation and accelerating the loss of the
electric shovel (Akter and Basak, 2022). Secondly, since most of bucket teeth
are made of high manganese steel or alloy steel, the hardness is much higher
than that of ore and other materials. Once the broken bucket teeth are
unloaded into the crusher together with ore, it will cause crusher failure
and affect the whole mine crushing production line (Singla et al., 2014).
The looseness or fracture of the bucket teeth is not detected in a timely
and accurate manner, which will lead to serious safety hazards. Therefore,
it is necessary to propose a high-precision, high real-time bucket teeth
target detection algorithm.</p>
      <p id="d1e99">In recent years, domestic and foreign scholars and research institutions
have been trying to use traditional image processing algorithms to realise
intelligent detection of the bucket teeth on existing video monitoring
systems. For example, He et al. (2011) applied support vector machine (SVM)
on the image's gray gradient histogram and applied structural feature
constraints to the detection results based on the positional relationships
between bucket teeth to improve the accuracy of object detection. Alma'aitah
and Hassanein (2014) used Sprouts sensor network platform to detect bucket teeth,
and embedded wireless power transmission units and auxiliary antennas in
bucket teeth to locate broken bucket teeth. Lim et al. (2016) determined the
position of bucket teeth through the image samples of the bucket teeth
movement process, used the optical flow method and frame difference method to
track and correct the bucket teeth position, and finally used the image grayscale
feature to judge the lack of bucket teeth. However, traditional algorithms
are often based on the surface layer information of images, and need
continuous parameter updates and template matching to obtain optimal
features. Each feature template defines a specific task, which cannot meet
the real-time and accurate detection requirement on the bucket teeth
falling off.</p>
      <p id="d1e102">With the continuous development of deep learning, breakthroughs have been
made in the field of computer vision. The convolutional neural network has
become an important means of image processing, and it has been introduced
into the field of target detection, making the real-time and accurate
detection of the bucket teeth falling off possible (Sun et al., 2022; Shen
et al., 2022; Ji et al., 2022). Ji et al. (2021) proposed an intelligent
monitoring system for missing bucket teeth, and used Faster R-CNN for bucket
teeth target detection, which improved the detection accuracy of bucket
teeth. Based on the DeepLabV3<inline-formula><mml:math id="M1" display="inline"><mml:mo>+</mml:mo></mml:math></inline-formula> detection algorithm, Liu et al. (2021)
completed the semantic segmentation of the image and obtained the profile
information of the bucket teeth by improving the loss function and adding
the attention mechanism. However, due to the deep network model and the
large amount of parameters, it is difficult to deploy to embedded mobile
devices. Therefore, under the premise of ensuring detection accuracy, it is
very important to compress and lighten the convolutional neural network
model (Zheng et al., 2022).</p>
      <p id="d1e112">At present, the development of compression methods and lightweight models of
convolutional neural networks are rapid. The model compression methods
mainly include knowledge distillation, low-rank decomposition, quantisation,
and channel pruning. Li et al. (2019) introduced knowledge distillation into
convolutional neural networks to achieve model compression and distilled
redundant network models to simple models; Denton et al. (2014) used
approximate matrix products to replace the weight matrix of the
convolutional neural networks, and decomposed the convolutional neural
networks with low-rank, so as to reduce the amount of parameters. Liu et al. (2017) used the channel pruning strategy to prune redundant parameters of
the convolutional layer, which effectively reduced the amount of computation
and parameters while ensuring the integrity of the convolutional neural
network. For lightweight models, single-shot multi-box detector (SSD) network structure based on deep feature fusion was proposed (Bai et al., 2022; Huang et al., 2022), which enhances
the complementarity of high-level features and improves the detection
performance of the SSD network for targets of different scales. Redmon et
al. (2016) proposed the YOLO (you only look once) network, which takes the
object detection process as a regression task and predicts multiple bounding
boxes directly from the input image. In response to the problem that YOLOv1
only supports the same resolution of the input image with the training
image, YOLOv2 (Redmon and Farhadi, 2017) improves the network structure and
position prediction mechanism, which can adapt the model to multi-scale
image input. YOLOv3 (J. Huang et al., 2021) uses a logical classifier to
achieve multi-label classification but due to the insufficient feature
extraction capability of the network itself, the accuracy of small target
recognition is low. The subsequent improved YOLOv4 (Tan et al., 2021) on
this basis uses the feature pyramid network structure to integrate the deep
features and shallow features to solve the problem of small target
recognition. In 2021, Ge et al. (2021) proposed a high-performance detection
model YOLOX, which does not use anchor and dynamically matches samples for
targets of different sizes. In the previous version of YOLO, the decoupling
head was integrated; that is, classification and regression were implemented
in a <inline-formula><mml:math id="M2" display="inline"><mml:mrow><mml:mn mathvariant="normal">1</mml:mn><mml:mo>×</mml:mo><mml:mn mathvariant="normal">1</mml:mn></mml:mrow></mml:math></inline-formula> convolution (L. Huang et al., 2021). YOLOX, it was believed,
adversely affected the identification of the network, so in YOLOX, YoloHead
was divided into two parts and implemented separately and finally
integrated during the prediction phase. Compared with YOLOv3 and YOLOv4, the
detection accuracy and speed of YOLOX model have been improved, the
end-to-end deployment is more flexible, and the model size and memory
occupation are more suitable for embedded mobile devices. However, the
overall detection performance of YOLOX algorithm needs to be further
improved.</p>
      <p id="d1e128">In order to achieve real-time and accurate detection of bucket teeth falling
off, an improved YOLOX algorithm is proposed in this paper. The main
contributions and innovations of this paper are as follows: (1) firstly, the
dilated convolution attention mechanism is added to enhance the saliency of
the target in the complex background; secondly, the deep separable
convolution is used to replace traditional convolution in the feature
pyramid network; finally, the model compression strategy is used to prune
redundant channels in the network, reduce model volume, and improve detection
speed. (2) WK-10 and WK-35 electric shovel datasets were collected under real working conditions for training deep learning models. (3) Comparing
multiple sets of experiments and evaluating the detection results, the mean
average precision of our algorithm can reach 95.26 %, the detection speed
is 50.8 fps, and the model volume is 28.42 MB. The experimental results on the
migration dataset show that the improved model has strong generalisation
ability.</p>
</sec>
<sec id="Ch1.S2">
  <label>2</label><title>Bucket teeth intelligent monitoring system</title>
      <p id="d1e139">The bucket teeth intelligent monitoring system is mainly composed of image
acquisition module, Ethernet communication module, detection module, and
display module. As shown in Fig. 1, firstly, a top-view camera is installed
under the crown wheel of the shovel boom and a front camera is installed on
the rotary platform, and thus the image information of bucket teeth is
collected to the greatest extent by using the complementary method of two
monitoring spots. Then, it is transmitted to the embedded Xavier processor
through Ethernet to detect bucket teeth based on the deep learning
algorithm. Finally, the detection results are transmitted to the industrial
touch all-in-one machine through UDP for display.</p>
      <p id="d1e142">In view of the bad working environment of the electric shovel and the poor
performance of the embedded equipment, a target detection algorithm that
meets the requirements of accuracy and speed is selected, and the problem of
bucket teeth falling off is converted into the problem of detecting the
type, number, and position of the input image target.</p>

      <?xmltex \floatpos{t}?><fig id="Ch1.F1" specific-use="star"><?xmltex \currentcnt{1}?><?xmltex \def\figurename{Figure}?><label>Figure 1</label><caption><p id="d1e147">Intelligent monitoring system of bucket teeth.</p></caption>
        <?xmltex \igopts{width=497.923228pt}?><graphic xlink:href="https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022-f01.png"/>

      </fig>

</sec>
<sec id="Ch1.S3">
  <label>3</label><title>YOLOX algorithm</title>
      <p id="d1e164">YOLOX is an end-to-end target detection algorithm which adds many
optimisation strategies on the basis of YOLOv3 to greatly improve the
detection performance. Its network structure is shown in Fig. 2. YOLOX is
divided into four main parts which are input, feature extraction backbone
network, neck, and prediction module. The input uses two methods to enhance
features: mosaic data enhancement and mixUP. The former one aims to improve
the model detection capacity on small objects in the image. The input data
are sliced by Focus before entering the feature extraction backbone network.
The Focus structure performs self-copying and slicing operations on the
image to obtain a down-sampling feature map with twice the information (Wang
et al., 2022). The CBS (composed of convolution, batch normalisation, and
SiLU activation function) module is used for feature extraction, and the cross-stage partial (CSP) structure is used to optimise the gradient information in the network,
reduce the amount of inference calculation, and speed up the calculation of
the network. The spatial pyramid pooling (SPP) layer performs feature extraction on the maximum pooling of different pooling kernel sizes, improving the receptive field of
the network, and enhancing the nonlinear expression ability of the network.
Neck improves the feature pyramid network <inline-formula><mml:math id="M3" display="inline"><mml:mo>+</mml:mo></mml:math></inline-formula> path aggregation network (FPN <inline-formula><mml:math id="M4" display="inline"><mml:mo>+</mml:mo></mml:math></inline-formula> PAN) structure to enhance the feature fusion of the network and the transmission of inference information in the network (Bello
et al., 2021). Prediction replaces the original classification and
regression layer couple head with decouple head, adopting anchor-free frame
mechanism and simOTA label assignment method to solve the optimal
transmission problem and reduce the amount of model parameters, which
significantly improves the regression speed and accuracy of the network.</p>

      <?xmltex \floatpos{t}?><fig id="Ch1.F2" specific-use="star"><?xmltex \currentcnt{2}?><?xmltex \def\figurename{Figure}?><label>Figure 2</label><caption><p id="d1e183">YOLOX network structure.</p></caption>
        <?xmltex \igopts{width=426.791339pt}?><graphic xlink:href="https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022-f02.png"/>

      </fig>

<?xmltex \hack{\newpage}?>
</sec>
<sec id="Ch1.S4">
  <label>4</label><title>Improved YOLOX algorithm</title>
      <p id="d1e202">The improved YOLOX network structure is shown in Fig. 3. First, the images
are fed into the backbone network, and the feature maps are obtained after
several convolution operations and pooling operations. Then, the obtained
feature map is fed into the feature pyramid structure, and the feature
extraction is enhanced by using the FPN structure and PAN structure, and the
up-sampling and down-sampling operations are completed at the same time to
combine the shallow feature map semantic information with the deep feature
map semantic information. Finally, the target detection work is completed in
the YoloHead structure.</p>

      <?xmltex \floatpos{t}?><fig id="Ch1.F3" specific-use="star"><?xmltex \currentcnt{3}?><?xmltex \def\figurename{Figure}?><label>Figure 3</label><caption><p id="d1e207">Improved YOLOX network structure.</p></caption>
        <?xmltex \igopts{width=497.923228pt}?><graphic xlink:href="https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022-f03.png"/>

      </fig>

<sec id="Ch1.S4.SS1">
  <label>4.1</label><title>Dilated CBAM (convolutional block attention module)</title>
      <p id="d1e223">When the electric shovel is working, the background often has uneven
illumination, which makes the target features inconspicuous. By adding an
attention mechanism, the interference caused by the environment can be
alleviated and the detection accuracy can be improved. At present, most of
the existing attention mechanisms are channel or spatial attention
mechanisms, which cannot make the network pay attention to the most
contributing channel and the most contributing space at the same time. To
take both into account, this paper adds a channel and spatial CBAM
(convolutional block attention module) (Zhang et al., 2020). This module
takes up less computation and can significantly improve the detection
performance. However, in order to increase the network receptive field and
improve the connection between target and background information, we change
the convolutional layer to a dilated convolutional layer. However, Yu et al. (2017) pointed out that dilated convolution can cause gridding artifacts
when the frequency of feature maps is higher than the sampling frequency of
dilated convolution. To remove the gridding artifacts, two convolutional
layers with a smaller dilation rate are added after the convolutional layer
with dilation rate of 4; the structure is shown in Fig. 4. Among them, a
dilated CBAM module is added after each CSP module, and up-sampling and
down-sampling to calculate the weight information of the feature map in the
channel and spatial position. According to the weight distribution, the
ability of the network to distinguish target from background is improved,
and a good detection result is achieved in practical applications.</p>

      <?xmltex \floatpos{t}?><fig id="Ch1.F4" specific-use="star"><?xmltex \currentcnt{4}?><?xmltex \def\figurename{Figure}?><label>Figure 4</label><caption><p id="d1e228">Dilated CBAM module.</p></caption>
          <?xmltex \igopts{width=455.244094pt}?><graphic xlink:href="https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022-f04.png"/>

        </fig>

      <p id="d1e237">The dilated CBAM module first sends the input features <inline-formula><mml:math id="M5" display="inline"><mml:mi>F</mml:mi></mml:math></inline-formula> into the
channel attention module, which obtains the information of each channel
through average pooling and max pooling, superimposes the obtained
parameters through a MLP (multi-layer perceptron) to generate different
channel attention map, and activates by the sigmoid function to the final
channel attention map. The calculation formula is shown as Eq. (1):
<?xmltex \hack{\newpage}?>
            <disp-formula id="Ch1.E1" content-type="numbered"><label>1</label><mml:math id="M6" display="block"><mml:mrow><?xmltex \hack{\hbox\bgroup\fontsize{9.5}{9.5}\selectfont$\displaystyle}?><mml:mtable class="split" rowspacing="0.2ex" displaystyle="true" columnalign="right left"><mml:mtr><mml:mtd><mml:mrow><mml:msub><mml:mi>M</mml:mi><mml:mi mathvariant="normal">c</mml:mi></mml:msub><mml:mfenced open="(" close=")"><mml:mi>F</mml:mi></mml:mfenced></mml:mrow></mml:mtd><mml:mtd><mml:mrow><mml:mo>=</mml:mo><mml:mi mathvariant="italic">σ</mml:mi><mml:mfenced open="(" close=")"><mml:mrow><mml:mi mathvariant="normal">MLP</mml:mi><mml:mfenced open="(" close=")"><mml:mrow><mml:mi mathvariant="normal">AvgPool</mml:mi><mml:mfenced close=")" open="("><mml:mi>F</mml:mi></mml:mfenced></mml:mrow></mml:mfenced><mml:mo>+</mml:mo><mml:mi mathvariant="normal">MLP</mml:mi><mml:mfenced close=")" open="("><mml:mrow><mml:mi mathvariant="normal">MaxPool</mml:mi><mml:mfenced close=")" open="("><mml:mi>F</mml:mi></mml:mfenced></mml:mrow></mml:mfenced></mml:mrow></mml:mfenced></mml:mrow></mml:mtd></mml:mtr><mml:mtr><mml:mtd/><mml:mtd><mml:mrow><mml:mo>=</mml:mo><mml:mi mathvariant="italic">σ</mml:mi><mml:mfenced open="(" close=")"><mml:mrow><mml:msub><mml:mi>W</mml:mi><mml:mn mathvariant="normal">1</mml:mn></mml:msub><mml:mfenced open="(" close=")"><mml:mrow><mml:msub><mml:mi>W</mml:mi><mml:mn mathvariant="normal">0</mml:mn></mml:msub><mml:mfenced close=")" open="("><mml:mrow><mml:msubsup><mml:mi>F</mml:mi><mml:mi mathvariant="normal">avg</mml:mi><mml:mi mathvariant="normal">c</mml:mi></mml:msubsup></mml:mrow></mml:mfenced></mml:mrow></mml:mfenced><mml:mo>+</mml:mo><mml:msub><mml:mi>W</mml:mi><mml:mn mathvariant="normal">1</mml:mn></mml:msub><mml:mfenced open="(" close=")"><mml:mrow><mml:msub><mml:mi>W</mml:mi><mml:mn mathvariant="normal">0</mml:mn></mml:msub><mml:mfenced open="(" close=")"><mml:mrow><mml:msubsup><mml:mi>F</mml:mi><mml:mo>max⁡</mml:mo><mml:mi mathvariant="normal">c</mml:mi></mml:msubsup></mml:mrow></mml:mfenced></mml:mrow></mml:mfenced></mml:mrow></mml:mfenced><mml:mo>,</mml:mo></mml:mrow></mml:mtd></mml:mtr></mml:mtable><?xmltex \hack{$\egroup}?></mml:mrow></mml:math></disp-formula>
          where <inline-formula><mml:math id="M7" display="inline"><mml:mrow><mml:mi mathvariant="italic">σ</mml:mi><mml:mo>(</mml:mo><mml:mo>⋅</mml:mo><mml:mo>)</mml:mo></mml:mrow></mml:math></inline-formula> is the sigmoid nonlinear activation function;
<inline-formula><mml:math id="M8" display="inline"><mml:mrow><mml:msub><mml:mi>W</mml:mi><mml:mn mathvariant="normal">0</mml:mn></mml:msub></mml:mrow></mml:math></inline-formula> and <inline-formula><mml:math id="M9" display="inline"><mml:mrow><mml:msub><mml:mi>W</mml:mi><mml:mn mathvariant="normal">1</mml:mn></mml:msub></mml:mrow></mml:math></inline-formula> are the hidden layer weights and output layer weights,
respectively; and <inline-formula><mml:math id="M10" display="inline"><mml:mrow><mml:msubsup><mml:mi>F</mml:mi><mml:mi mathvariant="normal">avg</mml:mi><mml:mi mathvariant="normal">c</mml:mi></mml:msubsup></mml:mrow></mml:math></inline-formula> and <inline-formula><mml:math id="M11" display="inline"><mml:mrow><mml:msubsup><mml:mi>F</mml:mi><mml:mo>max⁡</mml:mo><mml:mi mathvariant="normal">c</mml:mi></mml:msubsup></mml:mrow></mml:math></inline-formula> are the average pooling and
max pooling operations of the channel attention module, respectively.</p>
      <p id="d1e417">After the feature map <inline-formula><mml:math id="M12" display="inline"><mml:mrow><mml:msub><mml:mi>F</mml:mi><mml:mi>x</mml:mi></mml:msub></mml:mrow></mml:math></inline-formula> is sent to the spatial attention module, the
spatial information is gathered along the channel dimension through average
pooling and max pooling to generate a spatial feature map. Then, through the
dilated convolution calculations with dilation rates of 4, 2, and 1, and
activation of the sigmoid function, the spatial attention feature is
obtained. The calculation formula is shown as Eq. (2):
            <disp-formula id="Ch1.E2" content-type="numbered"><label>2</label><mml:math id="M13" display="block"><mml:mtable class="split" rowspacing="0.2ex" displaystyle="true" columnalign="right left"><mml:mtr><mml:mtd><mml:mrow><mml:msub><mml:mi>M</mml:mi><mml:mi mathvariant="normal">s</mml:mi></mml:msub><mml:mfenced close=")" open="("><mml:mrow><mml:msub><mml:mi>F</mml:mi><mml:mi>x</mml:mi></mml:msub></mml:mrow></mml:mfenced></mml:mrow></mml:mtd><mml:mtd><mml:mrow><mml:mo>=</mml:mo><mml:mi mathvariant="italic">σ</mml:mi><mml:mo mathsize="2.0em">(</mml:mo><mml:msubsup><mml:mi>f</mml:mi><mml:mn mathvariant="normal">4</mml:mn><mml:mrow><mml:mn mathvariant="normal">3</mml:mn><mml:mo>×</mml:mo><mml:mn mathvariant="normal">3</mml:mn></mml:mrow></mml:msubsup><mml:mo mathsize="1.1em">(</mml:mo><mml:msubsup><mml:mi>f</mml:mi><mml:mn mathvariant="normal">2</mml:mn><mml:mrow><mml:mn mathvariant="normal">3</mml:mn><mml:mo>×</mml:mo><mml:mn mathvariant="normal">3</mml:mn></mml:mrow></mml:msubsup><mml:mo mathsize="1.1em">(</mml:mo><mml:msubsup><mml:mi>f</mml:mi><mml:mn mathvariant="normal">1</mml:mn><mml:mrow><mml:mn mathvariant="normal">3</mml:mn><mml:mo>×</mml:mo><mml:mn mathvariant="normal">3</mml:mn></mml:mrow></mml:msubsup><mml:mo mathsize="1.1em">(</mml:mo><mml:mi mathvariant="normal">AvgPool</mml:mi><mml:mfenced open="(" close=")"><mml:mrow><mml:msub><mml:mi>F</mml:mi><mml:mi>x</mml:mi></mml:msub></mml:mrow></mml:mfenced><mml:mo>;</mml:mo></mml:mrow></mml:mtd></mml:mtr><mml:mtr><mml:mtd/><mml:mtd><mml:mrow><mml:mspace width="0.25em" linebreak="nobreak"/><mml:mspace linebreak="nobreak" width="1em"/><mml:mi mathvariant="normal">MaxPool</mml:mi><mml:mfenced close=")" open="("><mml:mrow><mml:msub><mml:mi>F</mml:mi><mml:mi>x</mml:mi></mml:msub></mml:mrow></mml:mfenced><mml:mo mathsize="1.1em">)</mml:mo><mml:mo mathsize="1.1em">)</mml:mo><mml:mo mathsize="1.1em">)</mml:mo><mml:mo mathsize="2.0em">)</mml:mo></mml:mrow></mml:mtd></mml:mtr><mml:mtr><mml:mtd/><mml:mtd><mml:mrow><mml:mo>=</mml:mo><mml:mi mathvariant="italic">σ</mml:mi><mml:mfenced open="(" close=")"><mml:mrow><mml:msubsup><mml:mi>f</mml:mi><mml:mn mathvariant="normal">4</mml:mn><mml:mrow><mml:mn mathvariant="normal">3</mml:mn><mml:mo>×</mml:mo><mml:mn mathvariant="normal">3</mml:mn></mml:mrow></mml:msubsup><mml:mfenced open="(" close=")"><mml:mrow><mml:msubsup><mml:mi>f</mml:mi><mml:mn mathvariant="normal">2</mml:mn><mml:mrow><mml:mn mathvariant="normal">3</mml:mn><mml:mo>×</mml:mo><mml:mn mathvariant="normal">3</mml:mn></mml:mrow></mml:msubsup><mml:mfenced close=")" open="("><mml:mrow><mml:msubsup><mml:mi>f</mml:mi><mml:mn mathvariant="normal">1</mml:mn><mml:mrow><mml:mn mathvariant="normal">3</mml:mn><mml:mo>×</mml:mo><mml:mn mathvariant="normal">3</mml:mn></mml:mrow></mml:msubsup><mml:mfenced close=")" open="("><mml:mrow><mml:msubsup><mml:mi>F</mml:mi><mml:mi mathvariant="normal">avg</mml:mi><mml:mi mathvariant="normal">s</mml:mi></mml:msubsup><mml:mo>;</mml:mo><mml:msubsup><mml:mi>F</mml:mi><mml:mo>max⁡</mml:mo><mml:mi mathvariant="normal">s</mml:mi></mml:msubsup></mml:mrow></mml:mfenced></mml:mrow></mml:mfenced></mml:mrow></mml:mfenced></mml:mrow></mml:mfenced><mml:mo>,</mml:mo></mml:mrow></mml:mtd></mml:mtr></mml:mtable></mml:math></disp-formula>
          where <inline-formula><mml:math id="M14" display="inline"><mml:mrow><mml:msubsup><mml:mi>f</mml:mi><mml:mn mathvariant="normal">4</mml:mn><mml:mrow><mml:mn mathvariant="normal">3</mml:mn><mml:mo>×</mml:mo><mml:mn mathvariant="normal">3</mml:mn></mml:mrow></mml:msubsup></mml:mrow></mml:math></inline-formula>, <inline-formula><mml:math id="M15" display="inline"><mml:mrow><mml:msubsup><mml:mi>f</mml:mi><mml:mn mathvariant="normal">2</mml:mn><mml:mrow><mml:mn mathvariant="normal">3</mml:mn><mml:mo>×</mml:mo><mml:mn mathvariant="normal">3</mml:mn></mml:mrow></mml:msubsup></mml:mrow></mml:math></inline-formula>, and <inline-formula><mml:math id="M16" display="inline"><mml:mrow><mml:msubsup><mml:mi>f</mml:mi><mml:mn mathvariant="normal">1</mml:mn><mml:mrow><mml:mn mathvariant="normal">3</mml:mn><mml:mo>×</mml:mo><mml:mn mathvariant="normal">3</mml:mn></mml:mrow></mml:msubsup></mml:mrow></mml:math></inline-formula> are the dilated convolution calculations with the convolution kernel size
of <inline-formula><mml:math id="M17" display="inline"><mml:mrow><mml:mn mathvariant="normal">3</mml:mn><mml:mo>×</mml:mo><mml:mn mathvariant="normal">3</mml:mn></mml:mrow></mml:math></inline-formula> and dilation rates of 4, 2, and 1, respectively; and
<inline-formula><mml:math id="M18" display="inline"><mml:mrow><mml:msubsup><mml:mi>F</mml:mi><mml:mi mathvariant="normal">avg</mml:mi><mml:mi mathvariant="normal">s</mml:mi></mml:msubsup></mml:mrow></mml:math></inline-formula> and <inline-formula><mml:math id="M19" display="inline"><mml:mrow><mml:msubsup><mml:mi>F</mml:mi><mml:mo>max⁡</mml:mo><mml:mi mathvariant="normal">s</mml:mi></mml:msubsup></mml:mrow></mml:math></inline-formula> are the average pooling and max pooling
operations of the spatial attention module, respectively. In addition, a
dilated CBAM module is added after CSP unit in the CSPX structure; its
structure is shown in Fig. 5.</p>

      <?xmltex \floatpos{t}?><fig id="Ch1.F5"><?xmltex \currentcnt{5}?><?xmltex \def\figurename{Figure}?><label>Figure 5</label><caption><p id="d1e702">Improvement of CSPX. <bold>(a)</bold> CSPX of the original model. <bold>(b)</bold> CSPX of
the algorithm in this paper.</p></caption>
          <?xmltex \igopts{width=213.395669pt}?><graphic xlink:href="https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022-f05.png"/>

        </fig>

</sec>
<sec id="Ch1.S4.SS2">
  <label>4.2</label><title>Deep separable convolution</title>
      <p id="d1e726">Deep separable convolution (Buiu et al., 2020; Yun et al., 2022) is the core
of the lightweight network MobileNet, which can achieve the same feature
extraction effect as traditional convolution. Replacing the traditional
convolution in PANet (path aggregation network) with the deep separable
convolution can improve the problem that the running speed cannot be
improved due to the high amount of computation in the traditional
convolution, and at the same time reduce the model volume and network
parameters. Its structure is shown in Fig. 6.</p>

      <?xmltex \floatpos{t}?><fig id="Ch1.F6"><?xmltex \currentcnt{6}?><?xmltex \def\figurename{Figure}?><label>Figure 6</label><caption><p id="d1e731">Deep separable convolution structure.</p></caption>
          <?xmltex \igopts{width=236.157874pt}?><graphic xlink:href="https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022-f06.png"/>

        </fig>

      <p id="d1e740">The deep separable convolution decomposes the convolution into one
channel-by-channel convolution and one point-by-point convolution, which
reduces the computation and model volume of the network and significantly
improves the detection speed of the network. It is assumed that the input
feature map size is <inline-formula><mml:math id="M20" display="inline"><mml:mrow><mml:msub><mml:mi>L</mml:mi><mml:mi>x</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:msub><mml:mi>L</mml:mi><mml:mi>y</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:mi>M</mml:mi></mml:mrow></mml:math></inline-formula>, the output feature map
size is<inline-formula><mml:math id="M21" display="inline"><mml:mrow><mml:mspace width="0.125em" linebreak="nobreak"/><mml:msub><mml:mi>L</mml:mi><mml:mi>x</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:msub><mml:mi>L</mml:mi><mml:mi>y</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:mi>N</mml:mi></mml:mrow></mml:math></inline-formula>, and the convolution kernel size is <inline-formula><mml:math id="M22" display="inline"><mml:mrow><mml:msub><mml:mi>D</mml:mi><mml:mi>k</mml:mi></mml:msub></mml:mrow></mml:math></inline-formula>. Among them, <inline-formula><mml:math id="M23" display="inline"><mml:mrow><mml:msub><mml:mi>L</mml:mi><mml:mi>x</mml:mi></mml:msub></mml:mrow></mml:math></inline-formula> and <inline-formula><mml:math id="M24" display="inline"><mml:mrow><mml:msub><mml:mi>L</mml:mi><mml:mi>y</mml:mi></mml:msub></mml:mrow></mml:math></inline-formula> are the width and height of the
feature map,  <inline-formula><mml:math id="M25" display="inline"><mml:mi>M</mml:mi></mml:math></inline-formula> and <inline-formula><mml:math id="M26" display="inline"><mml:mi>N</mml:mi></mml:math></inline-formula> are the number of input and output channels, and the
calculation of the traditional convolution is shown as Eq. (3):
            <disp-formula id="Ch1.E3" content-type="numbered"><label>3</label><mml:math id="M27" display="block"><mml:mrow><mml:msub><mml:mi>L</mml:mi><mml:mi>x</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:msub><mml:mi>L</mml:mi><mml:mi>y</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:mi>M</mml:mi><mml:mo>×</mml:mo><mml:mi>N</mml:mi><mml:mo>×</mml:mo><mml:msub><mml:mi>D</mml:mi><mml:mi>k</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:msub><mml:mi>D</mml:mi><mml:mi>k</mml:mi></mml:msub><mml:mo>.</mml:mo></mml:mrow></mml:math></disp-formula>
          The calculations of channel-by-channel convolution and point-by-point
convolution are

                <disp-formula specific-use="gather" content-type="numbered"><mml:math id="M28" display="block"><mml:mtable displaystyle="true"><mml:mlabeledtr id="Ch1.E4"><mml:mtd><mml:mtext>4</mml:mtext></mml:mtd><mml:mtd><mml:mrow><mml:mstyle displaystyle="true" class="stylechange"/><mml:msub><mml:mi>L</mml:mi><mml:mi>x</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:msub><mml:mi>L</mml:mi><mml:mi>y</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:mi>M</mml:mi><mml:mo>×</mml:mo><mml:msub><mml:mi>D</mml:mi><mml:mi>k</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:msub><mml:mi>D</mml:mi><mml:mi>k</mml:mi></mml:msub></mml:mrow></mml:mtd></mml:mlabeledtr><mml:mlabeledtr id="Ch1.E5"><mml:mtd><mml:mtext>5</mml:mtext></mml:mtd><mml:mtd><mml:mrow><mml:mstyle displaystyle="true" class="stylechange"/><mml:msub><mml:mi>L</mml:mi><mml:mi>x</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:msub><mml:mi>L</mml:mi><mml:mi>y</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:mi>M</mml:mi><mml:mo>×</mml:mo><mml:mi>N</mml:mi><mml:mo>.</mml:mo></mml:mrow></mml:mtd></mml:mlabeledtr></mml:mtable></mml:math></disp-formula>

            Combining Eqs. (4) and (5), the deep separable convolution is
calculated as
            <disp-formula id="Ch1.E6" content-type="numbered"><label>6</label><mml:math id="M29" display="block"><mml:mrow><mml:msub><mml:mi>L</mml:mi><mml:mi>x</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:msub><mml:mi>L</mml:mi><mml:mi>y</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:mi>M</mml:mi><mml:mo>×</mml:mo><mml:msub><mml:mi>D</mml:mi><mml:mi>k</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:msub><mml:mi>D</mml:mi><mml:mi>k</mml:mi></mml:msub><mml:mo>+</mml:mo><mml:msub><mml:mi>L</mml:mi><mml:mi>x</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:msub><mml:mi>L</mml:mi><mml:mi>y</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:mi>M</mml:mi><mml:mo>×</mml:mo><mml:mi>N</mml:mi><mml:mo>.</mml:mo></mml:mrow></mml:math></disp-formula>
          The reduction ratio of the deep separable convolution compared to
traditional convolution is
            <disp-formula id="Ch1.E7" content-type="numbered"><label>7</label><mml:math id="M30" display="block"><mml:mtable rowspacing="0.2ex" class="split" displaystyle="true" columnalign="right left"><mml:mtr><mml:mtd/><mml:mtd><mml:mrow><mml:mstyle displaystyle="true"><mml:mfrac style="display"><mml:mrow><mml:msub><mml:mi>L</mml:mi><mml:mi>x</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:msub><mml:mi>L</mml:mi><mml:mi>y</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:mi>M</mml:mi><mml:mo>×</mml:mo><mml:msub><mml:mi>D</mml:mi><mml:mi>k</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:msub><mml:mi>D</mml:mi><mml:mi>k</mml:mi></mml:msub><mml:mo>+</mml:mo><mml:msub><mml:mi>L</mml:mi><mml:mi>x</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:msub><mml:mi>L</mml:mi><mml:mi>y</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:mi>M</mml:mi><mml:mo>×</mml:mo><mml:mi>N</mml:mi></mml:mrow><mml:mrow><mml:msub><mml:mi>L</mml:mi><mml:mi>x</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:msub><mml:mi>L</mml:mi><mml:mi>y</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:mi>M</mml:mi><mml:mo>×</mml:mo><mml:mi>N</mml:mi><mml:mo>×</mml:mo><mml:msub><mml:mi>D</mml:mi><mml:mi>k</mml:mi></mml:msub><mml:mo>×</mml:mo><mml:msub><mml:mi>D</mml:mi><mml:mi>k</mml:mi></mml:msub></mml:mrow></mml:mfrac></mml:mstyle></mml:mrow></mml:mtd></mml:mtr><mml:mtr><mml:mtd/><mml:mtd><mml:mrow><mml:mspace width="1em" linebreak="nobreak"/><mml:mo>=</mml:mo><mml:mstyle displaystyle="true"><mml:mfrac style="display"><mml:mn mathvariant="normal">1</mml:mn><mml:mi>N</mml:mi></mml:mfrac></mml:mstyle><mml:mo>+</mml:mo><mml:mstyle displaystyle="true"><mml:mfrac style="display"><mml:mn mathvariant="normal">1</mml:mn><mml:mrow><mml:msubsup><mml:mi>D</mml:mi><mml:mi>k</mml:mi><mml:mn mathvariant="normal">2</mml:mn></mml:msubsup></mml:mrow></mml:mfrac></mml:mstyle><mml:mo>.</mml:mo></mml:mrow></mml:mtd></mml:mtr></mml:mtable></mml:math></disp-formula></p>
</sec>
<sec id="Ch1.S4.SS3">
  <label>4.3</label><title>Network layer model compression</title>
      <p id="d1e1149">Due to the large volume of the current convolutional neural network models,
they suffer from the problems of high computing power cost and large delay
in many practical projects. Using model compression, the redundant parts of
the network and unimportant channels can be automatically identified and
pruned in the training process, so as to obtain a compact model with high
accuracy (Yu et al., 2021).</p>
<sec id="Ch1.S4.SS3.SSS1">
  <label>4.3.1</label><title>Sparse training</title>
      <p id="d1e1159">Model channel sparse training can distinguish important channels and
unimportant channels as a preparation for pruning unimportant channels. An
indicator needs to be set for determining the importance of the channel, and
BN (batch normalisation) layer has the functions of moving parameters and
channel scaling, so the parameters in the BN layer can be used as the evaluation
indicator. Chen et al. (2020) pointed out that the parameter <inline-formula><mml:math id="M31" display="inline"><mml:mi mathvariant="italic">γ</mml:mi></mml:math></inline-formula> in the
BN layer directly affects the output <inline-formula><mml:math id="M32" display="inline"><mml:mi>y</mml:mi></mml:math></inline-formula>, and the smaller the value of
<inline-formula><mml:math id="M33" display="inline"><mml:mi mathvariant="italic">γ</mml:mi></mml:math></inline-formula>, the less important the channel information is. Therefore, this
paper takes the parameter <inline-formula><mml:math id="M34" display="inline"><mml:mi mathvariant="italic">γ</mml:mi></mml:math></inline-formula> as the quantification indicator of
channel importance and calls it the channel scaling factor. The loss
function of the channel pruning algorithm for introducing the scaling
factor <inline-formula><mml:math id="M35" display="inline"><mml:mi mathvariant="italic">γ</mml:mi></mml:math></inline-formula> is shown as Eq. (8):
              <disp-formula id="Ch1.E8" content-type="numbered"><label>8</label><mml:math id="M36" display="block"><mml:mrow><mml:mi>L</mml:mi><mml:mo>=</mml:mo><mml:munder><mml:mo movablelimits="false">∑</mml:mo><mml:mrow><mml:mo>(</mml:mo><mml:mi>x</mml:mi><mml:mo>,</mml:mo><mml:mi>y</mml:mi><mml:mo>)</mml:mo></mml:mrow></mml:munder><mml:mi>l</mml:mi><mml:mfenced close=")" open="("><mml:mrow><mml:mi>f</mml:mi><mml:mfenced close=")" open="("><mml:mrow><mml:mi>x</mml:mi><mml:mo>,</mml:mo><mml:mi>W</mml:mi></mml:mrow></mml:mfenced><mml:mo>,</mml:mo><mml:mi>y</mml:mi></mml:mrow></mml:mfenced><mml:mo>+</mml:mo><mml:mi>s</mml:mi><mml:munder><mml:mo movablelimits="false">∑</mml:mo><mml:mrow><mml:mi mathvariant="italic">γ</mml:mi><mml:mo>∈</mml:mo><mml:mi mathvariant="normal">Γ</mml:mi></mml:mrow></mml:munder><mml:mi>g</mml:mi><mml:mo>(</mml:mo><mml:mi mathvariant="italic">γ</mml:mi><mml:mo>)</mml:mo><mml:mo>,</mml:mo></mml:mrow></mml:math></disp-formula>
            where the first term is the training loss of YOLOX, <inline-formula><mml:math id="M37" display="inline"><mml:mi>x</mml:mi></mml:math></inline-formula> and <inline-formula><mml:math id="M38" display="inline"><mml:mi>y</mml:mi></mml:math></inline-formula> are
the input and output of training, respectively, <inline-formula><mml:math id="M39" display="inline"><mml:mi>W</mml:mi></mml:math></inline-formula> is the training
parameter of the network, and the second term is the L1 regular constraint term
of the BN layer parameter <inline-formula><mml:math id="M40" display="inline"><mml:mi mathvariant="italic">γ</mml:mi></mml:math></inline-formula>, which is the penalty caused by the
sparsity of the scaling factor. It takes <inline-formula><mml:math id="M41" display="inline"><mml:mrow><mml:mi>g</mml:mi><mml:mfenced close=")" open="("><mml:mi mathvariant="italic">γ</mml:mi></mml:mfenced><mml:mo>=</mml:mo><mml:mfenced close="|" open="|"><mml:mi mathvariant="italic">γ</mml:mi></mml:mfenced></mml:mrow></mml:math></inline-formula> in this paper, and <inline-formula><mml:math id="M42" display="inline"><mml:mi>s</mml:mi></mml:math></inline-formula> balances these two terms as a
penalty factor set in this experiment.</p>
</sec>
<sec id="Ch1.S4.SS3.SSS2">
  <label>4.3.2</label><title>Channel pruning and fine-tuning</title>
      <p id="d1e1325">After model channel sparse training, a model with many scaling factors close
to zero is obtained, as shown in Fig. 7. Assuming that the threshold value
is set to 0.1, the convolution layers below the threshold value are shown in
yellow and the convolution layers higher than the threshold value are shown in
blue. By pruning the input and output connections of the yellow convolution
layers and the corresponding weights, the model volume can be reduced and
the model integrity can be guaranteed. However, when the set threshold value
is too large, the accuracy of the pruned model will be greatly reduced, so
it is necessary to use fine-tuning to restore the accuracy of the pruned
network. In many cases of model compression, the fine-tuned narrow network
can even obtain higher accuracy than the original unpruned network.</p>

      <?xmltex \floatpos{t}?><fig id="Ch1.F7" specific-use="star"><?xmltex \currentcnt{7}?><?xmltex \def\figurename{Figure}?><label>Figure 7</label><caption><p id="d1e1330">Channel pruning strategy.</p></caption>
            <?xmltex \igopts{width=369.885827pt}?><graphic xlink:href="https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022-f07.png"/>

          </fig>

</sec>
</sec>
</sec>
<sec id="Ch1.S5">
  <label>5</label><title>Experiment and result analysis</title>
      <p id="d1e1350">The software and hardware configurations used in the experiment are shown in
Table 1.</p>

<?xmltex \floatpos{t}?><table-wrap id="Ch1.T1"><?xmltex \currentcnt{1}?><label>Table 1</label><caption><p id="d1e1356">Software and hardware configurations.</p></caption><oasis:table frame="topbot"><oasis:tgroup cols="2">
     <oasis:colspec colnum="1" colname="col1" align="justify" colwidth="3.5cm"/>
     <oasis:colspec colnum="2" colname="col2" align="justify" colwidth="3.5cm"/>
     <oasis:thead>
       <oasis:row rowsep="1">
         <oasis:entry colname="col1">Configurations</oasis:entry>
         <oasis:entry colname="col2">Version parameters</oasis:entry>
       </oasis:row>
     </oasis:thead>
     <oasis:tbody>
       <oasis:row rowsep="1">
         <oasis:entry colname="col1">Operating system</oasis:entry>
         <oasis:entry colname="col2">Ubuntu18.04LTS</oasis:entry>
       </oasis:row>
       <oasis:row rowsep="1">
         <oasis:entry colname="col1">GPU</oasis:entry>
         <oasis:entry colname="col2">NVIDIA GeForce RTX <?xmltex \hack{\hfill\break}?>3080 Ti</oasis:entry>
       </oasis:row>
       <oasis:row rowsep="1">
         <oasis:entry colname="col1">CPU</oasis:entry>
         <oasis:entry colname="col2">Intel Core <?xmltex \hack{\hfill\break}?>i5-10400F@2.90GHz <?xmltex \hack{\hfill\break}?> <inline-formula><mml:math id="M43" display="inline"><mml:mo>×</mml:mo></mml:math></inline-formula> 6 CPUs</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">Deep learning framework</oasis:entry>
         <oasis:entry colname="col2">Pytorch</oasis:entry>
       </oasis:row>
     </oasis:tbody>
   </oasis:tgroup></oasis:table></table-wrap>

<sec id="Ch1.S5.SS1">
  <label>5.1</label><title>Dataset preparation</title>
      <p id="d1e1439">In this paper, industrial cameras are used to collect samples, which have
more light input than ordinary surveillance cameras. In the open-air
low-light environment, those cameras can also present clear pictures. At the
same time, angle-adjustable camera brackets and mixed fill lights are
equipped, which is convenient to adjust the recording angle and fill light
at night, ensuring that the best image can be obtained even in bad working
conditions. A total of 1200 key frame images are collected in this
experiment including 500 at night and 700 during the day. The image
resolution is <inline-formula><mml:math id="M44" display="inline"><mml:mrow><mml:mn mathvariant="normal">1280</mml:mn><mml:mo>×</mml:mo><mml:mn mathvariant="normal">960</mml:mn></mml:mrow></mml:math></inline-formula>. The sample categories cover bucket teeth,
bucket, pedestrian, dump truck, patrol vehicle, and other targets in the
working scene. Traditional data enhancement methods are used for image
rotation, mosaic data enhancement, and Gaussian filtering, and thus the
dataset is expanded to 3600 images. Among them, 70 % of 3600 images are
used as training sample set, 20 % are used as validation sample set, and
10 % are used as the test sample set. Some dataset examples are shown in
Fig. 8.</p>

      <?xmltex \floatpos{t}?><fig id="Ch1.F8" specific-use="star"><?xmltex \currentcnt{8}?><?xmltex \def\figurename{Figure}?><label>Figure 8</label><caption><p id="d1e1456">Dataset examples.</p></caption>
          <?xmltex \igopts{width=369.885827pt}?><graphic xlink:href="https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022-f08.jpg"/>

        </fig>

</sec>
<sec id="Ch1.S5.SS2">
  <label>5.2</label><title>Model evaluation metrics</title>
      <p id="d1e1474">In order to verify the performance of the model, this experiment uses three
indicators – mAP (mean average precision), model volume, and FPS (frames per
second) – as the evaluation standard. The calculation formula of  mAP is
shown as Eq. (9):
            <disp-formula id="Ch1.E9" content-type="numbered"><label>9</label><mml:math id="M45" display="block"><mml:mrow><mml:mi mathvariant="normal">mAP</mml:mi><mml:mo>=</mml:mo><mml:mstyle displaystyle="true"><mml:mfrac style="display"><mml:mrow><mml:munderover><mml:mo movablelimits="false">∑</mml:mo><mml:mrow><mml:mi>n</mml:mi><mml:mo>=</mml:mo><mml:mn mathvariant="normal">1</mml:mn></mml:mrow><mml:mi>n</mml:mi></mml:munderover><mml:mi mathvariant="normal">AP</mml:mi><mml:mo>(</mml:mo><mml:mi>n</mml:mi><mml:mo>)</mml:mo></mml:mrow><mml:mi>n</mml:mi></mml:mfrac></mml:mstyle><mml:mo>×</mml:mo><mml:mn mathvariant="normal">100</mml:mn><mml:mspace width="0.125em" linebreak="nobreak"/><mml:mi mathvariant="italic">%</mml:mi><mml:mo>,</mml:mo></mml:mrow></mml:math></disp-formula>
          where <inline-formula><mml:math id="M46" display="inline"><mml:mi>n</mml:mi></mml:math></inline-formula> (taken as 9) is the number of detected categories and mAP is the
average value of detection accuracy of all categories. The larger the value,
the higher the detection accuracy is. Among them, AP is numerically equal
to the area enclosed by the <inline-formula><mml:math id="M47" display="inline"><mml:mi>P</mml:mi></mml:math></inline-formula>–<inline-formula><mml:math id="M48" display="inline"><mml:mi>R</mml:mi></mml:math></inline-formula> curve and the coordinate axis.</p>
      <p id="d1e1542"><?xmltex \hack{\newpage}?>The calculation formulas of  <inline-formula><mml:math id="M49" display="inline"><mml:mi>P</mml:mi></mml:math></inline-formula> and <inline-formula><mml:math id="M50" display="inline"><mml:mi>R</mml:mi></mml:math></inline-formula> are as follows:

                <disp-formula specific-use="gather" content-type="numbered"><mml:math id="M51" display="block"><mml:mtable displaystyle="true"><mml:mlabeledtr id="Ch1.E10"><mml:mtd><mml:mtext>10</mml:mtext></mml:mtd><mml:mtd><mml:mrow><mml:mstyle class="stylechange" displaystyle="true"/><mml:mi>P</mml:mi><mml:mo>=</mml:mo><mml:mstyle displaystyle="true"><mml:mfrac style="display"><mml:mi mathvariant="normal">TP</mml:mi><mml:mrow><mml:mi mathvariant="normal">TP</mml:mi><mml:mo>+</mml:mo><mml:mi mathvariant="normal">FP</mml:mi></mml:mrow></mml:mfrac></mml:mstyle><mml:mo>×</mml:mo><mml:mn mathvariant="normal">100</mml:mn><mml:mspace width="0.125em" linebreak="nobreak"/><mml:mi mathvariant="italic">%</mml:mi></mml:mrow></mml:mtd></mml:mlabeledtr><mml:mlabeledtr id="Ch1.E11"><mml:mtd><mml:mtext>11</mml:mtext></mml:mtd><mml:mtd><mml:mrow><mml:mstyle displaystyle="true" class="stylechange"/><mml:mi>R</mml:mi><mml:mo>=</mml:mo><mml:mstyle displaystyle="true"><mml:mfrac style="display"><mml:mi mathvariant="normal">TP</mml:mi><mml:mrow><mml:mi mathvariant="normal">TP</mml:mi><mml:mo>+</mml:mo><mml:mi mathvariant="normal">FN</mml:mi></mml:mrow></mml:mfrac></mml:mstyle><mml:mo>×</mml:mo><mml:mn mathvariant="normal">100</mml:mn><mml:mspace width="0.125em" linebreak="nobreak"/><mml:mi mathvariant="italic">%</mml:mi><mml:mo>,</mml:mo></mml:mrow></mml:mtd></mml:mlabeledtr></mml:mtable></mml:math></disp-formula>

            where <inline-formula><mml:math id="M52" display="inline"><mml:mi>P</mml:mi></mml:math></inline-formula> is the precision, <inline-formula><mml:math id="M53" display="inline"><mml:mi>R</mml:mi></mml:math></inline-formula> is the recall rate, and TP, FP, and FN is
the true positive, false positive, and false negative, respectively.</p>
</sec>
<sec id="Ch1.S5.SS3">
  <label>5.3</label><title>Model improvement effectiveness</title>
<sec id="Ch1.S5.SS3.SSS1">
  <label>5.3.1</label><?xmltex \opttitle{Experimental analysis of YOLOX-CD (YOLOX-dilated cbam\,$+$\,deep separable convolution)}?><title>Experimental analysis of YOLOX-CD (YOLOX-dilated cbam <inline-formula><mml:math id="M54" display="inline"><mml:mo>+</mml:mo></mml:math></inline-formula> deep separable convolution)</title>
      <p id="d1e1663">The initial learning rate is set to 0.001, the momentum coefficient is
0.949, the number of iterations is 4000, and the learning rate is changed to
0.0001 and 0.00001 at 3200 and 3600 iterations, respectively. The loss
function value change curves of model training are shown in Figs. 9 and 10.
It can be seen from the figure that the loss value of YOLOX gradually
decreased to 1.2 after 3300 iterations, and finally stabilised to about 1.
However, the loss value of YOLOX-CD is lower and the convergence speed is
faster, the loss value continued to stabilise to about 0.7 at 2800
iterations, and the model shows a convergence trend.</p>

      <?xmltex \floatpos{t}?><fig id="Ch1.F9"><?xmltex \currentcnt{9}?><?xmltex \def\figurename{Figure}?><label>Figure 9</label><caption><p id="d1e1668">Loss function value change curve of YOLOX.</p></caption>
            <?xmltex \igopts{width=213.395669pt}?><graphic xlink:href="https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022-f09.png"/>

          </fig>

      <?xmltex \floatpos{t}?><fig id="Ch1.F10"><?xmltex \currentcnt{10}?><?xmltex \def\figurename{Figure}?><label>Figure 10</label><caption><p id="d1e1680">Loss function value change curve of YOLOX-CD.</p></caption>
            <?xmltex \igopts{width=213.395669pt}?><graphic xlink:href="https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022-f10.png"/>

          </fig>

      <p id="d1e1689">After the model training is completed, the test sample set is input into
each model for testing, and the test result of YOLOX (baseline), YOLOX-C
with the dilated convolution attention, YOLOX-D with the deep separable
convolution introduced in PANet, and YOLOX-CD model with both are compared,
so as to verify the effectiveness of the model improvement. The experimental
results are shown in Table 2.</p>

<?xmltex \floatpos{t}?><table-wrap id="Ch1.T2"><?xmltex \currentcnt{2}?><label>Table 2</label><caption><p id="d1e1695">Comparison of experimental results.</p></caption><oasis:table frame="topbot"><oasis:tgroup cols="4">
     <oasis:colspec colnum="1" colname="col1" align="left"/>
     <oasis:colspec colnum="2" colname="col2" align="right"/>
     <oasis:colspec colnum="3" colname="col3" align="right"/>
     <oasis:colspec colnum="4" colname="col4" align="right"/>
     <oasis:thead>
       <oasis:row rowsep="1">
         <oasis:entry colname="col1">Model name</oasis:entry>
         <oasis:entry colname="col2">mAP (%)</oasis:entry>
         <oasis:entry colname="col3">Model volume (MB)</oasis:entry>
         <oasis:entry colname="col4">FPS</oasis:entry>
       </oasis:row>
     </oasis:thead>
     <oasis:tbody>
       <oasis:row>
         <oasis:entry colname="col1">YOLOX</oasis:entry>
         <oasis:entry colname="col2">95.59</oasis:entry>
         <oasis:entry colname="col3">96.46</oasis:entry>
         <oasis:entry colname="col4">38.9</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">YOLOX-C</oasis:entry>
         <oasis:entry colname="col2">97.61</oasis:entry>
         <oasis:entry colname="col3">99.27</oasis:entry>
         <oasis:entry colname="col4">35.2</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">YOLOX-D</oasis:entry>
         <oasis:entry colname="col2">94.44</oasis:entry>
         <oasis:entry colname="col3">87.69</oasis:entry>
         <oasis:entry colname="col4">44.5</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">YOLOX-CD</oasis:entry>
         <oasis:entry colname="col2">96.42</oasis:entry>
         <oasis:entry colname="col3">90.12</oasis:entry>
         <oasis:entry colname="col4">41.3</oasis:entry>
       </oasis:row>
     </oasis:tbody>
   </oasis:tgroup></oasis:table></table-wrap>

      <p id="d1e1793">It can be clearly seen from Table 2 that after introducing both the dilated
convolution attention mechanism and the deep separable convolution, the
performances are significantly improved compared with the original model.
Among them, the mAP is increased by 0.83 %, the detection speed is
increased by 2.4 fps, and the model volume is reduced by 6.34 MB.</p><?xmltex \hack{\newpage}?>
</sec>
<sec id="Ch1.S5.SS3.SSS2">
  <label>5.3.2</label><title>Experimental analysis of model compression</title>
      <p id="d1e1805">Under the same conditions as the initial training environment, the penalty
factor <inline-formula><mml:math id="M55" display="inline"><mml:mi>s</mml:mi></mml:math></inline-formula> is set as 0.0001, 0.001, 0.005, and 0.01, respectively, and the
optimal penalty factor is selected by comparing the model performance during
the sparse training process. Figure 11 shows the mAP change curve under
different penalty factors. It can be seen from the figure that the detection
performance of the YOLOX-CD reaches the best when the penalty factor is
<inline-formula><mml:math id="M56" display="inline"><mml:mrow><mml:mi>s</mml:mi><mml:mo>=</mml:mo><mml:mn mathvariant="normal">0.001</mml:mn></mml:mrow></mml:math></inline-formula>. Therefore, <inline-formula><mml:math id="M57" display="inline"><mml:mrow><mml:mi>s</mml:mi><mml:mo>=</mml:mo><mml:mn mathvariant="normal">0.001</mml:mn></mml:mrow></mml:math></inline-formula> is selected as the penalty factor for the
final channel sparse training in this paper.</p>

      <?xmltex \floatpos{t}?><fig id="Ch1.F11"><?xmltex \currentcnt{11}?><?xmltex \def\figurename{Figure}?><label>Figure 11</label><caption><p id="d1e1841">mAP curve under different penalty factors.</p></caption>
            <?xmltex \igopts{width=199.169291pt}?><graphic xlink:href="https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022-f11.png"/>

          </fig>

      <p id="d1e1850">After sparse training, YOLOX-CD is pruned, and the model is named YOLOX-CDP.
During the pruning process, the mAP of YOLOX-CDP will decrease, which needs
to be fine-tuned. Among them, the pruning rate is set to 0.5, 0.7, and 0.9,
respectively, and 20 training epochs are carried out in fine-tuning. The
model performance with different pruning rates are shown in Table 3.</p>

<?xmltex \floatpos{t}?><table-wrap id="Ch1.T3"><?xmltex \currentcnt{3}?><label>Table 3</label><caption><p id="d1e1857">Model performance comparison with different pruning rates.</p></caption><oasis:table frame="topbot"><oasis:tgroup cols="4">
     <oasis:colspec colnum="1" colname="col1" align="left"/>
     <oasis:colspec colnum="2" colname="col2" align="right"/>
     <oasis:colspec colnum="3" colname="col3" align="right"/>
     <oasis:colspec colnum="4" colname="col4" align="right"/>
     <oasis:thead>
       <oasis:row rowsep="1">
         <oasis:entry colname="col1">Model name</oasis:entry>
         <oasis:entry colname="col2">mAP (%)</oasis:entry>
         <oasis:entry colname="col3">Model volume (MB)</oasis:entry>
         <oasis:entry colname="col4">FPS</oasis:entry>
       </oasis:row>
     </oasis:thead>
     <oasis:tbody>
       <oasis:row>
         <oasis:entry colname="col1">YOLOX</oasis:entry>
         <oasis:entry colname="col2">95.59</oasis:entry>
         <oasis:entry colname="col3">96.46</oasis:entry>
         <oasis:entry colname="col4">38.9</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">YOLOX-CD-0.5</oasis:entry>
         <oasis:entry colname="col2">96.35</oasis:entry>
         <oasis:entry colname="col3">47.73</oasis:entry>
         <oasis:entry colname="col4">45.1</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">YOLOX-CD-0.7</oasis:entry>
         <oasis:entry colname="col2">95.26</oasis:entry>
         <oasis:entry colname="col3">28.42</oasis:entry>
         <oasis:entry colname="col4">50.8</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">YOLOX-CD-0.9</oasis:entry>
         <oasis:entry colname="col2">56.72</oasis:entry>
         <oasis:entry colname="col3">8.34</oasis:entry>
         <oasis:entry colname="col4">57.2</oasis:entry>
       </oasis:row>
     </oasis:tbody>
   </oasis:tgroup></oasis:table></table-wrap>

      <p id="d1e1955">In Table 3, as the pruning rate increases, both the model volume and the
prediction time decrease continuously. Compared with the YOLOX model,
although the mAP of YOLOX-CD-0.7 decreased by 0.33 %, the detection speed
increased by 11.9 fps, and the model volume decreased to 29.46 % of the
original YOLOX. The experimental results show that the model compression
method can reduce the model volume and improve the detection speed under the
premise of ensuring high precision. To more intuitively reflect the effect
of YOLOX-CD-0.7 model bucket teeth falling off real-time and accurate
detection, samples are selected by classification in the test set to carry
out multiple groups of comparative experiments. The experimental results are
shown in Fig. 12.</p>

      <?xmltex \floatpos{t}?><fig id="Ch1.F12" specific-use="star"><?xmltex \currentcnt{12}?><?xmltex \def\figurename{Figure}?><label>Figure 12</label><caption><p id="d1e1960">The detection results of YOLOX-CD-0.7 on the test sample set.</p></caption>
            <?xmltex \igopts{width=369.885827pt}?><graphic xlink:href="https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022-f12.png"/>

          </fig>

      <?xmltex \floatpos{t}?><fig id="Ch1.F13" specific-use="star"><?xmltex \currentcnt{13}?><?xmltex \def\figurename{Figure}?><label>Figure 13</label><caption><p id="d1e1971">Comparison of migration test results.</p></caption>
            <?xmltex \igopts{width=369.885827pt}?><graphic xlink:href="https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022-f13.png"/>

          </fig>

      <p id="d1e1980">Figure 12a shows the detection results of the top and bottom video images
in the normal working scene of the electric shovel, including top-view
excavation, bottom-view excavation, top-view loading, and bottom-view
loading. It can be seen from the detection results that the YOLOX-CD-0.7
model can effectively detect targets such as bucket teeth, bucket, dump
truck, etc. and the accuracy of the framed range is high. Figure 12b
shows the detection results of dangerous scenarios that may be encountered
during the operation of the electric shovel, including occlusion detection,
broken teeth detection, vehicle intrusion, and personnel intrusion. As can be
seen from the figure,  the algorithm proposed in this paper can not only
judge whether the bucket teeth fall off but also identify the invading
vehicles and personnel, which can effectively prevent the occurrence of
dangerous accidents. Figure 12c shows the image detection results in
complex and illuminated scenes. As can be seen from the figure, even under
complex and uneven lighting conditions, the algorithm in this paper can
still detect small, weak, and occluded targets well.</p>
      <p id="d1e1984">This paper takes the dataset of the WK-35 electric shovel as an example to carry
out the migration experiment. Samples under different scenarios are randomly
selected from the dataset for testing; the test results are shown in Fig. 13. It can be seen from the experimental results that for the untrained new
sample data, the original model cannot accurately identify the target in
complex scenes, while the improved model can complete the real-time and
accurate detection of targets in various complex scenes, which proves that
improved YOLOX has good generalisation ability.</p>
</sec>
</sec>
<sec id="Ch1.S5.SS4">
  <label>5.4</label><title>Comparison experiment of different algorithms</title>
      <p id="d1e1996">In order to further verify the advantages of the algorithm in this paper, it
is compared with four classic target detection algorithms of Faster R-CNN,
YOLOv3, YOLOv4, and YOLOv5 in the same dataset and training environment. The
comparison results are shown in Fig. 14.</p>

      <?xmltex \floatpos{t}?><fig id="Ch1.F14"><?xmltex \currentcnt{14}?><?xmltex \def\figurename{Figure}?><label>Figure 14</label><caption><p id="d1e2001"><inline-formula><mml:math id="M58" display="inline"><mml:mi>P</mml:mi></mml:math></inline-formula>–<inline-formula><mml:math id="M59" display="inline"><mml:mi>R</mml:mi></mml:math></inline-formula> curves of different algorithms.</p></caption>
          <?xmltex \igopts{width=213.395669pt}?><graphic xlink:href="https://ms.copernicus.org/articles/13/979/2022/ms-13-979-2022-f14.png"/>

        </fig>

      <p id="d1e2023">It can be seen from the figure that the <inline-formula><mml:math id="M60" display="inline"><mml:mi>P</mml:mi></mml:math></inline-formula>–<inline-formula><mml:math id="M61" display="inline"><mml:mi>R</mml:mi></mml:math></inline-formula> curve of our algorithm is
located at the top of all the curves, and the performance of detecting
bucket teeth is better than that of the other four algorithms. The above
experimental results are listed in Table 4. It can be seen that the
algorithm in this paper has improved mAP by 1.92 %, 5.53 %, 3.2 %, and
0.44 %, respectively, compared with Faster R-CNN, YOLOv3, YOLOv4, and YOLOv5,
with smaller model volume and faster detection speed. Comprehensive
performance of all aspects show that improved YOLOX can meet the
requirements of real-time and accurate detection of bucket teeth falling off
during the working process of the electric shovel.</p>

<?xmltex \floatpos{t}?><table-wrap id="Ch1.T4"><?xmltex \currentcnt{4}?><label>Table 4</label><caption><p id="d1e2044">Performance comparison of different algorithms.</p></caption><oasis:table frame="topbot"><oasis:tgroup cols="4">
     <oasis:colspec colnum="1" colname="col1" align="left"/>
     <oasis:colspec colnum="2" colname="col2" align="right"/>
     <oasis:colspec colnum="3" colname="col3" align="right"/>
     <oasis:colspec colnum="4" colname="col4" align="right"/>
     <oasis:thead>
       <oasis:row rowsep="1">
         <oasis:entry colname="col1">Model name</oasis:entry>
         <oasis:entry colname="col2">mAP (%)</oasis:entry>
         <oasis:entry colname="col3">Model volume (MB)</oasis:entry>
         <oasis:entry colname="col4">FPS</oasis:entry>
       </oasis:row>
     </oasis:thead>
     <oasis:tbody>
       <oasis:row>
         <oasis:entry colname="col1">Faster R-CNN</oasis:entry>
         <oasis:entry colname="col2">93.34</oasis:entry>
         <oasis:entry colname="col3">522.91</oasis:entry>
         <oasis:entry colname="col4">13.0</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">YOLOv3</oasis:entry>
         <oasis:entry colname="col2">89.73</oasis:entry>
         <oasis:entry colname="col3">236.20</oasis:entry>
         <oasis:entry colname="col4">31.2</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">YOLOv4</oasis:entry>
         <oasis:entry colname="col2">92.06</oasis:entry>
         <oasis:entry colname="col3">245.00</oasis:entry>
         <oasis:entry colname="col4">35.7</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">YOLOv5</oasis:entry>
         <oasis:entry colname="col2">94.82</oasis:entry>
         <oasis:entry colname="col3">89.54</oasis:entry>
         <oasis:entry colname="col4">43.0</oasis:entry>
       </oasis:row>
       <oasis:row>
         <oasis:entry colname="col1">OURS</oasis:entry>
         <oasis:entry colname="col2">95.26</oasis:entry>
         <oasis:entry colname="col3">28.42</oasis:entry>
         <oasis:entry colname="col4">50.8</oasis:entry>
       </oasis:row>
     </oasis:tbody>
   </oasis:tgroup></oasis:table></table-wrap>

</sec>
</sec>
<sec id="Ch1.S6" sec-type="conclusions">
  <label>6</label><title>Conclusions</title>
      <p id="d1e2166">We propose a real-time and accurate detection algorithm of bucket teeth
falling off based on improved YOLOX. Through the field experiment on the
working scene of electric shovel in a large open pit mine, the following
conclusions are drawn.</p>
      <p id="d1e2169">Adding the dilated CBAM attention mechanism solves the problem of weak
target saliency in complex environments and improves the target detection
accuracy. The deep separable convolution is used to replace the traditional
convolution in PANet, which reduces the overall computation of the network.
The channel pruning strategy is used to compress the model, which
significantly improves the network detection speed and reduces the model
volume on the basis of ensuring the integrity of the model. The experimental
results show that the improved YOLOX detection accuracy is 95.26 %, only
0.33 % lower, while the detection speed is 50.8 fps, 11.9 fps higher, and
the model volume is 28.42 MB, which is reduced to 29.46 % of the original.</p>
      <p id="d1e2172">Compared with Faster R-CNN, YOLOv3, YOLOv4, and YOLOv5 on the self-built
WK-10 electric shovel dataset, the improved YOLOX algorithm has great
advantages in detection accuracy, speed, and model volume, and is suitable
for deployment in embedded mobile devices for detection and calculation.</p>
      <p id="d1e2175">The migration experiment results show that the improved model has strong
generalisation ability and can complete the real-time and accurate detection
of targets of different types of electric shovels in complex scenarios,
which provides a theoretical basis and technical support for the development
of intelligent mines and mining intelligence.</p>
      <p id="d1e2179">The working environment of the open pit mine is harsh, which will produce
obvious fog phenomenon, affecting the detection performance of the algorithm
in this paper. In future work, image defogging module is introduced to
improve the detection accuracy of our algorithm in a fog environment.</p><?xmltex \hack{\newpage}?>
</sec>

      
      </body>
    <back><notes notes-type="dataavailability"><title>Data availability</title>

      <p id="d1e2187">All the code/data used in this paper can be obtained upon request to the
corresponding author.</p>
  </notes><notes notes-type="authorcontribution"><title>Author contributions</title>

      <p id="d1e2193">JL provided the ideas, reviewed the overall process for the work, and
revised the paper. YL provided the program, made the figures, and wrote the
majority of the paper.</p>
  </notes><notes notes-type="competinginterests"><title>Competing interests</title>

      <p id="d1e2199">The contact author has declared that neither of the authors has any competing interests.</p>
  </notes><notes notes-type="disclaimer"><title>Disclaimer</title>

      <p id="d1e2205">Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.</p>
  </notes><ack><title>Acknowledgements</title><p id="d1e2211">The authors would like to thank the National Natural Science Foundation of
China for supporting this research. We also greatly appreciate the efforts
of the reviewers and our colleagues.</p></ack><notes notes-type="financialsupport"><title>Financial support</title>

      <p id="d1e2216">This research has been supported by the National Natural Science Foundation of China (grant nos. 51774162 and 51874158).</p>
  </notes><notes notes-type="reviewstatement"><title>Review statement</title>

      <p id="d1e2222">This paper was edited by Zi Bin and reviewed by three anonymous referees.</p>
  </notes><ref-list>
    <title>References</title>

      <ref id="bib1.bib1"><label>1</label><?label 1?><mixed-citation>Akter, A. and Basak, H.: Design and analysis of biomimetics based excavator
bucket and tooth, P. I. Mech. Eng. E-J. Pro, 236, 1167–1175,
<ext-link xlink:href="https://doi.org/10.1177/09544089211057645" ext-link-type="DOI">10.1177/09544089211057645</ext-link>, 2022.</mixed-citation></ref>
      <ref id="bib1.bib2"><label>2</label><?label 1?><mixed-citation>Alma'aitah, A. and Hassanein, H.: Utilizing Sprouts WSN platform for
equipment detection and localization in harsh environments, in: 39th Annual
IEEE Conference on Local Computer Networks Workshops,
8–11 September 2014, Edmonton, Canada, 777–783, <ext-link xlink:href="https://doi.org/10.1109/lcnw.2014.6927734" ext-link-type="DOI">10.1109/lcnw.2014.6927734</ext-link>,
2014.</mixed-citation></ref>
      <ref id="bib1.bib3"><label>3</label><?label 1?><mixed-citation>Bai, D., Sun, Y., Tao, B., Tong, X., Xu, M., Jiang, G., Chen, B., Cao, Y.,
Sun, N., and Li, Z.: Improved single shot multibox detector target detection
method based on deep feature fusion, Concurr. Comp.-Pract.
E., 34, e6614, <ext-link xlink:href="https://doi.org/10.1002/CPE.6614" ext-link-type="DOI">10.1002/CPE.6614</ext-link>, 2022.</mixed-citation></ref>
      <ref id="bib1.bib4"><label>4</label><?label 1?><mixed-citation>Bello, R., Mohamed, A., and Talib, A.: Enhanced Mask R-CNN for herd
segmentation, Int. J. Agr. Biol. Eng., 14, 238–244,
<ext-link xlink:href="https://doi.org/10.25165/j.ijabe.20211404.6398" ext-link-type="DOI">10.25165/j.ijabe.20211404.6398</ext-link>, 2021.</mixed-citation></ref>
      <ref id="bib1.bib5"><label>5</label><?label 1?><mixed-citation>Buiu, C., Danaila, V., and Raduta, C.: MobileNetV2 ensemble for cervical
precancerous lesions classification, Processes, 8, 595,
<ext-link xlink:href="https://doi.org/10.3390/pr8050595" ext-link-type="DOI">10.3390/pr8050595</ext-link>, 2020.</mixed-citation></ref>
      <ref id="bib1.bib6"><label>6</label><?label 1?><mixed-citation>Chen, S., Zhan, R., Wang, W., and Zhang, J.: Learning Slimming SAR Ship
Object Detector Through Network Pruning and Knowledge Distillation, IEEE
Journal of Selected Topics in Applied Earth Observations and Remote Sensing,
14, 1267–1282, <ext-link xlink:href="https://doi.org/10.1109/jstars.2020.3041783" ext-link-type="DOI">10.1109/jstars.2020.3041783</ext-link>, 2020.</mixed-citation></ref>
      <ref id="bib1.bib7"><label>7</label><?label 1?><mixed-citation>Chen, W., Li, X., He, H., and Wang, L.: A Review of Fine-Scale Land Use and
Land Cover Classification in Open-Pit Mining Areas by Remote Sensing
Techniques, Remote Sens., 10, 15, <ext-link xlink:href="https://doi.org/10.3390/rs10010015" ext-link-type="DOI">10.3390/rs10010015</ext-link>, 2017.</mixed-citation></ref>
      <ref id="bib1.bib8"><label>8</label><?label 1?><mixed-citation>
Denton, E., Zaremba, W., Bruna, J., LeCun, Y., and Fergus, R.: Exploiting
Linear Structure Within Convolutional Networks for Efficient Evaluation, in:
Advances in Neural Information Processing Systems, 8–13
December 2014, Montreal, Canada,  https://arxiv.org/abs/1404.0736, 2014.</mixed-citation></ref>
      <ref id="bib1.bib9"><label>9</label><?label 1?><mixed-citation>Ge, Z., Liu, S., Wang, F., Li, Z., and Sun, J.: YOLOX: Exceeding yolo series
in 2021, arXiv [preprint],
<ext-link xlink:href="https://doi.org/10.48550/arxiv.2107.08430" ext-link-type="DOI">10.48550/arxiv.2107.08430</ext-link>, 6 August 2021.</mixed-citation></ref>
      <ref id="bib1.bib10"><label>10</label><?label 1?><mixed-citation>He, L., Wang, H., and Zhang, H.: Object detection by parts using appearance,
structural and shape features, in: 2011 IEEE International Conference on
Mechatronics and Automation, Beijing, China, 7–10 August 2011, 489–494,
<ext-link xlink:href="https://doi.org/10.1109/icma.2011.5985611" ext-link-type="DOI">10.1109/icma.2011.5985611</ext-link>, 2011.</mixed-citation></ref>
      <ref id="bib1.bib11"><label>11</label><?label 1?><mixed-citation>Huang, J., Zhang, H., Wang, L., Zhang, Z., and Zhao, C.: Improved YOLOv3
Model for miniature camera detection, Opt. Laser Technol., 142, 107133,
<ext-link xlink:href="https://doi.org/10.1016/j.optlastec.2021.107133" ext-link-type="DOI">10.1016/j.optlastec.2021.107133</ext-link>, 2021.</mixed-citation></ref>
      <ref id="bib1.bib12"><label>12</label><?label 1?><mixed-citation>Huang, L., Fu, Q., He, M., Jiang, D., and Hao, Z.: Detection algorithm of
safety helmet wearing based on deep learning, Concurr. Comp.-Pract. E., 33, e6234, <ext-link xlink:href="https://doi.org/10.1002/cpe.6234" ext-link-type="DOI">10.1002/cpe.6234</ext-link>,
2021.</mixed-citation></ref>
      <ref id="bib1.bib13"><label>13</label><?label 1?><mixed-citation>Huang, L., Chen, C., Yun, J., Sun, Y., Tian, J., Hao, Z., Yu, H., and Ma,
H.: Multi-Scale Feature Fusion Convolutional Neural Network for Indoor Small
Target Detection, Frontiers in Neurorobitics, 16, 881021,
<ext-link xlink:href="https://doi.org/10.3389/fnbot.2022.881021" ext-link-type="DOI">10.3389/fnbot.2022.881021</ext-link>, 2022.</mixed-citation></ref>
      <ref id="bib1.bib14"><label>14</label><?label 1?><mixed-citation>Ji, S., Li, W., Zhang, B., Zhou, L., and Duan, C.: Bucket Teeth Detection
Based on Faster Region Convolutional Neural Network, IEEE Access, 9,
17649–17661, <ext-link xlink:href="https://doi.org/10.1109/access.2021.3054436" ext-link-type="DOI">10.1109/access.2021.3054436</ext-link>, 2021.</mixed-citation></ref>
      <ref id="bib1.bib15"><label>15</label><?label 1?><mixed-citation>Ji, Y., Lu, Q., and Yao, Q.: Short communication: A case study of stress monitoring with non-destructive stress measurement and deep learning algorithms, Mech. Sci., 13, 291–296, <ext-link xlink:href="https://doi.org/10.5194/ms-13-291-2022" ext-link-type="DOI">10.5194/ms-13-291-2022</ext-link>, 2022.</mixed-citation></ref>
      <ref id="bib1.bib16"><label>16</label><?label 1?><mixed-citation>Li, Z., Xu, X., Xie, L., and Su, H.: Learning Slimming SSD through Pruning
and Knowledge Distillation, in: 2019 Chinese Automation Congress, 22–24 November 2019, Hangzhou,
China, 2701–2705,
<ext-link xlink:href="https://doi.org/10.1109/cac48633.2019.8996995" ext-link-type="DOI">10.1109/cac48633.2019.8996995</ext-link>, 2019.</mixed-citation></ref>
      <ref id="bib1.bib17"><label>17</label><?label 1?><mixed-citation>Lim, S., Soares, J., and Zhou, N.: Tooth guard: A vision system for
detecting missing tooth in rope mine shovel, in: 2016 IEEE Winter Conference
on Applications of Computer Vision, 7–10 March
2016, New York, United States, 1–7, <ext-link xlink:href="https://doi.org/10.1109/wacv.2016.7477594" ext-link-type="DOI">10.1109/wacv.2016.7477594</ext-link>, 2016.</mixed-citation></ref>
      <ref id="bib1.bib18"><label>18</label><?label 1?><mixed-citation>Liu, X., Qi, X., and Jiang, Y.: Electric Shovel Teeth Missing Detection
Method Based on Deep Learning, Comput. Intel. Neurosc.,
2021, 6503029, <ext-link xlink:href="https://doi.org/10.1155/2021/6503029" ext-link-type="DOI">10.1155/2021/6503029</ext-link>, 2021.</mixed-citation></ref>
      <ref id="bib1.bib19"><label>19</label><?label 1?><mixed-citation>Liu, Z., Li, J., Shen, Z., Huang, G., Yan, S., and Zhang, C.: Learning
efficient convolutional networks through network slimming, in: 2017 IEEE
International Conference on Computer Vision, 22–29 October
2017, Venice, Italy, 2736–2744, <ext-link xlink:href="https://doi.org/10.1109/iccv.2017.298" ext-link-type="DOI">10.1109/iccv.2017.298</ext-link>, 2017.</mixed-citation></ref>
      <ref id="bib1.bib20"><label>20</label><?label 1?><mixed-citation>Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.: You Only Look Once:
Unified, Real-Time Object Detection, in: 2016 IEEE Conference on Computer
Vision and Pattern Recognition, 27–30 June 2016, Washington, United States,
779–788, <ext-link xlink:href="https://doi.org/10.1109/cvpr.2016.91" ext-link-type="DOI">10.1109/cvpr.2016.91</ext-link>, 2016.</mixed-citation></ref>
      <ref id="bib1.bib21"><label>21</label><?label 1?><mixed-citation>Redmon, J. and Farhadi, A.: YOLO9000: Better, Faster, Stronger, in: 30th
IEEE Conference on Computer Vision and Pattern Recognition, 21–26 July 2017, Honolulu, United
States, 6517–6525, <ext-link xlink:href="https://doi.org/10.1109/cvpr.2017.690" ext-link-type="DOI">10.1109/cvpr.2017.690</ext-link>,
2017.</mixed-citation></ref>
      <ref id="bib1.bib22"><label>22</label><?label 1?><mixed-citation>Shen, X., Huang, Q., and Xiong, G.: Modelling and predictive investigation on the vibration response of a propeller shaft based on a convolutional neural network, Mech. Sci., 13, 485–494, <ext-link xlink:href="https://doi.org/10.5194/ms-13-485-2022" ext-link-type="DOI">10.5194/ms-13-485-2022</ext-link>, 2022.</mixed-citation></ref>
      <ref id="bib1.bib23"><label>23</label><?label 1?><mixed-citation>Singla, S., Kang, A., Grewal, J., and Cheema, G.: Wear behavior of weld
overlays on excavator bucket teeth, Proc. Mat. Sci., 5, 256–266,
<ext-link xlink:href="https://doi.org/10.1016/j.mspro.2014.07.265" ext-link-type="DOI">10.1016/j.mspro.2014.07.265</ext-link>, 2014.</mixed-citation></ref>
      <ref id="bib1.bib24"><label>24</label><?label 1?><mixed-citation>Sun, Y., Zhao, Z., Jiang, D., Tong, X., Tao, B., Jiang, G., Kong, J., Yun,
J., Liu, X., Zhao, G., and Fang, Z.: Low-illumination image enhancement
algorithm based on improved multi-scale Retinex and ABC algorithm
optimization, Frontiers in Bioengineering and Biotechnology, 10, 865820,
<ext-link xlink:href="https://doi.org/10.3389/fbioe.2022.865820" ext-link-type="DOI">10.3389/fbioe.2022.865820</ext-link>, 2022.
</mixed-citation></ref><?xmltex \hack{\newpage}?>
      <ref id="bib1.bib25"><label>25</label><?label 1?><mixed-citation>Tan, L., Lv, X., Lian, X., and Wang, G.: YOLOv4_Drone: UAV
image target detection based on an improved YOLOv4 algorithm, Comput.
Electr. Eng., 93, 107261, <ext-link xlink:href="https://doi.org/10.1016/j.compeleceng.2021.107261" ext-link-type="DOI">10.1016/j.compeleceng.2021.107261</ext-link>,
2021.</mixed-citation></ref>
      <ref id="bib1.bib26"><label>26</label><?label 1?><mixed-citation>Wang, G., Zheng, H., and Zhang, X.: A robust checkerboard corner detection
method for camera calibration based on improved YOLOX, Front. Phys.,
9, 819019, <ext-link xlink:href="https://doi.org/10.3389/fphy.2021.819019" ext-link-type="DOI">10.3389/fphy.2021.819019</ext-link>, 2022.</mixed-citation></ref>
      <ref id="bib1.bib27"><label>27</label><?label 1?><mixed-citation>Wang, X., Sun, W., Li, E., and Song, X.: Energy-minimum optimization of the
intelligent excavating process for large cable shovel through trajectory
planning, Struct. Multidiscip. O., 58, 2219–2237,
<ext-link xlink:href="https://doi.org/10.1007/s00158-018-2011-6" ext-link-type="DOI">10.1007/s00158-018-2011-6</ext-link>, 2018.</mixed-citation></ref>
      <ref id="bib1.bib28"><label>28</label><?label 1?><mixed-citation>Yu, F., Koltun, V., and Funkhouser, T.: Dilated residual networks, in: 30th
IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, United
States, 21–26 July 2017, 636–644, <ext-link xlink:href="https://doi.org/10.1109/cvpr.2017.75" ext-link-type="DOI">10.1109/cvpr.2017.75</ext-link>,
2017.</mixed-citation></ref>
      <ref id="bib1.bib29"><label>29</label><?label 1?><mixed-citation>Yu, Z., Shen, Y., and Shen, C.: A real-time detection approach for bridge
cracks based on YOLOv4-FPM, Automat. Constr., 122, 103514,
<ext-link xlink:href="https://doi.org/10.1016/j.autcon.2020.103514" ext-link-type="DOI">10.1016/j.autcon.2020.103514</ext-link>, 2021.</mixed-citation></ref>
      <ref id="bib1.bib30"><label>30</label><?label 1?><mixed-citation>Yun, J., Jiang, D., Liu, Y., Sun, Y., Tao, B., Kong, J., Tian, J., Tong, X.,
Xu, M., and Fang, Z.: Real-time target detection method based on lightweight
convolutional neural network, Frontiers in Bioengineering and Biotechnology,
10, 861286, <ext-link xlink:href="https://doi.org/10.3389/fbioe.2022.861286" ext-link-type="DOI">10.3389/fbioe.2022.861286</ext-link>, 2022.</mixed-citation></ref>
      <ref id="bib1.bib31"><label>31</label><?label 1?><mixed-citation>Zhang, Y., Huang, J., and Cai, F.: On Bridge Surface Crack Detection Based
on an Improved YOLO v3 Algorithm, IFAC-PapersOnLine, 53, 8205–8210,
<ext-link xlink:href="https://doi.org/10.1016/j.ifacol.2020.12.1994" ext-link-type="DOI">10.1016/j.ifacol.2020.12.1994</ext-link>, 2020.</mixed-citation></ref>

  </ref-list></back>
    <!--<article-title-html>A real-time and accurate detection approach for bucket teeth falling off based on improved YOLOX</article-title-html>
<abstract-html/>
<ref-html id="bib1.bib1"><label>1</label><mixed-citation>
Akter, A. and Basak, H.: Design and analysis of biomimetics based excavator
bucket and tooth, P. I. Mech. Eng. E-J. Pro, 236, 1167–1175,
<a href="https://doi.org/10.1177/09544089211057645" target="_blank">https://doi.org/10.1177/09544089211057645</a>, 2022.
</mixed-citation></ref-html>
<ref-html id="bib1.bib2"><label>2</label><mixed-citation>
Alma'aitah, A. and Hassanein, H.: Utilizing Sprouts WSN platform for
equipment detection and localization in harsh environments, in: 39th Annual
IEEE Conference on Local Computer Networks Workshops,
8–11 September 2014, Edmonton, Canada, 777–783, <a href="https://doi.org/10.1109/lcnw.2014.6927734" target="_blank">https://doi.org/10.1109/lcnw.2014.6927734</a>,
2014.
</mixed-citation></ref-html>
<ref-html id="bib1.bib3"><label>3</label><mixed-citation>
Bai, D., Sun, Y., Tao, B., Tong, X., Xu, M., Jiang, G., Chen, B., Cao, Y.,
Sun, N., and Li, Z.: Improved single shot multibox detector target detection
method based on deep feature fusion, Concurr. Comp.-Pract.
E., 34, e6614, <a href="https://doi.org/10.1002/CPE.6614" target="_blank">https://doi.org/10.1002/CPE.6614</a>, 2022.
</mixed-citation></ref-html>
<ref-html id="bib1.bib4"><label>4</label><mixed-citation>
Bello, R., Mohamed, A., and Talib, A.: Enhanced Mask R-CNN for herd
segmentation, Int. J. Agr. Biol. Eng., 14, 238–244,
<a href="https://doi.org/10.25165/j.ijabe.20211404.6398" target="_blank">https://doi.org/10.25165/j.ijabe.20211404.6398</a>, 2021.
</mixed-citation></ref-html>
<ref-html id="bib1.bib5"><label>5</label><mixed-citation>
Buiu, C., Danaila, V., and Raduta, C.: MobileNetV2 ensemble for cervical
precancerous lesions classification, Processes, 8, 595,
<a href="https://doi.org/10.3390/pr8050595" target="_blank">https://doi.org/10.3390/pr8050595</a>, 2020.
</mixed-citation></ref-html>
<ref-html id="bib1.bib6"><label>6</label><mixed-citation>
Chen, S., Zhan, R., Wang, W., and Zhang, J.: Learning Slimming SAR Ship
Object Detector Through Network Pruning and Knowledge Distillation, IEEE
Journal of Selected Topics in Applied Earth Observations and Remote Sensing,
14, 1267–1282, <a href="https://doi.org/10.1109/jstars.2020.3041783" target="_blank">https://doi.org/10.1109/jstars.2020.3041783</a>, 2020.
</mixed-citation></ref-html>
<ref-html id="bib1.bib7"><label>7</label><mixed-citation>
Chen, W., Li, X., He, H., and Wang, L.: A Review of Fine-Scale Land Use and
Land Cover Classification in Open-Pit Mining Areas by Remote Sensing
Techniques, Remote Sens., 10, 15, <a href="https://doi.org/10.3390/rs10010015" target="_blank">https://doi.org/10.3390/rs10010015</a>, 2017.
</mixed-citation></ref-html>
<ref-html id="bib1.bib8"><label>8</label><mixed-citation>
Denton, E., Zaremba, W., Bruna, J., LeCun, Y., and Fergus, R.: Exploiting
Linear Structure Within Convolutional Networks for Efficient Evaluation, in:
Advances in Neural Information Processing Systems, 8–13
December 2014, Montreal, Canada,  https://arxiv.org/abs/1404.0736, 2014.
</mixed-citation></ref-html>
<ref-html id="bib1.bib9"><label>9</label><mixed-citation>
Ge, Z., Liu, S., Wang, F., Li, Z., and Sun, J.: YOLOX: Exceeding yolo series
in 2021, arXiv [preprint],
<a href="https://doi.org/10.48550/arxiv.2107.08430" target="_blank">https://doi.org/10.48550/arxiv.2107.08430</a>, 6 August 2021.
</mixed-citation></ref-html>
<ref-html id="bib1.bib10"><label>10</label><mixed-citation>
He, L., Wang, H., and Zhang, H.: Object detection by parts using appearance,
structural and shape features, in: 2011 IEEE International Conference on
Mechatronics and Automation, Beijing, China, 7–10 August 2011, 489–494,
<a href="https://doi.org/10.1109/icma.2011.5985611" target="_blank">https://doi.org/10.1109/icma.2011.5985611</a>, 2011.
</mixed-citation></ref-html>
<ref-html id="bib1.bib11"><label>11</label><mixed-citation>
Huang, J., Zhang, H., Wang, L., Zhang, Z., and Zhao, C.: Improved YOLOv3
Model for miniature camera detection, Opt. Laser Technol., 142, 107133,
<a href="https://doi.org/10.1016/j.optlastec.2021.107133" target="_blank">https://doi.org/10.1016/j.optlastec.2021.107133</a>, 2021.
</mixed-citation></ref-html>
<ref-html id="bib1.bib12"><label>12</label><mixed-citation>
Huang, L., Fu, Q., He, M., Jiang, D., and Hao, Z.: Detection algorithm of
safety helmet wearing based on deep learning, Concurr. Comp.-Pract. E., 33, e6234, <a href="https://doi.org/10.1002/cpe.6234" target="_blank">https://doi.org/10.1002/cpe.6234</a>,
2021.
</mixed-citation></ref-html>
<ref-html id="bib1.bib13"><label>13</label><mixed-citation>
Huang, L., Chen, C., Yun, J., Sun, Y., Tian, J., Hao, Z., Yu, H., and Ma,
H.: Multi-Scale Feature Fusion Convolutional Neural Network for Indoor Small
Target Detection, Frontiers in Neurorobitics, 16, 881021,
<a href="https://doi.org/10.3389/fnbot.2022.881021" target="_blank">https://doi.org/10.3389/fnbot.2022.881021</a>, 2022.
</mixed-citation></ref-html>
<ref-html id="bib1.bib14"><label>14</label><mixed-citation>
Ji, S., Li, W., Zhang, B., Zhou, L., and Duan, C.: Bucket Teeth Detection
Based on Faster Region Convolutional Neural Network, IEEE Access, 9,
17649–17661, <a href="https://doi.org/10.1109/access.2021.3054436" target="_blank">https://doi.org/10.1109/access.2021.3054436</a>, 2021.
</mixed-citation></ref-html>
<ref-html id="bib1.bib15"><label>15</label><mixed-citation>
Ji, Y., Lu, Q., and Yao, Q.: Short communication: A case study of stress monitoring with non-destructive stress measurement and deep learning algorithms, Mech. Sci., 13, 291–296, <a href="https://doi.org/10.5194/ms-13-291-2022" target="_blank">https://doi.org/10.5194/ms-13-291-2022</a>, 2022.
</mixed-citation></ref-html>
<ref-html id="bib1.bib16"><label>16</label><mixed-citation>
Li, Z., Xu, X., Xie, L., and Su, H.: Learning Slimming SSD through Pruning
and Knowledge Distillation, in: 2019 Chinese Automation Congress, 22–24 November 2019, Hangzhou,
China, 2701–2705,
<a href="https://doi.org/10.1109/cac48633.2019.8996995" target="_blank">https://doi.org/10.1109/cac48633.2019.8996995</a>, 2019.
</mixed-citation></ref-html>
<ref-html id="bib1.bib17"><label>17</label><mixed-citation>
Lim, S., Soares, J., and Zhou, N.: Tooth guard: A vision system for
detecting missing tooth in rope mine shovel, in: 2016 IEEE Winter Conference
on Applications of Computer Vision, 7–10 March
2016, New York, United States, 1–7, <a href="https://doi.org/10.1109/wacv.2016.7477594" target="_blank">https://doi.org/10.1109/wacv.2016.7477594</a>, 2016.
</mixed-citation></ref-html>
<ref-html id="bib1.bib18"><label>18</label><mixed-citation>
Liu, X., Qi, X., and Jiang, Y.: Electric Shovel Teeth Missing Detection
Method Based on Deep Learning, Comput. Intel. Neurosc.,
2021, 6503029, <a href="https://doi.org/10.1155/2021/6503029" target="_blank">https://doi.org/10.1155/2021/6503029</a>, 2021.
</mixed-citation></ref-html>
<ref-html id="bib1.bib19"><label>19</label><mixed-citation>
Liu, Z., Li, J., Shen, Z., Huang, G., Yan, S., and Zhang, C.: Learning
efficient convolutional networks through network slimming, in: 2017 IEEE
International Conference on Computer Vision, 22–29 October
2017, Venice, Italy, 2736–2744, <a href="https://doi.org/10.1109/iccv.2017.298" target="_blank">https://doi.org/10.1109/iccv.2017.298</a>, 2017.
</mixed-citation></ref-html>
<ref-html id="bib1.bib20"><label>20</label><mixed-citation>
Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.: You Only Look Once:
Unified, Real-Time Object Detection, in: 2016 IEEE Conference on Computer
Vision and Pattern Recognition, 27–30 June 2016, Washington, United States,
779–788, <a href="https://doi.org/10.1109/cvpr.2016.91" target="_blank">https://doi.org/10.1109/cvpr.2016.91</a>, 2016.
</mixed-citation></ref-html>
<ref-html id="bib1.bib21"><label>21</label><mixed-citation>
Redmon, J. and Farhadi, A.: YOLO9000: Better, Faster, Stronger, in: 30th
IEEE Conference on Computer Vision and Pattern Recognition, 21–26 July 2017, Honolulu, United
States, 6517–6525, <a href="https://doi.org/10.1109/cvpr.2017.690" target="_blank">https://doi.org/10.1109/cvpr.2017.690</a>,
2017.
</mixed-citation></ref-html>
<ref-html id="bib1.bib22"><label>22</label><mixed-citation>
Shen, X., Huang, Q., and Xiong, G.: Modelling and predictive investigation on the vibration response of a propeller shaft based on a convolutional neural network, Mech. Sci., 13, 485–494, <a href="https://doi.org/10.5194/ms-13-485-2022" target="_blank">https://doi.org/10.5194/ms-13-485-2022</a>, 2022.
</mixed-citation></ref-html>
<ref-html id="bib1.bib23"><label>23</label><mixed-citation>
Singla, S., Kang, A., Grewal, J., and Cheema, G.: Wear behavior of weld
overlays on excavator bucket teeth, Proc. Mat. Sci., 5, 256–266,
<a href="https://doi.org/10.1016/j.mspro.2014.07.265" target="_blank">https://doi.org/10.1016/j.mspro.2014.07.265</a>, 2014.
</mixed-citation></ref-html>
<ref-html id="bib1.bib24"><label>24</label><mixed-citation>
Sun, Y., Zhao, Z., Jiang, D., Tong, X., Tao, B., Jiang, G., Kong, J., Yun,
J., Liu, X., Zhao, G., and Fang, Z.: Low-illumination image enhancement
algorithm based on improved multi-scale Retinex and ABC algorithm
optimization, Frontiers in Bioengineering and Biotechnology, 10, 865820,
<a href="https://doi.org/10.3389/fbioe.2022.865820" target="_blank">https://doi.org/10.3389/fbioe.2022.865820</a>, 2022.

</mixed-citation></ref-html>
<ref-html id="bib1.bib25"><label>25</label><mixed-citation>
Tan, L., Lv, X., Lian, X., and Wang, G.: YOLOv4_Drone: UAV
image target detection based on an improved YOLOv4 algorithm, Comput.
Electr. Eng., 93, 107261, <a href="https://doi.org/10.1016/j.compeleceng.2021.107261" target="_blank">https://doi.org/10.1016/j.compeleceng.2021.107261</a>,
2021.
</mixed-citation></ref-html>
<ref-html id="bib1.bib26"><label>26</label><mixed-citation>
Wang, G., Zheng, H., and Zhang, X.: A robust checkerboard corner detection
method for camera calibration based on improved YOLOX, Front. Phys.,
9, 819019, <a href="https://doi.org/10.3389/fphy.2021.819019" target="_blank">https://doi.org/10.3389/fphy.2021.819019</a>, 2022.
</mixed-citation></ref-html>
<ref-html id="bib1.bib27"><label>27</label><mixed-citation>
Wang, X., Sun, W., Li, E., and Song, X.: Energy-minimum optimization of the
intelligent excavating process for large cable shovel through trajectory
planning, Struct. Multidiscip. O., 58, 2219–2237,
<a href="https://doi.org/10.1007/s00158-018-2011-6" target="_blank">https://doi.org/10.1007/s00158-018-2011-6</a>, 2018.
</mixed-citation></ref-html>
<ref-html id="bib1.bib28"><label>28</label><mixed-citation>
Yu, F., Koltun, V., and Funkhouser, T.: Dilated residual networks, in: 30th
IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, United
States, 21–26 July 2017, 636–644, <a href="https://doi.org/10.1109/cvpr.2017.75" target="_blank">https://doi.org/10.1109/cvpr.2017.75</a>,
2017.
</mixed-citation></ref-html>
<ref-html id="bib1.bib29"><label>29</label><mixed-citation>
Yu, Z., Shen, Y., and Shen, C.: A real-time detection approach for bridge
cracks based on YOLOv4-FPM, Automat. Constr., 122, 103514,
<a href="https://doi.org/10.1016/j.autcon.2020.103514" target="_blank">https://doi.org/10.1016/j.autcon.2020.103514</a>, 2021.
</mixed-citation></ref-html>
<ref-html id="bib1.bib30"><label>30</label><mixed-citation>
Yun, J., Jiang, D., Liu, Y., Sun, Y., Tao, B., Kong, J., Tian, J., Tong, X.,
Xu, M., and Fang, Z.: Real-time target detection method based on lightweight
convolutional neural network, Frontiers in Bioengineering and Biotechnology,
10, 861286, <a href="https://doi.org/10.3389/fbioe.2022.861286" target="_blank">https://doi.org/10.3389/fbioe.2022.861286</a>, 2022.
</mixed-citation></ref-html>
<ref-html id="bib1.bib31"><label>31</label><mixed-citation>
Zhang, Y., Huang, J., and Cai, F.: On Bridge Surface Crack Detection Based
on an Improved YOLO v3 Algorithm, IFAC-PapersOnLine, 53, 8205–8210,
<a href="https://doi.org/10.1016/j.ifacol.2020.12.1994" target="_blank">https://doi.org/10.1016/j.ifacol.2020.12.1994</a>, 2020.
</mixed-citation></ref-html>--></article>
