0% found this document useful (0 votes)
5 views5 pages

A Comparative Study of Deep Learning Models for Guava Leaf Disease Detection

This study compares various deep learning models for detecting guava leaf diseases, emphasizing the importance of early diagnosis for improving agricultural productivity. The DenseNet169 model achieved the highest accuracy of 96.12% in classifying diseases such as Canker, Dot, Mummification, and Rust using a dataset of guava leaf images. The research highlights the effectiveness of convolutional neural networks and transfer learning in automating disease detection in plants.

Uploaded by

erpramodphd
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
5 views5 pages

A Comparative Study of Deep Learning Models for Guava Leaf Disease Detection

This study compares various deep learning models for detecting guava leaf diseases, emphasizing the importance of early diagnosis for improving agricultural productivity. The DenseNet169 model achieved the highest accuracy of 96.12% in classifying diseases such as Canker, Dot, Mummification, and Rust using a dataset of guava leaf images. The research highlights the effectiveness of convolutional neural networks and transfer learning in automating disease detection in plants.

Uploaded by

erpramodphd
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

A Comparative Study of Deep Learning Models

for Guava Leaf Disease Detection


2023 Third International Conference on Advances in Electrical, Computing, Communication and Sustainable Technologies (ICAECT) | 978-1-6654-9400-7/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICAECT57570.2023.10117860

Rajasekaran Thangaraj* Mohan M Moulik M


Department of Computer Science and Department of Computer Science and Department of Computer Science and
Engineering, Centre for IoT and AI Engineering Engineering
KPR Institute of Engineering and KPR Institute of Engineering and KPR Institute of Engineering and
Technology Technology Technology
Coimbatore, India Coimbatore, India Coimbatore, India
[email protected] [email protected] [email protected]

Logeshwari A Loganathan T Koushika P


Department of Computer Science and Department of Computer Science and Department of Computer Science and
Engineering Engineering Engineering
KPR Institute of Engineering and KPR Institute of Engineering and KPR Institute of Engineering and
Technology Technology Technology
Coimbatore, India Coimbatore, India Coimbatore, India
[email protected] [email protected] [email protected]

Abstract— Plant diseases are regarded as one of the major thought to have great therapeutic value as well as other health
variables affecting food production and reducing production advantages. It is primarily recognized for its antioxidant
losses. Early identification and disease diagnosis of fruit crops abilities. One of the best sources of dietary fiber is found in
are essential for the food industry to increase production. The it. Guavas are more reminiscent of citrus than other vitamin
development of the healthy fruit sector is highly dependent on C-rich fruits. It has various health advantages, including the
plant diseases. Guava fruit is a good source of fiber, calcium,
pectins, and vitamin C. Guava disease has become a major issue
ability to control metabolism, which promotes weight loss. It
for guava production, which has a negative impact on farmers' supports fertility and nourishes the skin and also it stops the
socioeconomic growth. Infections that harm guava plants and spread of cancerous cells. It halts the progression of diabetes.
fruits pose the greatest obstacles, although they can be avoided
by early identification. Deep learning has recently been widely Canker and algal spots are the two main ailments that can
used to automatically identify a variety of diseases that cannot damage guava trees. The quantity and quality of the crop are
be seen with the human eye. In this study, convolution neural significantly improved by early disease identification. Now-
models are suggested to use leaf images to detect guava disease. a-days, it is very challenging to identify the disease in its early
The presented models can be used to categorize the main guava stages. Subsequently, the regular monitoring of the crops in
leaf diseases, including Canker, Dot, Healthy, Mummification
and Rust. The kaggle dataset is used to gather images of guava
the agriculture field is also difficult. Moreover, professionals
with the disease Canker, Dot, Healthy, Mummification and are involved in the identification of plant disease with the
Rust. In our work, CNN models like DenseNet121, DenseNet169, naked eye, this method provides ridiculous results due to a
InceptionV3, and Xception are used to categorize the diseases lack of competence and excessive processing time. In order
that affect guava leaves. The experimental findings indicate to minimize a significant loss, automatic detection is
that, with an accuracy of 96.12%, the DenseNet169 model proposed to detect the diseases in the beginning stage using
outperformed other models. When the model's effectiveness is leaf images. The diagnosis of plant diseases, the prediction of
assessed using performance metrics like precision, recall, and F1 plant development, the detection of pests, etc. are just a few
score, the results demonstrate that DenseNet169 gives excellent of the agricultural domains where Deep Learning (DL), a
results.
subset of Machine Learning (ML), has excelled recently. The
primary focus of DL is to automate the manual process of
Keywords—Plant leave disease, Guava leave disease, Deep
learning, Convolutional Neural Network, Transfer Learning, Pre- extracting high-level features from raw data. In the
trained models. agriculture domain, the collection of a huge number of the
image is a tedious task that imposes to adapt the pre-trained
I. INTRODUCTION DL models that utilize the weights of ImageNet[ 2]. Transfer
Guava is a well-liked fruit that is grown for commercial Learning(TL) is one of the ML methods that guarantees a
purposes in many nations because of its great nutritional and solution with a limited dataset and applies the information
therapeutic worth. In general, the guava plant's leaves and gained from solving one problem to others that are similar.
fruits provide amazing nutrition and minerals that are
beneficial to human health [1]. However, the diseases of the Transfer Learning (TL), one of the deep learning algorithms,
guava plant have a big impact on the economy, diminishing guarantees a solution with a limited dataset and applies the
the quantity and quality of production. However, physically information gained from solving one problem to others that
recognizing and diagnosing guava leaf infections is the are similar. Deep feature extraction and fine-tuning are the
challenging part. Guava, a popular tropical fruit, is often only two strategies that are frequently used in the TL. The pre-
available in India in May and June of each year. Guava is trained models' pre-learned features are used to train the new
output layer in the feature extraction phase. When a model

Authorized licensed use limited to: Manipal University. Downloaded on February 27,2024 at 16:49:23 UTC from IEEE Xplore. Restrictions apply.
has to be fine-tuned, it is retrained using new images, and the encouraging results with an accuracy of 98.06% for the three
weights of the model's top layer are revised in accordance types of maize leaf disease.
with the unseen images. Additionally, the number of classes
in the model's top layer corresponds to the number of classes Several pre-trained deep learning models with optimized
in the new dataset[3 ].In order to identify guava disease using model parameters are utilized by Tu et al[14] to categorize
leaf images, this research compares and contrasts the most and diagnose plant diseases. Maximum test accuracy of
recent deep learning models. Canker, Dot, Mummification, 99.75% is attained with the DenseNet model. With an
and rust are the different forms of guava leaf diseases from accuracy rate of 99.97% using EfficientNetB4 on the
which the DL models have been trained to identify. benchmark dataset, Atila et al. [15] demonstrated that
EfficientNet outperforms the conventional CNN model such
The paper is divided into the following sections: Section II as AlexNet, VGG, etc. by identifying sickness in plants. Deep
describes the pertinent research projects that were carried out learning techniques are more accurate in identifying plant
to discover plant diseases. In Section III, the dataset's diseases than more conventional machine learning techniques
description is given. The procedures used to diagnose the and have a better learning capacity. However, in the future,
guava leaf disease are described in Section IV. Sections V several images are needed to train the model, and image
and VI, respectively, discuss evaluation metrics and gathering is a crucial component. The results of the study
experimental findings. Section VII contains a discussion of showed that early diagnosis of plant disease resulted in an
the outcomes. The conclusion is presented in Section VIII. increase in production. The present approach mentioned in
the aforementioned paper made use of a number of deep
II. RELATED WORKS learning models to enhance the prediction of plant leaf
Recent work on deep learning-based plant disease detection disease. A transfer learning technique is also incorporated to
is provided in this section. The most often used traditional improve classification accuracy.
machine learning methods for automatically identifying plant
leaf diseases are KNN, SVM, and random forests[4]. When III. DATASET DESCRIPTION
recognizing guava leaf disease, Islam et al.[5] used image The guava dataset used in this investigation is obtained from
processing approaches that obtained 95% accuracy. For the the kaggle[16] database. Nearly,2300 images for the guava
experiment, the author used guava leaf images from the leaf diseases are used to assess the model's performance. Fig.
PlantVillage dataset. Song et al. introduced an SVM classifier 1 shows the sample images from the guava dataset for each
in [6] that accurately detects disease in maize leaves with an class. The entire dataset is split into these two parts, 80% for
89.6% accuracy rate. In order to classify three different guava training and 20% for testing. Table 1 lists the number of
leaf diseases, Rothe et al.[7] used image processing images utilized for the training, validation, and testing. In
techniques including pattern recognition, which had an order to meet the dimensional requirements of the model, the
accuracy rate of 89.60%. SVM and K-nearest neighbour (K- images are also scaled down to 299 x 299 pixels for
NN) classifiers were recommended for diagnosing guava leaf InceptionV3 and Xception. Similarly, the images are resized
diseases by Abirami et al. in [8]. Five different types of leaf to 224 x 224 for DenseNet models. The learning rate is set to
diseases are employed in this piece, which includes 125 0.001, the batch size is 32 and the number of epochs is 25 for
images.The results show that SVM and K-NN achieved a all the models during training. The optimizer also uses
classification accuracy of 97.2% and 92%, respectively. Adaptive Moment Estimation (Adam). Cross-entropy is the
loss function.
Mohanty et al. [9] found 26 illnesses in 14 distinct plant
species using AlexNet and GoogleNet. For the purpose of
training the model, they gave equal weight to strategies for
transfer learning and learning from scratch. They reached the
highest accuracy of 99.34% using GoogleNet. Ferentinos
[10] uses five pre-trained models to diagnose disease,
including VGG, AlexNet, AlexNetOWTBn, Overfeat, and
GoogleNet. Gitaramani and Pandian [11] used a 9-step deep
neural network to identify plant disease and attained an
accuracy level of 96.46%. Liu et al. [12] created a model to
recognize four distinct apple leaf diseases by substituting the
fully connected AlexNet layer with the initial layer and
attaining accuracy of 97.62%. They were inspired by AlexNet
and GoogleLeNet architecture. Pandey, H. M. et al [13] The
adoption of compact CNN (Compact Net) software for leaf
disease diagnosis and detection. The study thoroughly
examined the advantages and disadvantages of the present
neural network topologies within the setting of the study. To Fig. 1. Sample Images of Guava Leaf diseases
enable other researchers to re-implement the theoretical
model for future uses, a step-by-step overview of the model
has been supplied. The DenseNet model has shown

Authorized licensed use limited to: Manipal University. Downloaded on February 27,2024 at 16:49:23 UTC from IEEE Xplore. Restrictions apply.
TABLE I. DATASET SUMMARY
convolution neural network as well as traditional convolution
Classes Training Validation Testing are both used in the Xception model. The phrase "Separable
Convolutional" is used to describe a depth-wise separable
convolution. Because it is more cost-effective to compute,
Canker 166 55 57 this layer is meant to take the role of the traditional
convolutional layer. Point-wise convolution is performed
after acting on the depth-wise separable convolution firstly in
Dot 106 35 37
the initial depth-wise separable convolutional layer. The
Xception model employs a modified depth-wise detachable
Healthy 773 257 259 convolution layer and performs a 1 1 convolution on each
channel before to performing a depth-wise convolution
operator.The pre-trained ImageNet model used in this study
Mummification 165 55 55 is named Xception.
V. EVALUATION METRICS
Rust 167 55 57 The effectiveness of the models in detecting guava leaf
diseases is assessed using metrics such as accuracy, precision,
recall, and F1 score. The formulas for the various
measurement indicators are listed below
IV. STATE-OF-THE-ART DEEP LEARNING IMAGE CLASSIFIERS TP + TN
Accuracy = (1)
A. Inception V3 TP + FP + TN + FN
The GoogLeNet model was first presented by [17], and the
InceptionV3 is an updated version of that model. Factorizing
TP
convolution reduces the amount of parameters without Pr ecision = (2)
sacrificing the model's effectiveness, and it has 42 layers with TP + FP
fewer parameters. Convolutions of various sizes, including TP
Re call = (3)
1x1, 3x3, and 5x5, can be found in this model's inception
TP + FN
module. Additionally, the 5x5 convolution layer in
InceptionV3 is split into two 3x3 layers using a factorization
feature.The parameter reduction enhances accuracy and 𝑃𝑟 𝑐𝑐𝑖𝑠𝑖𝑜𝑛 𝑋 𝑅𝑒 𝑐𝑎𝑙𝑙
reduces overfitting. By deleting the previous classification 𝑅𝑒 𝑐 𝑎𝑙𝑙 = 2 𝑋 (4)
𝑃𝑟 𝑒𝑐𝑖𝑠𝑖𝑜𝑛 +𝑅𝑒 𝑐𝑎𝑙𝑙
layers and including a new softmax classifier, InceptionV3
has been improved. The number of classes in the animal
dataset matches the classes in the top layer of the model. When Where TP→True Positive, FP→False Positive,
fine-tuning, training using the new dataset updates the weight TN→ True Negative, FN→ False Negative
of the new softmax layer.
VI. EXPERIMENTAL RESULT
B. DenseNet
In this study, the effectiveness of convolutional neural
The DenseNet family, the most widely used image
network models for guava disease diagnosis from leaf images
classification model, includes 121 layers, and the
is evaluated. To maximize model efficiency and yield the best
DenseNet121 model is one of them. The models are created
results, all models have been run for 25 epochs. As the
originally in the Torch framework and then translated into the
classification problem is multiclass, the batch size of the
Caffe format. All DenseNet models begin with the ImageNet
image for each epoch is set to 32, and Adam is used as the
database as their pre-training set.Similarly, DenseNet169, is
optimizer. The learning rate is a crucial consideration for
a well known architecture mainly used for classification
deep learning models, and whilst a low learning rate could
problems. Convolutional layers, maxpool layers, dense
result in slower convergence to the ideal solution, a high
layers, and a transition layer are all parts of the architecture.
learning rate yields longer jumps for the solutions that escape
The architecture of the model comprises the ReLU activation
the local optima. The model is assessed for each experiment
function, and the last layer incorporates softmax activation.
using the accuracy measure and categorical cross-entropy
The maxpool layer reduces the dimensionality of its inputs
loss. The accuracy curves for all the models are shown in Fig.
while the convolutional layer retrieves image characteristics.
2, and Fig.3 presents the loss value obtained by each model
Following the flatten layer, which presents as an artificial
in various epochs. Apparently, the DenseNet169 model
neural network with a single array input, is the fully
achieves the highest accuracy and low loss value compared
connected layer[18].
to other models. The accuracy achieved by each CNN model
C. Xception for guava leaf disease detection is shown in Fig.4. As the
The Xception [19] model, an extension of the inception result, DenseNet169 model achieved the highest
network, uses separable convolution layers as opposed to classification accuracy of 96.12% followed by Xception
regular convolution layers. The 36 convolution layers in it are which is 90.54%. Moreover, DenseNet121 and InceptionV3
used to extract the input's features. A depth-wise separable shows similar performance with small variations in accuracy

Authorized licensed use limited to: Manipal University. Downloaded on February 27,2024 at 16:49:23 UTC from IEEE Xplore. Restrictions apply.
89.89% and 89.46% respectively. The model efficiency is
further assessed by measuring the precision, recall, F1-score Accuracy(%)
values achieved by each model, and its values are depicted in
Table II. The results confirm that in our case smaller models
produce better results compared to the deeper models with 100 96.12
more parameters. The comparison of parameters for each 89.89 89.46 90.54
95
model is shown in Table II.
90
85
80
75
70
65
60

(a) (b)

DenseNet169 DenseNet121
InceptionV3 Xception
Fig. 4. Test Accuracy Comparison of the models

TABLE. II. COMPARISON OF PERFORMANCE METRICS

Accuracy F1 Parameters
Models (%) Precision Recall score (Millions)

(c) (d) DenseNet169 96.12 0.96 0.94 0.94 25


Fig. 2. Comparison of Training and Validation Accuracy- (a)DenseNet169,
(b) DenseNet121, (c) Inception V3, (d) Xception
DenseNet121 89.89 0.89 0.88 0.88 23

InceptionV3
89.46 0.89 0.89 0.88 138

Xception 90.54 0.90 0.89 0.88 144

VII. DISCUSSION
Examining advanced deep learning models for
diagnosing plant diseases is advancing research. The
InceptionV3, Xception, and DenseNets are the
(a) (b) networks utilized in our work. As evidenced by the
test result in Fig.4, DenseNet169, achieves a far better
result than Xception, InceptionV3 and DenseNet121.
This experiment shows that smaller models
perform better than deeper networks. Also, compared
to the deeper models such as Xception and
InceptionV3, DenseNet has fewer parameters and it
shows significant performance. Among the smaller
models, DenseNet121 has the least parameter than
DenseNet169 but its accuracy achieved is less
compared to the DenseNet169 model. Xception has
more learning parameters than other models even
(c) (d) though the model shows the least performance
Fig. 3. Comparison of Training and Validation Loss-
(a) DenseNet169,(b)DenseNet121,(c)InceptionV3,(d) Xception compared to DenseNet169. Compared to
DenseNet121 and InceptionV3 the performance of
Xception is better. Analyzing the overall results
smaller model is the best fit for identifying guava leaf
diseases than the deeper models.

Authorized licensed use limited to: Manipal University. Downloaded on February 27,2024 at 16:49:23 UTC from IEEE Xplore. Restrictions apply.
VIII. CONCLUSION [6] Sanjeev S Sannakki, Vijay S Rajpurohit, VB Nargund, and
Pallavi Kulkarni. Diagnosis and classification of grape leaf
The agricultural sector has seen numerous changes diseases using neural networks. In 2013 Fourth International
recently. To boost agricultural productivity, it is also critical Conference on Computing, Communications and Networking
to recognize plant diseases. With a growing population and Technologies (ICCCNT), pages 1–5. IEEE, 2013.
[7] M. T. S. Abirami, “Application of image processing in
rising demand for its fruit, guava is a crucial plant to keep an diagnosing guava leaf diseases,” International Journal of
eye on. Guava diseases have grown to be a serious issue since Scientific Research and Management, vol. 5, no. 7, pp. 5927–
they may cause a sharp decline in product quality and quantity, 5933, 2017.
which would undermine agricultural sustainability. For [8] S. R. Dubey, P. Dixit, N. Singh, and J. P. Gupta, “Infected fruit
part detection using k-means clustering segmentation technique,”
farmers, the improper diagnosis might result in significant Ijimai, vol. 2, no. 2, pp. 65–72, 2013.
financial losses. According to this study's identification [9] Yang Lu, Shujuan Yi, Nianyin Zeng, Yurong Liu, and Yong
accuracy, the proposed DCNN model is more efficient and Zhang. Identification of rice diseases using deep convolutional
offers a better solution to detect and treat guava leaf disease. neural networks. Neurocomputing, 267:378–384, 2017.
[10] Sharada P Mohanty, David P Hughes, and Marcel Salathe. Using
The most recent DL models, such as DenseNet121, deep ´ learning for image-based plant disease detection. Frontiers
DenseNet169, InceptionV3, and Xception, are used in this in plant science, 7:1419, 2016.
study to identify guava leaf diseases. Additionally, this work [11] Geetharamani, G., & Pandian, A. (2019). Identification of plant
suggests that transfer learning can be used to identify diseases leaf diseases using a nine-layer deep convolutional neural
network. Computers & Electrical Engineering, 76, 323-338.
with little to no data. The results of the experiment [12] Bin Liu, Yun Zhang, DongJian He, and Yuxiang Li.
demonstrate that, when compared to the other models, the Identification of apple leaf diseases based on deep convolutional
DenseNet model with the transfer learning strategy had the neural networks. Symmetry, 10(1):11, 2017.{17}
highest recognition accuracy of 96.12%. All of the models' [13] Waheed, A., Goyal, M., Gupta, D., Khanna, A., Hassanien, A. E.,
& Pandey, H. M. (2020). An optimized dense convolutional
performance comparisons demonstrate that DenseNet with neural network model for disease recognition and classification
fine-tuning exhibits the best performance. The model's in corn leaf. Computers and Electronics in Agriculture, 175,
performance makes it clear that deep learning models based 105456.
on transfer learning are ideal for automatically identifying leaf [14] Timo Ojala, Matti Pietikainen, and Topi Maenpaa.
Multiresolution grayscale and rotation invariant texture
diseases. classification with local binary patterns. IEEE Transactions on
pattern analysis and machine intelligence, 24(7):971–987, 2002.
REFERENCES [15] S. Lawrence, C. L. Giles, A. C. Tsoi, and A. D. Back, “Face
recognition: A convolutional neural-network approach,” IEEE
[1] M Thilagavathi S Abirami. Application of image processing in transactions on neural networks, vol. 8, no. 1, pp. 98–113, 1997.
diagnosing guava leaf diseases. International Journal of ScientificResearch [16] https://www.kaggle.com/datasets/omkarmanohardalvi/guava-
and Management, 5(7):5927–5933, 2017. disease-dataset-4-types
[2] Alvaro Fuentes, Sook Yoon, Sang Cheol Kim, and Dong Sun Park. A [17] Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z.
robust deep-learning-based detector for real-time tomato plant diseases (2016). Rethinking the inception architecture for computer
and pests recognition. Sensors, 17(9):2022, 2017. vision. In Proceedings of the IEEE conference on computer
[3] Thangaraj, R., Anandamurugan, S., & Kaliappan, V. K. (2021). vision and pattern recognition (pp. 2818-2826).
Automated tomato leaf disease classification using transfer learning- [18] G. Huang, Z. Liu, L. van der Maaten and K. Q. Weinberger,
based deep convolution neural network. Journal of Plant Diseases and
"Densely connected convolutional networks", Proc. CVPR, vol.
Protection, 128(1), 73-86.
1, pp. 2261-2269, Jul. 2017
[4] Monzurul Islam, Anh Dinh, Khan Wahid, and Pankaj Bhowmik. [19] F.Chollet, Xception: Deep learning with depthwise separable
Detection of guava diseases using image segmentation and convolutions. In: Proceedings of the IEEE conference on
multiclass support vector machine. In Electrical and Computer computer vision and pattern recognition,(2017) pp. 1251-1258.
Engineering (CCECE), 2017 IEEE 30th Canadian Conference on,
pages 1–4. IEEE, 2017.
[5] Youssef Es-saady, Ismail El Massi, Mostafa El Yassa, Driss
Mammass, and Abdeslam Benazoun. Automatic recognition of
plant leaves diseases based on serial combination of two svm
classifiers. In Electrical and Information Technologies (ICEIT),
2016 International Conference on, pages 561–566. IEEE, 2016.

Authorized licensed use limited to: Manipal University. Downloaded on February 27,2024 at 16:49:23 UTC from IEEE Xplore. Restrictions apply.

You might also like