A Report on

DEEP LEARNING-ASSISTED FRAMEWORK FOR PLANT DISEASE

IDENTIFICATION WITH CNN MODELS

Submitted in partial fulfillment of the requirements for the award of the degree of
BACHELOR OF TECHNOLOGY
In
ELECTRONICS AND COMMUNICATION ENGINEERING
By
G. HARINI 21001A0443
T. SAI JAHNAVI 21001A0446
M. DEEPSIKA REDDY 21001A0433
I. ESWAR NAIK 21001A0441
N. HARIKA 21001A0439

Under the Esteemed guidance of

MR. M. SREEDHAR M. Tech, (Ph.D.)

Assistant Professor Adhoc,

Department of E.C.E

Submitted to

DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING

JNTUA COLLEGE OF ENGINEERING(Autonomous) ANANTAPUR
JAWAHARLAL NEHRU TECHNOLOGICAL UNIVERSITY
ANANTAPUR-515002
ANDHRA PRADESH
2024-2025
 JAWAHARLAL NEHRU TECHNOLOGICAL UNIVERSITY
ANANTHAPURAMU-515002A.P., INDIA.

DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING

CERTIFICATE

This is to certify that the project work entitled “A Deep Learning-Assisted Framework
for Plant Disease Identification with CNN Models”, being submitted in partial fulfillment of
requirements for the award of degree of “Bachelor of Technology” in Electronics and
Communication Engineering, Jawaharlal Nehru Technological University College of
Engineering Ananthapuramu and the bonafide record of project work carried out under my
supervision. The result provide in this report have not been submitted to any other University or
Institute for the award of any Degree

G. HARINI 21001A0443
T. SAI JAHNAVI 21001A0446
M. DEEPSIKA REDDY 21001A0433
I. ESWAR NAIK 21001A0441
N. HARIKA 21001A0439

Signature of Project Guide Signature of Head of the Department

Mr. M. SREEDHAR, M.Tech, (Ph.D.)., Dr. G.MAMATHA M.Tech,Ph.D.,

Assistant Professor Adhoc Head & Associate Professor

Department of E.C.E. Department of E.C.E.

J.N.T.U.A College of Engineering J.N.T.U.A College of Engineering

Ananthapuramu-515002 Ananthapuramu-515002
 DECLARATION
We here by declare that the dissertation entitled “A DEEP LEARNING-

ASSISTED FRAMEWORK FOR PLANT DISEASE IDENTIFICATION WITH

CNN MODELS ” submitted by us to JNTUA College of Engineering, Anantapur in

partial fulfillment of the requirement of the degree of Bachelor of Technology (B.Tech.)

in Electronics and Communication Engineering (ECE) is a record of project work carried

out under guidance of Mr. M. SREEDHAR, Assistant Professor Adhoc, Department of

Electronics and Communication Engineering.

We further declare that the work reported by this project has been submitted and

will not be submitted, either in part or full, for the award of any other degree or diploma in

this institute or any other institute or university.

Place: Anantapur
Date:

G. HARINI 21001A0443
T. SAI JAHNAVI 21001A0446
M. DEEPSIKA REDDY 21001A0433
I. ESWAR NAIK 21001A0441
N. HARIKA 21001A0439
 ACKNOWLEDGEMENT
The satisfaction and euphoria that accompany successful completion of any task would
be incomplete without the mention of people who made it possible, whose constant guidance
and encouragement crowned our efforts with success.

We are highly indebted to Mr. M. SREEDHAR, Assistant Professor of Electronics and

Communication Engineering Department, JNTUA College of Engineering, Anantapur for his
guidance and constant supervision as well as for providing necessary information regarding the
project. His support has helped us in the completion of the project.
We would like to express special thanks to Dr. G. MAMATHA, Head of the
Department, Electronics and Communication Engineering JNTUA College of Engineering,
Anantapur for her wide support, knowledge and enthusiastic encouragement has impressed us
to better involvement into our project and technical design.
We want to express our gratitude to Prof. P. CHENNA REDDY, Principal & Professor
of Computer Science Engineering in JNTUA College of Engineering Anantapuramu, for his
co-operation and his timely help in the successful completion of the project. We express our
sincere thanks to the project committee members, faculty and staff of Electronics and
Communication Engineering, JNTUA College of Engineering, Anantapur, for their valuable
guidance and technical support. Last but far from least, we also thank our family members and
our friends for their moral support and constant encouragement. We are very thankful to one
and all who helped us for the successful completion of the project.

With Gratitude

G. HARINI 21001A0443
T. SAI JAHNAVI 21001A0446
M. DEEPSIKA REDDY 21001A0433
I. ESWAR NAIK 21001A0441
N. HARIKA 21001A0439
 ABSTRACT
The rapid spread of plant diseases continues to be a major challenge in modern
agriculture, contributing to significant reductions in crop yield and quality. Early and
accurate diagnosis of plant diseases is critical to ensure timely intervention and
sustainable crop management. This project proposes an intelligent, automated system for
plant disease detection and affected area estimation by leveraging deep learning and
computer vision techniques. A lightweight yet powerful Convolutional Neural Network
(CNN) based on the MobileNetV2 architecture was trained on the extensive PlantVillage
dataset, encompassing 43,444 images across 38 distinct classes. The model achieved a
remarkable training accuracy of 98.02% and a validation accuracy of 97.82%, confirming
its effectiveness in identifying a wide range of plant diseases.

To further enhance the system’s diagnostic utility, a complementary image

processing module was integrated to assess the severity of infection. Using HSV color
space transformation, thresholding, and contour detection via Open CV, the model isolates
diseased regions and calculates the percentage of the leaf area that is affected. The final
output provides users with both qualitative (disease classification) and quantitative
(severity percentage) results, delivering a holistic evaluation of plant health. This dual-
function system not only empowers farmers with precise diagnostic insights but also
promotes data-driven decision-making in precision agriculture. The proposed solution is
scalable, adaptable for real-time deployment, and represents a significant advancement
toward intelligent and sustainable crop protection.
 TABLE OF CONTENTS

1 INTRODUCTION

1.1 Introduction 1

1.2 Problem Statement 4

1.3 Objectives 5

2 LITERATURE SURVEY 7

3 EXISTING METHODS

3.1 Manual Inspection 9

3.2 Classical Image Processing Techniques 9

3.3 Machine Learning Based Approaches 10

3.3.1 Support Vector Machines (SVM) 10

3.3.2 K Nearest Neighbors (K-NN) 11

3.3.3 Random Forest 11

3.3.4 Decision Trees (DT) 12

3.3.5 Naive Bayes 12

3.4 Limitations Of the Study 13

4 METHODOLOGY

4.1 Block Diagram 15

4.2 Data Augmentation Techniques Used 17

4.3 Software Implementation 18

5 RESULTS AND DISCUSSIONS

5.1 Results 20

5.2 Representation of Data 24

6 CONCLUSION AND FUTURE SCOPE

6.1 Conclusion 26

6.2 Future Scope 28

7 REFERENCES 30
 LIST OF FIGURES

Figure No Figure Name Page No

1 Bacterial Blight 2

2 Rust 2

3 Powdery Mildew 3

4 Browning 3

5 Necrosis 3

6 Block Diagram of plant Disease Detection 16

7 Healthy Grape Image 23

8 Diseased Apple Image 24

9 Graph of Training Accuracy and Validation 25

Accuracy
 LIST OF TABLES

Table No Table Name Page No

1 Model Training and Validation Summary 26

2 Area Estimation results (Sample Outputs) 27

 Plant Disease Detection
CHAPTER – 1
INTRODUCTION

1.1 INTRODUCTION

Agriculture plays a crucial role in the sustenance of human life and global economic
development. However, one of the major challenges in agriculture is the timely identification and
control of plant diseases, which significantly impact crop yield, quality, and farmer income.
Traditionally, plant disease detection relies on manual inspection, which is subjective, slow, and often
inaccurate due to limited expertise and environmental factors. The need for a reliable, automated
system has led to increased interest in artificial intelligence and computer vision techniques for solving
this problem.

This project focuses on developing a comprehensive system for plant disease detection and
affected area estimation using deep learning and image processing techniques[1]. The goal is to classify
plant leaf images into their respective disease categories and to quantify the percentage of the leaf area
that is affected bythe disease — providing bothdiagnosis and severityestimation.

The project began by downloading and preparing the Plant Village dataset, which contains
over 43,000 labeled images of healthy and diseased leaves across 38 classes. The dataset was
organized into training and validation folders. Image preprocessing steps such as resizing to a uniform
shape (128×128 or 64×64), normalization, and augmentation were applied to increase model
generalization and reduce overfitting.

A Convolutional Neural Network (CNN) was then designed and trained using TensorFlow and
Keras. Later, the CNN was enhanced using MobileNetV2, a pre-trained model, to achieve faster and
more accurate classification with fewer computational resources[1]. The training process
yielded excellent performance, reaching 98.02% training accuracy and 97.82% validation
accuracy after several epochs. Data was fed using TensorFlow’s image_dataset_from_directory
method, and overfitting was minimized through real-time augmentation and caching.

After successfully classifying diseases, the next major component was to calculate the
percentage of diseased area on the leaf. This was achieved through HSV-based image segmentation
using OpenCV[2]. Diseased regions were highlighted by defining color thresholds and
applying morphological operations to clean up noise. The contours of the diseased parts were
extracted, and their area was compared with the total image area to estimate the severity percentage.

DEPARTMENT OF ECE, JNTUACEA 1

 Plant Disease Detection
Throughout the project, Jupyter Notebook was used as the development environment, and
key Python libraries such as TensorFlow, OpenCV, NumPy, and Matplotlib were utilized. The
trained model was saved in both .h5 and .keras formats for reuse, and additional features such as
prediction functions and image visualization were developed for ease of testing.

By combining image classification with visual area analysis, this project offers a powerful
tool for farmers, agronomists, and researchers[3]. It not only identifies the disease accurately but
also informs the user of how much of the plant is affected, enabling early diagnosis, better
decision- making, and targeted treatment. This system can be further extended into mobile or real-
time IoT- based platforms for broader agricultural use.

1.1.1 PLANT DISEASES

Figure 1.1 - Bacterial Blight

Bacterial blight is a plant disease caused by various bacterial pathogens that leads to
symptoms like leaf spots, wilting, and stem lesions, ultimately affecting the plant's health
and yield. The bacteria can spread through water, wind, insects, and contaminated tools,
thriving in favorable environmental conditions like high humidity and warm temperatures.

Figure 1.2- Rust

Rust is a fungal disease characterized by the appearance of powdery pustules in

shades of yellow, orange, red, brown, or black on various plant parts, most commonly
leaves and stems. These rust-like spots are masses of spores that can hinder photosynthesis
and overall plant health, sometimes leading to reduced vigor or premature leaf drop.

DEPARTMENT OF ECE, JNTUACEA 2

 Plant Disease Detection

Figure 1.3- Powdery Mildew

Powdery mildew is a common fungal diseasecharacterized by a white or grayish

powdery growth on the surface of leaves, stems, and sometimes flowers and fruits. While
often considered a cosmetic issue, severe infections can lead to stunted growth, yellowing,
and premature leaf drop.

Figure 1.4- Browning

Browning in plants is a common symptom of various diseases, often indicating
tissue death due to fungal, bacterial, or viral infections, or even environmental stresses. This
discoloration can manifest on leaves, stems, or roots, and its pattern and progression can
help in diagnosing the underlying cause.

Figure 1.5- Necrosis

Necrosis in plants describes the localized death of plant cells and tissues, often
appearing as browning or blackening on leaves, stems, or other parts. This cell death can be
triggered by various factors including pathogen infections, nutrient deficiencies, and
environmental stresses.

DEPARTMENT OF ECE, JNTUACEA 3

 Plant Disease Detection

1.2 PROBLEM STATEMENT

Agriculture, being one of the most vital sectors for global sustenance and economic
development, is constantly threatened by plant diseases. These diseases not only reduce
crop yield but also compromise the quality of agricultural produce, leading to substantial
economic losses and food insecurity[1]. The early detection and accurate diagnosis of such
diseases are essential to prevent widespread damage and ensure timely intervention.
However, in current farming practices, disease identification is primarily based on manual
visual inspection, which is highly subjective, time-consuming, and dependent on the
experience and expertise of the individual performing the inspection.
Moreover, traditional approaches often fail to provide a quantitative analysis of the
disease's severity, which is equally important for deciding the intensity and scope of
treatment. Simply knowing whether a plant is diseased is not sufficient; farmers also need
to understand how much of the plant is affected, which helps in determining whether a plant
can be salvaged, how much pesticide or fungicide should be applied, or whether isolation
is necessary to prevent further spread[2].
In recent years, various machine learning and deep learning techniques have
emerged for disease classification. While many studies have demonstrated the success of
Convolutional Neural Networks (CNNs) in identifying plant diseases with high accuracy,
very few systems provide the additional capability of estimating the affected area of the
plant. This limits their practical utility in real-world agricultural decision-making[3].
This project addresses the gap by proposing an integrated solution that not only
classifies the disease present on a plant leaf using a deep learning model trained on the
PlantVillage dataset but also uses image processing techniques (such as color space
transformation and morphological operations) to calculate the percentage of the leaf area
that is infected. This combination of classification and quantification delivers a
comprehensive plant health assessment tool that is both scalable and efficient.
The system aims to reduce the dependency on manual intervention, increase
diagnostic accuracy, and enable precise and timely disease management. This contributes
significantly to smart farming practices, helping farmers reduce losses, optimize resource
usage, and improve overall crop productivity[4].

DEPARTMENT OF ECE, JNTUACEA 4

 Plant Disease Detection

1.3 OBJECTIVES

The primary objective of this study is to design and implement a comprehensive,

AI-driven system capable of detecting plant diseases from leaf images and calculating the
severity of infection by estimating the percentage of the leaf area affected. In support of
this overarching goal, the study is guided by the following specific objectives:

To Develop an Automated Plant Disease Detection System

The study aims to build a reliable and scalable system that leverages deep learning
techniques to automatically detect and identify various plant diseases. This reduces the
dependency on manual inspection and enables early intervention, which is critical in
minimizing crop damage[5].

 To Train a CNN Model on the PlantVillage Dataset

The objective is to use a publicly available and extensively curated dataset—

PlantVillage—which includes over 43,000 images across 38 plant disease categories. A
Convolutional Neural Network (CNN) architecture, including pre-trained models like
MobileNetV2, is employed to improve the model's accuracy and computational efficiency.

 To Implement Preprocessing Techniques for Improved Model Performance

Proper image preprocessing is essential for robust model training. This includes
resizing all input images to a uniform dimension (e.g., 128×128), normalizing pixel values,
and applying data augmentation techniques to increase dataset variability and reduce
overfitting.

 To Estimate the Percentage of Leaf Area Affected by Disease

Beyond classification, the study seeks to quantify the extent of infection by

calculating how much of the leaf is actually affected. This is accomplished through image
processing techniques such as converting images to HSV color space, applying
thresholding to segment diseased regions, and using morphological operations to remove
noise. The ratio of diseased area to total leaf area is then computed[7].

 To Provide Visual Interpretation of Disease Spread

Another key objective is to enhance interpretability by overlaying contours or

masks on the original leaf image to visually display the diseased regions. This allows users
not only to receive a diagnosis but also to understand the physical spread of the disease.

DEPARTMENT OF ECE, JNTUACEA 5

 Plant Disease Detection

 To Evaluate Model Accuracy and Generalization

The system's effectiveness is validated through performance metrics such as

training and validation accuracy, loss curves, and confusion matrices. The goal is to achieve
high accuracy while ensuring that the model generalizes well across unseen data, avoiding
overfitting[8].

 To Create a Foundation for Scalable Agricultural Applications

The final objective is to design the system in a way that it can be extended into real-
world applications. This includes the potential integration into mobile apps or IoT
platforms for on-field diagnosis, making it a practical tool for farmers, agronomists, and
agricultural extension officers.

These objectives collectively contribute to building a smart, accurate, and user-

friendly solution for plant disease monitoring—ultimately promoting precision agriculture
and helping mitigate the impact of crop diseases on food production and the economy.

DEPARTMENT OF ECE, JNTUACEA 6

 Plant Disease Detection
CHAPTER-2
LITERATURE SURVEY

2.1 INTRODUCTION:

In recent years, the integration of artificial intelligence and image processing in

agriculture has become a significant area of research, especially for applications such as
plant disease detection. Traditional approaches to detecting plant diseases primarily relied
on machine learning algorithms like Support Vector Machines (SVM), k-Nearest Neighbors
(k-NN), and Decision Trees. These methods typically depended on manually extracted
features such as color, texture, and shape descriptors derived from leaf images. While these
approaches showed reasonable success, their performance was often hindered by variations
in lighting conditions, leaf orientation, and background noise, making them less reliable in
practical agricultural environments.
The emergence of deep learning, particularly Convolutional Neural Networks
(CNNs), revolutionized the field of image classification and provided a powerful alternative
to manual feature extraction. CNNs are capable of learning and generalizing complex
patterns directly from image data. In a notable study, researchers trained CNN models on
the PlantVillage dataset and achieved classification accuracies above 99% for multiple crops
and disease categories. These results established CNNs as a state-of-the-art solution for
plant disease detection and encouraged further exploration using deeper and more complex
architectures such as VGG16, ResNet, and InceptionNet, which further enhanced accuracy
and robustness.
With the growing need for real-time and resource-efficient solutions, the use of
lightweight, pre-trained models such as MobileNet, EfficientNet, and DenseNet became
increasingly popular. These models use transfer learning to adapt knowledge from large
datasets like ImageNet to plant disease datasets, reducing the need for extensive
computational resources while maintaining high performance. MobileNetV2, in particular,
has been widely used in mobile and edge computing applications, making it highly suitable
for deployment in agricultural settings with limited infrastructure.
Despite the advances in disease classification, most existing studies have overlooked
the importance of quantifying disease severity. Understanding how much of
the leaf is affected is essential for making informed decisions regarding treatment intensity
and resource allocation.

DEPARTMENT OF ECE, JNTUACEA 7

 Plant Disease Detection
segmentation using U-Net and Mask R-CNN, have also been explored, but their relianceon
pixel-level annotated data poses challenges for scalability.

The lack of integration between classification and severity analysis presents a clear
gap in the existing literature. Most systems provide only a disease label without indicating
the extent of damage, which limits their practical utility. This project addresses that
limitation by combining a CNN-based classification model with image processing
techniques to segment and calculate the affected area of the leaf. By doing so, it provides
not just a diagnosis, but also an estimate of disease progression, thus offering a more
comprehensive and actionable tool for modern agricultural practices.

DEPARTMENT OF ECE, JNTUACEA 8

 Plant Disease Detection

CHAPTER-3
EXISTING METHODS
Plant disease detection has traditionally relied on manual and semi-automated methods,
many of which present significant limitations in terms of accuracy, speed, and scalability.
Below is a summary of key existing approaches used prior to the emergence
of deep learning:

3.1 MANUAL INSPECTION:

Manual inspection is the oldest and most widely practiced method for detecting plant
diseases. It involves visual examination by farmers or agricultural experts who rely on
experience and knowledge to identify symptoms such as spots, discoloration, curling, or
wilting. While it may offer reasonably good detection in familiar scenarios, the accuracy is
highly dependent on the observer’s skill, environmental conditions, and clarity of the
symptoms. This method is especially unreliable for early-stage diseases or subtle symptoms,
which can be easily overlooked[9].
Moreover, manual diagnosis is time-consuming and labor-intensive, making it
unsuitable for large-scale or commercial farming operations. Inconsistent results across
different observers, fatigue, and human error further reduce the reliability of this approach.
With the increasing demand for real-time and precision farming, manual inspection is
proving insufficient as it lacks scalability, speed, and repeatability. Hence, modern solutions
are moving toward automation and machine intelligence[10].

3.2 CLASSICAL IMAGE PROCESSING TECHNIQUES

Classical image processing methods brought initial automation to plant disease
detection. These approaches use pixel-based operations such as color space conversion,
thresholding, edge detection, and morphological transformations. For instance, color
thresholding in RGB or HSV spaces helps to segment out areas of discoloration, while edge
detection algorithms like Canny or Sobel outline lesion boundaries. Morphological
operations are then used to clean and refine the segmented regions[11].
However, these techniques rely on fixed rules and thresholds, making them
extremely sensitive to changes in lighting, background clutter, and leaf orientation. The
inability to adapt to new environments limits their effectiveness in real-world agricultural
settings. While useful in controlled laboratory conditions, classical image processing fails
to generalize across plant species and varying symptoms, making them less viable for
practical field deployment[12].

DEPARTMENT OF ECE, JNTUACEA 9

 Plant Disease Detection

3.3 MACHINE LEARNING-BASED APPROACHES

Traditional machine learning methods improved upon image processing by
introducing automated classification using extracted features. The workflow usually
includes:Manual feature extraction (e.g., color histograms, texture features like GLCM, or
shape descriptors).
Classification using algorithms such as:
 Support Vector Machines (SVM)
 k-Nearest Neighbors (k-NN)
 Random Forests
 Decision Trees (DT)
 Naive Bayes
While these methods showed moderate accuracy and offered some level of
automation, their performance heavily depended on domain knowledge for feature
engineering. Furthermore, they could not handle large datasets efficiently or learn complex
patterns without human intervention.

3.3.1 SUPPORT VECTOR MACHINES (SVM)

Support Vector Machines (SVM) are one of the most popular supervised learning
algorithms used for classification tasks. In the context of plant disease detection, SVMs
classify images based on extracted features such as color, texture, and shape. The algorithm
works by finding an optimal hyperplane that separates different classes of data with the
maximum margin. SVMs are especially effective in high-dimensional spaces and are known
for their robustness in cases where the number of dimensions exceeds the number of
samples[1].
However, the performance of SVMs is heavily dependent on the quality of the
feature extraction process. Since SVMs do not learn features on their own, domain
knowledge is required to select appropriate input variables. Additionally, selecting the right
kernel function (linear, polynomial, RBF, etc.) and tuning hyperparameters is critical for
accurate results. While SVMs can yield good accuracy on well-preprocessed datasets, they
are not as scalable or adaptable as deep learning models when dealing with complex and
large image datasets[13].

3.3.2 K-NEAREST NEIGHBORS (K-NN)

k-Nearest Neighbors (k-NN) is a simple, instance-based learning algorithm that classifies
an image based on the majority class among its 'k' closest neighbors in the feature space.
For plant disease detection, once features like leaf color or texture are extracted, the k-NN
DEPARTMENT OF ECE, JNTUACEA 10
 Plant Disease Detection
algorithm calculates distances (usually Euclidean) between the test image and training
samples to determine its class. Its simplicity and interpretability make it a common baseline
for many classification tasks[14].
Despite its ease of implementation, k-NN has several drawbacks. It is
computationally expensive during the prediction phase because it must calculate distances
from all training samples for each new input. Moreover, its performance deteriorates in
high-dimensional spaces (known as the “curse of dimensionality”), and it is sensitive to
noisy or irrelevant features. In practical plant disease detection scenarios where real-time
processing is required, k-NN is often too slow and memory-intensive to be viable.

3.3.3 RANDOM FORESTS

Random Forests are an ensemble learning technique that combines the predictions
of multiple decision trees to improve classification accuracy and prevent overfitting. Each
tree is trained on a random subset of the data and features, and the final prediction is made
based on a majority vote. In plant disease detection, Random Forests have been used to
classify diseases based on feature vectors derived from image preprocessing techniques.
While Random Forests generally offer better accuracy and stability than individual
decision trees, they still rely on manually extracted features and lack the ability to learn
spatial hierarchies from raw image data. Furthermore, as the number of trees increases, the
model can become computationally intensive, particularly during prediction. Although
effective on structured datasets, Random Forests are less suitable for large-scale image
classification tasks compared to deep learning models, which inherently learn features from
data.

3.3.4 DECISION TREES (DT)

Decision Trees are hierarchical models that make decisions by splitting the data
based on specific feature thresholds. Each node in the tree represents a condition, and
branches represent the outcome of that condition, leading to a leaf node that represents a
class label. For plant disease detection, decision trees can be used to classify input features
such as texture contrast, mean pixel intensity, or shape descriptors into disease categories.
The primary advantage of Decision Trees is their interpretability and speed. They
are easy to visualize and understand, making them useful for initial experimentation.
However, they tend to overfit training data, especially with complex datasets, unless pruning
techniques are applied. Additionally, decision trees are sensitive to small variations in the
data and can produce different models with minor changes. As a result, while they serve as
a good baseline, they are rarely used as standalone models in modern, high-accuracy
DEPARTMENT OF ECE, JNTUACEA 11
 Plant Disease Detection
applications like plant disease diagnosis.

3.3.5 NAIVE BAYES

Naive Bayes classifiers are probabilistic models based on Bayes’ Theorem,
assuming independence among predictors. Despite this “naive” assumption, they perform
surprisingly well in many classification problems. In plant disease detection, the model
estimates the probability that a set of image features belongs to each disease class and
chooses the class with the highest posterior probability.
Naive Bayes is highly efficient and performs well on small datasets with clearly
separated features. However, its assumption of feature independence often doesn’t hold true
for image data, where pixel intensities and textures are interdependent. This limits its
effectiveness for complex visual tasks. Moreover, it struggles with large and noisy datasets,
which are common in real-world agricultural environments. As a result, Naive Bayes is
mostly used for academic purposes or as a comparative baseline rather than a production-
level classifier in image-based plant disease detection.

3.4 LIMITATIONS OF THE STUDY

 Controlled Dataset Environment
The model was trained on the PlantVillage dataset, which consists of images
captured in lab-like conditions with uniform backgrounds and lighting. This limits the
model’s ability to generalize to field images with varied lighting, complex backgrounds,
and environmental noise.
 Sensitivity to Shadows and Color Variations
The HSV-based image processing method used for estimating the affected area can
misclassify shadows, natural pigmentation, or moisture marks as diseased regions, leading
to inaccurate area calculations.
 Limited Segmentation Accuracy
The current approach does not use advanced segmentation techniques like U-Net or
Mask R-CNN, which could provide more precise disease region extraction. This limits the
accuracy of severity estimation in complex cases.
 Resource and Hardware Constraints
Although MobileNetV2 is lightweight, training and prediction can still be
computationally demanding, especially when deployed on low-end or mobile devices
without proper optimization.

DEPARTMENT OF ECE, JNTUACEA 12

 Plant Disease Detection

 No Temporal Disease Tracking

The model processes images individually and does not support analysis over time. This
limits its capability to monitor disease progression or the effectiveness of treatments.
 Limited to Known Classes
The model can only classify diseases that were included in the training data. It may not
recognize or correctly predict new or rare diseases not represented in the dataset.
 No Integration with External Agricultural Data
The system does not currently incorporate environmental parameters like temperature,
humidity, or soil data, which could enhance disease prediction and contextual decision-
making.

DEPARTMENT OF ECE, JNTUACEA 13

 Plant Disease Detection

CHAPTER - 4
METHODOLOGY
The methodology adopted in this project is structured to systematically address the
dual objectives of disease classification and diseased area estimation from plant leaf images.
The approach integrates deep learning for disease identification and image processing
techniques for quantifying the extent of infection. The entire process can be divided into
several key stages, as detailed below.
The first step involves data acquisition and preprocessing. The publicly available
PlantVillage dataset was used as the primary source for model training. This dataset contains
over 43,000 images of plant leaves categorized into 38 distinct classes, representing both
healthy and diseased conditions. The images were organized into training and validation
directories. To prepare the images for input into the model, preprocessing steps such as
resizing (to 128×128 or 64×64 pixels), normalization (scaling pixel values between 0 and
1), and data augmentation (including rotation, flipping, and zooming) were performed.
These steps ensured consistency across the dataset and helped enhance the model's ability
to generalize.
Following preprocessing, a Convolutional Neural Network (CNN) was implemented
for plant disease classification. Initially, a custom CNN architecture was trained, and later
the performance was optimized using MobileNetV2, a lightweight pre- trained model
known for its efficiency and accuracy. Transfer learning was employed by loading the base
MobileNetV2 model without its top classification layers and fine-tuning it on the
PlantVillage dataset. The model was compiled using the RMSprop optimizer, categorical
crossentropy loss function, and accuracy as the evaluation metric. It was trained over
multiple epochs until it achieved a training accuracy of 98.02% and validation accuracy of
97.82%. TensorFlow’s image_dataset_from_directory function was used to streamline the
loading and batching of image data, with caching and prefetching applied to boost training
efficiency.
Once the classification component was successfully developed and evaluated, the
project proceeded to the second major task: estimating the affected area of the diseased leaf.
For this purpose, image processing techniques using OpenCV were applied. The input image
was first converted from BGR to HSV color space, which is more effective for isolating
specific color ranges associated with diseased regions. A lower and upper HSV threshold
was defined to segment the infected areas based on color. Morphological operations such as
erosion, dilation, and closing were applied to remove noise and refine the segmented
regions. Using contour detection, the system identified distinct diseased patches and
calculated the sum of their areas.
DEPARTMENT OF ECE, JNTUACEA 14
 Plant Disease Detection
To estimate the percentage of the affected area, the total pixel area of all detected
contours was divided by the total pixel area of the leaf image and multiplied by 100. This
gave an accurate percentage value representing the extent of disease spread on the leaf. To
make the results more intuitive, the diseased regions were visually highlighted on the image
using contours, enabling users to verify the segmented output.
The final system includes both disease prediction and area estimation functionalities.
The user can input any image of a plant leaf from the dataset, and the system will output the
predicted disease name along with a visualization and percentage of the affected area. All
components were developed and tested using Python in Jupyter Notebook, with key libraries
including TensorFlow, OpenCV, NumPy, and Matplotlib.
In summary, the methodology adopted in this project effectively combines the
predictive power of deep learning with the analytical capability of image processing. This
two-stage approach offers a more comprehensive plant disease monitoring solution,
capable of supporting real-time diagnosis and facilitating better decision-making in
agricultural practices.

4.1 BLOCK DIAGRAM:

The block diagram below illustrates the complete workflow of the plant disease
detection and affected area estimation system. It outlines the sequential steps taken from
input image acquisition to the final output of disease type and infected area percentage. Each
block represents a major component of the system implemented through a combination of
deep learning and image processing techniques.
 Input Leaf Image: A digital image of a plant leaf is provided by the user.
 Image Preprocessing: The image is resized to a fixed dimension (e.g., 128x128),
normalized, and converted to a format suitable for model input.
 CNN-Based Classification: The processed image is passed through a trained
MobileNetV2-based CNN model to classify the leaf into one of the 38 possible disease
categories.
 Predicted Disease Class: The model outputs the most likely disease label for the given
leaf.
 Image Processing using OpenCV: The same image undergoes HSV transformation and
color thresholding to isolate potentially infected areas.
 Contour Detection & Area Estimation: The system calculates the area of detected
contours and computes the percentage of diseased area relative to the total leaf size.
 Visual Output: The system overlays contours on the original image and displays the
percentage of affected area along with the predicted disease label.
DEPARTMENT OF ECE, JNTUACEA 15
 Plant Disease Detection

Input Leaf Image

Image Preprocessing (Resize,

Normalize, etc)

CNN Based Classification

(MobileNet, V2 Model)

Predicted Disease Class

Image Processing
-Convert to HSV
-Color Thresholding
-Morphological Operations

Contour Detection and Area

Estimation
(Calculate % of infected)

Visual Output
(Original + Contour Image)

Figure 4.1- Block Diagram Of Plant Disease detection

DEPARTMENT OF ECE, JNTUACEA 16

 Plant Disease Detection

4.2 DATA AUGMENTATION TECHNIQUES USED

To enhance the robustness and generalization ability of the plant disease classification
model, various data augmentation techniques were applied during training. Data
augmentation artificially increases the size and variability of the training dataset by applying
transformations to existing images. This helps the model learn to recognize disease patterns
across different orientations, lighting conditions, and leaf positions—conditions that are
commonly encountered in real-world agricultural fields.
The augmentation processes used in this project include:
 Rotation
Images were randomly rotated within a certain range (e.g., ±30 degrees) to simulate
different leaf orientations. This ensures the model remains invariant to the direction in
which a leaf appears.
 Zooming
Random zoom was applied to simulate varying distances between the camera and the
plant. This helps the model learn features at different scales and resolutions.
 Width and Height Shifts
Horizontal and vertical translations were used to shift the image slightly along the x
and y axes. This helps the model handle off-centered leaves or misaligned frames.

 Horizontal Flipping
Random horizontal flipping was used to reflect the image along the vertical axis. Since
leaf structures are often symmetrical, this technique is effective in doubling the dataset
variation without changing the disease characteristics.
 Brightness Adjustment
Minor changes in image brightness were applied to simulate different lighting
conditions such as natural daylight or shaded environments. This ensures that the model
does not misclassify based on lighting alone.
 Rescaling
All images were rescaled by dividing pixel values by 255 to normalize them to a range
between 0 and 1. This improves convergence during training and standardizes input data.
These augmentation strategies not only improve model generalization and accuracy,
but also significantly reduce overfitting by ensuring that the model does not memorize
specific patterns or features from the original dataset. This is particularly valuable in
agricultural scenarios where input conditions can vary widely in the field.

DEPARTMENT OF ECE, JNTUACEA 17

 Plant Disease Detection

4.3 Software Implementation

The software implementation of this project involves the integration of deep learning
for plant disease classification and image processing techniques for quantifying the diseased
area on plant leaves. The implementation was carried out in a modular and systematic
manner, ensuring clarity, maintainability, and scalability of the code. The entire
development was conducted using Python in the Jupyter Notebook environment, which
allowed for interactive coding, visualization, and testing.
The first phase of implementation was focused on dataset preparation and
preprocessing. The PlantVillage dataset, which contains over 43,000 labeled images across
38 classes, was downloaded, extracted, and organized into train and val directories. Each
class folder contained corresponding disease or healthy leaf images. File paths were loaded
using Python’s os and pathlib libraries, and image augmentation techniques such as rotation,
zooming, flipping, and contrast adjustments were applied using TensorFlow’s
ImageDataGenerator and image_dataset_from_directory functions. This helped to
artificially expand the dataset and make the model robust against overfitting.
Next, the model training module was implemented using the TensorFlow and Keras
libraries. Initially, a basic CNN was constructed with convolutional, pooling, and dense
layers. However, to improve performance and speed, the project transitioned to using
MobileNetV2, a pre-trained lightweight model suitable for resource-constrained
environments. Transfer learning was employed by loading MobileNetV2 without the top
classification layer and adding custom dense layers tailored for the 38-class problem. The
model was compiled with the RMSprop optimizer and categorical_crossentropy as the loss
function, and trained for several epochs with real-time training accuracy and loss
visualization using Matplotlib. After training, the model was saved in both .keras and .h5
formats for future inference.
Following successful classification training, attention shifted to disease area
estimation. This was implemented using the OpenCV library for image processing. The
selected test image was read using cv2.imread(), and converted to HSV color space using
cv2.cvtColor() to facilitate color-based segmentation. HSV thresholds were defined to
isolate colors typically associated with diseased regions (brown, yellow, and dark patches).
A binary mask was created using cv2.inRange() and cleaned with morphological operations
such as opening and closing to reduce noise.
Contours representing diseased regions were extracted using cv2.findContours() and
drawn on the original image using cv2.drawContours() for visual confirmation. The area of
each contour was calculated and summed, and then divided by the total image area to
compute the percentage of affected area. This percentage was displayed alongside a side-
DEPARTMENT OF ECE, JNTUACEA 18
 Plant Disease Detection
The system also included a prediction pipeline, where a new image could be passed
through preprocessing, prediction, and area estimation steps in a single workflow. This
modular structure made the system highly usable and testable. Several checkpoints and
error-handling mechanisms were included to verify if the input path was valid, the image
was readable, and the results were within expected bounds.
The complete software stack included:
Python – primary programming language
Jupyter Notebook – development and testing environment
TensorFlow/Keras – deep learning model development
OpenCV – image processing and area estimation
Matplotlib – data visualization and plotting
NumPy – numerical operations and matrix handling
scikit-learn – (optional) for metrics and performance evaluation
In conclusion, the software implementation effectively bridges the gap between
plant disease identification and severity estimation. It is designed in a modular fashion,
allowing each component—data loading, classification, area calculation, and
visualization—to be developed, tested, and improved independently. The result is a reliable
and user-friendly tool with the potential for real-world deployment in agriculture.

DEPARTMENT OF ECE, JNTUACEA 19

 Plant Disease Detection

CHAPTER -5

RESULTS AND DISCUSSIONS

The results of the project are evaluated in terms of the system’s accuracy in detecting
plant diseases, its ability to generalize across multiple classes, and the effectiveness of
estimating the diseased area on plant leaves. The implementation combined deep learning
for classification and image processing techniques for quantification, and both components
were tested thoroughly to assess their performance.

The Convolutional Neural Network (CNN) model, based on the MobileNetV2

architecture, was trained using the PlantVillage dataset containing 43,444 images across 38
disease and healthy classes. The training and validation process showed promising
performance over 10 epochs. The model started with an initial training accuracy of 75.32%
and gradually improved to 98.02%. Simultaneously, the validation accuracy rose from
92.16% to 97.82%, indicating excellent generalization and minimal overfitting. The training
and validation loss decreased consistently, with final values of 0.0686 and 0.0912,
respectively, suggesting that the model successfully learned distinguishing features of
different plant diseases.

The training and validation accuracy/loss graphs were analyzed, and both sets of
curves showed a converging trend. The close proximity of the curves confirms that the
model does not suffer from significant overfitting or underfitting. These results demonstrate
that the use of a lightweight pretrained model like MobileNetV2, in combination with
transfer learning and data augmentation, can yield high performance even with limited
computational resources.

To evaluate the system’s prediction capability, several test images from the
validation set and a few manually selected images from the dataset were passed through the
model. The model was able to correctly classify the majority of samples, even when there
were variations in lighting, orientation, and background. This confirms that the model has
learned robust features that generalize well across different inputs.

In addition to classification, the project included the estimation of the affected area
on the leaf, which is a novel feature not present in many traditional systems. For this, the
OpenCV library was used to process the image and segment the infected region. The HSV
color space was particularly effective for isolating disease-like discolorations. Threshold
values were selected based on visual experimentation, and morphological operations helped
refine the segmented areas.

DEPARTMENT OF ECE, JNTUACEA 20

 Plant Disease Detection
Contour detection was performed on the binary mask of the diseased region, and the
area of all detected contours was calculated and expressed as a percentage of the total leaf
area. This value provided a practical estimation of disease severity. The affected area was
then visualized by overlaying green contours on the original leaf image. In most cases, the
estimation was accurate; however, in some cases, non-diseased regions such as shadows or
naturally darker areas were mistakenly identified as affected. This remains a known
limitation of color-based segmentation techniques.

Despite minor false positives in area estimation, the system offers valuable insights
by combining qualitative classification with quantitative severity measurement. This dual
output can assist farmers not only in diagnosing plant health but also in determining the
urgency and intensity of the required treatment.

Overall, the integration of CNN-based classification with contour-based area

estimation provides a comprehensive plant disease detection tool. The results confirm that
the system is capable of achieving high classification accuracy while also estimating the
severity of infection, making it a practical solution for digital agriculture and precision
farming.

FIGURE 5.1 - HEALTHY GRAPE IMAGE

The figure above illustrates a case where the model has accurately identified a
healthy grape leaf with no signs of disease. The left panel displays the original image of the
leaf, which visually appears free from any discoloration, spots, or lesions. The right panel
shows the analysis output where the model confirms the absence of any infected regions,
indicating a diseased area of 0.00%. No segmentation or contour markings are present,
which aligns with the predicted class label: Grape healthy. This outcome demonstrates
DEPARTMENT OF ECE, JNTUACEA 21
 Plant Disease Detection
the model's robustness in distinguishing between healthy and diseased leaves, ensuring that
false positives are minimized. It also highlights the model's capability to not only detect
disease when present but also accurately validate plant health, making it a reliable tool for
automated plant monitoring in agricultural settings.

FIGURE 5.2 - DISEASED APPLE IMAGE

The figure above demonstrates the effectiveness of the proposed model in detecting
plant disease and calculating the affected area on an apple leaf. The left panel displays the
original image of the leaf, which shows visible symptoms of infection. The right panel
presents the processed output where the diseased region, identified as Apple Black Rot, is
accurately segmented and highlighted with a bright green contour. The model successfully
classifies the disease type using a convolutional neural network trained on a comprehensive
plant disease dataset. In addition to disease classification, the model quantifies the severity
of infection, estimating that approximately 3.50% of the leaf area is affected. This
percentage is calculated by analyzing pixel data corresponding to the infected region. The
figure highlights the model’s dual capability to not only identify the disease but also provide
an accurate assessment of its impact, aiding in timely and informed decision-making for
crop protection.

DEPARTMENT OF ECE, JNTUACEA 22

 Plant Disease Detection

FIGURE 5.3- GRAPH OF TRAINING ACCURACY AND VALIDATION

ACCURACY
The graph above represents the training and validation accuracy trends of the plant
disease detection model over 10 epochs. The blue line indicates the training accuracy, while
the orange line represents the validation accuracy. Initially, the model's training accuracy
started around 83.5% and showed a consistent upward trend, ultimately reaching
approximately 96.8%, demonstrating effective learning from the training data. The
validation accuracy also increased steadily, beginning at about 92% and stabilizing around
94–95%, which suggests that the model is able to generalize well on unseen data. The
closeness of the two curves and the absence of significant divergence between them
indicates that the model avoids overfitting and achieves a good balance between bias and
variance. This performance underscores the model’s reliability and robustness for accurate
classification in real-world scenarios.

DEPARTMENT OF ECE, JNTUACEA 23

 Plant Disease Detection

5.2 REPRESENTATION OF DATA

MODEL TRAINING AND VALIDATION SUMMARY

Epoch Training Validation Training Loss Validation

Accuracy(%) Accuracy(%) Loss
1 75.32 92.16 0.8996 0.2387
5 94.87 94.60 0.1484 0.1700
10 98.02 97.82 0.0686 0.0912

The graph presented illustrates the training and validation accuracy of the plant
disease detection model over a span of 10 epochs. The blue curve denotes the training
accuracy, while the orange curve represents the validation accuracy. As the training
progresses, the training accuracy shows a steady improvement from approximately 83.5%
to nearly 96.8%, reflecting the model’s ability to effectively learn patterns from the training
dataset. Simultaneously, the validation accuracy also increases consistently, starting from
around 92% and stabilizing at approximately 94–95%, indicating good generalization on
unseen data.
The proximity of the training and validation accuracy curves suggests that the model
is not overfitting, maintaining a healthy balance between learning the training data and
performing well on the validation set. This steady convergence and the eventual flattening
of both curves imply that the model is reaching its optimal learning potential. Such a
performance outcome validates the robustness of the model architecture and the
effectiveness of the preprocessing, augmentation, and training strategies employed. This
accuracy progression supports the reliability of the system for practical applications in
detecting plant diseases with high precision.

DEPARTMENT OF ECE, JNTUACEA 24

 Plant Disease Detection

AREA ESTIMATION RESULTS (SAMPLE OUTPUTS)

Image Name Predicted Class Estimated

Affected Area(%)
Tomato_late_blight_1234_jpg Tomato_late_blight 18.56%
Apple_cedar_apple_rust_0164_jpg Apple_cedar_apple_rust 12.45%
Grape_black_rot_3438_jpg Grape_black_rot 23.78%
Potato_healthy2_jpg Potato_healthy2 0.00%(healthy leaf)

The table above presents a sample output of the model’s dual functionality—disease
classification and affected area estimation—on various plant leaf images. Each row
corresponds to a different test image, detailing the predicted disease class and the percentage
of the leaf area determined to be infected.
For the image Tomato Late_blight_1234.JPG, the model correctly identified the
disease as Tomato Late Blight and estimated that approximately 18.56% of the leaf area is
affected.
In the case of Apple Cedar_apple_rust_0164.JPG, the predicted class was Apple
Cedar Apple Rust, with an affected area of 12.45%, highlighting a moderate level of
infection.
For Grape Black_rot_3438.JPG, the model predicted Grape Black Rot with a
higher severity level of 23.78%, indicating more extensive damage.
Lastly, Potato Healthy2.JPG was identified as Potato Healthy, with an affected area of
0.00%, confirming that the leaf is free from any visible disease symptoms.
These results demonstrate the system’s ability to not only recognize the type of
disease with high accuracy but also to quantify the severity by calculating the infected
portion of the leaf. This combination of qualitative and quantitative analysis is essential for
supporting early diagnosis, guiding treatment decisions, and enabling precision agriculture
practices.

DEPARTMENT OF ECE, JNTUACEA 25

 Plant Disease Detection

CHAPTER-6
CONCLUSION AND FUTURE SCOPE
6.1 CONCLUSION
The study successfully demonstrates the effectiveness of integrating deep learning and
image processing techniques to automate plant disease detection and estimate the severity
of infection. With agriculture forming the backbone of many economies and food systems,
the need for intelligent, data-driven, and scalable solutions to combat crop loss has never
been more pressing. This project addresses that need by developing a dual-function system
capable of classifying plant diseases and quantifying the percentage of the leaf area affected.
A robust CNN-based classification model was implemented using the MobileNetV2
architecture, which was fine-tuned on the PlantVillage dataset. The model achieved
impressive accuracy—98.02% on the training set and 97.82% on the validation set—
demonstrating its capability to learn meaningful patterns across 38 distinct disease and
healthy classes. Data augmentation, image normalization, and transfer learning were
essential in achieving high generalization performance while optimizing resource usage.
Beyond classification, the system incorporates an image processing pipeline using
OpenCV to analyze the physical spread of disease on a given leaf. By converting the image
to HSV color space, applying adaptive thresholding, and using morphological
transformations, the infected regions were effectively isolated and quantified. The resulting
area was expressed as a percentage of the total leaf surface, offering valuable insight into
the severity of the condition. Although the area detection approach occasionally
misinterprets shadows or natural discolorations as disease, it provides a solid foundation for
further refinement.
Overall, the project demonstrates how modern AI technologies can contribute to
precision agriculture by enhancing early disease diagnosis and providing actionable
insights. The developed system is scalable, user-friendly, and can be further integrated into
mobile or IoT-based platforms for real-time field usage. By empowering farmers with
timely and accurate disease detection, this work has the potential to minimize crop losses,
reduce pesticide misuse, and promote sustainable farming practices.

DEPARTMENT OF ECE, JNTUACEA 26

 Plant Disease Detection

6.2 FUTURE SCOPE:

While the current system effectively addresses plant disease detection and infected
area estimation using deep learning and image processing, there remain several
opportunities to enhance its capabilities, scalability, and real-world usability. The future
scope of this project can be categorized into technological, operational, and functional
advancements.
One significant direction for future development is the integration of real-time
mobile and edge computing applications. Although the model has been successfully tested
in a controlled environment using Jupyter Notebook, deploying the system onto mobile
devices or embedded platforms (e.g., Raspberry Pi or NVIDIA Jetson Nano) will make it
more accessible to farmers and agricultural workers in remote areas. Using lightweight
models like MobileNetV2, further optimized with TensorFlow Lite or ONNX, will support
real-time predictions without reliance on internet connectivity.
Another promising enhancement involves the use of advanced segmentation
techniques. While the current system uses HSV thresholding and contour detection to
estimate affected area, future iterations can incorporate deep learning-based semantic
segmentation models such as U-Net, DeepLabV3+, or Mask R-CNN. These models offer
pixel-level accuracy and can differentiate between multiple types of lesions, even in
complex background scenarios. With sufficient annotated data, this would significantly
improve the precision of severity estimation and reduce false positives caused by shadows
or discolorations.
The inclusion of a disease progression tracking system could also be introduced. By
capturing images of the same plant over time, the model could analyze disease evolution,
quantify progression rates, and help schedule treatments more effectively. This time-series
analysis could be further supported by integrating external factors such as humidity,
temperature, and soil conditions to provide predictive diagnostics.
To expand the system’s applicability, the dataset can be diversified by incorporating
images taken in natural field conditions with varying lighting, angles, and occlusions. Real-
world scenarios often differ from curated datasets like PlantVillage, which are typically
captured in controlled environments. Building a custom, annotated
dataset through field collaboration will help make the model more resilient and accurate in
practical applications.
Additionally, the system could be scaled to support multi-plant detection within a
single image or perform object detection across large farm fields using drone imagery. This
would allow for large-scale disease mapping and precision agriculture at the farm level.
Integration with GIS (Geographic Information Systems) and farm management systems can
DEPARTMENT OF ECE, JNTUACEA 27
 Plant Disease Detection
offer farmers a holistic platform to monitor crop health and take timely interventions.
From an operational standpoint, building a user interface (UI) or a mobile app where
users can upload images and receive immediate feedback—including disease name,
confidence score, affected area, and suggested remedies—can make the tool more user-
friendly and impactful.
In conclusion, this project serves as a strong foundation for intelligent crop disease
management, and its future scope offers numerous possibilities for real-world deployment,
technical refinement, and interdisciplinary integration. With advancements in data
collection, model optimization, and user engagement, this system can evolve into a
comprehensive plant health monitoring solution for smart and sustainable agriculture.

DEPARTMENT OF ECE, JNTUACEA 28

 Plant Disease Detection

CHAPTER-7
REFERENCE

[1] https://ieeexplore.ieee.org/document/93992

[2] https://ieeexplore.ieee.org/document/71551

[3] https://ieeexplore.ieee.org/document/84385
[4] Mohanty, S.P., Hughes, D.P., & Salathé, M. (2016). Using deep learning for image-
based plant disease detection. Frontiers in Plant Science, 7, 1419.
https://doi.org/10.3389/fpls.2016.01419
[5] Ferentinos, K. P. (2018). Deep learning models for plant disease detection and diagnosis.
Computers and Electronics in Agriculture, 145, 311–318.
https://doi.org/10.1016/j.compag.2018.01.009
[6] PlantVillage Dataset. PlantVillage - Disease Classification Dataset. Available at:
https://www.kaggle.com/emmarex/plantdisease
[7] Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., & Chen, L.C. (2018). MobileNetV2:
Inverted Residuals and Linear Bottlenecks. Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), 4510–4520.
[8] Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional Networks for
Biomedical Image Segmentation. In International Conference on Medical Image
Computing and Computer-Assisted Intervention (pp. 234–241). Springer.
[9] OpenCV Documentation. OpenCV: Open Source Computer Vision Library. Available
at: https://docs.opencv.org/
[10] TensorFlow Developers. (2023). TensorFlow API Documentation. Available at:
https://www.tensorflow.org/api_docs
[11] Simonyan, K., & Zisserman, A. (2015). Very Deep Convolutional Networks for Large-
Scale Image Recognition. arXiv preprint arXiv:1409.1556.
[12] Sibiya, M., & Sumbwanyambe, M. (2019). A survey on deep learning techniques for
image-based plant disease detection. Computers and Electronics in Agriculture, 172, 105-
130.
[13] K. P. Ferentinos, “Deep learning models for plant disease detection and diagnosis,” Computers
and Electronics in Agriculture, vol. 145, pp. 311–318, Feb. 2018. DOI:
10.1016/j.compag.2018.01.009
[14] S. P. Mohanty, D. P. Hughes, and M. Salathé, “Using deep learning for image-based plant
disease detection,” Frontiers in Plant Science, vol. 7, pp. 1–10, Sep. 2016. DOI:
10.3389/fpls.2016.01419

DEPARTMENT OF ECE, JNTUACEA 29

 Plant Disease Detection

DEPARTMENT OF ECE, JNTUACEA 30