agronomy

Article
Plant Disease Recognition Model Based on Improved YOLOv5
Zhaoyi Chen 1 , Ruhui Wu 2 , Yiyan Lin 1 , Chuyu Li 1 , Siyu Chen 1 , Zhineng Yuan 2 , Shiwei Chen 2
and Xiangjun Zou 1,3, *

1 College of Engineering, South China Agricultural University, Guangzhou 510642, China;

[email protected] (Z.C.); [email protected] (Y.L.); [email protected] (C.L.);
[email protected] (S.C.)
2 Guangdong Agribusiness Tropical Agricultrue Institute Co., Ltd., Guangzhou 511365, China;
[email protected] (R.W.); [email protected] (Z.Y.); [email protected] (S.C.)
3 Foshan-Zhongke Innovation Research Institute of Intelligent Agriculture, Foshan 528010, China
* Correspondence: [email protected]

Abstract: To accurately recognize plant diseases under complex natural conditions, an improved
plant disease-recognition model based on the original YOLOv5 network model was established. First,
a new InvolutionBottleneck module was used to reduce the numbers of parameters and calculations,
and to capture long-distance information in the space. Second, an SE module was added to improve
the sensitivity of the model to channel features. Finally, the loss function ‘Generalized Intersection
over Union’ was changed to ‘Efficient Intersection over Union’ to address the former’s degeneration
into ‘Intersection over Union’. These proposed methods were used to improve the target recognition
effect of the network model. In the experimental phase, to verify the effectiveness of the model,
sample images were randomly selected from the constructed rubber tree disease database to form
training and test sets. The test results showed that the mean average precision of the improved
YOLOv5 network reached 70%, which is 5.4% higher than that of the original YOLOv5 network. The



precision values of this model for powdery mildew and anthracnose detection were 86.5% and 86.8%,


respectively. The overall detection performance of the improved YOLOv5 network was significantly
Citation: Chen, Z.; Wu, R.; Lin, Y.; Li,
better compared with those of the original YOLOv5 and the YOLOX_nano network models. The
C.; Chen, S.; Yuan, Z.; Chen, S.; Zou,
X. Plant Disease Recognition Model
improved model accurately identified plant diseases under natural conditions, and it provides a
Based on Improved YOLOv5. technical reference for the prevention and control of plant diseases.
Agronomy 2022, 12, 365.
https://doi.org/10.3390/ Keywords: plant diseases recognition; YOLOv5; InvolutionBottleneck; SE module; EIOU
agronomy12020365

Academic Editor: Imre J. Holb

Received: 24 December 2021 1. Introduction

Accepted: 29 January 2022 Agricultural production is an indispensable part of a nation’s economic development.
Published: 31 January 2022 Crops are affected by climate, which may make them susceptible to pathogen infection
Publisher’s Note: MDPI stays neutral during the growth period, resulting in reduced production. In severe cases, the leaves
with regard to jurisdictional claims in fall off early and the plants die. To reduce the economic losses caused by diseases, it is
published maps and institutional affil- necessary to properly diagnose plant diseases. Currently, two methods are used: expert
iations. diagnoses and pathogen analyses. The former refers to plant protection experts, with
years of field production and real-time investigatory experience, diagnosing the extent of
plant lesions. This method relies highly on expert experience and is prone to subjective
differences and a low accuracy [1]. The latter involves the cultivation and microscopic
Copyright: © 2022 by the authors. observation of pathogens. This method has a high diagnostic accuracy rate, but it is time
Licensee MDPI, Basel, Switzerland.
consuming, and the operational process is cumbersome, making it not suitable for field
This article is an open access article
detection [2,3].
distributed under the terms and
In recent years, the rapid development of machine vision and artificial intelligence has
conditions of the Creative Commons
accelerated the process of engineering intelligence in various fields, and machine vision
Attribution (CC BY) license (https://
technology has also been rapidly improved in industrial, agricultural and other complex
creativecommons.org/licenses/by/
4.0/).
scene applications [4–9]. In response to the plant disease detection problem, disease

Agronomy 2022, 12, 365. https://doi.org/10.3390/agronomy12020365 https://www.mdpi.com/journal/agronomy

Agronomy 2022, 12, 365 2 of 14

detection methods based on visible light and near-infrared spectroscopic digital images
have been widely used. Near-infrared spectroscopic and hyperspectral images contain
continuous spectral information and provide information on the spatial distributions
of plant diseases. Consequently, they have become the preferred technologies of many
researchers [10–13]. However, the equipment for acquiring spectral images is expensive
and difficult to carry; therefore, this technology cannot be widely applied. The acquisition
of visible light images is relatively simple and can be achieved using various ordinary
electronic devices, such as digital cameras and smart phones, which greatly reduces the
challenges of visible light image-recognition research [14,15].
Because of the need for real-time monitoring and sharing of crop growth information,
visible light image recognition has been successfully applied to the field of plant disease
detection in recent years [16–20]. A variety of traditional image-processing methods have
been applied. First, the images are segmented, then the characteristics of plant diseases are
extracted and, finally, the diseases are classified. Shrivastava et al. [21] proposed an image-
based rice plant disease classification approach using color features only, and it successfully
classifies rice plant diseases using a support vector machine classifier. Alajas et al. [22]
used a hybrid linear discriminant analysis and a decision tree to predict the percentage of
damaged leaf surface on diseased grapevines, with an accuracy of 97.79%. Kianat et al. [23]
proposed a hybrid framework based on feature fusion and selection techniques to classify
cucumber diseases. They first used the probability distribution-based entropy approach
to reduce the extracted features, and then, they used the Manhattan distance-controlled
entropy technique to select strong features. Mary et al. [24] used the merits of both the
Gabor filter and the 2D log Gabor filter to construct an enhanced Gabor filter to extract
features from the images of the diseased plant, and then, they used the k-nearest neighbor
classifier to classify banana leaf diseases. Sugiarti et al. [25] combined the grey-level
co-occurrence matrix extraction function with the naive Bayes classification to greatly
improved the classification accuracy of apple diseases. Mukhopadhyay et al. [26] proposed
a novel method based on image-processing technology, and they used the non-dominated
sorting genetic algorithm to detect the disease area on tea leaves, with an average accuracy
of 83%. However, visible light image-recognition based on traditional image processing
technologies requires the artificial preprocessing of images and the extraction of disease
features. The feature information is limited to shallow learning, and the generalization
ability of new data sets needs to be improved.
However, deep learning methods are gradually being applied to agricultural research
because they can automatically learn the deep feature information of images, and their
speed and accuracy levels are greater than those of traditional algorithms [27–30]. Deep
learning has also been applied to the detection of plant diseases from visible light im-
ages. Abbas et al. [31] proposed a deep learning-based method for tomato plant disease
detection that utilizes the conditional generative adversarial network to generate synthetic
images of tomato plant leaves. Xiang et al. [32] established a lightweight convolutional
neural network-based network model with channel shuffle operation and multiple-size
modules that achieved accuracy levels of 90.6% and 97.9% on a plant disease severity
and PlantVillage datasets, respectively. Tan et al. [33] compared the recognition effects
of deep learning networks and machine learning algorithms on tomato leaf diseases and
found that the metrics of the tested deep learning networks are all better than those of
the measured machine learning algorithms, with the ResNet34 network obtaining the
best results. Alita et al. [34] used the EfficientNet deep learning model to detect plant leaf
disease, and it was superior to other state-of-the-art deep learning models in terms of
accuracy. Mishra et al. [35] developed a sine-cosine algorithm-based rider neural network
and found that the detection performance of the classifier improved, achieving an accuracy
of 95.6%. In summary, applying deep learning to plant disease detection has achieved good
results.
As a result of climatic factors, rubber trees may suffer from a variety of pests and
diseases, most typically powdery mildew and anthracnose, during the tender leaf stage.
Agronomy 2022, 12, 365 3 of 14

Rubber tree anthracnose is caused by Colletotrichum gloeosporioides and Colletotrichum acuta-

tum infections, whereas rubber tree powdery mildew is caused by Oidiumheveae [36,37]. The
lesion features of the two diseases are highly similar, making them difficult to distinguish,
which has a certain impact on the classification results of the network model. Compared
with traditional image processing technology, deep convolutional neural networks have
greater abilities to express abstract features and can obtain semantic information from
complex images.Target detection algorithms based on deep learning can be divided into
two categories, one-stage detection algorithms (such as the YOLO series) and two-stage
detection algorithms (such as FasterR-CNN). The processing speeds of the former are faster
than those of the latter, which makes them more suitable for the real-time detection of plant
diseases in complex field environments.
In this paper, we report our attempts to address the above issues, as follows: First, we
used convolutional neural networks to automatically detect rubber tree powdery mildew
and anthracnose in visible light images, which has some practical benefits for the prevention
and control of rubber tree diseases. Second, we focused on solving the existing difficulties
in detecting rubber tree diseases using YOLOv5, and we further improved the detection
accuracy of the model. Consequently, a rubber tree disease recognition model based on the
improved YOLOv5 was established, with the aim of achieving the accurate classification
and recognition of rubber tree powdery mildew and anthracnose under natural light
conditions. The main contributions of our work are summarized below:
(1) In the backbone network, the Bottleneck module in the C3 module was replaced
with the InvolutionBottleneck module that reduced the number of calculations in the
convolutional neural network;
(2) The SE module was added to the last layer of the backbone network to fuse disease
characteristics in a weighted manner;
(3) The existing loss function Generalized Intersection over Union (GIOU) in YOLOv5
was replaced by the loss function Efficient Intersection over Union (EIOU), which
takes into account differences in target frame width, height and confidence;
(4) The proposed model can realize the accurate and automatic identification of rubber
tree diseases in visible light images, which has some significance for the prevention
and control of rubber tree diseases.
The remainder of this article is organized as follows: In Section 2, we give a brief
review of the original YOLOv5 model, and the improved YOLOv5 model is proposed. In
Section 3, we list the experimental materials and methods. Experiments and analyses of the
results are covered in Section 4. Finally, the conclusions are summarized in Section 5.

2. Principle of the Detection Algorithm

2.1. YOLOv5 Network Module
YOLOv5 [38] is a one-stage target recognition algorithm proposed by Glenn Jocher in
2020. On the basis of differences in network depth and width, YOLOv5 can be divided into
four network model versions: YOLOv5s, YOLOv5m, YOLOv5l and YOLOv5x. Among
them, the YOLOv5s network has the fastest calculation speed, but the average precision is
the lowest, whereas the YOLOv5x network has the opposite characteristics. The model size
of the YOLOv5 network is approximately one-tenth that of the YOLOv4 network. It has
faster recognition and positioning speeds, and the accuracy is no less than that of YOLOv4.
The YOLOv5 network is composed of three main components: Backbone, Neck and Head.
After the image is inputted, Backbone aggregates and forms image features on different
image granularities. Then, Neck stitches the image features and transmits them to the
prediction layer, and Head predicts the image features to generate bounding boxes and
predicted categories. The YOLOv5 network uses GIOU as the network loss function, as
shown in Equation (1).
|C − ( A ∪ B)|
GIOU = IOU − (1)
|C |
Agronomy 2022, 12, 365 4 of 14

where A, B ⊆ S ⊆ Rn represent two arbitrary boxes, C represents the smallest convex box,
C ⊆ S ⊆ Rn , enclosing both A and B and IOU = | A ∩ B|/| A ∪ B|.
When the input network predicts image features, the optimal target frame is filtered
by combining the loss function GIOU and the non-maximum suppression algorithm.

2.2. Improved YOLOv5 Network Construction

2.2.1. InvolutionBottleneck Module Design
In the Backbone, the Bottleneck module in the C3 module was replaced with the
InvolutionBottleneck module. The two inherent principles of standard convolution kernels
are spatial-agnostic and channel-specific, whereas those of involution [39] are the opposite.
Convolutional neural networks usually increase the receptive field by stacking convolution
kernels of different sizes. Using different kernel calculations for each channel causes a sub-
stantial increase in the number of calculations. In the Backbone, the InvolutionBottleneck
module was used to replace the Bottleneck module, which alleviated the kernel redundancy
by sharing the involution kernel along the channel dimension, and this is beneficial for
capturing the long-distance information of the spatial range and reducing the number of
network parameters. The output feature map Y of the Involution–convolution operation is
defined as shown in Equations (2) and (3).
 
Hi,j = φ Xψi,j (2)

Yi,j,k = ∑ Hi,j,u+bK/2c,v+bK/2c,dkG/Ce Xi+u,j+v,k (3)

(u,v)∈∆k

where C represents the calculation channel, φ represents the generation function of the
convolution kernel and ψi,j represents the index to the pixel set Hi,j . The convolution kernel
Hi,j,...,g ∈ RK ×K ( g = 1, 2, . . . , G ) was specifically customized for the pixel Xi,j ∈ RC located
at the corresponding coordinates (i, j), but it is shared on the channel. G represents the
number of groups sharing the same convolution kernel. The size of the convolution kernel
depends on the size of the input feature map.

2.2.2. SE Module Design

The squeeze-and-excitation network [40] is a network model proposed by Hu et al.
(2017) that focuses on the relationship between channels. It aims to learn each image feature
according to the loss function, increase the weight of effective image features and reduce
the weight of invalid or ineffective image features, thereby training the network model to
produce the best results. The SE modules with different structures are shown in Figure 1.
The SE module is a calculation block that can be built on the transformation between
the input feature vector X and the output feature map u, and the transformation relationship
is shown in Equation (4):
C0
uc = vc ∗ X = ∑ vsc ∗ xs (4)
s =1
h 0
i h i
where ∗ represents convolution,vc = v1c , v2c , . . . , vC , X = X 1 , X 2 , . . . , X C 0 and u ∈
c c
R H ×W . vsc represents a 2D spatial kernel, which denotes a singlechannel of vc that acts on
the corresponding channel of X.
In this paper, the SE module was added to the last layer of the Backbone, allowing it
to merge the image features of powdery mildew and anthracnose in a weighted manner,
thereby improving the network performance at a small cost.
12, x FOR PEER REVIEW 5 of 14
Agronomy 2022, 12, 365 5 of 14

Figure 1. SE modules with different structures. (a) SE module with Inception structure; (b) SE module
Figure 1. SE modules with different structures. (a) SE module with Inception structure; (b) SE
with Residual structure.
module with Residual structure.
2.2.3. Loss Function Design
The SE module Theisloss
a calculation
function was block
changedthat can
from GIOUbe to
built
EIOUon[41].
theThetransformation
GIOU function be-
was
proposed on the basis of the IOU function. It solves the problem
tween the input feature vector X and the output feature map U, and the transformation of the IOU not being able
to reflect how the two boxes intersect. However, if the anchor and target boxes are part
relationship is shown in Equation (4):
of a containment relationship, then GIOU will still degenerate into IOU. Therefore, we
changed the loss function GIOU to EIOU. EIOU was obtained on the basis of complete-
IOU loss (CIOU), and it not only takes into account the central point distance and (4)the
𝑢 =𝑣 ∗𝑋 = 𝑣 ∗𝑥
aspect ratio, but also the true discrepancies in the target and anchor boxes’ widths and
heights. The EIOU loss function directly minimizes these discrepancies and accelerates
where ∗ model convergence.
represents 𝑣 =loss
The EIOU
convolution, 𝑣 function
, 𝑣 , … , 𝑣is shown
, 𝑋 =in 𝑋 , 𝑋 , …(5).
Equation ,𝑋 and 𝑢 ∈
𝑅 × .𝑣 represents a 2D spatial kernel, which denotes b gt


ρ2 w, w gt of ρ𝑣2 h,that
ρ2 ab, singlechannel


h gt acts


on the corresponding L EIOU
channel of+𝑋.
= L IOU Ldis + L asp = 1 − IOU +
C2
+
Cw2
+
Ch2
(5)
In this paper, the SE module was added to the last layer of the Backbone, allowing
imageCwfeatures
it to merge the where and Ch represent
of powderythe width and height,
mildew andrespectively,
anthracnose of the
in smallest enclosing
a weighted box
man-
covering the two boxes; b and b gt represent the central points of the predicted and target
ner, thereby improving the network performance at a small cost.
boxes, respectively; ρ represents the Euclidean distance; C represents the diagonal length
of the smallest enclosing box covering the two boxes. The loss function EIOU is divided
2.2.3. Loss Function Design
into three parts: the IOU loss L IOU , the distance loss Ldis and the aspect loss L asp .
Combined with the InvolutionBottleneck and the SE modules, the whole improved
The loss function was changed from GIOU to 𝐸𝐼𝑂𝑈 [41]. The GIOU function was
YOLOv5 network model framework is constructed, as shown in Figure 2.
proposed on the basis of the IOU function. It solves the problem of the IOU not being
able to reflect how the two boxes intersect. However, if the anchor and target boxes are
part of a containment relationship, then GIOU will still degenerate into IOU. Therefore,
we changed the loss function GIOU to 𝐸𝐼𝑂𝑈. 𝐸𝐼𝑂𝑈 was obtained on the basis of com-
plete-IOU loss (CIOU), and it not only takes into account the central point distance and
the aspect ratio, but also the true discrepancies in the target and anchor boxes’ widths
and heights. The 𝐸𝐼𝑂𝑈 loss function directly minimizes these discrepancies and accel-
 and target boxes, respectively; 𝜌 represents the Euclidean distance; 𝐶 represents the
diagonal length of the smallest enclosing box covering the two boxes. The loss function
𝐸𝐼𝑂𝑈 is divided into three parts: the 𝐼𝑂𝑈 loss 𝐿 , the distance loss 𝐿 and the as-
pect loss 𝐿 .
Agronomy 2022, 12, 365 Combined with the InvolutionBottleneck and the SE modules, the whole 6improved
of 14
YOLOv5 network model framework is constructed, as shown in Figure 2.

Figure
Figure 2. The
2. The improvedYOLOv5
improved YOLOv5 network
networkmodel
modelstructure.
structure.
3. Materials and Methods
3. Materials and Methods
3.1. Experimental Materials
3.1. Experimental
The imagesMaterials
of rubber tree diseases were collected from a rubber plantation in Shengli
State
TheFarm,
imagesMaoming City,tree
of rubber China. It is located
diseases at 22◦ 60 N,from
were collected 110◦a 0 E, with an altitude of
80rubber plantation in Shen-
34–69 m, an average annual precipitation of 1698.1 mm and an annual average temperature
gli State Farm,◦Maoming City, China. It is located at 22°6′ N, 110°80′ E, with an altitude of
of 19.9–26.5 C. The high humidity and warm climate are conducive to widespread epi-
34–69 m, an average annual precipitation of 1698.1 mm and an annual average temper-
demics of powdery mildew and anthracnose. To ensure the representativeness of the image
ature
set,of 19.9–26.5
they °C. Theunder
were collected highnatural
humidity lightand warm climate
conditions. A Sonyare conducive
digital ILCE-7m3to camera
widespread
epidemics
was used to photograph powdery mildew and anthracnose of rubber leaves at differentof the
of powdery mildew and anthracnose. To ensure the representativeness
image set,with
angles, theyanwere
imagecollected
resolutionunder
of 6000natural light
× 4000 conditions.
pixels. A Sony
There were digital in
2375 images ILCE-7m3
the
camera
rubberwas
tree used
diseasetodatabase,
photograph powdery
including mildewmildew
1203 powdery and anthracnose of rubber
images and 1172 leaves at
anthracnose
images,angles,
different which were
with used for theresolution
an image training and of testing
6000 × of disease
4000 recognition
pixels. models.
There were 2375We images
in identified
the rubber these two
tree diseasesdatabase,
disease with the guidance
including of plant
1203 protection
powderyexperts.
mildewImages
imagesof these
and 1172
rubber treeimages,
anthracnose diseases which
are shownwerein Figure
used for 3. the training and testing of disease recognition
models. We identified these two diseases with the guidance of plant protection experts.
Images of these rubber tree diseases are shown in Figure 3.
Agronomy 2022, 12,
Agronomy x FOR
2022, PEER REVIEW
12, 365 7 of 14
7 of 14

Agronomy 2022, 12, x FOR PEER REVIEW 7 of 14

Figure 3. Rubber
Figure
Figure tree
3. Rubber
3. treediseases
diseases images. (a)Powdery
images. (a)
images. Powderymildew
mildew image;
image; (b) (b)
(b) Anthracnose
Anthracnose
Anthracnose image.
image.
image.

3.2.3.2.
Data
3.2. Data Preprocessing
Preprocessing
Data Preprocessing
Before
Before
Before the
the images
theimages were
were inputted
imageswere inputted into
inputted intothe
into theimproved
the improved
improved YOLOv5
YOLOv5
YOLOv5 network
network
network model,
model, the
the the
model,
mosaic
mosaic data
data enhancement
enhancement method
method was
was used
used to
to expand
expand the
the image
image set.
set. The
The images
images were
were
mosaic data enhancement method was used to expand the image set. The images were
spliced
spliced using several
usingseveral methods,
methods, such
severalmethods, such asas random
random scaling,
scaling, random
random cropping
cropping and
and random
random
spliced using
arrangement, such astherandom scaling, random cropping and random
arrangement, which not only expanded the image set, but also improved the detection of
which not only expanded image set, but also improved the detection of
arrangement,
small which not only expanded the image set, but also improved the detection of
small targets.
targets.In
Inaddition,
addition,before
beforetraining
trainingthethemodel,
model, adaptive scaling
adaptive andand
scaling filling operations
filling opera-
small
weretargets. In addition,
performed on the beforeoftraining the model, adaptive scaling and filling opera-
tions were performed onimages
the images rubber
of rubbertree diseases,
tree diseases,andand the
theinput
inputimage
imagesize size was
was
tions were
normalizedperformed
to 640 × on
640 the images
pixels. The of rubber
preprocessing tree diseases,
results are and
shown
normalized to 640 × 640 pixels. The preprocessing results are shown in Figure 4. the
in input
Figure image
4. size was
normalized to 640 × 640 pixels. The preprocessing results are shown in Figure 4.

Figure 4.
Figure 4. Image
Image preprocessing
preprocessing result.
result.

3.3. Experimental Equipment

Figure 4.AImage preprocessing
desktop computer wasresult.
used as the processing platform, the operating system was
Ubuntu
Ubuntu 18.04 and the Pytorch framework
framework and the YOLOv5 environment
environment were
were built
built in the
3.3.Anaconda3
Experimental
Anaconda3 EquipmentThe
environment.
environment. The program
program was
was written
written in
in Python
Python 3.8,
3.8, and
and the CUDA version
was
A 10.1.
was 10.1. For
For hardware,
desktop computerthe
hardware, the
wasprocessor
used aswas
processor the an
was an Intel
Intel Core
processing i3-4150,
i3-4150, the
Coreplatform, the main
main
the frequency
frequency
operating was
was was
system
3.5
3.5 GHz,
GHz, the
the memory
memory was
was 3G
3G and
and the
the graphics
graphics card
card was
was a
a GeForce
GeForce GTX
GTX 1060
1060 6G.
6G. The
Thein the
Ubuntu 18.04 and the Pytorch framework and the YOLOv5 environment were built
specific
specific configurations
configurations are
are provided
provided in
in Table
Table 1.1.
Anaconda3 environment. The program was written in Python 3.8, and the CUDA version
was 10.1. For hardware, the processor was an Intel Core i3-4150, the main frequency was
3.5 GHz, the memory was 3G and the graphics card was a GeForce GTX 1060 6G. The
specific configurations are provided in Table 1.
Agronomy 2022, 12, x FOR PEER REVIEW 8 of 14

Table 1. Test environment setting.

Agronomy 2022, 12, 365 8 of 14
Parameter Configuration
Operating system Ubuntu 18.04
Deep
Table learning
1. Test framework
environment setting. Pytorch 1.8
Programming language Python 3.8
GPU accelerated Parameter
environment CUDA Configuration
10.1
GPU
Operating system GeForce GTX 1060 18.04
Ubuntu 6G
Deep learning framework
CPU Pytorch
Intel(R) Core(TM) i3-4150 CPU1.8@ 3.50GHz
Programming language Python 3.8
GPU accelerated environment CUDA 10.1
3.4. Experimental Process GPU GeForce GTX 1060 6G
First, the manual labeling
CPU method was used to mark each
Intel(R) rubberi3-4150
Core(TM) diseaseCPU
image for
@ 3.50GHz
powdery mildew or anthracnose to obtain training label images, and then the disease
image 3.4.
set was divided Process
Experimental at a 4:1:1 ratio into training, validation and test sets. The training
set was inputted into the improved
First, the manual labeling YOLOv5
methodnetwork
was used oftodifferent
mark eachstructures
rubberfor training.
disease image for
The training process was divided into 80 batches, with each batch
powdery mildew or anthracnose to obtain training label images, and then containing 96 images.
the disease
The Stochastic
image set Gradient
was dividedDescent algorithm
at a 4:1:1 was
ratio into used to
training, optimize
validation andthetest
network
sets. Themodel
training set
duringwastheinputted
traininginto
process, and the YOLOv5
the improved optimal network
network of weight wasstructures
different obtainedfor after the The
training.
training was completed. Subsequently, the performance of the network model
training process was divided into 80 batches, with each batch containing 96 images. The was de-
termined using the
Stochastic test setDescent
Gradient and compared
algorithm with
was the testtoresults
used of the
optimize theoriginal
networkYOLOv5
model during
and thethe training process, and the optimal network weight was obtained was
YOLOX_nano networks. The network model with the best result afterselected as was
the training
the rubber tree disease
completed. recognition
Subsequently, the model. The test
performance process
of the is shown
network modelinwasFigure 5.
determined using the
test set and compared with the test results of the original YOLOv5 and the YOLOX_nano
networks. The network model with the best result was selected as the rubber tree disease
recognition model. The test process is shown in Figure 5.

Figure 5. Test flow chart.

Figure 4. Results
5. Test flowand Analysis
chart.
4.1. Convergence Results of the Network Model
The training and verification sets were inputted into the network for training. After
80 batches of training, the loss function value curves of the training and verification sets
were determined (Figure 6), and they included the detection frame loss, the detection object
loss and the classification loss.
 4. Results and Analysis
4.1. Convergence Results of the Network Model
The training and verification sets were inputted into the network for training. After
80 batches of training, the loss function value curves of the training and verification sets
Agronomy 2022, 12, 365 were determined (Figure 6), and they included the detection frame loss, the detection 9 of 14
object loss and the classification loss.

Figure 6. Convergence of the loss functions of training and validation sets.

Figure 6. Convergence of the loss functions of training and validation sets.
The loss of the detection frame indicates whether an algorithm can locate the center
point The
of loss of thewell
an object detection frame indicates
and whether whether
the detection targetan is algorithm
covered bycan thelocate the center
predicted bound-
point of an object well and whether the detection target is covered
ing box. The smaller the loss function value, the more accurate the prediction frame. by the predictedThe
bounding
object lossbox. The issmaller
function the aloss
essentially function
measure value,
of the the more
probability that accurate the target
the detection prediction
exists
frame.
in the The
regionobject loss function
of interest. is essentially
The smaller the value a measure
of the lossoffunction,
the probability
the higherthatthe
theaccuracy.
detec-
tion target exists in the region of interest. The smaller the value of the
The classification loss represents the ability of the algorithm to correctly predict a given loss function, the
higher
object the accuracy.
category. The The classification
smaller loss represents
the loss value, the ability
the more accurate theof the algorithm to cor-
classification.
rectly Aspredict a given object category. The smaller the
shown in Figure 6, the loss function value had a downward loss value, the more
trendaccurate
duringthethe
classification.
training process, the Stochastic Gradient Descent algorithm optimized the network and
the As shownweight
network in Figure 6, the parameters
and other loss function value
were had a downward
constantly trend the
updated. Before during the
training
training process, the Stochastic Gradient Descent algorithm optimized
batch reached 20, the loss function value dropped rapidly, and the accuracy, recall rate the network and
the
andnetwork
averageweight
accuracy andrapidly
other parameters
improved. were constantly
The network updated.toBefore
continued theWhen
iterate. training
the
batch reached
training batch20, the loss
reached function value
approximately dropped
20, the rapidly,
decrease in theandlossthe accuracy,
function valuerecall rate
gradually
and average
slowed. accuracy
Similarly, rapidly improved.
the increases in parameters Thesuch
network continued
as average accuracyto iterate. WhenWhen
also slowed. the
training batch reached approximately 20, the decrease in the loss function
the training batch reached 80, the loss curves of the training and validation sets showed value gradu-
ally slowed.
almost Similarly, the
no downward increases
trends, in parameters
and other index valuessuchalsoas average
tended to accuracy also slowed.
have stabilized. The
When
networkthe model
training batch reached
basically reached 80, the the loss curves
convergence of the
state, andtraining and network
the optimal validation sets
weight
showed almostatno
was obtained thedownward trends, and other index values also tended to have stabi-
end of training.
lized. The network model basically reached the convergence state, and the optimal net-
4.2. Verification
work weight wasofobtained
the Network Model
at the end of training.
To evaluate the detection performance of the improved YOLOv5 network, it was
crucial to use appropriate evaluation metrics for each problem. The precision, recall,
average precision and mean average precision were used as the evaluation metrics, and
they were respectively defined as follows:

TP
P= (6)
TP + FP
 To evaluate the detection performance of the improved YOLOv5 network, it was
crucial to use appropriate evaluation metrics for each problem. The precision, recall, av-
erage precision and mean average precision were used as the evaluation metrics, and
they were respectively defined as follows:
Agronomy 2022, 12, 365 10 of 14
𝑇𝑃
𝑃= (6)
𝑇𝑃
𝐹𝑃
𝑇𝑃
𝑅 = TP
R= (7)
(7)
𝑇𝑃
TP + FN𝐹𝑁
Z 1
AP𝐴𝑃
i =
= P(𝑃R𝑅
)d(𝑑R)𝑅 (8)
(8)
0
1 1 N
mAP 𝑚𝐴𝑃= = ∑i=1 AP 𝐴𝑃i (9)
(9)
N𝑁
whereTP
where 𝑇𝑃represents
representsthe number
the number of of
positive samples
positive samplesthatthat
are correctly detected,
are correctly FP rep-
detected, 𝐹𝑃
resents the number
represents the numberof negative samples
of negative thatthat
samples are falsely detected
are falsely FN represents
and and
detected the
𝐹𝑁 represents
number of positive
the number samples
of positive that are
samples thatnot
aredetected.
not detected.
In
In total, 200 powdery mildew imagesand
total, 200 powdery mildew images and200 200anthracnose
anthracnoseimages
images were
wererandomly
randomly
selected as the test set and inputted into the improved YOLOv5 network for
selected as the test set and inputted into the improved YOLOv5 network for testing. testing. The test
The
results were compared with those of the original YOLOv5 and theYOLOX_nano
test results were compared with those of the original YOLOv5 and theYOLOX_nano networks.
The comparison
networks. results are shown
The comparison resultsinareFigure
shown 7. in Figure 7.

Figure7.7.Performance
Figure Performancecomparison
comparison ofof
allall
thethe network
network models.
models. (a) Powdery
(a) Powdery mildew
mildew recognition
recognition re-
results;
sults; (b) Anthracnose recognition results; (c) Mean average precision; (d) Processing times
(b) Anthracnose recognition results; (c) Mean average precision; (d) Processing times per photo. per
photo.
As shown in Figure 7, the detection performance of the improved YOLOv5 network
As shown
was better than in Figure
that of the7,original
the detection
YOLOv5performance
network forof the
eachimproved YOLOv5
of the tested network
two diseases
was better than that of the original YOLOv5 network for each of the tested
of rubber trees. For the detection of powdery mildew, precision increased by 8.7% two diseases
and
average precision increased by 1%; however, recall decreased by 1.5%. For the detectionand
of rubber trees. For the detection of powdery mildew, precision increased by 8.7% of
average precision
anthracnose, increased
average byincreased
precision 1%; however, recall
by 9.2% anddecreased by 1.5%.byFor
recall increased the however,
9.3%; detection
precision decreased by 5.2%. Overall, the mean average precision increased by 5.4%. The
improved YOLOV5 network achieved 86.5% and 86.8% precision levels for the detection
of powdery mildew and anthracnose, respectively. In summary, the improved YOLOv5
network’s performance was greatly enhanced compared with that of the original YOLOv5
network; consequently, it achieved more accurate rubber tree disease identification and
location functions. As shown in Figure 7, the performance of the improved YOLOv5
network was better than those of the original YOLOv5 and YOLOX_nano networks for
the detection of the two diseases of rubber trees. Compared with the original YOLOv5
 5.4%. The improved YOLOV5 network achieved 86.5% and 86.8% precision levels for the
detection of powdery mildew and anthracnose, respectively. In summary, the improved
YOLOv5 network’s performance was greatly enhanced compared with that of the origi-
nal YOLOv5 network; consequently, it achieved more accurate rubber tree disease iden-
Agronomy 2022, 12, 365 tification and location functions.As shown in Figure 7, the performance of the improved 11 of 14
YOLOv5 network was better than those of the original YOLOv5 and YOLOX_nano net-
works for the detection of the two diseases of rubber trees. Compared with the original
YOLOv5 network,
network, the precision
the precision of powdery of powdery mildew detection
mildew detection increasedincreased
by 8.7% and by 8.7% and the
the average
average precision increased by 1%; however, recall decreased by 1.5%.
precision increased by 1%; however, recall decreased by 1.5%. The average precision The average pre-
of
cision of anthracnose detection increased by 9.2% and recall increased by
anthracnose detection increased by 9.2% and recall increased by 9.3%; however, precision 9.3%; however,
precision decreased
decreased by 5.2%.the
by 5.2%. Overall, Overall, the mean
mean average averageincreased
precision precisionby increased by 5.4%.
5.4%. Compared
Compared with the YOLOX_nano network, the precision of powdery
with the YOLOX_nano network, the precision of powdery mildew detection increased mildew detection
increased
by 3.7% and by 3.7% and theprecision
the average average precision
increasedincreased
by 0.3%;by 0.3%; however,
however, recall decreased
recall decreased by 2%.
by 2%. The precision of anthracnose detection increased by 4.4% and recall
The precision of anthracnose detection increased by 4.4% and recall increased by 3.8%; increased by
3.8%; however, precision decreased by 4.4%. Overall, themean average
however, precision decreased by 4.4%. Overall, themean average precision increased by precision in-
creased
1.4%. Thebyimproved
1.4%. TheYOLOv5
improved YOLOv5
network network
achieved achieved
86.5% 86.5%
and 86.8% and for
levels 86.8%the levels
detection for
the detection of powdery mildew and anthracnose, respectively.
of powdery mildew and anthracnose, respectively. In summary, the improved YOLOv5 In summary, the im-
proved YOLOv5
network’s network’s
performance performance
was greatly enhancedwascompared
greatly enhanced
with thosecompared with YOLOv5
of the original those of
the original
and YOLOX_nano YOLOv5 and YOLOX_nano
networks; consequently,networks; consequently,
it more accurately locatesitand
more accurately
identifies rubberlo-
catesdiseases.
tree and identifies rubber tree diseases.

4.3. Comparison of Recognition Results

The original YOLOv5, the YOLOX_nano and the improved YOLOv5 networks were
used toto detect
detecttwo
twokinds
kindsofofdiseases
diseases
of of rubber
rubber trees
trees to verify
to verify the the actual
actual classification
classification and
and recognition effects of the improved network. A comparison of test results
recognition effects of the improved network. A comparison of test results is shown in is shown
in Figure
Figure 8. 8.

Figure 8.
Figure 8. Comparison
Comparison of of the
the recognition
recognition effects
effects of
of all
all the
the network
network models.
models. (a–c)
(a–c) Powdery
Powdery mildew
mildew
recognition effects of the (a) original YOLOv5; (b)YOLOX_nano; and (c) improved YOLOv5 net-
recognition effects of the (a) original YOLOv5; (b) YOLOX_nano; and (c) improved YOLOv5 network
work models; (d–f) Anthracnose recognition effects of the (d) original YOLOv5; (e) YOLOX_nano;
models; (d–f) Anthracnose recognition effects of the (d) original YOLOv5; (e) YOLOX_nano; and
and (f) improved YOLOv5 network models.
(f) improved YOLOv5 network models.

As shown
As shown ininFigure
Figure8, comparedwith
8, compared withthe
the other
other networks,
networks, thethe improved
improved network
network sig-
significantly
nificantly improved
improved the detection
the detection of powdery
of powdery mildew,
mildew, including
including obscuredobscured diseased
diseased leaves.
leaves. Additionally,
Additionally, the recognition
the recognition effect
effect of the of the YOLOX_nano
YOLOX_nano network fornetwork
powdery formildew
powdery is
better than that of the original YOLOv5 network. For the detection of anthracnose, the
recognition effects of the three networks were similar, with all three effectively detecting
anthracnose. Therefore, the effectiveness of the improved network for diseased leaves detec-
tion is generally better than those of the original YOLOv5 and the YOLOX_nano networks.

5. Conclusions
The detection and location of plant diseases in the natural environment are of great
significance to plant disease control. In this paper, a rubber tree disease recognition model
based on the improved YOLOv5 network was established. We replaced the Bottleneck
Agronomy 2022, 12, 365 12 of 14

module with the InvolutionBottleneck module to achieve channel sharing within the group
and reduce the number of network parameters. In addition, the SE module was added to
the last layer of the Backbone for feature fusion, which improved network performance at
a small cost. Finally, the loss function was changed from GIOU to EIOU to accelerate the
convergence of the network model. According to the experimental results, the following
conclusions can be drawn:
(1) The model performance verification experiment showed that the rubber tree disease
recognition model based on the improved YOLOv5 network achieved 86.5% precision
for powdery mildew detection and 86.8% precision for anthracnose detection. In gen-
eral, the mean average precision reached 70%, which is an increase of 5.4% compared
with the original YOLOv5 network. Therefore, the improved YOLOv5 network more
accurately identified and classified rubber tree diseases, and it provides a technical
reference for the prevention and control of rubber tree diseases.
(2) A comparison of the detection results showed that the performance of the improved
YOLOv5 network was generally better than those of the original YOLOv5 and the
YOLOX_nano networks, especially in the detection of powdery mildew. The problem
of the missing obscured diseased leaves was improved.
Although the improved YOLOv5 network, as applied to rubber tree disease detection,
achieved good results, the detection accuracy still needs to be improved. In future research,
the network model structure will be further optimized to improve the network performance
of the rubber tree disease recognition model.

Author Contributions: Conceptualization, Z.C. and X.Z.; methodology, Z.C.; software, Z.C.; valida-
tion, Z.C.; formal analysis, Z.C. and X.Z.; investigation, Z.C., X.Z., R.W., Y.L., C.L. and S.C. (Siyu
Chen); resources, Z.C., X.Z., R.W., Z.Y. and S.C. (Shiwei Chen); data curation, Z.C.; writing—original
draft preparation, Z.C.; writing—review and editing, Z.C. and X.Z.; visualization, Z.C.; supervision,
X.Z.; project administration, X.Z. All authors have read and agreed to the published version of
the manuscript.
Funding: The paper is funded by the No. 03 Special Project and the 5G Project of Jiangxi Province
under Grant 20212ABC03A27 and the Key-area Research and Development Program of Guangdong
Province under Grant 2019B020223003.
Data Availability Statement: The data presented in this study are available on request from the cor-
responding author. The data are not publicly available due to the privacy policy of the organization.
Acknowledgments: The authors would like to thank the anonymous reviewers for their critical
comments and suggestions for improving the manuscript.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Lin, G.; Tang, Y.; Zou, X.; Xiong, J.; Fang, Y. Color-, depth-, and shape-based 3D fruit detection. Precis. Agric. 2020, 21, 1–17.
[CrossRef]
2. Joshi, R.C.; Kaushik, M.; Dutta, M.K.; Srivastava, A.; Choudhary, N. VirLeafNet: Automatic analysis and viral disease diagnosis
using deep-learning in Vigna mungoplant. Ecol. Inform. 2021, 61, 101197. [CrossRef]
3. Buja, I.; Sabella, E.; Monteduro, A.G.; Chiriacò, M.S.; De Bellis, L.; Luvisi, A.; Maruccio, G. Advances in Plant Disease Detection
and Monitoring: From Traditional Assays to In-Field Diagnostics. Sensors 2021, 21, 2129. [CrossRef] [PubMed]
4. Liu, S.; Liu, D.; Srivastava, G.; Połap, D.; Woźniak, M. Overview and methods of correlation filter algorithms in object tracking.
Complex Intell. Syst. 2020, 7, 1895–1917. [CrossRef]
5. Tang, Y.; Chen, M.; Wang, C.; Luo, L.; Li, J.; Lian, G.; Zou, X. Recognition and localization methods for vision-based fruit picking
robots: A review. Front. Plant Sci. 2020, 11, 510. [CrossRef]
6. Li, J.; Tang, Y.; Zou, X.; Lin, G.; Wang, H. Detection of fruit-bearing branches and localization of litchi clusters for vision-based
harvesting robots. IEEE Access 2020, 8, 117746–117758. [CrossRef]
7. Wu, F.; Duan, J.; Chen, S.; Ye, Y.; Ai, P.; Yang, Z. Multi-Target Recognition of Bananas and Automatic Positioning for the
Inflorescence Axis Cutting Point. Front. Plant Sci. 2021, 12, 705021. [CrossRef]
8. Wang, C.; Tang, Y.; Zou, X.; Luo, L.; Chen, X. Recognition and matching of clustered mature litchi fruits using binocular
charge-coupled device (CCD) color cameras. Sensors 2017, 17, 2564. [CrossRef]
Agronomy 2022, 12, 365 13 of 14

9. Luo, L.; Liu, W.; Lu, Q.; Wang, J.; Wen, W.; Yan, D.; Tang, Y. Grape Berry Detection and Size Measurement Based on Edge Image
Processing and Geometric morphology. Machines 2021, 9, 233. [CrossRef]
10. Gui, J.; Fei, J.; Wu, Z.; Fu, X.; Diakite, A. Grading method of soybean mosaic disease based on hyperspectral imaging technology.
Inf. Process. Agric. 2021, 8, 380–385. [CrossRef]
11. Luo, L.; Chang, Q.; Wang, Q.; Huang, Y. Identification and Severity Monitoring of Maize Dwarf Mosaic Virus Infection Based on
Hyperspectral Measurements. Remote Sens. 2021, 13, 4560. [CrossRef]
12. Appeltans, S.; Pieters, J.G.; Mouazen, A.M. Detection of leek white tip disease under field conditions using hyperspectral proximal
sensing and supervised machine learning. Comput. Electron. Agric. 2021, 190, 106453. [CrossRef]
13. Fazari, A.; Pellicer-Valero, O.J.; Gómez-Sanchıs, J.; Bernardi, B.; Cubero, S.; Benalia, S.; Zimbalatti, G.; Blasco, J. Application of
deep convolutional neural networks for the detection of anthracnose in olives using VIS/NIR hyperspectral images. Comput.
Electron. Agric. 2021, 187, 106252. [CrossRef]
14. Shi, Y.; Huang, W.; Luo, J.; Huang, L.; Zhou, X. Detection and discrimination of pests and diseases in winter wheat based on
spectral indices and kernel discriminant analysis. Comput. Electron. Agric. 2017, 141, 171–180. [CrossRef]
15. Phadikar, S.; Sil, J.; Das, A.K. Rice diseases classification using feature selection and rule generation techniques. Comput. Electron.
Agric. 2013, 90, 76–85. [CrossRef]
16. Ahmed, N.; Asif, H.M.S.; Saleem, G. Leaf Image-based Plant Disease Identification using Color and Texture Features. arXiv 2021,
arXiv:2102.04515.
17. Singh, S.; Gupta, S.; Tanta, A.; Gupta, R. Extraction of Multiple Diseases in Apple Leaf Using Machine Learning. Int. J. Image
Graph. 2021, 2140009. [CrossRef]
18. Gadade, H.D.; Kirange, D.K. Machine Learning Based Identification of Tomato Leaf Diseases at Various Stages of Development.
In Proceedings of the 2021 5th International Conference on Computing Methodologies and Communication (ICCMC), Erode,
India, 8–10 April 2021.
19. Almadhor, A.; Rauf, H.; Lali, M.; Damaševičius, R.; Alouffi, B.; Alharbi, A. AI-Driven Framework for Recognition of Guava
Plant Diseases through Machine Learning from DSLR Camera Sensor Based High Resolution Imagery. Sensors 2021, 21, 3830.
[CrossRef]
20. Kundu, N.; Rani, G.; Dhaka, V.S.; Gupta, K.; Nayak, S.C.; Verma, S.; Ijaz, M.F.; Woźniak, M. IoT and Interpretable Machine
Learning Based Framework for Disease Prediction in Pearl Millet. Sensors 2021, 21, 5386. [CrossRef]
21. Shrivastava, V.K.; Pradhan, M.K. Rice plant disease classification using color features: A machine learning paradigm. J. Plant
Pathol. 2020, 103, 17–26. [CrossRef]
22. Alajas, O.J.; Concepcion, R.; Dadios, E.; Sybingco, E.; Mendigoria, C.H.; Aquino, H. Prediction of Grape Leaf Black Rot Damaged
Surface Percentage Using Hybrid Linear Discriminant Analysis and Decision Tree. In Proceedings of the 2021 International
Conference on Intelligent Technologies (CONIT), Hubli, India, 25–27 June 2021.
23. Kianat, J.; Khan, M.A.; Sharif, M.; Akram, T.; Rehman, A.; Saba, T. A joint framework of feature reduction and robust feature
selection for cucumber leaf diseases recognition. Optik 2021, 240, 166566. [CrossRef]
24. Mary, N.A.B.; Singh, A.R.; Athisayamani, S. Classification of Banana Leaf Diseases Using Enhanced Gabor Feature Descriptor. In
Inventive Communication and Computational Technologies; Springer: Berlin/Heidelberg, Germany, 2020; pp. 229–242.
25. Sugiarti, Y.; Supriyatna, A.; Carolina, I.; Amin, R.; Yani, A. Model Naïve Bayes Classifiers For Detection Apple Diseases. In
Proceedings of the 2021 9th International Conference on Cyber and IT Service Management (CITSM), Bengkulu, Indonesia,
22–23 September 2021.
26. Mukhopadhyay, S.; Paul, M.; Pal, R.; De, D. Tea leaf disease detection using multi-objective image segmentation. Multimed. Tools
Appl. 2021, 80, 753–771. [CrossRef]
27. Chen, M.; Tang, Y.; Zou, X.; Huang, K.; Huang, Z.; Zhou, H.; Wang, C.; Lian, G. Three-dimensional perception of orchard banana
central stock enhanced by adaptive multi-vision technology. Comput. Electron. Agric. 2020, 174, 105508. [CrossRef]
28. Li, Q.; Jia, W.; Sun, M.; Hou, S.; Zheng, Y. A novel green apple segmentation algorithm based on ensemble U-Net under complex
orchard environment. Comput. Electron. Agric. 2021, 180, 105900. [CrossRef]
29. Cao, X.; Yan, H.; Huang, Z.; Ai, S.; Xu, Y.; Fu, R.; Zou, X. A Multi-Objective Particle Swarm Optimization for Trajectory Planning
of Fruit Picking Manipulator. Agronomy 2021, 11, 2286. [CrossRef]
30. Anagnostis, A.; Tagarakis, A.C.; Asiminari, G.; Papageorgiou, E.; Kateris, D.; Moshou, D.; Bochtis, D. A deep learning approach
for anthracnose infected trees classification in walnut orchards. Comput. Electron. Agric. 2021, 182, 105998. [CrossRef]
31. Abbas, A.; Jain, S.; Gour, M.; Vankudothu, S. Tomato plant disease detection using transfer learning with C-GAN synthetic
images. Comput. Electron. Agric. 2021, 187, 106279. [CrossRef]
32. Xiang, S.; Liang, Q.; Sun, W.; Zhang, D.; Wang, Y. L-CSMS: Novel lightweight network for plant disease severity recognition. J.
Plant Dis. Prot. 2021, 128, 557–569. [CrossRef]
33. Tan, L.; Lu, J.; Jiang, H. Tomato Leaf Diseases Classification Based on Leaf Images: A Comparison between Classical Machine
Learning and Deep Learning Methods. AgriEngineering 2021, 3, 542–558. [CrossRef]
34. Atila, Ü.; Uçar, M.; Akyol, K.; Uçar, E. Plant leaf disease classification using EfficientNet deep learning model. Ecol. Inform. 2021,
61, 101182. [CrossRef]
35. Mishra, M.; Choudhury, P.; Pati, B. Modified ride-NN optimizer for the IoT based plant disease detection. J. Ambient Intell.
Humaniz. Comput. 2021, 12, 691–703. [CrossRef]
Agronomy 2022, 12, 365 14 of 14

36. Liu, X.; Li, B.; Cai, J.; Zheng, X.; Feng, Y.; Huang, G. Colletotrichum species causing anthracnose of rubber trees in China. Sci. Rep.
2018, 8, 10435. [CrossRef] [PubMed]
37. Wu, H.; Pan, Y.; Di, R.; He, Q.; Rajaofera, M.J.N.; Liu, W.; Zheng, F.; Miao, W. Molecular identification of the powdery mildew
fungus infecting rubber trees in China. Forest Pathol. 2019, 49, e12519. [CrossRef]
38. Jocher, G.; Stoken, A.; Borovec, J.; Christopher, S.T.; Laughing, L.C. Ultralytics/yolov5: V4.0-nn.SILU() Activations, Weights &
Biases Logging, Pytorch Hub Integration. Zenodo 2021. [CrossRef]
39. Li, D.; Hu, J.; Wang, C.; Li, X.; She, Q.; Zhu, L.; Zhang, T.; Chen, Q. Involution: Inverting the inherence of convolution for visual
recognition. arXiv 2021, arXiv:2103.06255.
40. Hu, J.; Shen, L.; Sun, G. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, Salt Lake City, UT, USA, 18–22 June 2018; pp. 7132–7141.
41. Zhang, Y.-F.; Ren, W.; Zhang, Z.; Jia, Z.; Wang, L.; Tan, T. Focal and Efficient IOU Loss for Accurate Bounding Box Regression.
arXiv 2021, arXiv:2101.08158.