agriculture

Article
Tomato Leaf Disease Diagnosis Based on Improved
Convolution Neural Network by Attention Module
Shengyi Zhao, Yun Peng, Jizhan Liu * and Shuo Wu

Key Laboratory of Modern Agricultural Equipment and Technology, Ministry of Education, Jiangsu University,
Zhenjiang 212013, China; [email protected] (S.Z.); [email protected] (Y.P.);
[email protected] (S.W.)
* Correspondence: [email protected]; Tel.: +86-511-88797338

Abstract: Crop disease diagnosis is of great significance to crop yield and agricultural production.
Deep learning methods have become the main research direction to solve the diagnosis of crop
diseases. This paper proposed a deep convolutional neural network that integrates an attention
mechanism, which can better adapt to the diagnosis of a variety of tomato leaf diseases. The network
structure mainly includes residual blocks and attention extraction modules. The model can accurately
extract complex features of various diseases. Extensive comparative experiment results show that the
proposed model achieves the average identification accuracy of 96.81% on the tomato leaf diseases
dataset. It proves that the model has significant advantages in terms of network complexity and
real-time performance compared with other models. Moreover, through the model comparison
experiment on the grape leaf diseases public dataset, the proposed model also achieves better



results, and the average identification accuracy of 99.24%. It is certified that add the attention

 module can more accurately extract the complex features of a variety of diseases and has fewer
Citation: Zhao, S.; Peng, Y.; Liu, J.; parameters. The proposed model provides a high-performance solution for crop diagnosis under the
Wu, S. Tomato Leaf Disease Diagnosis real agricultural environment.
Based on Improved Convolution
Neural Network by Attention Keywords: tomato leaf disease; deep learning; convolutional neural network (cnn); attention mecha-
Module. Agriculture 2021, 11, 651. nism; classification
https://doi.org/10.3390/agriculture
11070651

Academic Editors: Sarah 1. Introduction

J. Pethybridge and Frank Hay
Tomato is an important vegetable crop in the world, with a per capita consumption
about 20 kg per year, accounting for around 15% of the total vegetable consumption [1].
Received: 22 June 2021
Accepted: 10 July 2021
The global annual output of fresh tomato exceeds 170 million tons, ranking first in vegetable
Published: 11 July 2021
crop production [2]. The United States, India, Turkey, Egypt and China are the main
producers of tomatoes [3]. According to the survey data of the Food and Agriculture
Publisher’s Note: MDPI stays neutral
Organization of the United Nations, tomato disease is the main reason for the decrease in
with regard to jurisdictional claims in
global tomato production, with an annual loss rate of as high as 8%–10% [4]. However,
published maps and institutional affil- most tomato diseases start from the leaves and then spread to the entire plant [5]. Automatic
iations. identification of tomato leaf diseases accurately can help to improve the management of
tomato production and provides a good growth environment.
Traditional expert diagnosis on tomato leaf disease has a high cost and has a subjective
misjudgment risk. With the rapid development of computer technology, computer vision,
Copyright: © 2021 by the authors.
machine learning, and deep learning are widely used in crop disease detection [6,7].
Licensee MDPI, Basel, Switzerland.
Traditional machine vision methods segment the RGB images of crop diseases by color,
This article is an open access article
texture, or shape features. However, the characteristics of different diseases are similar,
distributed under the terms and so it is difficult to judge the types of diseases, and the accuracy of disease recognition is
conditions of the Creative Commons poor in a complex natural environment. The Convolutional Neural Network (CNN) is a
Attribution (CC BY) license (https:// high-performance deep learning network; it abandons complex image preprocessing and
creativecommons.org/licenses/by/ feature extraction operations, and adopts an end-to-end structure, which greatly simplifies
4.0/). the recognition process compared to its learning [8–10]. Nowadays, CNN is widely used in

Agriculture 2021, 11, 651. https://doi.org/10.3390/agriculture11070651 https://www.mdpi.com/journal/agriculture

Agriculture 2021, 11, 651 2 of 15

crop disease recognition for real agriculture environments [11–14]; the automatic detection
of tomato leaf diseases combined with CNN is conducive to improving the accuracy of
diagnosis and reducing labor costs.
Several studies have been carried out to use deep learning technology to improve
the survival rate of vegetables, fruits, and field crops through early disease detection and
subsequent disease management. Wang et al. [15] apply transfer learning to the original
Alex Net network, and the average recognition rate of 10 categories of tomato leaves is
better. Rangarajan et al. [16] use the original AlexNet, VGG16 network structure, combined
with migration learning to obtain an accuracy of about 97% on the seven segmented tomato
diseased leaves. The effects of weight, deviation, and learning rate on the accuracy and
speed of disease detection are analyzed. Alcaro et al. [17] use cameras with different
resolutions to capture images of 9 tomato diseases and insect pests, and use Faster R-
CNN, R-FCN, and SSD for training. Long et al. [18] trained AlexNet and GoogleNet
networks by using transfer learning technology for Camellia oleifera diseases identification.
Kaur et al. [12] use a pre-trained ResNet network to classify 7 tomato diseases with an
accuracy rate of 98.8%. Karthik et al. [19] proposed a deep detection model structure for
tomato leaf diseases, optimized and improved the residual network, and used transfer
learning to obtain important disease classification features. Although transfer learning
can achieve better recognition results, the original AlexNet and VGG16 networks have
complex structures and numerous parameters, which cannot meet the actual application
and deployment of the model.
Based on the previous investigations, CNNs have more potential than the traditional
feature extraction methods. Razavi et al. [20] used the improved CNN network to train
the disease detection model for the open source disease dataset, and compared them
with traditional classifiers such as SVM, LBP, and GIST, which proved that the model
is higher than other classifications in terms of classification accuracy. Yang et al. [21]
used the saliency analysis of the image to locate the pests in tea gardens, reduced the
number of network layers and convolution kernels for AlexNet, and combined with the
Dropout model optimization algorithm to improve the accuracy. The optimized model is
effective against 23 pests in tea gardens. The average recognition accuracy rate reaches
88.1%. Sun et al. [22] improved the Alex Net network model by reducing the size of the
convolution kernel, which improved the disease accuracy and reduced the parameters
required for the model. Liu et al. [7] improved a CNN model based on AlexNet to identify
four apple leaf diseases, and the model could achieve an average recognition accuracy
of 97.62%. Grinblat et al. [23] developed a powerful neural network for the successful
identification of three different legume species based on the morphological patterns of the
leaves’ veins. In the real agriculture environment, the above mentioned study provides a
lot of reference for the diagnosis of tomato leaf diseases.
In recent years, due to the characteristic of extracting discriminative features of the
area of interest, the attention network began to be widely used in machine translation,
generative adversarial, and so on [24,25]. However, it is still in the exploratory stage in
the field of agriculture diseases detection. Tang et al. [26] added the attention module
into ShuffleNet, which improved the recognition rate of grape diseases in the PlantVillage
dataset to 99.14%. Zhong et al. [27] based on the ResNet18 added a group attention module,
and the pixel accuracy of semantic segmentation of cucumber diseased leaves in the natural
environment reached 93.9%.
For tomato leaf disease diagnosis, the diseased area only occupies a part of the leaf
image size. So, this study adds an attention module to the original CNN network model to
automatically extract important disease feature information from a complex environment.
The feature extraction is focused on the disease feature channel, and the invalid feature
channel information is eliminated. In this paper, an improved CNN network model is
proposed to diagnosis tomato multiple leaf disease accurately.
The main contributions of this paper are summarized as follows:
Agriculture 2021, 11, x FOR PEER REVIEW 3 of 17

Agriculture 2021, 11, 651 3 of 15

The main contributions of this paper are summarized as follows:

• In order to meet the diagnosis requirements of various tomato leaf diseases in the
• In orderenvironment,
natural to meet the diagnosis
this paperrequirements of various
constructs a dataset of 9tomato
tomatoleaf
leafdiseases
diseasesinand
the
natural environment, this paper constructs a dataset of 9 tomato leaf
healthy leaves. Furthermore, through data enhancement methods, the generalization diseases and
healthy
ability leaves.
and Furthermore,
adaptability of the through
model indata enhancement
practical methods,
applications the generalization
are improved.
• ability and adaptability of the model in practical applications are improved.
This paper proposes a multi-scale CNN network structure for the diagnosis of tomato
• leafThis paper proposes a multi-scale CNN network structure for the diagnosis of tomato
diseases. Based on the residual block, a multi-scale feature extraction module is
leaf diseases. Based on the residual block, a multi-scale feature extraction module is
added. The SE module is deeply integrated into the ResNet-50 network model.
added. The SE module is deeply integrated into the ResNet-50 network model.
• This paper established a multi-dimensional dependency relationship between the
• This paper established a multi-dimensional dependency relationship between the
three dimensions (C, H, W) of the extracted tomato leaf disease feature map and used
three dimensions (C, H, W) of the extracted tomato leaf disease feature map and
channel and spatial information with a small amount of calculation. In this way, ef-
used channel and spatial information with a small amount of calculation. In this way,
fective features of lesions can be obtained in a complex background, and contextual
effective features of lesions can be obtained in a complex background, and contextual
information can be discriminated.
information can be discriminated.
The rest of the paper is organized in the following manner. Section 2 introduces to-
The rest of the paper is organized in the following manner. Section 2 introduces tomato
mato leaf disease dataset augmentation and tomato leaf disease diagnosis model improve-
leaf disease dataset augmentation and tomato leaf disease diagnosis model improvement.
ment. Section 3 conducted comparative experiments on the proposed model performance
Section 3 conducted comparative experiments on the proposed model performance and
and verified
verified the applicability
the applicability of model
of the the model on other
on other cropcrop disease
disease datasets.
datasets. Section
Section 4 com-
4 compares
pares the model
the model results
results of thisofstudy
this study with other
with other researchers
researchers in detail.
in detail. Finally,
Finally, the conclu-
the conclusion is
sion is provided in
provided in Section 5.Section 5.

2.2.Materials
Materialsand
andMethods
Methods
2.1.Build
2.1. Buildthe
theDataset
Dataset
The image data
The dataofoftomato
tomato leafleaf
health andand
health disease in thisinpaper
disease comes from
this paper comes thefrom
PlantVil-
the
lage open source
PlantVillage open database [22]. The
source database database
[22]. containscontains
The database a large anumber of plantofdisease
large number plant
imagesimages
disease and is and
the world’s largest
is the world’s crop crop
largest database. After
database. initially
After acquiring
initially acquiringthetheimage
image of
tomato leaves, the image data needed for the research is manually screened
of tomato leaves, the image data needed for the research is manually screened to avoid to avoid prob-
lems suchsuch
problems as image duplication
as image andand
duplication classification errors
classification in the
errors dataset.
in the Finally,
dataset. a dataset
Finally, a da-
contained
taset 45854585
contained tomato leaf leaf
tomato images
imagesis obtained,
is obtained,andandthethesizesize
of each
of eachpicture
picture is is
fixed
fixedat
224
at × 224. The
224×224. The dataset
dataset contains
contains aa total
total of
of 10
10 tomato
tomatoleaf
leafcategories,
categories,suchsuchasasbacterial
bacterialspot,
spot,
early blight,
early blight, healthy,
healthy, late
lateblight,
blight,leaf
leafmold,
mold, mosaic
mosaicvirus, septoria
virus, leafleaf
septoria spot, target
spot, spot,spot,
target two-
spotted spider mite, and yellow leaf curl virus. The images of tomato
two-spotted spider mite, and yellow leaf curl virus. The images of tomato leaves in 10leaves in 10 categories
are shownare
categories in Figure
shown 1. in Figure 1.

(a) (b)

(c) (d)
Figure 1. Cont.
Agriculture 2021, 11, 651 4 of 15
Agriculture 2021, 11, x FOR PEER REVIEW 4 of 17

(e) (f)

(g) (h)

(i) (j)
Figure
Figure1.1.Tomato
Tomatoleaf
leafdataset:
dataset:(a)
(a)bacterial
bacterialspot;
spot;(b)
(b)early
earlyblight;
blight; (c)(c)
healthy; (d)(d)
healthy; late blight;
late blight;(e)(e)
leaf
leaf
mold;
mold;(f)
(f)mosaic
mosaicvirus;
virus;(g)
(g)septoria
septorialeaf
leafspot;
spot;(h)
(h)target
targetspot;
spot;(i)(i)two-spotted
two-spottedspider
spidermite; (j)(j)
mite; yellow
yellow
leaf
leafcurl
curlvirus.
virus.

2.2.
2.2.Data
DataAugmentation
Augmentation
In
Indeep
deep learning, diversityofofthe
learning, the diversity thedata
datasetset can
can enhance
enhance thethe generalization
generalization ability
ability and
robustness
and robustnessof the model
of the model[28].[28].
Therefore, thisthis
Therefore, paper usedused
paper a variety of image
a variety enhancement
of image enhance-
techniques,
ment and enhanced
techniques, and enhanced imageimage
data indata
combination with OpenCV
in combination under the
with OpenCV Pytorch
under the
framework.
Pytorch framework.
◦ , 90◦ , 180◦ , and 270◦ will not change the
1.1. Spin:
Spin:Rotated
Rotated the
the picture
picture randomly
randomly byby0°,090°, 180°, and 270° will not change the rel-
relative
ative position
position of the
of the diseased
diseased spot
spot and andthethe healthypart,
healthy part,simulated
simulatedthetherandomness
randomness
ofofthe
theshooting
shootingangleangleunder
undernatural
naturalconditions.
conditions.
2.2. Zoom:
Zoom:Reduced
Reduced anan image
image according
according to ato a certain
certain ratioratio
helpshelps to identify
to identify targets targets
at mul-at
multiple scales. For the zoomed image, the resolution of the image
tiple scales. For the zoomed image, the resolution of the image is expanded to 224×224 is expanded to
224 ×by
pixels 224 pixels
filling by filling 0 pixels.
0 pixels.
3.3. Add
Addnoise:
noise:Added
Addedsalt saltand
andpepper
peppernoise
noiseor orgaussian
gaussiannoise
noisetotothe
theimage
imageto tosimulate
simulate
imageswith
images withdifferent
differentdefinitions
definitionstaken
takenininthethenatural
naturalenvironment.
environment.
Colorjitter:
4.4. Color jitter:Changed
Changedthe thebrightness,
brightness,saturation,
saturation,and andcontrast
contrastof ofthe
theimage
imageto tosimulate
simulate
the image difference caused by the light intensity when shooting in the naturalnatural
the image difference caused by the light intensity when shooting in the envi-
environment.
ronment.
Usingthe
Using theabove
abovedata
data enhancement
enhancement method,
method, thethe number
number of samples
of samples in each
in each cate-
category
gory was expanded by 4 times, and the enhanced tomato leaf disease data set
was expanded by 4 times, and the enhanced tomato leaf disease data set contained 22,925 contained
22,925 images.
images. The trainset
The trainset and validationset
and validationset are randomly
are randomly divideddivided
into 8:2.into
The8:2. The detailed
detailed infor-
information of the dataset is shown in
mation of the dataset is shown in Table 1. Table 1.
Agriculture 2021, 11, x FOR PEER REVIEW 5 of 17

Agriculture 2021, 11, 651 5 of 15

Table 1. Tomato leaf diseases dataset.

Type Class Origin Images Augmentation Images

Table 1. Tomato leaf diseases dataset.
Trainset
bacterial spot 0 425 2125 1700
earlyType
blight Class
1 Origin 480
Images Augmentation
2400 Images Trainset
1920
healthy
bacterial spot 02 481
425 2405
2125 1924
1700
early blight
late blight 13 480
463 2400
2315 1920
1852
healthy 2 481 2405 1924
leaf mold 4 470 2350 1880
late blight 3 463 2315 1852
mosaic virus
leaf mold 45 448
470 2240
2350 1792
1880
septoria
mosaicleaf
virusspot 56 436
448 2180
2240 1744
1792
septoria leaf
target spot spot 6 7 436
457 2180
2285 1744
1828
target spot
two-spotted spider mite 7 8 457
435 2285
2175 1828
1740
two-spotted spider mite 8 435 2175 1740
yellow leaf curl virus
yellow leaf curl virus 9
9 490
490
2450
2450
1960
1960
Total 4585 22925 18340
Total 4585 22,925 18,340

2.3. Deep Learning Models

2.3. Deep Learning Models
2.3.1. The Feature Extraction Network
2.3.1. The Feature Extraction Network
Feature extraction is the key to deep learning, the different feature extraction net-
Feature extraction is the key to deep learning, the different feature extraction networks
works have different parameters, speeds, and performance. Nowadays, some wide range
have different parameters, speeds, and performance. Nowadays, some wide range of
of convolutional neural network models have been proposed, such as AlexNet [29], VGG-
convolutional neural network models have been proposed, such as AlexNet [29], VG-
Net [30], and GoogleNet [31]. However, these CNN models reduced the speed of training
GNet [30], and GoogleNet [31]. However, these CNN models reduced the speed of training
and detection due to the large numbers of parameters and computational operations [26].
and detection due to the large numbers of parameters and computational operations [26].
He et
He et al.
al. [32]
[32]proposed
proposedaaresidual
residualnetwork
networkwithwith101101layers to to
layers effectively
effectivelysolve thethe
solve problem
prob-
of gradient degradation, and won the 2015 ImageNet Large-scale
lem of gradient degradation, and won the 2015 ImageNet Large-scale Visual Recognition Visual Recognition
Challenge. Compared
Challenge. Compared with with AlexNet,
AlexNet, VGGNet,
VGGNet, and and GoogLeNet,
GoogLeNet, ResNet
ResNet has has less compu-
less compu-
tation and higher performance. In this paper, ResNet-50 with less
tation and higher performance. In this paper, ResNet-50 with less calculation and better calculation and better
performance is used as the feature extraction
performance is used as the feature extraction network. network.
In Figure
In Figure2,2,thethetomato
tomatoleaf leafdisease
disease image
image is input
is input into
into thethe ResNet-50
ResNet-50 network
network struc-
structure,
ture, first passes through convolutional layer, BN layer, and activation
first passes through convolutional layer, BN layer, and activation layer, and then the layer, and then the
obtained feature
obtained feature mapmap is is maximized
maximized pooling.
pooling. TheThe ResNet50
ResNet50 model
model mainly
mainly includes
includes Stage
Stage
1–4, and each stage consists of 1 sampling module and multiple identity
1–4, and each stage consists of 1 sampling module and multiple identity mapping modules. mapping mod-
ules. The output feature map undergoes an AVG pooling operation,
The output feature map undergoes an AVG pooling operation, and then passes through the and then passes
throughlayer
Flatten the Flatten
to make layer
the to makemulti-dimensional
output the output multi-dimensional features one-dimensional,
features one-dimensional, and finally
output through the fully connected layer. When deepening the numberthe
and finally output through the fully connected layer. When deepening number layers,
of network of net-
work
if layers, ifcharacteristics
the internal the internal characteristics
of the network of the
have network
reachedhave reached level
the optimal the optimal level
in a certain
in a certain
layer, layer, the subsequent
the subsequent superimposed superimposed
network layersnetwork willlayers will not
not change thechange the char-
characteristics.
acteristics.
The residual The residual
module module
(Figure 3) in(Figure 3) in can
ResNet-50 ResNet-50 can solve
effectively effectively solve the
the problem of problem
identity
of identity[32]
mapping mapping
and can[32] alsoand can also
reduce reduce
network network parameters
parameters and calculations.
and calculations.

Figure 2.
Figure 2. The structure
structure of
of the
the ResNet-50.
ResNet-50.
culture 2021, 11, x FOR PEER REVIEW 6 of 17
ure 2021, 11, x FOR PEER REVIEW 6 of 17
lture 2021, Agriculture
11, x FOR PEER
2021,REVIEW
11, 651 6 of 17 6 of 15

Figure 3. The residual block.

Figure 3. The residual block.
Figure 3. The residual block.
Figure 3. The residual block.
2.3.2. Attention Module
2.3.2. Attention Module
Multiple2.3.2.
2.3.2. Attention small Attention
Module Modulecan occur on the tomato leaf, which are usually of var-
disease lesions
Multiple small disease lesions can occur on the tomato leaf, which are usually of var-
iousMultiple
shapes. small
Using channellesions
Multiple
disease dependency
small disease
can occur islesions
an
onimportant
thecan occur
tomato wayon towhich
improve
the tomato CNN leaf,model
which per-
are usually of
ious shapes. Using channel dependency is an important way leaf,
to improve are
CNN usually
modelofper-var-
formance.
ious shapes. To boost
various the performance
shapes. Using of
channel existing state-of-the-art
dependency is an models
important with
way slight
to com-
improve
per- CNN model
formance. To Using
boost channel dependency
the performance is an important
of existing way to improve
state-of-the-art models with CNNslightmodel com-
putation
formance. cost.
To As As
boost shown
performance. in
ToFigure
boost
the performance 4,
theHu
of et et al.
performance
existing [33] mentioned
of existing the Squeeze-and-Excitation
state-of-the-art models with slight com-
putation cost. shown in Figure 4, Hu al. [33]state-of-the-art
mentioned the models with slight com-
Squeeze-and-Excitation
Networks
putation cost. in the CVPR
putation 2017
cost. As ImageNet
shown in Workshop
Figure 4, Huspeech.
et al. The
[33] weights
mentioned of different
the chan-
Squeeze-and-Excitation
Networks in theAs shown
CVPR in Figure
2017 ImageNet 4, Hu et al. [33]speech.
Workshop mentioned The the Squeeze-and-Excitation
weights of different chan-
nels are
Networks Networks
trained
in the through
CVPR the in the
the
2017 CVPR
cost 2017
function,
ImageNet ImageNet
and the Workshop
weight speech.
coefficients The
of weights
ofeach feature of different channels
chan-
nels are trained through cost function,Workshop
and the weightspeech. The weights
coefficients of each different
feature chan-chan-
nelare
aretrained are trained
automatically through
obtained. the cost function,
Then, according to and the
the size weight coefficients
of the weight of
coefficient each feature channel
of each
nels
nel are automaticallythrough the cost
obtained. function,
Then, and the
according weight
to the coefficients
size of the weight ofcoefficient
each featureof chan-
each
feature are automatically
channel, the obtained. obtained.
effective feature Then,
channel according and
istoenhanced, to the thesize of the
invalid weight
feature coefficient of each
channel
nel
featurearechannel,
automatically
the effective Then,channel
feature according the sizeand
is enhanced, of thetheweight
invalidcoefficient of
feature channel each
is suppressed. feature channel, the effective feature channel is enhanced, and the invalid feature channel
isfeature channel, the effective feature channel is enhanced, and the invalid feature channel
suppressed. is suppressed.
is suppressed.

Figure 4. The Squeeze-and-Excitation Networks.

Figure 4. The Squeeze-and-Excitation Networks.
Figure 4. The Squeeze-and-Excitation Networks.
Figure 4. The Squeeze-and-Excitation Networks.
2.4. Tomato Diagnosis
2.4. Tomato Model of ResNet Fused of the SE Module
2.4. Tomato Diagnosis ModelDiagnosis
of ResNetModelFusedofofResNet
the SE Fused
Module of the SE Module
Due to the
2.4. Tomato Diagnosis flexibility
Due Model of
toofthe the SE Module,
ResNet Fused it
of the can
SESE Moduleit canapplied
be directly to existing network
Due to the flexibility theflexibility
SE Module, of the
it can Module,
be directly be directly
applied applied
to existing to existing network
network
architectures.
Due to the In this
architectures.paper, SENet
In this is added
paper, SENetto the
is original
added tomodel
the structure
original of
model ResNet-50
network of ResNet-
structure
architectures. In flexibility
this paper,ofSENet the SEisModule,
added to it the
can original
be directly applied
model to existing
structure of ResNet-50
to obtain the50SE-ResNet50
to obtain mode.
the The
SE-ResNet50 networkmode. architectures
The network of improved ResNet-50
architectures are
of improved ResNet-
toarchitectures. In this paper,
obtain the SE-ResNet50 SENet
mode. is added
The network to architectures
the original model structure
of improved of ResNet-50
ResNet-50 are
depicted in Figure
50 are 5. In the
depicted SE-ResNet50
in Figure network
5. In the structure,
SE-ResNet50 SENet-block
networkuses uses global
structure, av-
to obtaininthe
depicted SE-ResNet50
Figure 5. In mode.
the SE-ResNet50 The network
network architectures
structure, of improved
SENet-block global SENet-block
ResNet-50 are
av- uses
erage pooling
depicted in to compress
global
Figure 5. average
In the feature
pooling
SE-ResNet50 maps.
to Connect
compress
network the two
feature
structure, fully
maps. connected
Connect
SENet-block the layers
usestwo together
fully
global connected
av- layers
erage pooling to compress feature maps. Connect the two fully connected layers together
to form a modular
together structure
to form toa express
modular the correlation
structure toand and
express dependence
the between
correlation and the char-
dependence between
toerage
formpooling
a modular to compress
structure feature
to expressmaps. theConnect
correlation the two fully connected
dependence layers
between thetogether
char-
acteristic channels
to form achannelsthe
modularand and keep
characteristic
structure the number
channels
to express of
and characteristic
keep the channels
number of unchanged
characteristic at the
channelsinputunchanged at
acteristic keep the number the correlation and
of characteristic dependence
channels between
unchanged theinput
at the char-
and output
acteristic of
thethe two
input fully
and connected
output of layer
the two modules.
fully connected layer modules.
and outputchannels
of the two andfully
keepconnected
the number of characteristic
layer modules. channels unchanged at the input
and output of the two fully connected layer modules.

Figure 5. The SE-ResNet50 module.

Figure 5. The SE-ResNet50 module.
Figure 5. The SE-ResNet50 module.
The upper
Figure 5. The SE-ResNet50 parts of the frameworks shown in Figure 5 are the SE module. When the
module.
The upper parts of the frameworks shown in Figure 5 are layer,
the SEthe
module. When the is reduced
The upperinput
partsfeature map reaches shown
of the frameworks the firstinfully connected
Figure 5 are the feature
SE module. Whendimension
the
input feature map reaches
to parts
1/R of
ofthe the
input.first fully connected
Then connected
it shown
is processed layer,
bythe the
the feature
ReLu dimension
theactivation is reduced
function, and
the then input to
input The upper
feature map reaches the
the frameworks
first fully inlayer,
Figure 5 feature
are SE module.
dimension isWhen
reduced
to 1/R of themap
input. Thenfully
it is first
processed by the ReLu activation function, andisthen input to
to 1/R of the input. Then it is processed by the ReLu activation function, and then input to the number of
input feature the second
reaches the connected
fully layer.
connected At this
layer,time,
the the zoom
feature ratio
dimension R,
is that is,
reduced
feature
to 1/R of the input. channels
Then is changed
it is processed by back to theactivation
the ReLu (1 × 1 ×and
input sizefunction, c). then input to
Agriculture 2021, 11, x FOR PEER REVIEW 7 of 17

Agriculture 2021, 11, 651 7 of 15

the second fully connected layer. At this time, the zoom ratio is R, that is, the number of
feature channels is changed back to the input size (1 × 1 × 𝑐).
Global average pooling is used in the compression stage, and the H × W spatial di-
Global average pooling is used in the compression stage, and the H × W spatial
mensions of the whole image are shrunk to F∈Rc. Theccompression process follows Equa-
dimensions of the whole image are shrunk to F ∈ R . The compression process follows
tion (1):
Equation (1):
HH WW
FsqF(uc()u=) = 1 ∑ ∑ uuc ((i,j)
1
i, j) (1)
(1)
sq c H × W
H×W i=1 j=1 c
i=1 j=1
After compression, a feature map of 1 × 1 × c2 is obtained, a parameter W is in-
After compression, a feature map of 1 × 1 × 𝑐2 is obtained, a parameter W is intro-
troduced, and a weight is generated for each feature channel through W. The different
duced, and a weight is generated for each feature channel through W. The different pa-
parameters represent the different importance of the characteristic channel, which is the
rameters represent the different importance of the characteristic channel, which is the core
core of the entire SENet module. These weights are allocated to the input feature maps.
of the entire SENet module. These weights are allocated to the input feature maps. This
This process is called feature recalibration, namely and gating mechanism. The excitation
process is called feature recalibration, namely and gating mechanism. The excitation pro-
process follows Equation (2):
cess follows Equation (2):
Fex (Fz, W ) = σ( g(z, W )) = σ(W δ(W z))
ex (z,W) = σ(g(z,W)) = σ(W22 δ(W11 z))
(2)
(2)
∗C *C C C
where z is the
where z isresult of theofcompression
the result process;
the compression σ is the
process; ReLU
σ is function,
the ReLU W1 ∈ R
function, Wr1∈Rand
r
C
W 2 ∈ RC∗ r :∗Cra dimensionality-reduction layer with parameters W 1 with reduction ratio r,
and W2∈R : a dimensionality-reduction layer with parameters W1 with reduction ratio
a ReLU and then a dimensionality-increasing layer with parameters W . Reweight is a
r, a ReLU and then a dimensionality-increasing layer with parameters W22. Reweight is a
re-calibration process, which uses the output weight of excitation as the importance of each
re-calibration process, which uses the output weight of excitation as the importance of
feature channel after feature selection. According to the degree of importance, the channel
each feature channel after feature selection. According to the degree of importance, the
is added to the original feature through the Equation (3) and keeping the number of feature
channel is added to the original feature through the Equation (3) and keeping the number
channels unchanged and not introducing new feature dimensions.
of feature channels unchanged and not introducing new feature dimensions.
x=
x =Fscale (u(u c ) = uc × sc
c , s,s (3)
Fscale c c ) = uc × sc (3)
where x and Fscale
where⎯x Fscale refers
referstotothe
thechannel
channelwise
wisemultiplication
multiplication between the scalar sc and the
scalar sc the
feature map u ∈ R H×W .
feature map ucc∈R . H×W

The above
The above isis the
thecomplete
completestructure
structureand
andoperation
operationprocess
processof SENet module.
of SENet ThisThis
module. sub-
network structure is embedded in ResNet50. The combination of characteristic
sub-network structure is embedded in ResNet50. The combination of characteristic chan- channel
recalibration
nel strategy
recalibration strategyandand residual network
residual networkcancan
effectively improve
effectively improve thethe
network
network perfor-
per-
mance, and thus does not need to increase the computational cost much.
formance, and thus does not need to increase the computational cost much. Through fea- Through feature
refinement,
ture the learning
refinement, abilityability
the learning of complex disease
of complex features
disease is enhanced.
features The entire
is enhanced. Thetomato
entire
leaf disease diagnosis network structure is shown in Figure
tomato leaf disease diagnosis network structure is shown in Figure 6. 6.

Figure 6. The structure of tomato leaf disease diagnosis.

Agriculture 2021, 11, 651 8 of 15

2.5. Experiment Setup

The operating platform for this experiment is a Dell T7920 graphics workstation,
the operating environment is Windows 10, the CPU is two Intel Xeon Gold 6248R, and the
GPU is two NVIDIA Quadro RTX 5000, 64GRAM, 1T solid state drive. The training
environment is created by Anaconda3, and the environment configuration is Python 3.6.13
and Pytorch 1.4.0, torchvision 0.5.0 artificial neural network library. At the same time
CUDA 10.1 deep neural network acceleration library is used.
The weight value of the feature extraction network uses the parameters of the pre-
trained ImageNet classification model. This method can greatly reduce the model calcula-
tion cost and calculation time. After each training, the validationset is tested and the model
is saved, and the model with the highest accuracy is selected.

2.6. The Evaluation Index

In order to evaluate the performance, the proposed network is compared with several
famous CNN networks: VGG-19, Xception, ResNet-101, and GoogleNet. The average
accuracy evaluation index recognized in the field of image classification is used to evaluate
the classification results, including Precision (PPV), Recall (TPR), F1 Score (F1), and Detection
speed (TA ).
TP
PPV = (4)
TP + FP
TP
TPR = (5)
TP + FN
PPV × TPR
F1 = 2 × (6)
PPV + TPR
where TP (true positive) is the number of positive samples predicted as positive samples,
FP (false positive) is the number of negative samples considered to be positive samples, and
FN (false negative) is the number of negative samples considered to be negative samples.

T
TA = (7)
N
where T is the total detection time for validationset and N is the total number for the
validationset.

3. Results
3.1. Comparison of Various Convolution Neural Networks
The comparison of various CNN model test accuracy curves of different networks
is REVIEW
Agriculture 2021, 11, x FOR PEER shown
in Figure 7. The training iteration epochs are plotted on the X9 axis
of 17 and the

corresponding training accuracy is plotted on the Y axis.

(a) (b)

Figure 7.Figure 7. The training

The training accuracy
accuracy curves:
curves: (a) (a) SE-ResNet50accuracy
SE-ResNet50 accuracy curves;
curves;(b)(b)
Comparison model
Comparison accuracy
model curves. curves.
accuracy
The evaluation results of different approaches on the tomato leaf disease are obtained
in Table 2. Under the same experimental conditions, the SE-ResNet50 model proposed in
this paper has the highest average accuracy, with an accuracy of 96.81%. Compared with
GoogleNet, ResNet-101, Xception, and VGG-19 models, the average accuracy is 9.54%,
6.68%, 8.65%, and 6.39% higher, respectively, significantly ahead of the 4 mainstream
Agriculture 2021, 11, 651 9 of 15

The evaluation results of different approaches on the tomato leaf disease are obtained
in Table 2. Under the same experimental conditions, the SE-ResNet50 model proposed
in this paper has the highest average accuracy, with an accuracy of 96.81%. Compared
with GoogleNet, ResNet-101, Xception, and VGG-19 models, the average accuracy is 9.54%,
6.68%, 8.65%, and 6.39% higher, respectively, significantly ahead of the 4 mainstream CNN
networks. At the same time, it can be seen from Figure 7 that the SE-ResNet50 model
starts to converge after 150 iteration epochs, the convergence rate is the fastest among all
models. Also, the model tends to be stable after convergence, and the fluctuation range is
smaller. Moreover, the SE-ResNet50 model proposed in this paper has the fastest average
diagnosis time for a single disease image, which is only 31.68 ms. Compared with the
second-ranked Xception model, the time is reduced by 1.23 ms, which meets the needs of
real-time diagnosis of tomato leaf diseases. Synthesizing the above analysis, the proposed
model achieves the best performance in terms of accuracy and convergence speed.

Table 2. The evaluation results.

Model Input PPV TPR F1 TA (ms) Accuracy (%)

GoogleNet 224 0.8716 0.8709 0.8712 33.56 87.27
ResNet-101 224 0.8995 0.9013 0.9004 39.53 90.13
Xception 224 0.8825 0.8814 0.8819 32.91 88.16
VGG-19 224 0.9039 0.9047 0.9243 40.26 90.42
SE-ResNet50 224 0.9677 0.9681 0.9679 31.68 96.81

Agriculture 2021, 11, x FOR PEER REVIEW 10 of 17

Figure 8 shows the confusion matrixes for 9 tomato leaf diseases and health leaf using
our SE-ResNet50 model.

Figure 8.
Figure Confusion matrixes
8. Confusion matrixes for
for tomato
tomato leaf
leaf diseases.
diseases.

The SE-ResNet50
The SE-ResNet50 model
model proposed
proposed in in this
this paper
paper has
has an
an accuracy
accuracy ofof over
over 97%
97% for
for the
the
diagnosis of healthy tomato leaves, and over 98% for the three diseases of bacterial
diagnosis of healthy tomato leaves, and over 98% for the three diseases of bacterial spot, spot,
mosaic virus,
mosaic virus, and
and yellow
yellow leaf
leaf curl
curl virus.
virus. The
The diagnosis
diagnosis accuracy
accuracy ofof early
early blight,
blight, target
target spot
spot
and two-spotted spider mite is low, but they have reached 93%, 94%, and 94%
and two-spotted spider mite is low, but they have reached 93%, 94%, and 94% respec- respectively,
which meets the accuracy requirements in actual diagnosis operations.
tively, which meets the accuracy requirements in actual diagnosis operations.
To better understand the learning capacity of the channel-wise mechanism, the vi-
sualizations with several tomato leaf disease feature maps of proposed SE-ResNet50 are
Agriculture 2021, 11, 651 10 of 15

Agriculture 2021, 11, x FOR PEER REVIEW 11 of 17

shown in Figure 9. The proposed model can retain more image details due to important
feature reuse.

(a) (b)

(c) (d)
Figure 9.
Figure 9. Visualization
Visualization of
of tomato
tomato leaf
leaf network
network feature
feature map:
map: (a) healthy; (b)
(a) healthy; early blight;
(b) early blight; (c)
(c) target
target spot;
spot; (d)
(d) two-spotted
two-spotted
spider mite.
spider mite.

3.2. Comparison of Diagnosisthe

To better understand Performance with Attention
learning capacity of the Module
channel-wise mechanism, the visu-
alizations with
In order several
to prove thetomato leaf
effect of disease
adding feature
attention maps of proposed
mechanism SE-ResNet50
on model accuracy, are
keeping
shown in Figure 9.conditions
the experimental The proposed model can retain
and parameters moreaimage
consistent, detailsexperiment
comparison due to important
of the
feature reuse.of the SE-ResNet50 and ResNet50 models was carried out. The results of the
performance
comparative experiment of the proposed model and the ResNet50 model without attention
3.2. Comparison
module of Diagnosis
on tomato Performance
leaf disease with Attention
are represented in TableModule
3.
In order to prove the effect of adding attention mechanism on model accuracy, keep-
Table
ing The results of SE-ResNet50
3. experimental
the conditions compared with ResNet50
and parameters without
consistent, attention module.
a comparison experiment of
the performance of the SE-ResNet50 and ResNet50 models was carried out. The results of
Model PPV TPR F1 TA (ms) Accuracy (%)
the comparative experiment of the proposed model and the ResNet50 model without at-
ResNet-50
tention 0.9251 leaf disease
module on tomato 0.9256 0.9253 in Table 3.33.85
are represented 92.56
SE-ResNet50 0.9677 0.9681 0.9679 31.68 96.81
Table 3. The results of SE-ResNet50 compared with ResNet50 without attention module.
It can be seen from Table 3 that the results of the model are improved after adding the
Model PPV TPR F1 TA (ms) Accuracy (%)
attention module. Adding the attention mechanism, the accuracy of the model is increased
by ResNet-50
4.25%, the average 0.9251 0.9256
detection time 0.9253
of a single 33.85
disease is shortened 92.56and the
by 2.17 ms,
SE-ResNet50 0.9677 0.9681 0.9679 31.68
model parameters are only a little bit. It can be concluded that the proposed network 96.81
is effective.
It can be
Based on seen from results,
the above Table 3itthat
can the results
be seen thatofthe
the model are improved
SE-ResNet50 after adding
model proposed in this
the
paperattention module.
can well Adding
complete theofattention
the task mechanism,
tomato leaf the accuracy
disease diagnosis of the
and has highmodel is in-
robustness
creased by 4.25%, the average detection time of a single disease is shortened by
and accuracy. This model can be a very useful detection tool in the field of crop diseases. 2.17 ms,
and the model parameters are only a little bit. It can be concluded that the proposed net-
work is effective.
 Agriculture 2021, 11, x FOR PEER REVIEW 12 of 17

Agriculture 2021, 11, 651 Based on the above results, it can be seen that the SE-ResNet50 model proposed in 11 of 15
this paper can well complete the task of tomato leaf disease diagnosis and has high ro-
bustness and accuracy. This model can be a very useful detection tool in the field of crop
diseases.
3.3. The SE-ResNet50 Effectiveness on Other Corp Disease Dataset
InSE-ResNet50
3.3. The order to verify the practical
Effectiveness application
on Other performance
Corp Disease Dataset of the SE-ResNet50 model
proposed in this paper, we conducted experiments on the public dataset ofmodel
In order to verify the practical application performance of the SE-ResNet50 grape leaf
proposed in this paper, we conducted experiments on the public dataset of grape leaf dis-black
diseases. The public dataset contains 2750 grape leaf disease images, including
measles,
eases. black rot,
The public brown
dataset spots,
contains healthy,
2750 anddisease
grape leaf leaf blight. Sample
images, images
including black are shown in
measles,
Figure 10.
black rot, brown spots, healthy, and leaf blight. Sample images are shown in Figure 10.

(a) (b)

(c) (d)
Figure
Figure10.
10.Grape
Grapeleaf dataset:
leaf (a)(a)
dataset: black measles;
black (b) brown
measles; spots;
(b) brown (c) healthy;
spots; (d) leaf
(c) healthy; (d)blight.
leaf blight.

Under
Underthe
thesame
sameexperimental
experimental conditions,this
conditions, this paper selected
paper Googlenet,
selected Resnet-50,
Googlenet, Resnet-50,
and
andXception
Xceptionfor
forcomparative
comparative experiments
experiments on on
grape leafleaf
grape diseases. As shown
diseases. in Figure
As shown 11, 11,
in Figure
the
theconvergence
convergencetimes
timesof of
thethefour models
four modelsareare
similar, but but
similar, the the
finalfinal
convergence accuracy
convergence accuracy
of
ofthe
theSE-ResNet50
SE-ResNet50 model
model is is
higher than
higher thatthat
than of the GoogleNet,
of the GoogleNet, ResNet-50, and Xception
ResNet-50, and Xception
models. At the same time, the SE-ResNet50 model proposed in this
models. At the same time, the SE-ResNet50 model proposed in this paper has paper has a small con-
a small
Agriculture 2021, 11, x FOR PEER REVIEW 13 of 17
vergence accuracy fluctuation range. The convergence accuracy of GoogleNet,
convergence accuracy fluctuation range. The convergence accuracy of GoogleNet, ResNet- ResNet-
50,
50,and
andXception
Xceptionmodels
models has a large
has fluctuation
a large fluctuationrange.
range.

Figure11.
Figure Thetraining
11.The trainingaccuracy
accuracy curves
curves forfor grape
grape leafleaf diseases.
diseases.

The evaluation results of approaches on the grape leaf disease are obtained in Table
4. The SE-ResNet50 model proposed in this paper has an average diagnostic accuracy of
99.24% for the four classifications of grape leaves. Compared with tomato leaf diseases,
the accuracy of diagnosis is increased by 2.43%, mainly due to the decrease of 6 kinds of
Agriculture 2021, 11, 651 12 of 15

The evaluation results of approaches on the grape leaf disease are obtained in Table 4.
The SE-ResNet50 model proposed in this paper has an average diagnostic accuracy of
99.24% for the four classifications of grape leaves. Compared with tomato leaf diseases,
the accuracy of diagnosis is increased by 2.43%, mainly due to the decrease of 6 kinds
of diseases. Compared with ResNet-50, GoogleNet and Xception models, the average
accuracy is 5.33%, 6.46% and 6.61% higher, respectively. Moreover, the SE-ResNet50 model
proposed in this paper has the fastest average diagnosis time for a grape leaf image, which is
only 31.42 ms. Compared with the second-ranked Xception model, the time is reduced
by 0.55 ms. Synthesizing the above analysis, the proposed model for grape leaf diseases
diagnosis also achieves the best performance in terms of accuracy and convergence speed.

Table 4. The results of SE-ResNet50 compared with ResNet50 without the attention module.

Grape Input PPV TPR F1 TA (ms) Accuracy (%)

SE-ResNet50 256 0.9918 0.9924 0.9921 31.42 99.24
ResNet-50 256 0.9388 0.9392 0.9390 33.27 93.91
GoogleNet 256 0.9274 0.9269 0.9271 32.89 92.78
Xception 256 0.9261 0.9263 0.9262 31.97 92.63

The identification result is represented by confusion matrix in Figure 12; the diagnostic
accuracy of black measles exceeds 98%, the diagnostic accuracy of brown spots and leaf
blight exceeds 99%, and the diagnostic accuracy of healthy leaves is 100%. It is proven that
Agriculture 2021, 11, x FOR PEER REVIEW 14 of 17
the proposed method has a wide range of applicability and has better performance relative
to deep based methods on other crop public datasets.

Figure12.
Figure Confusionmatrixes
12.Confusion matrixesfor
forgrape
grapeleaf
leafdiseases.
diseases.

4. Discussion
4. Discussion
Crop diseases are a major threat to global vegetable supply security, and the latest
Crop diseases
technologies need toare
beaapplied
major threat to global vegetable
to the agriculture supplydiseases.
field to control security,Due
andtothe
thelatest
long-
technologies need to be applied to the agriculture field to control diseases. Due
term continuous operations, ease of data acquiring, good robustness, and quick computing to the long-
term continuous operations,
of deep-learning-based ease
disease of data acquiring,
detection, it is widelygood robustness, and quick compu-
investigated.
ting of deep-learning-based disease detection, it is widely investigated.
According to the complex characteristics of tomato leaf diseases, this study designed a
According
diagnosis modeltofor
themulti-scale
complex characteristics of tomato
extraction of disease leaf diseases,
features. this study
In this study, designed
the dataset was
adivided
diagnosis
intomodel forclassification
10-class multi-scale extraction of disease
(bacterial spot, features.
early blight, In this
healthy, study,
late the
blight, dataset
leaf mold,
was divided
mosaic virus,into 10-class
septoria leaf classification (bacterial
spot, target spot, spot, early
two-spotted blight,
spider mite, healthy,
and yellowlateleaf
blight,
curl
leaf mold, mosaic virus, septoria leaf spot, target spot, two-spotted
virus). The SE-ResNet50 model proposed in this paper reaches an average detectionspider mite, and yel-
low leaf curl virus). The SE-ResNet50 model proposed in this paper reaches an average
detection accuracy of 96.81%, which is 4.25% higher than the original ResNet50 network
accuracy. The diagnostic accuracy of this model for the four diseases exceeds 97%, and the
detection accuracy for early blight is the worst, but it also exceeds 93%. The average diag-
nosis time of a single disease image is only 31.68 ms, and the diagnosis speed is faster to
Agriculture 2021, 11, 651 13 of 15

accuracy of 96.81%, which is 4.25% higher than the original ResNet50 network accuracy.
The diagnostic accuracy of this model for the four diseases exceeds 97%, and the detection
accuracy for early blight is the worst, but it also exceeds 93%. The average diagnosis time
of a single disease image is only 31.68 ms, and the diagnosis speed is faster to meet the
needs of real-time operations.
The results in this study are compared with study results as summarized in Table 5.
As shown in Table 5, Durmuş et al. [13], Wang et al. [15], Agarwal et al. [30], and Tm et al. [34]
used the same dataset as this study. The accuracy of all these studies is lower than the
model proposed in this paper. Even the accuracy of model designed by Guo et al. [29] for
eight-class tomato leaves is 4.11% lower than our study. It is found that the proposed model
accuracy of Kaur et al.[12], Rangarajan et al. [16], Karthik et al. [19], and Kaushik et al. [35]
are higher than our results, which are attributed to the less classification classes of diseases
(at most 7 classes). Overall, our model has good general performance and high diagnostic
accuracy for tomato leaf diseases.

Table 5. Results in the paper compared with other state-of-the-art results.

Paper Dataset Model Classification Precision Recall F1-Score Accuracy

Durmuş et al. [13] Plant Village AlexNet 10-class — — — 95.65%
Wang et al. [15] Plant Village AlexNet 10-class — — — 95.62%
Agarwal et al. [30] Plant Village Vgg-16 10-class 90% 92% 91% 91.2%
Tm et al. [34] Plant Village LeNet 10-class 94.81% 94.78% 94.8% 94%
ResNet-
Our model Plant Village 10-class 96.77% 96.81% 96.79% 96.81%
50+SeNet
Kaur et al.[12] Plant Village ResNet-101 7-class 98.8% 98.8% 98.8% 98.8%
Rangarajan et al. [16] Plant Village AlexNet 7-class — — — 97.49
ResNet +
Karthik et al. [19] Plant Village 4-class — — — 98%
Dense
Guo et al. [29] Plant Village AlexNet 8-class — — — 92.7%
Kaushik et al. [35] Plant Village ResNet-50 6-class — — — 97.01%

5. Conclusions
In this work, we successfully developed a multi-scale feature extraction model for
tomato leaf diseases diagnosis. The model deeply integrates the residual block and the
attention module and is trained for the detection of healthy and different unhealthy tomato
leaf images. The obtained results show that our model outperforms some recent deep learn-
ing studies by using the most popular publicly available PlantVillage dataset.It was also
found that the SE-ResNet50 was best suited for the diagnosis of tomato leaf diseases com-
pared to other model. Besides, the performance of the SE-ResNet50 model generally further
improved when trained with more different environment images. The trained models can
be used in the early automatic diagnosis of tomato and other crop diseases. Thus, this work
can be beneficial in early and automatic disease diagnosis of tomato crops enabled by the
latest technologies such as smartphones, drone cameras, and robotic platforms.
In the next step, we will deploy the proposes model to the greenhouse inspection robot
independently developed by the team to realize the automatic identification of tomato leaf
diseases in the real agricultural environment. At the same time, we will also establish a
dataset of tomato leaf diseases in a real agricultural environment to improve the diagnostic
performance of the inspection robot. It can help farmers accurately identify diseases, carry
out corresponding agricultural tasks such as applying pesticides and fertilizing according to
the types of diseases, and successfully realize agricultural modernization and intelligence.

Author Contributions: Conceptualization, S.Z. and J.L.; methodology, S.Z. and J.L. and Y.P.; software,
S.Z.; validation, S.Z. and J.L.; formal analysis, S.Z., J.L. and Y.P.; investigation, S.Z. and S.W.; resources,
J.L.; data curation, S.Z. and S.W.; writing—original draft preparation, S.Z.; writing—review and
editing, S.Z., J.L. and Y.P. All authors have read and agreed to the published version of the manuscript.
Agriculture 2021, 11, 651 14 of 15

Funding: The work was supported by Graduate Research and Innovation Projects of Jiangsu Province
(No. KYCX20_3034), Primary Research & Developement Plan of Changzhou City (Modern Agri-
culture) (No. CE20202021), Primary Research & Developement Plan of Jiangsu Province-Modern
Agriculture (No. BE2020383), Priority Academic Program Development of Jiangsu Higher Education
Institutions (No. PAPD-2018-87).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Chowdhury, M.E.; Rahman, T.; Khandakar, A.; Ayari, M.A.; Khan, A.U.; Khan, M.S.; Al-Emadi, N.; Reaz, M.B.I.; Islam, M.T.;
Ali, S.H.M. Automatic and Reliable Leaf Disease Detection Using Deep Learning Techniques. AgriEngineering 2021, 3, 294–312.
[CrossRef]
2. Valenzuela, M.E.M.; Restović, F. Valorization of Tomato Waste for Energy Production. In Tomato Chemistry, Industrial Processing
and Product Development; Royal Society of Chemistry: London, UK, 2019; pp. 245–258.
3. Elnaggar, S.; Mohamed, A.M.; Bakeer, A.; Osman, T.A. Current status of bacterial wilt (Ralstonia solanacearum) disease in major
tomato (Solanum lycopersicum L.) growing areas in Egypt. Arch. Agric. Environ. Sci. 2018, 3, 399–406. [CrossRef]
4. Zhu, X.K. Research on Tomato Disease Identification Based on Convolutional Neural Network; Beijing University of Technology: Beijing,
China, 2020.
5. Sardogan, M.; Tuncer, A.; Ozen, Y. Plant leaf disease detection and classification based on CNN with LVQ algorithm. In Pro-
ceedings of the 2018 3rd International Conference on Computer Science and Engineering (UBMK), Sarajevo, Herzegovina,
20–23 September 2018; pp. 382–385.
6. Pooja, V.; Das, R.; Kanchana, V. Identification of plant leaf diseases using image processing techniques. In Proceedings of the
2017 IEEE Technological Innovations in ICT for Agriculture and Rural Development (TIAR), Chennai, India, 7–8 April 2017;
pp. 130–133.
7. Liu, B.; Zhang, Y.; He, D.; Li, Y. Identification of apple leaf diseases based on deep convolutional neural networks. Symmetry 2018,
10, 11. [CrossRef]
8. Boulent, J.; Foucher, S.; Théau, J.; St-Charles, P.-L. Convolutional neural networks for the automatic identification of plant diseases.
Front. Plant Sci. 2019, 10, 941. [CrossRef] [PubMed]
9. Ma, J.; Zheng, F.; Zhang, L.; Sun, Z. Disease recognition system for greenhouse cucumbers based on deep convolutional neural
network. Trans. Chin. Soc. Agric. Eng. 2018, 34, 186–192.
10. Chao, X.; Sun, G.; Zhao, H.; Li, M.; He, D. Identification of apple tree leaf diseases based on deep learning models. Symmetry 2020,
12, 1065. [CrossRef]
11. Arya, S.; Singh, R. A Comparative Study of CNN and AlexNet for Detection of Disease in Potato and Mango leaf. In Proceedings
of the 2019 International Conference on Issues and Challenges in Intelligent Computing Techniques (ICICT), Ghaziabad, India,
27–28 September 2019; pp. 1–6.
12. Kaur, M.; Bhatia, R. Development of an improved tomato leaf disease detection and classification method. In Proceedings of the
2019 IEEE Conference on Information and Communication Technology, Baghdad, Iraq, 15–16 April 2019; pp. 1–5.
13. Durmuş, H.; Güneş, E.O.; Kırcı, M. Disease detection on the leaves of the tomato plants by using deep learning. In Proceedings of
the 2017 6th International Conference on Agro-Geoinformatics, Fairfax, VA, USA, 7–10 August 2017; pp. 1–5.
14. Liu, Y.; Feng, Q.; Wang, S. Plant disease identification method based on lightweight CNN and its mobile application. J. Agric. Eng.
2019, 35, 194–204.
15. Wang, Y.; Zhang, H.; Liu, Q.; Zhang, Y. Image classification of tomato leaf diseases based on transfer learning. J. China Agric. Univ.
2019, 24, 124–130.
16. Rangarajan, A.K.; Purushothaman, R.; Ramesh, A. Tomato crop disease classification using pre-trained deep learning algorithm.
Procedia Comput. Sci. 2018, 133, 1040–1047. [CrossRef]
17. Fuentes, A.; Yoon, S.; Kim, S.C.; Park, D.S. A robust deep-learning-based detector for real-time tomato plant diseases and pests
recognition. Sensors 2017, 17, 2022. [CrossRef] [PubMed]
18. Long, M.; Ouyang, C.; Liu, H.; Fu, Q. Image recognition of Camellia oleifera diseases based on convolutional neural network &
transfer learning. Trans. Chin. Soc. Agricult. Eng. 2018, 34, 194–201.
19. Karthik, R.; Hariharan, M.; Anand, S.; Mathikshara, P.; Johnson, A.; Menaka, R. Attention embedded residual CNN for disease
detection in tomato leaves. Appl. Soft Comput. 2020, 86, 105933.
20. Razavi, S.; Yalcin, H. Using convolutional neural networks for plant classification. In Proceedings of the 2017 25th Signal
Processing and Communications Applications Conference (SIU), Antalya, Turkey, 15–18 May 2017; pp. 1–4.
21. Yang, G.; Bao, Y.; Liu, Z. Localization and recognition of pests in tea plantation based on image saliency analysis and convolutional
neural network. Trans. Chin. Soc. Agric. Eng. 2017, 33, 156–162.
Agriculture 2021, 11, 651 15 of 15

22. Sun, J.; Tan, W.; Mao, H.; Wu, X.; Chen, Y.; Wang, L. Recognition of multiple plant leaf diseases based on improved convolutional
neural network. Trans. Chin. Soc. Agric. Eng. 2017, 33, 209–215.
23. Grinblat, G.L.; Uzal, L.C.; Larese, M.G.; Granitto, P.M. Deep learning for plant identification using vein morphological patterns.
Comput. Electron. Agric. 2016, 127, 418–424. [CrossRef]
24. Luong, M.-T.; Pham, H.; Manning, C.D. Effective approaches to attention-based neural machine translation. arXiv 2015,
arXiv:1508.04025.
25. Zhang, H.; Goodfellow, I.; Metaxas, D.; Odena, A. Self-attention generative adversarial networks. In Proceedings of the
International Conference on Machine Learning, Long Beach, CA, USA, 9–15 June 2019; pp. 7354–7363.
26. Tang, Z.; Yang, J.; Li, Z.; Qi, F. Grape disease image classification based on lightweight convolution neural networks and
channelwise attention. Comput. Electron. Agric. 2020, 178, 105735. [CrossRef]
27. Zhong, C.; Hu, Z.; Yang, X.; Li, H.; Liu, F.; Li, M. Triple Stream Segmentation Network for Plant Disease Segmentation.
In Proceedings of the 2021 IEEE 5th Advanced Information Technology, Electronic and Automation Control Conference (IAEAC),
Chongqing, China, 13–14 March 2021; pp. 496–501.
28. Pulgar, F.J.; Rivera, A.J.; Charte, F.; del Jesus, M.J. On the impact of imbalanced data in convolutional neural networks performance.
In Proceedings of the International Conference on Hybrid Artificial Intelligence Systems, La Rioja, Spain, 21–23 June 2017;
pp. 220–232.
29. Guo, X.; Fan, T.; Shu, X. Tomato leaf diseases recognition based on improved multi-scale AlexNet. Trans. Chin. Soc. Agricult. Eng.
2019, 35, 162–169.
30. Agarwal, M.; Gupta, S.K.; Biswas, K. Development of Efficient CNN model for Tomato crop disease identification. Sustain.
Comput. Inform. Syst. 2020, 28, 100407. [CrossRef]
31. Shin, J.; Chang, Y.K.; Heung, B.; Nguyen-Quang, T.; Price, G.W.; Al-Mallahi, A. A deep learning approach for RGB image-based
powdery mildew disease detection on strawberry leaves. Comput. Electron. Agric. 2021, 183, 106042. [CrossRef]
32. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 770–778.
33. 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, 18–23 June 2018; pp. 7132–7141.
34. Tm, P.; Pranathi, A.; SaiAshritha, K.; Chittaragi, N.B.; Koolagudi, S.G. Tomato leaf disease detection using convolutional neural
networks. In Proceedings of the 2018 Eleventh International Conference on Contemporary Computing (IC3), Noida, India,
2–4 August 2018; pp. 1–5.
35. Kaushik, M.; Prakash, P.; Ajay, R.; Veni, S. Tomato Leaf Disease Detection using Convolutional Neural Network with Data
Augmentation. In Proceedings of the 2020 5th International Conference on Communication and Electronics Systems (ICCES),
Coimbatore, India, 10–12 June 2020; pp. 1125–1132.