(IJACSA) International Journal of Advanced Computer Science and Applications,

Vol. 14, No. 6, 2023

A Novel Classification Approach for Grape Leaf

Disease Detection Based on Different Attention
Deep Learning Techniques

S Phani Praveen1 , Rajeswari Nakka2 , Anuradha Chokka3 ,

Venkata Nagaraju Thatha4 , Sai Srinivas Vellela5 , Uddagiri Sirisha6
Department of Computer Science & Engineering, Prasad V Potluri Siddhartha Institute of Technology, Andhra Pradesh, India1
Department of Computer Science & Engineering, Seshadri Rao Gudlavalleru Engineering College,
Gudlavalleru, Andhra Pradesh, India2
Department of Computer Science & Engineering, Koneru Lakshmaiah Education Foundation,
Vaddeswaram, Andhra Pradesh, India3
Department of Information Technology, MLR Institute of Technology, Hyderabad, Telangana, India4
Department of Computer Science & Engineering, Chalapathi Institute of Technology, Guntur, Andhra Pradesh, India5
School of Computer Science & Engineering, VIT-AP, Andhra Pradesh, India6

Abstract—Preventing and controlling grape diseases is essen- manage diseases effectively. Early detection enables timely
tial for a good grape harvest. With the help of “single shot multi- interventions, minimizing potential damage and crop losses.
box detectors”, “faster region based convolutional neural net- Different grape leaf diseases require specific treatments, and
works”, & “You only look once-X,” the study improved grape leaf accurate identification helps growers implement targeted con-
disease detection accuracy with effective attention mechanisms, trol measures. This optimizes the use of pesticides, reduces
which includes convolutional block attention module, squeeze &
excitation networks, & efficient channel attention. The various
environmental impact, and ensures effective disease manage-
attention techniques helped to emphasize important features ment.
while reducing the impact of irrelevant ones, which ultimately
Grape leaf diseases can significantly impact the yield and
improved the precision of the models and allowed for real-time
performance. As a result of examining the optimal models from quality of grapevine production. Some diseases cause defolia-
the three types, it was found that the Faster (R-CNN) model had tion, reducing the vine’s ability to photosynthesize and produce
a lower precision value, while You only look once-X and SSD energy, leading to decreased fruit quality, delayed ripening,
with various attention techniques required the fewest parameters and reduced yield. Early disease detection enables growers to
with the highest precision, with the best real-time performance. In protect the crop and implement measures to minimize yield
addition to providing insights into grape diseases & symptoms in losses. Early identification of grape leaf diseases is essential
automated agricultural production, this study provided valuable for preventing their spread within vineyards. Prompt isolation
insights into grape leaf disease detection. and treatment of infected vines help prevent diseases from af-
Keywords—Grape leaves; faster region-based convolutional fecting healthy plants. Additionally, preventive measures such
neural networks; you only look once (x); single shot detection as pruning, canopy management, and cultural practices can be
attention techniques implemented to reduce the likelihood of disease occurrence
and spread. Economically, grapevines are valuable crops, and
detecting diseases in grape leaves allows growers to make in-
I. I NTRODUCTION
formed decisions on disease management, optimizing resource
Preventing and controlling crop diseases is crucial for utilization, and reducing unnecessary costs. This helps preserve
producing safe and healthy vegetables, minimizing losses, and the economic viability of vineyards and sustain profitability in
reducing the use of pesticides in the production of crops [1]. grape production.
Thus, early detection & prevention of diseases are crucial.
Grape plants can be affected by various diseases, such as Efficient disease detection and management practices also
powdery mildew, brown blotch, and anthracnose, which can contribute to sustainable agriculture. Early identification min-
significantly impact the yield and quality of the fruit. Tradi- imizes the use of broad-spectrum pesticides, reducing their
tional methods of detecting grape diseases rely on the expe- negative impact on the environment and non-target organ-
rience of the growers or the guidance of experts, which can isms. Targeted treatments based on accurate disease detection
be slow, inefficient, and lack real-time performance. Images of help reduce chemical inputs, promote ecological balance, and
grape leaves are used to detect, identify, and provide guidance support sustainable cultivation practices for grapevines. In
about diseases infected with grape leaves [2] because disease- summary, grape leaf disease detection is vital for crop health
infected grape leaves often have visible spots. monitoring, disease management, yield protection, disease pre-
vention, economic considerations, and sustainable agriculture.
Grape leaf disease detection is crucial for several rea- Early detection allows for timely interventions, optimization
sons. Firstly, it allows growers to monitor the health of of disease control measures, minimization of crop losses, and
their grapevines and take appropriate actions to prevent or the long-term sustainability of grapevine production.
www.ijacsa.thesai.org 1199 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 6, 2023

Due to the rapid development of artificial intelligence paper are to enhance the accuracy and speed of grape leaf
technologies, a wide variety of vision approaches are utilised disease detection, improve the efficiency of feature extraction
in the processing of photos for various crop diseases [3][4][5]. networks, validate the performance improvements on a grape
Research into classifying agricultural diseases uses a wide disease dataset, and provide a foundation for future grape
range of approaches, including “genetic algorithms” [6], “sup- disease control measures.
port vector machines” [7],“K-means clustering” [8], “ensemble
learning” [9], “Bayesian classification” [10], “radial basis However, the existing literature lacks research on incor-
functions” [11], & “filter segmentation” techniques [12]. Un- porating attention mechanisms, such as the ”convolutional
fortunately, conventional approaches to crop disease classifica- block attention module,” ”efficient channel attention,” and
tion and identification rely on labour-intensive, environment- ”squeeze & excitation attention,” into grape detector models
dependent manual feature selection. In particular, the develop- like ”Faster(R-CNN),” ”SSD,” and ”YOLO-X.” There is a
ment of deep learning’s Convolutional Neural Network (CNN) gap in knowledge regarding the potential impact of attention
has led to vast improvements in the field of autonomous mechanisms on improving detection accuracy and processing
detection and identification of agricultural diseases. speed for grape leaf diseases. The main objectives of the paper
are:
An object detection system that uses a convolutional neural
network (CNN) has made great strides recently. Several appli- • Enhance the accuracy and speed of grape leaf disease
cations make use of this technique, including recognition of detection: The purpose of this work is to enhance
faces [13], navigation [14], detection of road obstacles [15], the efficiency of previously developed grape detection
detection of pedestrians, abnormal activity recognition[16], models by incorporating attention mechanisms such
monitoring of physical activity[17][?],[18] detection of fruits, as ”convolutional block attention module,” ”efficient
and detection of weeds [19]. Despite complex backdrops, crop channel attention,” and ”squeeze & excitation atten-
leaf diseases can be detected using object detection algorithms tion” into the models of ”Faster(R-CNN),” ”SSD,” and
due to CNN’s ability to extract high-dimensional properties ”YOLO-X.” The objective is to achieve higher detec-
from object images. tion accuracy and faster processing speeds, addressing
the limitations of slow operation and low detection
As a result, scientists in China and others have studied precision in existing models.
object detection algorithms to develop models for detecting
crop diseases. For instance, Some authors have applied various • Improve the efficiency of feature extraction networks:
models for object detection to the tomato disease dataset, By integrating attention methods into the models, the
including the Faster(R-CNN), and the Single Shot Multibox paper aims to enhance the efficiency of the feature
Detector. Faster (R-CNN) as well as VGG16, produced the extraction networks. The attention mechanisms help to
best disease detection results. Dynamic identification of grape prioritize relevant features and emphasize health issues
leaf illnesses was accomplished by using Faster (R-CNN) related to grape leaf diseases, leading to more effective
on time-series images of grape leaves. Using an enhanced and accurate detection.
Faster (R-CNN) model, the authors of [20] detected diseases • Validate the performance improvements on a grape
in bitter gourd leaves with excellent results. Using an in-house disease dataset: The research conducts experiments
dataset, The authors of [21] trained the SSD model to identify using a dataset specifically focused on grape diseases.
agricultural diseases with an overall accuracy of 83.90%. By evaluating the models based on diverse attention
An enhanced model based on MobileNetv2 & YOLOv3 was mechanisms, such as ”Faster(R-CNN),” ”SSD,” and
proposed by the authors [22], which allowed for the early ”YOLO-X,” the paper aims to demonstrate significant
detection of grey speck disease in tomatoes. This refined improvements in detection accuracy and operation per-
model benefits from a number of desirable characteristics, formance. The experiments involve minimal parameter
including a low memory size, outstanding detection accuracy, tweaks, ensuring that the observed enhancements are
and lightning-fast identification. primarily attributed to integrating attention mecha-
nisms.
Previous studies have shown that using object detection
technology to detect grape leaf diseases is feasible. Existing • Provide a foundation for future grape disease control
grape detector models, however, operate slowly and have low measures: The findings of this study serve as a basis
detection precision, which severely limits their application. for future work on grape disease control measures.
This research included the attention methods of “convolu- By demonstrating the effectiveness of attention mech-
tional block attention module,” “efficient channel attention,” & anisms in improving detection accuracy and speed, the
“squeeze &excitation attention” into the models of “Faster(R- paper offers valuable insights and guidance for the
CNN),” “SSD,” & “YOLO-X” to boost their accuracy and development of advanced and efficient techniques for
speed. The goal was to boost the feature extraction network’s managing and controlling grape leaf diseases.
efficiency and put more emphasis on health issues. Experi-
ments were run on a plant village dataset of grape diseases, and II. R ELATED W ORK
the findings revealed that models based on diverse attention
mechanisms, such as “Faster(R-CNN),” “SSD,” & “YOLO- Detecting plant diseases in a timely manner is crucial
X,” significantly improved detection accuracy and operation for effectively managing plant losses. However, relying on
performance with only little parameter tweaks. The findings manual diagnosis by humans is a time-consuming process
of this study can be used as a foundation for future work on that is prone to errors and can be costly. To address these
grape disease control measures. The main objectives of the challenges, researchers have been actively exploring automated
www.ijacsa.thesai.org 1200 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 6, 2023

techniques for disease detection and classification in plants.

The utilization of automated equipment and methods has
emerged as a promising approach for monitoring crop fields.
In this section, we will delve into the specifics of computer
vision methods employed to identify and diagnose diseases in
plant leaves.
The authors of [23] conducted a study focusing on detect-
ing black rot in grape leaves using a YOLOv3-SPP-based deep
learning method. The researchers employed a combination of
super-resolution image enhancement and convolutional neural
network techniques to identify the disease in grape leaves.
The initial step involved upsampling the input image through
bilinear interpolation. After enhancement, the processed inputs
were fed into the YOLOv3-spatial pyramid pooling model,
resulting in a remarkable detection accuracy of 95.79%. How-
ever, when tested in real field conditions, the precision of this
method dropped to 86.69%. In a separate study, The authors
of [24] proposed a deep learning approach specifically for
accurately detecting black rot spots on grape leaves. They Fig. 1. Sample images from dataset.
employed the DeepLab V3+ model, which incorporates feature
maps from different levels and utilizes ResNet 101 as the
backbone network. The test results demonstrated that the
improved DeepLab V3+ model outperformed conventional TABLE I. L IST OF PARAMETERS USED FOR I MAGE E NHANCEMENT
methods. Enhancement method Parameters
gaussian filter sigma-range(0.4,1.3)
The authors of [25], [26]developed a novel support vector mean filter kernel-size-range(3,5)
machine & image processing-enabled technique for identi- median filter kernel-size-range(2,5)
image acutance alpha-range(0,0,2)
fying and categorizing grape leaf disease. The authors of brightness gamma(2.0)
[27] employed a CNN-SVM-based approach to classifying contrast alpha(1.0)
five different species of grapevine leaves. They utilized the
MobileNetv2 CNN model for leaf-type classification. Initially,
features were extracted from the pre-trained MobileNet2 logits common classification, accounting for roughly 34% of the
layer, and classification was performed using SVM with vari- images, while Black Rot and Leaf Blight make up 29% and
ous kernels. The Chi-squares method was applied for feature 26% respectively. There are also 1076 Healthy images and
selection, resulting in an impressive classification accuracy of 10.4% Normal images in the collection.
97.60%. The use of feature selection techniques significantly
contributed to the improved accuracy of classification. B. Image Pre-processing and Augmentation
The authors of [28] focused on detecting grape The size of the dataset must be increased by using data
black measles disease. They utilized the ResNet-50-based augmentation techniques in order to prepare grape leaves
DeepLabV3 segmentation model in combination with fuzzy disease images for disease identification. Training the recog-
logic to determine the severity of the disease. The input image nition model in this way ensures that it will be more re-
provided region of interest features and the percentage of silient and can generalize more effectively. Using standard
infections. A fuzzy rule-based reference system was developed data augmentation methods, the experiment compared the
based on each feature, which was then used to grade the grape effectiveness of the data augmentation method proposed in
disease. The grading system allowed for the classification this study. The traditional methods included flipping the im-
of healthy, mild, medium, and severe cases, specifically for age horizontally and vertically, rotating the image, applying
measles disease. different types of filtering (Gaussian, mean, and median)
with a probability of 0.2, enhancing image contrast, sharp-
III. M ATERIALS AND M ETHODS ening images with a probability of 0.3, and adjusting image
brightness. According to Table I, the parameters used for
A. Image Acquisition each image enhancement method are listed. Table II pro-
The plant-Village dataset provides 4,062 images of grape vides more details about the dataset and can be accessed
leaves displaying common symptoms. In this dataset, 1,180 at https://www.kaggle.com/datasets/rm1000/augmented-grape-
images were found to be affected by Black Rot, 1,383 by Esca disease-detection-dataset.
measles, 1,076 by Leaf spot, and 423 by healthy leaves, all
with a resolution of 256 × 256 pixels. A leaf with black rot, a C. Proposed Methodology for Grape Leaf Disease Detection
leaf with black measles, a leaf with blight, and a healthy leaf
is displayed in Fig. 1. This study focuses on the detection of grape leaf diseases
using three specific models: faster-rcnn, YOLOx, and SSD.The
The data set contains varying quantities of images for each training process for these models to detect diseases in grape
category, indicating significant imbalances. Esca is the most leaves is depicted in Fig. 2. The process begins with inputting
www.ijacsa.thesai.org 1201 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 6, 2023

TABLE II. I NFORMATION ABOUT THE DATASET Attention Module” attention mechanism is introduced to take
into account the importance of pixels in different places. All
Class No of images No of images with three attention methods contribute significantly to improving
without augmenta- augmentation
tion the model’s efficiency and precision.
Healthy 423 3000
Esca 1383 3000 2) ECA Attention Module: It uses local cross channel
measles interaction methods without reducing the magnitude of the
Leaf spot 1076 3000
Black rot 1180 3000
dimensionality can be accomplished without using reduced-
Total 4062 12000 dimension SE. The functionality of the attention module is
enhanced while its complexity is decreased thanks to this
mechanism. In Fig. 4 we can observe the construction of the
the selected grape leaf disease images. Next, classification efficient channel attention mechanism.
features are extracted from the input images. Output is then
3) CBAM Attention Module: The “CBAM Spatial Atten-
derived from the findings of disease identification using the
tion Module” is made up of 2 modules, first one is the
faster-rcnn, YOLOx, and SSD models.
“spatial attention module”, second one is the “channel attention
A loss function is used throughout to quantify the degree module” and is designed to optimize input feature maps by in-
to which the projected disease species deviates from the true ferring attention maps on both channel and spatial dimensions.
disease species. This enables the models to learn and improve These attention maps are then multiplied with the input fea-
their detection accuracy over time. The optimization of the ture map, resulting in self-adaptive feature optimization. The
final output result is achieved through the utilization of the CBAM mechanism is effective in enhancing useful features
Adam optimizer, a widely used optimization algorithm in deep while suppressing those that are not useful, making it a popular
learning. tool in practical applications. Fig. 5 illustrates the network
structure of CBAM.
By following this approach, the study aims to leverage the
capabilities of faster-rcnn, YOLO:x & SSD models to detect
grape leaf diseases effectively. The training flow chart provides E. Dection Models for Disease Detection in Grape Leaves
a systematic framework for the feature extraction and disease with Attention Mechanism
detection process, facilitating the accurate identification of CNN-based object detection can be categorized into two
different disease species in grape leaves. main types. The first type uses a regional proposal to detect
objects. This involves identifying candidate regions in the
D. Attention Mechanism Models image, which are then divided to detect objects. This two-
stage approach is exemplified by methods such as “R—CNN”,
The study utilizes three attention mechanisms: “Squeeze & “Fast(R-CNN)”, & “Faster(R-CNN)”. The second type of
Excitation”, “efficient channel attention”, and “Convolutional object detection does not use a regional proposal and is referred
Block Attention” spatial attention mechanism. We chose the to as one-stage object detection. An image is analyzed based on
SE attention mechanism because it is simple and adds only a CNN prediction of an object’s position & properties. There
a few new parameters. With ECA attention, models become are a variety of algorithms available for this type of detection,
more accurate without significantly increasing model complex- such as SSDs and YOLOs.
ity. It is an enhanced version of the SE attention mechanism.
Finally, the CBAM attention mechanism is useful because it The study used three models, namely the “Faster R–CNN
connects the spatial domain and the channel domain, leading model”, “YOLO-X model”, & “SSD model” for detecting
to more effective improvement in network performance. grape leaves disease. The input of the selected grape leaf
disease images, extraction of classification features, and use
1) Squeeze & Excitation Attention: In order to extract of the three disease detection models were involved in the
features, the SE channel attention mechanism employs the process. The output was an analysis of the disease detection
CNN channel. It requires re-calibrating features so that the results. For optimizing the final output, an Adam optimizer
model can pick up and remember relevant details from all of was used to predict the difference between reality and the
the available feature channels. Fig. 3 depicts the two steps prediction of disease species.
involved in this mechanism: squeezing and excitement. After
the feature image has been spatially compressed using the Researchers found that the ” Faster(R-CNN)” model boosts
squeeze technique, the feature channel’s relative relevance high detection accuracy and can detect targets end-to-end.
can be determined using the excitation technique; a model However, its running speed is relatively slow. On the other
is created based on the correlation between the channels. In hand, the “YOLO-X” model runs quickly, but it doesn’t detect
doing so, the original feature images are excited into matching small objects. The “SSD” technique has faster running speed
channels. The SE mechanism has few additional parameters and higher detection accuracy than the “YOLO-X” model, but
and is computationally simple. its training process heavily relies on prior experience, and its
performance in detecting small targets is not as good as the
The “efficient channel attention” attention mechanism is “Faster(R-CNN)” model. The characteristics of these models
utilized to enhance cross-channel interaction and reduce model are elaborated as follows:
complexity, while the Squeeze & Excitation attention mech-
anism is used to prioritize the most informative channel fea- 1) Grape Leaves Disease Detection using Faster (R-CNN)
tures for disease identification. For end-to-end training of the Model: This model is comprised of three main components:
grape leaf disease detection model, the “Convolutional Block the “Extraction of features”, the “Region Proposal Network”
www.ijacsa.thesai.org 1202 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 6, 2023

Fig. 2. Proposed attention model for grape disease detection.

Fig. 3. SE attention mechanism.

Fig. 4. ECA attention mechanism.

www.ijacsa.thesai.org 1203 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 6, 2023

Fig. 5. CBAM attention mechanism.

, and the “Region with Convolutional Neural Network Fea- attention module”, “efficient channel attention” and “squeeze
tures”. Fig. 6 depicts the Faster (R-CNN) model with attention & excitation attention mechanisms.
techniques. A Faster (R-CNN) method is used to detect grape
3) Grape Leaves Disease Detection using SSD Model::
leaf diseases in four primary steps: generating candidate dis-
Using a tiny convolution kernel and multi-dimensional feature
ease regions, extracting disease characteristics, categorizing the
prediction, the model combines the anchor mechanism of
disease, and performing bounding box regression. The Faster
Faster (R-CNN) with the regression mechanism of “YOLO”
(R-CNN) model utilizes convolutional neural networks for the
for fast and accurate detection. Fig. 8 depicts the SSD model
extraction of features and then generates feature maps for cor-
that includes attention mechanisms. The first component is
responding images. However, the convolution kernel’s inherent
an enhanced capability for disease detection based on the
locality means that only local information of disease images
deep learning network model used to collect baseline disease
is retained, leading to information loss and reduced detection
features. The multi-scale feature detection network is the
accuracy. To address this issue, the study introduced attention
second part, and it uses cascaded-neural-networks to categorize
mechanisms, namely SE, ECA, and CBAM, without changing
features at various scales in order to learn about the disease’s
the feature extraction network’s structure or backbone features.
category and location, as well as low-layer convolutional layer
As a result of forward propagation after the last identity block,
features to enhance detection precision and Non–Maximum
these mechanisms were introduced to enhance the model.
suppression to generate the final detection results. Using a
2) Grape Leaves Disease Detection using YOLO-X Model:: multi-dimensional prediction strategy, the SSD model is able
The YOLO-X with Darknet53 network is a model with high to distinguish between small and large objects; the front-end
operational speed and flexibility. It includes four primary deep-learning models are responsible for the former, while
components: the input end, Backbone network, Neck, and Pre- at the back-end multi-dimensional feature detection models
diction. Fig. 7 illustrates the YOLO-X model based on various handle them. Although the front-end network delivers precise
attention mechanisms. In the YOLO-X model, the YOLO Head coordinates and geometry, it has a limited range of perception
has been changed to a decoupled head in the prediction section, and isn’t great at representing abstract concepts. Whereas
the anchor-based approach has been replaced with an anchor- the frontal network has a narrow receptive field and poor
free method, and the SimOTA method has been introduced representational capacity for geometric data, the posterior net-
for dynamic matching with positive samples. The model’s work has a wide receptive field and excellent representational
detection accuracy and speed have both been enhanced by ability for semantic data. Because of this, the SSD model
these revisions, and the models’ parameter sizes have been may overlook some diseases or incorrectly identify others.
significantly decreased. The YOLO-X model is known for its Six feature images of varying sizes were collected from the
high detection speed and precision, but it has some limita- “SSD” model and supplied into the various attention modules
tions when applied directly for disease detection in different in order to better represent critical feature information and
environments. For instance, its backbone lacks the ability to identify disease object features. With this method, the SSD
extract features and integrate high-quality contextual feature model is better able to recognize diseased items.
information, leading to a reduction in the model’s detection
precision. Therefore, in this study, the Darknet53 network IV. R ESULTS AND E XPERIMENTS
structure of the YOLO-X model remained unchanged, allowing
A. Evaluation Metrics
pre-training weights to be directly loaded into model training.
The YOLO-X model can selectively strengthen key features Results were evaluated based on standard measures for
while suppressing irrelevant ones based on the branches of the evaluating target detection. One class of targets will be eval-
backbone network, namely “Darknet53”, “convolutional block uated using ”Precision,” ”Recall,” ”Average Precision,” and
www.ijacsa.thesai.org 1204 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 6, 2023

Fig. 6. Faster (R-CNN) model with attention techniques.

Fig. 7. YOLO-X model with attention mechanism.

Fig. 8. SSD model with attention mechanism.

”Mean Average Precision,” while all targets will be evaluated In Eq. 4, the variables P, TP, FP, R, FN, and F1 represent
using ”Mean Average Precision.” However, in this study, we various metrics used to evaluate the performance of a model.
evaluated the grape leaf disease detection model’s performance P is the precision, which measures the percentage of correct
on a wider set of metrics, including the mean absolute per- positive predictions. The probability that grape disease leaves
centage (mAP), the frame rate (FPS), the parameters, and the are accurately detected is denoted by true positives (’TP’),
precision (P) and recall (R) values. The Eq. 1,2 and 3 were whereas the probability that they are mistakenly categorised
used to calculate P, R, and F1. as positive is denoted by false positives (’FP’). Recall, or
the proportion of true positives that were correctly detected,
T rueP ositives
P recision = ∗ 100 (1) is denoted by the letter R. The likelihood of mislabeling a
T rueP ositives + F alseP ositives positive sample as negative is known as the ”Fasle negatives”
(T rueP ositives) (’FN’) rate. F1 is a measure of accuracy that is the harmonic
Recall = ∗ 100 (2) mean of two other metrics, recall and precision.
(T rueP ositives + F alseN egatives)
1
(2 ∗ P recision.Recall)
Z
F 1score = (3) P RdR (4)
(P recision + Recall) 0

www.ijacsa.thesai.org 1205 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 6, 2023

TABLE III. C OMPARISON A NALYSIS OF FASTER (R-CNN) M ODELS WITH TABLE IV. C OMPARISON A NALYSIS OF YOLO-X M ODELS WITH
D IFFERENT ATTENTION T ECHNIQUES FOR D ETECTING G RAPE D ISEASES D IFFERENT ATTENTION T ECHNIQUES FOR D ETECTING G RAPE D ISEASES
Model Precision Recall F1-Score mAP Model Precision Recall F1-Score mAP
Faster (R-CNN) model 75.06 74.42 74.74 79.12 YOLO-X model 82.35 74.85 78.42 83.22
Faster (R-CNN) with SE At- 79.80 84.23 81.96 85.39 YOLO-X with SE Attention 82.46 82.21 82.33 84.02
tention YOLO-X with ECA Attention 87.77 86.07 86.91 88.66
Faster (R-CNN) with ECA 76.54 78.71 77.61 81.93 YOLO-X with CBAM Attention 85.81 77.91 81.67 84.21
Attention YOLO-X with SE, ECA,CBAM Attention 89.77 86.97 85.91 88.96
Faster (R-CNN) with CBAM 75.75 75.89 75.82 79.65
Attention
Faster (R-CNN) with SE, 84.52 86.32 80.79 84.31
ECA,CBAM Attention speed of detection are achieved through the attention mech-
anism for grape leaves images. Among the various models,
“Faster (R-CNN) with SE, ECA, CBAM Attention” displayed
A higher value for TP indicates a more accurate prediction the best detection effect when compared with “Faster (R-CNN)
& better performance of the model. A model’s performance with SE Attention”. The “Faster (R-CNN) with SE Attention”
can be measured using mAP, which is a metric that averages model demonstrated a 3.26%, 5.52%, and 4.35% increase in
the average precision of all diseases. Eq. 5 defines mAP as P, R, and F1 values, respectively, with an increase of 3.46%
the average of all AP values. FPS stands for the number in mAP. In comparison with ” Faster (R-CNN) with CBAM
of pictures handled each second. The algorithm’s ability to Attention”, ” Faster (R-CNN) with SE Attention” increased P,
recognize items improves as the FPS increases. R, and F1 by respectively 4.05%, 8.34%, and 6.14%. When
N precision is considered, the ” Faster (R-CNN) with SE, ECA,
1 X and CBAM Attention” model shows optimal results. It focuses
mAP = AP (5)
N m=1 on channel features with the most significant information while
suppressing un-important features, making it ideal for detecting
A computer with 16 GB of RAM is used for this research, grape diseases in the dataset.
which runs Windows 10. Model parameters and hardware
configuration are considered in Pytorch 1.10.1. 2) YOLO-X Result Analysis: The YOLO-X model has been
enhanced with different attention mechanisms: SE, ECA, and
B. Experiment Results and Analysis CBAM. To compare their performance, all the models (includ-
ing the original YOLO-X model) were tested on the dataset
The grape disease dataset was utilized to compare the under the same configuration. The results are shown in Table
Faster(R-CNN), YOLO-X, and SSD models with the classical IV and in Fig. 10. Table IV shows that the “YOLO-X with
versions based on different attention mechanisms. The models SE Attention” model has improved performance compared to
were all trained and detected with the same configuration the YOLO-X model. Specifically, the precision, recall, and F1
information and training platform. values of the “YOLO-X with SE Attention” model increased
1) Faster (R-CNN) Result Analysis: The “Faster(R-CNN)” by 0.11 %, 7.36 %, and 3.91 %, respectively, while the
model can be combined with different attention mechanisms to mAP increased by 0.8%. Similarly, the “YOLO-X with ECA
create different versions. Also we have combined the three at- Attention” model also outperformed the YOLO-X model, with
tention mechanisms i.e. Faster (R-CNN) with SE, ECA,CBAM increases of 5.42%, 11.22%, and 8.49% in precision, recall,
Attention. To test their performance in detecting grape dis- and F1 values, respectively. The mAP also increased by 5.44%
eases, all these versions were used in the same experimental respectively. The “YOLO-X with CBAM Attention” model
setup, and the results are presented in Table III and in Fig. 9. also showed improvements, with increases of 3.46%, 3.06%,
Table III presents a comparison between the Faster (R-CNN) and 3.25% in precision, recall, and F1 values, respectively, and
model and four modified versions: “Faster (R-CNN) with SE a 0.99% increase in mAP. Based on the analysis above, it was
Attention”, “Faster (R-CNN) with ECA Attention”, and “Faster found that the detection performance of the YOLO-X model
(R-CNN) with CBAM Attention”. The results indicate that was improved with the introduction of attention mechanisms,
the Faster (R-CNN) with SE Attention model outperformed despite a slight increase in the parameters of the “YOLO-
the original model with an increase in P, R, and F1 values by X with SE Attention” and “YOLO-X with ECA Attention”
4.74%, 9.81%, and 7.22% respectively, and an increase in mAP models. Models were able to identify disease objects more
by 6.27%. Similarly, the Faster (R-CNN) with ECA Attention accurately due to the attention mechanisms that allowed them
model showed improvements over the original model with an to extract more comprehensive and rich features.
increase in P, R, and F1 values by 1.48%, 4.29%, and 2.87% Out of all the models, the “YOLO-X with SE, ECA,CBAM
respectively, and an increase in mAP by 2.81%. Finally, the Attention” model had the best detection performance. Com-
“Faster (R-CNN) with CBAM Attention” model showed slight pared to the “YOLO-X with SE Attention” model, the “YOLO-
improvements over the original model with an increase in P, X with ECA Attention” model had a 5.31%, 3.86%, and
R, and F1 values by 0.69%, 1.47%, and 1.08% respectively, 4.58% increase in P, R, and F1 values, respectively, a 4.64%
and an increase in mAP by 0.53%. increase in mAP, a 4.8 increase in FPS value, and a 0.49
MB expansion in parameters. Compared to the “YOLO-X with
Based on the analysis above, it is evident that the per- CBAM Attention” model, the “YOLO-X with ECA Attention”
formance of Faster (R-CNN) improved after the inclusion of model had a 1.96%, 8.16%, and 5.24% increase in P, R, and F1
attention mechanisms, despite a slight increase in parameters values, respectively, a 4.45% increase in mAP, a 1.8 increase in
for “Faster (R-CNN) with SE Attention and Faster (R-CNN) FPS value, and a 0.66 MB expansion in parameters. Compared
with CBAM Attention”. Enhanced precision and accelerated to other models YOLO-X with SE, ECA,CBAM Attention
www.ijacsa.thesai.org 1206 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 6, 2023

Fig. 9. Comparison analysis of Faster (R-CNN) models with different attention techniques.

Fig. 10. Comparison analysis of YOLO-X models with different attention techniques.

models had outperformed than previous models. In conclusion, Compared to the SSD model, the SSD with SE Attention
even though the “YOLO-X with SE, ECA,CBAM Attention” model showed significant improvements in P, R, and F1 values
model had more parameters than the other three models, it by 2.72%, 15.23%, and 9.45%, respectively.
achieved the best detection results with fast operation speed on
the grape disease dataset, partially due to its ability to achieve The SSD with ECA Attention model also showed improve-
cross-channel interaction. ments over the SSD model, but to a lesser degree. A relative
increase of 1.35%, 8.77%, and 5.47% in P, R, and F1 values
3) SSD Result Analysis: Under the same experimental was experienced, while a relative increase of 6.67% was seen
conditions, all the models were utilized to detect diseases on in mAP.
the plant village dataset, & the results of the experiment can
be found in Table V and in Fig. 11. The SSD with CBAM Attention model showed the smallest
Table V displays the results of various different models. The improvements over the SSD model. There is an increase
comparison is based on various metrics, including precision of 0.94 %, 3.61 %, and 2.48 % in P, R, and F1 values,
(P), recall (R), and F1 values, mean average precision (mAP). respectively, as well as a 4.91% increase in mAP and a
www.ijacsa.thesai.org 1207 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 6, 2023

TABLE V. C OMPARISON ANALYSIS OF SSD MODELS WITH DIFFERENT ”SSD” model was found to be ideal for monitoring field
ATTENTION TECHNIQUES FOR DETECTING GRAPE DISEASES grapes in real time. The ”YOLO-X” models demonstrated the
highest detection accuracy with the fewest parameters, and
Model Precision Recall F1-Score mAP
SSD model 80.74 68.87 74.33 76.23
they performed well while recognising both small objects and
SSD with SE Attention 83.46 84.10 83.78 86.96 items that were partially obscured.
SSD with ECA Attention 82.09 77.64 79.80 82.90
SSD with CBAM Attention 81.68 72.48 76.81 81.14
SSD with SE, ECA,CBAM Attention 85.46 84.90 84.78 83.96 R EFERENCES
[1] W. Baudoin, A. Nersisyan, A. Shamilov, A. Hodder, D. Gutierrez,
D. PASCALE S, S. Nicola, N. Gruda, L. Urban, J. Tanny et al., Good
3.38 MB increase in the model parameters. ”SSD with SE Agricultural Practices for greenhouse vegetable production in the South
East European countries-Principles for sustainable intensification of
Attention”, ”SSD with ECA Attention”, and ”SSD with CBAM smallholder farms. FAO, 2017, vol. 230.
Attention” models were all enhanced by the incorporation
[2] G. A. Carlson, “A decision theoretic approach to crop disease prediction
of attention modules in the network architecture. However, and control,” American Journal of Agricultural Economics, vol. 52,
the three models were able to effectively identify important no. 2, pp. 216–223, 1970.
information in feature images while filtering out irrelevant [3] U. Sirisha and B. S. Chandana, “Privacy preserving image encryption
information based on feature importance. As a result, the with optimal deep transfer learning based accident severity classification
detection performance of the three attention mechanisms with model,” Sensors, vol. 23, no. 1, p. 519, 2023.
SSD was superior to that of the SSD model. [4] U. Sirisha and S. C. Bolem, “Aspect based sentiment & emotion anal-
ysis with roberta, lstm,” International Journal of Advanced Computer
We have applied the different attention mechanism but, the Science and Applications, vol. 13, no. 11, 2022.
“SSD with SE, ECA,CBAM Attention model” demonstrated [5] U. Sirisha and B. Sai Chandana, “Semantic interdisciplinary evaluation
the best detection performance with significantly faster real- of image captioning models,” Cogent Engineering, vol. 9, no. 1, p.
2104333, 2022.
time processing than the other three models. Compared to the
“SSD with ECA Attention” model, the “SSD with SE Atten- [6] R. Ghaffari, J. Laothawornkitkul, D. Iliescu, E. Hines, M. Leeson,
R. Napier, J. P. Moore, N. D. Paul, C. N. Hewitt, and J. E. Taylor,
tion” model showed a 1.37%, 6.46%, and 3.98% improvement “Plant pest and disease diagnosis using electronic nose and support
in P, R, & F1 values, respectively. Compared to the “SSD with vector machine approach,” Journal of plant diseases and protection,
CBAM Attention” model, the “SSD with SE Attention” model vol. 119, pp. 200–207, 2012.
showed a 1.78%, 11.62%, and 6.97% improvement in P, R, and [7] D. Zhang, G. Chen, H. Zhang, N. Jin, C. Gu, S. Weng, Q. Wang,
F1 values, respectively. and Y. Chen, “Integration of spectroscopy and image for identifying
fusarium damage in wheat kernels,” Spectrochimica Acta Part A:
These experimental results demonstrate that the SE attention Molecular and Biomolecular Spectroscopy, vol. 236, p. 118344, 2020.
mechanism optimized feature images, resulting in signifi-
[8] Z. Wang, K. Wang, S. Pan, and Y. Han, “Segmentation of crop dis-
cantly better detection performance and real-time processing ease images with an improved k-means clustering algorithm,” Applied
compared to the other three models. Therefore, the “SSD engineering in agriculture, vol. 34, no. 2, pp. 277–289, 2018.
with SE, ECA,CBAM Attention” model can be effectively [9] R. Kamath, M. Balachandra, and S. Prabhu, “Crop and weed discrimi-
applied in the detection of various grape diseases with superior nation using laws’ texture masks,” International Journal of Agricultural
comprehensive performance. and Biological Engineering, vol. 13, no. 1, pp. 191–197, 2020.
[10] C. Bi and G. Chen, “Bayesian networks modeling for crop diseases,” in
4) Comparison Analysis : After screening, the three opti- Computer and Computing Technologies in Agriculture IV: 4th IFIP TC
mal disease detection models were compared to present their 12 Conference, CCTA 2010, Nanchang, China, October 22-25, 2010,
disease detection performance. The analysis above showed that Selected Papers, Part I 4. Springer, 2011, pp. 312–320.
“Faster(R-CNN)”, “YOLO-X”, and “SSD” models when com- [11] K. P. Ferentinos, “Deep learning models for plant disease detection
bined with multiple attention mechanisms were the optimal and diagnosis,” Computers and electronics in agriculture, vol. 145, pp.
311–318, 2018.
models of their respective detection methods. Fastest R-CNN
[12] F. Schroff, D. Kalenichenko, and J. Philbin, “Facenet: A unified
models exhibited the lowest overall detection accuracy, the embedding for face recognition and clustering,” in Proceedings of the
slowest operating speed, and the most parameters. The ”SSD” IEEE conference on computer vision and pattern recognition, 2015, pp.
models’ rapid operation speed and great accuracy made them 815–823.
ideal for near-instantaneous disease diagnosis in vineyards. [13] W. Yang, S. Fan, S. Xu, P. King, B. Kang, and E. Kim, “Autonomous
Strong robustness was demonstrated by the ”YOLO-X” mod- underwater vehicle navigation using sonar image matching based on
els, which achieved the maximum detection precision with the convolutional neural network,” IFAC-PapersOnLine, vol. 52, no. 21,
pp. 156–162, 2019.
fewest parameters and performed well while identifying both
[14] S. Sivaraman and M. M. Trivedi, “Active learning for on-road vehicle
small objects and items hidden by background clutter. detection: A comparative study,” Machine vision and applications,
vol. 25, pp. 599–611, 2014.
V. C ONCLUSION [15] L. Zhang, L. Lin, X. Liang, and K. He, “Is faster r-cnn doing
well for pedestrian detection?” in Computer Vision–ECCV 2016: 14th
After initial screening, three top disease detection models European Conference, Amsterdam, The Netherlands, October 11-14,
were selected and their performance was compared. The results 2016, Proceedings, Part II 14. Springer, 2016, pp. 443–457.
of the foregoing investigation demonstrated that the ” Faster(R- [16] U. Sirisha and B. S. Chandana, “Gitaar-git based abnormal activity
CNN),” ”YOLO-X,” and ”SSD” models, when enhanced with recognition on ucf crime dataset,” in 2023 5th International Conference
numerous attention mechanisms, provided the most accurate on Smart Systems and Inventive Technology (ICSSIT). IEEE, 2023, pp.
1585–1590.
detection results. Overall, ” Faster (R-CNN) ” models exhibited
[17] P. N. Srinivasu, G. JayaLakshmi, R. H. Jhaveri, and S. P. Praveen,
the lowest detection precision, the slowest operating speed, “Ambient assistive living for monitoring the physical activity of diabetic
and the most parameters of the three types of models. Due adults through body area networks,” Mobile Information Systems, vol.
to its excellent accuracy and quick processing speed, the 2022, pp. 1–18, 2022.

www.ijacsa.thesai.org 1208 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 14, No. 6, 2023

Fig. 11. Comparison analysis of SSD models with different attention techniques.

[18] N. R. Sai, B. S. Chandana, S. P. Praveen, S. S. Kumar et al., “Improving learning,” Frontiers in Plant Science, vol. 12, p. 695749, 2021.
performance of ids by using feature selection with ig-r,” in 2021 Fifth [24] H. Yuan, J. Zhu, Q. Wang, M. Cheng, and Z. Cai, “An improved deeplab
International Conference on I-SMAC (IoT in Social, Mobile, Analytics v3+ deep learning network applied to the segmentation of grape leaf
and Cloud)(I-SMAC). IEEE, 2021, pp. 1–8. black rot spots,” Frontiers in Plant Science, vol. 13, 2022.
[19] A. Fuentes, S. Yoon, S. C. Kim, and D. S. Park, “A robust deep- [25] A. S. Ansari, M. Jawarneh, M. Ritonga, P. Jamwal, M. S. Mohammadi,
learning-based detector for real-time tomato plant diseases and pests R. K. Veluri, V. Kumar, and M. A. Shah, “Improved support vector
recognition,” Sensors, vol. 17, no. 9, p. 2022, 2017. machine and image processing enabled methodology for detection and
[20] Z. Liu, X. Yuan, J. Weng, Y. Liao, and L. Xie, “Application of bitter classification of grape leaf disease,” Journal of Food Quality, vol. 2022,
gourd leaf disease detection based on faster r-cnn,” in Advancements 2022.
in Mechatronics and Intelligent Robotics: Proceedings of ICMIR 2020. [26] S. P. Praveen, T. B. Murali Krishna, C. Anuradha, S. R. Mandalapu,
Springer, 2021, pp. 191–198. P. Sarala, and S. Sindhura, “A robust framework for handling health
[21] J. Qi, X. Liu, K. Liu, F. Xu, H. Guo, X. Tian, M. Li, Z. Bao, and care information based on machine learning and big data engineering
Y. Li, “An improved yolov5 model based on visual attention mechanism: techniques,” International Journal of Healthcare Management, pp. 1–
Application to recognition of tomato virus disease,” Computers and 18, 2022.
electronics in agriculture, vol. 194, p. 106780, 2022. [27] M. Koklu, M. F. Unlersen, I. A. Ozkan, M. F. Aslan, and K. Sabanci,
[22] R. Polly and E. A. Devi, “A deep learning-based study of crop “A cnn-svm study based on selected deep features for grapevine leaves
diseases recognition and classification,” in 2022 Second International classification,” Measurement, vol. 188, p. 110425, 2022.
Conference on Artificial Intelligence and Smart Energy (ICAIS). IEEE, [28] M. Ji and Z. Wu, “Automatic detection and severity analysis of
2022, pp. 296–301. grape black measles disease based on deep learning and fuzzy logic,”
[23] J. Zhu, M. Cheng, Q. Wang, H. Yuan, and Z. Cai, “Grape leaf black Computers and Electronics in Agriculture, vol. 193, p. 106718, 2022.
rot detection based on super-resolution image enhancement and deep

www.ijacsa.thesai.org 1209 | P a g e
 © 2023. This work is licensed under
http://creativecommons.org/licenses/by/4.0/ (the “License”). Notwithstanding
the ProQuest Terms and Conditions, you may use this content in accordance
with the terms of the License.