Proceedings of the Second International Conference on Electronics and Sustainable Communication Systems (ICESC-2021)

IEEE Xplore Part Number: CFP21V66-ART; ISBN: 978-1-6654-2867-5

Leaf Disease Detection Using Deep Learning

Teenu Sahasra M
th
4 Year B.Tech, Department of IT Sai Kumari S
th
VR Siddhartha Engineering College 4 Year B.Tech, Department of IT
Vijayawada, India VR Siddhartha Engineering College
[email protected] Vijayawada, India
2021 Second International Conference on Electronics and Sustainable Communication Systems (ICESC) | 978-1-6654-2867-5/21/$31.00 ©2021 IEEE | DOI: 10.1109/ICESC51422.2021.9532909

Sai Meghana S
4th Year B.Tech, Department of IT
VR Siddhartha Engineering College
Vijayawada, India
[email protected] m

Mrs. P. Rama Devi

Professor, Department of IT’
VR Siddhartha Engineering College
Vijayawada, India
rama.polagani99 @gmail.com

Abstract— Economy contributes the most for the productivity production. So, to detect the disease we used a DL based
of the agriculture. In agricultural field, the disease in plants is approach. For disease recognition and classifying, we are
more common and the detection of disease in plants has become using a CNN based approach. In this pre-trained model, the
more feasible due to the above reason. These days’s plant disease CNN consists of convolution layers and max-pooling layers
detection has acquired enlarging scrutiny in surveilling crops of
and the later one is the one which is 2 fully connected layers.
large and various fields. Farmers undergo significant hassles in
chop and changing from one disease administer principle to a
The trial and error result has shown the [11] efficiency of the
different one. We can identify or spotting the tomato leaf diseases proposed model than VGG16 which is the pre-trained model.
for detection for surveillance and monitoring experts is the The accuracy of the classification differs from 75% to 100%
standard approach for detection. The plants get seriously concerning classes and the proposed model’s mean accuracy is
affected if the proper control hasn’t been take n and this 90.2% for 9 types of diseases and 1 healthy class.
represents the quality of the pants the production of the plants Agriculture has been one of the foremost predominant field s
will be affected. Detection of disease through some mechaniz ed of the Indian Economy. In the agricultural field, a serious role
technique and methodology is efficient and constructive because is being played by the disease detection in plants . The growth
it decreases an outsized toil of surveilling in the large cultivation.
of the plants is affected [14] because of the presence of the
In the premature phase we can detect the symptoms of the plant
diseases since their first appearance on their leaves of the plants.
diseases in plants. With the help of disease detection
By using this paper we can identify the algorithm which is used methodology and techniques the plant disease can be detected
for image segmentation and for automated classification used for at premature phases. In various roles of the plant such as
the detection of diseases of leaves in the plants. It also covers leaves, roots, etc and detects the traits of leaf diseases that are
distinct disease classification methods of working which is used observed. It uses images of the leaves for monotonous job
for the detection of diseases in plants. which can make the [3] detection and classification method of
the disease using images of the leaves automatic.
Keywords—CNN; VGG16; Deep learning; Pre-processing;
Keras; Tensorflow; Opencv II. BASIC T ERMINOLOGY
II-A) DEEP LEARNING
I. INT RODUCT ION (PROBLEM STATEMENT) Deep learning is used convert the human brain into process the
information and creating different design in artificial
Now a day's tomato has become the more welcomed crop
intelligence (AI) that is used for making conclusion. It is a
around the globe and it is often found in the kitchen, using it type of machine learning footing artificial neural networks
in various forms, ignoring the type of the cuisine. Following during which multiple layers of transform are won’t to extract
potato, tomato which is cultivated world widely. India stands
more and more high-level features from data.
in the 2nd position in the fabrication of tomatoes [2]. In tomato,
different forms of diseases come with the quality of the farm

978-1-6654-2867-5/21/$31.00 ©2021 IEEE 1797

Authorized licensed use limited to: KLE Technological University. Downloaded on November 19,2024 at 05:13:17 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Second International Conference on Electronics and Sustainable Communication Systems (ICESC-2021)
IEEE Xplore Part Number: CFP21V66-ART; ISBN: 978-1-6654-2867-5

II-B) VGG16 Our Work:

VGG16 is used in numerous deep learning image Leaf disease detection is used for detection of disease in th e
classification drawbacks; however, smaller network leaves by using CNN pre-trained models and also improves
architectures are often more worthwhile (such as SqueezeNet, the accuracy of the model. We used tomato leafs from the
GoogLeNet, etc.). But it's a superb building block for learning “Plant Village Dataset” for this project we also have different
purposes because it's straight forward to implement. diseases in them like Target spot, Early Blight etc. By using
II-C) KERAS CNN pre-trained model (VGG16) we can better accuracy and
Keras is an API library it is designed for humans . It is deep performance of the model. We have both advantages and
learning framework. Keras used to run on it. disadvantages for using this model like have more parameters
II-D) TENSORFLOW and model is somewhat expensive.VGG16 used for plant
Tensorflow is a deep learning framework. It is an end to end disease detection had 98.2% high accuracy and faster
open source platform used to train ML models. It mainly detection.
focuses on the related version that is suitable for running the
program. It is used in large scale for ML, DL projects in our
day to day life. IV. PROPOSED WORK
IV-A)PROJECT DESIGNM ETHODOLOGY:
III. LITERATURE REVIEW Use case Diagram:
We have referred to some of the latest conference papers In Unified Modelling Language (UML), a use case diagram to
which are related to our project on change detection and also build one, you'll use a group of specialized symbols and
observed some of the methods as well as procedures which are connectors. In figure-1.1, when we give a new input image
implemented and written those observations of each paper that first the module extracts the leaf features. Then it goes through
will be useful in development of our project. the CNN model. It then contrasts the attributes with an already
trained dataset. Then it goes through dense CNN and therefore
We have number of papers published for leaf disease detection the leaf attributes are extricated independently. Then the
and this is the Journal Selected Topics in [1] Applied Plant module will predict whether the plant leaf healthy or diseased.
Disease Detection (2020) and the methods proposed are Faster
RCNN, K-Means, and VGG16.We have different proposed
methods from different number of papers some of the uses
CNN, Support Vector Machine(SVM), VGGNet and other
pre-trained models likes AlexNet, DenseNet, VGG19 etc.
Disease in the plant can be identified in root, stem, leaf, fruit
etc. We have number of paper and journals for leaf disease
detection for tomato leafs, apple leafs, potato leafs and soon.
[5] They use different neural networks such as CNN, ANN etc
in the leaf disease detection. In [3] Detection of plant disease
we have use different techniques like image segmentation, soft
computing and confusion metrics.

By using confusion metrics we can find the accuracy of the

model and other variables like F1 score, Precision and recall.
Segmentation and soft computing are the techniques used for
the feature extraction [7] and detection in the images. These
CNN pre-trained models are used detect whether the leaf is
healthy or diseased. If the leaf is healthy it tell us that it is
healthy or tells us the leaf is diseased and tells us the name of Figure-1.1: Use case diagram for leaf disease detection
the disease and required remedies for it. [6] VGG16 is a pre- Class Diagram:
trained model of Convolution Neural Network it tells us about The class diagram is the building block of object-oriented
better performance and accuracy of the model depending on it modeling used for conceptual modeling of the structure of an
features in the model. [4] After the detection of the disease we appeal, and modeling explain the model into a code. Class
can also apply different techniques on it. diagrams also be used for data whittle. As we see the
normalization class comprises a raw image and it is fed to the
Information collected from the different paper we mainly CNN model which comprises dense and weight. The CNN
applied these on tomato, apple, potato, cherry, and corn leafs model systematizes and distinguishes by using the training
and frequently used model are [2] CNN pre-trained models. model. Leaf detection gets used to the features.

978-1-6654-2867-5/21/$31.00 ©2021 IEEE 1798

Authorized licensed use limited to: KLE Technological University. Downloaded on November 19,2024 at 05:13:17 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Second International Conference on Electronics and Sustainable Communication Systems (ICESC-2021)
IEEE Xplore Part Number: CFP21V66-ART; ISBN: 978-1-6654-2867-5

Figure-1.2: Class diagram for leaf disease detection

Activity Diagram:
The activity diagram is a behavioral diagram, it portray the
behavior of a system in UML diagram. An activity diagram is
similar to a flow chart that whittles the flow of one activity to
another activity. Activity Diagrams relate how activities are
interlinked to provide service that can be at conflicting levels
of ancestry. It is also appropriate for modeling. Collection of Figure 2: Leaf Disease Detection Architecture
use cases coordinate to constitute workflows.
IV-C) OVERVIEW OF ALGORITHM USED
CNN Algorithm:
A Convolutional Neural Network (CNN) is also a Deep
Learning algorithmic rule and functionality that can take
image as an input. We can extract the information from an
input image with the help of the multiple layers that are hidden
in the convolution neural network. There are four important
layers in CNN. They are Convolution layer (first layer), ReLU
layer (middle layer), Pooling layer (middle layer), Fully
connected layer (last layer). [8] The pixels that are present in
an image are recognized by CNN and that pixels of an image
are fed to the convolutional layer, then the convolutional layer
will perform some convolution operations then the following
result will be the convolution map. The rectified feature map
will be obtained when the convoluted map is given to the
ReLU layer. The image should be and is processed with
number of convolutions and ReLU layers [13] for locating the
features. Now to identify specific and particular parts of an
image we use different layers of pooling with various filters.
Figure-1.3: Activity diagram for leaf disease detection Now the featured map will be flattened and is fed to the last
layer, a fully connected layer to get the final output.
IV-B) ARCHITECTURE DIAGRAM VGG16:
In VGG16, it consists of 16 layers.
• Download the leaf disease dataset from the Kaggle website.
The input to VGG is based on ImageNet for RGB images.
Load the images from the dataset.
Preprocessing layer takes the RGB image with pixel values in
• After loading the dataset we have to Pre-processing the
the range of 0–255 and subtracts the mean image values which
plant's images.
are calculated over the entire ImageNet training set.
• Then we have to extract various features like color, shape,
and size from the leaf images.
• We have to train the leaf dataset and process them by using
RGB leaf images
• CNN algorithm is used for leaf disease detection. In that, we
use the model which is trained as VGG16 for classifying.
• The VGG16 model finds the accuracy and loss of it by
training and validating the model.
Figure-3: VGG16 Representation

978-1-6654-2867-5/21/$31.00 ©2021 IEEE 1799

Authorized licensed use limited to: KLE Technological University. Downloaded on November 19,2024 at 05:13:17 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Second International Conference on Electronics and Sustainable Communication Systems (ICESC-2021)
IEEE Xplore Part Number: CFP21V66-ART; ISBN: 978-1-6654-2867-5

In VGG16, it consists of 16 layers. VGG16 uses ImageNet for {(800, 200)}. It means 800 are trained images and 200 are
the classification of the images. Convolution layers of 3 filters, testing images.
maxpool layer of 2 filters. In the end it has 2 Fully connected
layers (2FC) and last layer is used as softmax for output. SOURCE: https://www.kaggle.com/emmarex/plantdisease
Convolutional Layer:
In a convolutional layer, a neuron is associated with the
specific extent of input neurons alternatively full-connection
in which the narrowing of the parameters that are to be learned
is done. As a result the growth of the network will be deeper
without any difficult parameters [9].
Pooling Layer:
These layers wont to minimize the dimension of the feature
maps. By reducing the parameters that are to be found and
used for the performed computation within the network. [12]
The features present during a region of the feature map can be Figure-4: Healthy and Diseased leaf images
summarized by the pooling layer, which are generated by a
convolution layer. TOOLS AND TECHNOLOGIES USED
Dense layer:
The neurons which are present are contributed by the previous 1. Python – 3.7(64bit):
layers input for [1] every particular neuron, during a layer so
they are connected densely. We can also s ay it as fully Python 3.0 was introduced in the year 2008. Python may be an
connected layer, meaning all neurons are connected during a application-oriented language and object-oriented scripting
layer for the upcoming layer language even through this version is used to be incompatible;
Working Procedure of the Algorithm: afterward many of the attributes are converted into compatible
1. Begin the process of VGG16 algorithm. with the help of current versions. This tells us about the gives
2. Take the leaf of different types of shape and size to adequate knowledge of the Python 3 version programming
generate their images.
3. Do the data pre-processing to normalize, shuffle, split 2. Anaconda3-64 bit:
the dataset implement different other operations like
feature extraction also. Anaconda is a package manager, clustering of many more
4. Pre-processing images [15] mainly depends on RGB open-source packages. It is free and easy to install required
images for the colour classification of the image. modules, libraries and packages which provide support from
5. This VGG16 models uses 80% of the data as trained the community. Anaconda is used from writing programs of
images and 20% of the data as testing images for Python, R and data science for classification. Open the
classification of the healthy and diseased leafs. Anaconda Cheat Sheet from the browser and download
6. In this model we used 9 diseased leaf groups and 1 Anaconda. Install conda and use it to ins tall required packages
healthy diseased leaf group for tomato leaf dataset. by using it you will get Miniconda.
7. Dataset is available at kaggle website of different leaf
images. VI. RESULTS & OBSERVATIONS OF THE PROJECT
8. VGG16 model is used to train/validate the dataset of
the leaf images. VI-A) DESCRIPTION OF RESULTS OBTAINED IN THE
9. After the training/validating the model we uses PROJECT
confusion matrix on leaves and find out f1 score, The VGG16 model is one of the deep learning techniques to
Accuracy, Confusion metrics etc. build our project. We trained the leaf dataset to reduce noises
10. By using the algorithm we display the disease name and disturbances that present in the leaf. The libraries that are
and the remedies for it and calculate the accuracy imported are tensorflow, keras, numpy, opencv.
rate, loss rate of the give model. The structure of files for the detection is as follows:

V. DATASET AND TOOLS

Plant Village Dataset:

Principally, the complete sets of images have been classified

among 2 classes i.e. healthy and diseased. The entire dataset
has been divided into 10 subject categories. We have collected
about 16,012 images of which contain 1,591 images of healthy
leaf and 14,421 images of the diseased leaf. Our target size is

978-1-6654-2867-5/21/$31.00 ©2021 IEEE 1800

Authorized licensed use limited to: KLE Technological University. Downloaded on November 19,2024 at 05:13:17 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Second International Conference on Electronics and Sustainable Communication Systems (ICESC-2021)
IEEE Xplore Part Number: CFP21V66-ART; ISBN: 978-1-6654-2867-5

A) Dataset – This file is made up of both train and val d. Segmentation, separate the background from the
(validation) images of leaves. Train file consists of 1 foreground images. The background is separated to
healthy leaf images and 9 diseased leaf images. Val get more clarity of the images.
file also consists of 1 healthy and 9 diseased leaf
images. e. Morphology, smoothing edges of the images. After
separating background from the images the edges are
B) finalLabel– This file is used for label binarizer. smoothened to remove any noisy and disturbances if
any that are present.
C) finalModel–This file is used for loading the images.
f. Training and testing dataset are created.
D) Main – This file consists of the code that is used to
run to detect the images. Training data take the input entry and gives the output by
creating the trained data.
E) myModel– This model is used for loading the model.
3. Training VGG16 model –
F) plant_disease_label_transform– This file is used for
loading the labels. The architecture (vgg16) deep learning models consist of 16
layers. Convolution layers have stride 1 and filer 3x3 and
The model is built by using the following procedure: consistently use zero padding and max pool layer have stride 2
1. We need to import required libraries, modules and the and filter 2x2. The last layer is dense layer. After the training
images are loaded from it– model of epochs is completed, we can only save the trained
model by finding the model's accuracy reached 100%.
Create a python file in jupyter and name it as main.py and
import required needed modules and libraries. Load the
required tomato images from the particular dataset.

VI-B) EXPERIMENT AND TEST SUIT RESULTS:

In our experiment and test suit, the images are loaded from the
dataset; the dataset contains 1 healthy and 9 diseased leaf
images. Any image can be loaded. We can see the loaded
image in the user interface. The model is already trained with
the related information that consists in the files and the load the
images related to the dataset, it verify the data and gives the
feedback according to the image dataset. The loaded image is
health, then it displays as healthy. If the loaded image is
2. Preprocessing the leaf disease dataset - diseased, it displays the disease name.

Pre-processing techniques are used due to the clash of working

with raw data for the system to easily understand. For this
model go through the following,
a. Reading the images – First the images are taken from
the dataset.
b. Resizing the images – The images taken from the
dataset that is each image in the dataset contains
different size and shape. In this step, it compresses
and reduces the size of each image into one particular
dimension. So that all the images contain same size.
From this it will be easy to classify.
c. Removing the noises in the images – After resizing
the images the images may contain some noisy data
like scratches in the image, blurred images and the
images with no clarity and so on. In this step, it
removes all such type of data. So that it will be easy to
classify the images.

978-1-6654-2867-5/21/$31.00 ©2021 IEEE 1801

Authorized licensed use limited to: KLE Technological University. Downloaded on November 19,2024 at 05:13:17 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Second International Conference on Electronics and Sustainable Communication Systems (ICESC-2021)
IEEE Xplore Part Number: CFP21V66-ART; ISBN: 978-1-6654-2867-5

After getting the output as whether the given leaf image is

healthy or diseased we will find the accuracy and loss.

The respective formula for accuracy and loss is

2. The following are the metrics for the leaf images. The first
one is healthy leaf of tomato and the second leaf is diseased
leaf of the tomato i.e., Tomato Early Blight. Here we can see
both color and gray images of the leaves. And the comparison
 Accuracy is defined as no. of images classified and difference in the colored and gray images can be seen.
correctly to the total no. of images. a. Tomato_ Healthy
 Loss is determined as the dissimilarity between the
value predicted by our model and the true value. The
formulas of accuracy and loss are represented above,
y i,j means the true value. 1 if the given sample
i be included in to the group j and otherwise it is 0.
p i,j represents the probability that our model
predicted of the given sample i be included in to the
category j.

VI-C) METRICS VALUES, RESULTS AND GRAPH b. Tomato_Early_Blight

1. The following are the metrics that are obtained for this
project.

3. The below table shows us the attributes values represented in

the table which used for calculating the accuracy of the VGG16
model that occurred at the time of each epoch.

978-1-6654-2867-5/21/$31.00 ©2021 IEEE 1802

Authorized licensed use limited to: KLE Technological University. Downloaded on November 19,2024 at 05:13:17 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Second International Conference on Electronics and Sustainable Communication Systems (ICESC-2021)
IEEE Xplore Part Number: CFP21V66-ART; ISBN: 978-1-6654-2867-5

References
[1] V. S. Babu, R. Satheesh Kumar, R. Sunder – “A
Epoch Train_loss Valid_loss Accuracy Error_rate Time comparative Study on Disease Detection of Plants Using
Machine Learning Techniques”, 2021 7th International
0 0.9598 0.0707 0.9997 0.0003 04:37 Conference on Advanced Computing and Communicating
1 0.9820 0.0380 0.9998 0.0002 04:36 Systems (ICACCS 2021)
2 0.9900 0.0639 0.9999 0.0001 04:36 [2] Yang Zhang, Chenglong Song, and Dongwen Zhang ,
3 0.9820 0.2217 0.9997 0.0003 04:37 “Deep Learning-Based Object Detection Improvement for
Tomato Disease”, IEEE, 2020
[3] Jun Sun, Yu Yang, Xiaofei He, and Xiaohong Wu,
“Northern Maize Leaf Blight Detection under Complex Field
4. The below graph is drawn from the above table. Environment Based on Deep Learning”, IEEE, 2020
[4] Peng Jiang, Yeuhan Chen, Bin Liu, Dongiian He,
Chunquan Liang, “Real-Time Detection of Apple Leaf
Diseases Using Deep Learning Approach Based on Improved
Convolutional Neural Networks”. IEEE, 2019
[5] Nikita Jadhav, HimaliKasar Prof. SumitaChandak, Shivani
Machha, “Crop Leaf Disease Diagnosis using Convolutional
Neural Network”, Published 2020 Biology International
Journal of Trend in Scientific Research and Development
[6] G. Geetha, S. Samundeswari, Saranya Gangadhara
Moorthy, Meenakshi K., M. Nithya – “Plant Leaf Disease
Classification and Detection System Using Machine
Learning” – December 2020 Journal of Physics Conference
Series (ICCPET 2020)
[7] M. Francis, C. Deisy, “Disease Detection and
Classification in Agricultural Plants Using Convolutional
VI. CONCLUSION & FUTURE WORK Neural Networks”, - A Visual Understanding Computer
Science 2019 6th International Conference on Signal
Processing and Integrated Networks (SPIN) 2019
VII-A) CONCLUSION: [8] S. Hari, M. Siva Kumar, P. Renuga, S. Karthikeyan, S.
The conclusion on distinct diseases categorization techniques Suriya, - “Detection of Plant Disease by Leaf Image Using
used for leaf disease detection and the algorithm used for the Convolutional Neural Network”, Computer Science 2019
image segmentation and pre-processing technique that is used International Conference on Vision Towards Emerging Trends
for automatic leaf diseases detection which is well known as in Communication and Networking (ViTECoN) 2019
classification. Apple and tomato types are used for testing the [9] Md Rahmat Ullah, Nagifa Anjum, Abdus Sattar – “Plant
suggested algorithms. Therefore, allied diseases are taken Disease Recognition Using Machine Learning”, 2019 8th
from the plants for recognition. Estimating the slighter efforts International Conference System Modeling and Advancement
of the superlative outcome which are produced, it expresses in Research Trenda (SMART 2019)
the competence of the suggested algorithm to identify and [10] MelikeSardogan, AdemTuncer, YunnusOzen, “Plant Leaf
classification for the tomato leaf diseases. To provide better Disease Detection and Classification Based on CNN with
perform [16] for identification rate in the categorization LVQ Algorithm”, 3rd International Conference on Computer
process convolutional algorithms can also be used. Science and Engineering (UBMK), Posted 2018
[11] Shima Ramesh Maniyath, Vinod P V, Niveditha M, P. R,
P. N, Shashank N, R.Hebbar – “Plant Disease Detection Using
VII-B) FUTURE WORK:
Machine Learning” – 2018 International Conference on
Our future scope is to create a mobile app which will be useful Design Innovations for 3C’s Compute Communicate Control
for the farmers as a proper guide to do agriculture. In future, (ICDI3C)
we can do the disease detection techniques using various parts [12] P. Maheswari, P. Raja, N. M. Ghangoankar – “Intelligent
of the crops or plants like stem, flower, and root. The project Disease Detection System for Early Blight of Tomato Using
mainly focuses on the enlargement the given [10] Foldscope: A Pilot Study, 2018 IEEE 4th International
methodology used to recognition and grouping for tomato leaf Symposium in Robotics and Manufacturing Automation
plant disease symptoms affected on leaves. However, the (ROMA)
systems have a poor performance in Leaf miner and Septoria [13] Yow-Wen Tian, Peng-Hui Zheng, Rui-Yao Shi, “The
leaf spot. Future works may include enhancing the visual Detection System for Greenhouse Tomato Disease Degree
characterization and an additional evaluation . Based on Android Platform”, 3rd International Conference on
Information Science and Control Engineering, Posted: 2016

978-1-6654-2867-5/21/$31.00 ©2021 IEEE 1803

Authorized licensed use limited to: KLE Technological University. Downloaded on November 19,2024 at 05:13:17 UTC from IEEE Xplore. Restrictions apply.
 Proceedings of the Second International Conference on Electronics and Sustainable Communication Systems (ICESC-2021)
IEEE Xplore Part Number: CFP21V66-ART; ISBN: 978-1-6654-2867-5

[14] C Ravi, Jibu Shinde, C Mathew, Prof C Y Patil, [16] U. Mokhtar, Mona A. S. Ali, Aboul Ella Hassenian, H.
“Segmentation Technique for Soybean Leaves Disease Hefny – “Tomato Leaves Diseases Detection Approach Based
Detection”, International Journal of Advanced Research, on Support Vector Machines ”, 2015 11th International
volume 3, issue 5, Posted: 2015 Computer Engineering Conference (ICENCO 2015)
[15] M Ramakrishna and A SahayaAnselin Nisha, “Groundnut
Leaf Disease Detection and Classification by using Nack
Propagation Algorithm”, IEEE International Conference,
Melmaruvathur 2015

978-1-6654-2867-5/21/$31.00 ©2021 IEEE 1804

Authorized licensed use limited to: KLE Technological University. Downloaded on November 19,2024 at 05:13:17 UTC from IEEE Xplore. Restrictions apply.