Medical Leave Images

Classification Of Indian Medicinal Leaves Using Transfer Learning Based Convolutional Neural
Networks
CLASSIFICATION OF INDIAN MEDICINAL

LEAVES USING TRANSFER LEARNING
BASED CONVOLUTIONAL NEURAL
NETWORKS
A Project work submitted to
Acharya Nagarjuna University
Department of Computer Science and Engineering
In partial fulfillment of the requirements for

The award of the degree of
Master of Computer Applications

by
ULLAMPARTHI DILEEP KUMAR

Regd. No. Y23MC20060
Under the guidance of
Dr. G. NEELIMA
Assistant Professor
Department of computer science & engineering
Acharya Nagarjuna University
DEPARTMENT OF COMPUTER SCIENCE & ENGINEERING

ACHARYA NAGARJUNA UNIVERSITY
Nagarjuna Nagar, Guntur,
Andhra Pradesh, India
March 2024
Networks
DECLARATION
I hereby declare that the entire thesis work entitled "
CLASSIFICATION OF INDIAN MEDICINAL LEAVES USING
TRANSFER LEARNING BASED CONVOLUTIONAL NEURAL
NETWORKS" is being submitted to the Department of Computer
Science and Engineering, University College of Sciences, Acharya
Nagarjuna University, in partial fulfillment of the requirement for the
award of the degree of Master of Computer Applications (MCA) is a
bon-fide work of my own, carried out under the supervision of Dr. G.
NEELIMA, Assistant Professor, Department of Computer Science &
Engineering, Acharya Nagarjuna University.
I further declare that the Project, either in part or full, has not been submitted
earlier by me or others for the award of any degree in any University.
ULLAMPARTHI DILEEP KUMAR
Reg. No. Y23MC20060
i
Networks
ACHARYA NAGARJUNA UNIVERSITY

NAGARJUNA NAGAR, GUNTUR
Department of Computer Science & Engineering.
CERTIFICATE
This is to certify that this project entitled “CLASSIFICATION OF
INDIAN MEDICINAL LEAVES USING TRANSFER LEARNING
BASED CONVOLUTIONAL NEURAL NETWORKS” is a Bon-fide
record of the project work done and submitted by
ULLAMPARTHI DILEEP KUMAR (Y23MC20060) during the year
2022 - 2024 in partial fulfillment of the requirements for the award of
degree of Master of Computer Applications (MCA) in the department
of Computer Science & Engineering. I certify that he carries this
project as an independent project under my guidance.
Head of the Department Project Guide

(Prof. K. Gangadhara Rao) (Dr. G .Neelima)
ii
Networks
External Examiner
iii
Networks
ACKNOWLEDGEMENTS
Undertaking this Project has been a truly life-changing experience for me and it
would not have been possible to do without the support and guidance that I received
from many people.
I would like to first say a very big thank you to my supervisor Dr. G. NEELIMA for
all the support and encouragement he gave me. Her friendly guidance and expert
advice have been invaluable throughout all stages of the work. Without her guidance
and constant feedback this Project work not have been achievable.
I would also wish to express my gratitude to Prof. K. Gangadhara Rao for extended
discussions and valuable suggestions which have contributed greatly to the
improvement of the thesis.
I am thankful to and fortunate enough to get constant encouragement, support and
guidance from all Teaching staffs of Department which helped us in successfully
completing our project work. Also, I would like to extend our sincere regards to all
the non-teaching staff of the department for their timely support.
I must also thank my parents and friends for the immense support and help during this
project. Without their help, completing this project would have been very difficult.
iv
Networks
ABSTRACT
There’s a rapid increment in the utilization of herbal-based items over the world.
Medicinal herbs play a major part in creating ayurvedic drugs and homegrown items.
Herbal medications are regarded as a substitute for synthetic drugs in both developing
and developed countries, owing to their lack of adverse effects. Many medicinal
species are extant in India. Identifying those medicinal plants/leaves over a bunch of
different kinds of plant species is necessary as they are many health benefits with
medicinal plants. Classification of medicinal leaves can be automated based on their
respective textures, shape, and edge or margin of a leaf. In this paper, Indian
medicinal leaves classifying utilizing transfer learning based convolutional neural
networks has been proposed. In order to make the classification more accurate, the
data augmentation technique had been utilized to enlarge the data which avoids
insufficiency of data. The regularization technique was utilized to avoid the menace
of overfitting. The dataset consists of 1835 images and thirty different types of
beneficial Indian medicinal herbs such as Tulsi, Jasmine, etc. Data has been trained
using different architectures such as VGG16, Xception, and efficieNetv2. The
preliminary results of the Indian medicinal herbs dataset showed that the Xception
architecture performed better than EfficientNetv2 and VGG16.
v
Networks
TABLE OF CONTENTS
TITLE PAGE NO
DECLARATION i
CERTIFICATE iii
ACKNOWLEDGEMENT iv
ABSTRACT v
TABLE OF CONTENTS vii
LIST OF FIGURES xi
Chapter 1: INTRODUCTION
1.1 Overview 2
1.2 Problem Statement 5
1.3 Objectives of the Project 6
Chapter 2: LITERATURE SURVEY
2 Literature Survey 9
Chapter 3: SYSTEM ANALYSIS
3.1 Problem statement 14
3.2 Existing System 14
3.3 Proposed System 16
3.4 Modules 17
vi
Networks
Chapter 4: SYSTEM REQUIREMENTS
4.1 Functional Requirements 20
4.2 Non-Functional Requirements 23
4.3 Performance Requirements 24
4.4 Feasibility Study 25
Chapter 5: METHODOLOGY
5.1 Dataset 29
5.2 Data Preparation 30
5.3 Model Development 32
5.3.1 VGG16 34
5.3.2 Xception 37
5.3.3 EfiecientNetV2 40
5.4 Performance Comparison Charts 44
5.5 Training and Evalution 45
5.5.1 Model Training 46
5.5.2 Model Evalution 46
5.5.3 Computational Efficiency and Scalability 47
5.6 Algorithms 47
vii
Networks
Chapter 6: SYSTEM DESIGN
6.1 System Architecture 57
6.2 UML diagrams 59
6.2.1 Use case diagram 59
6.2.2 Sequence diagram 62
6.2.3 Activity Diagram 65
6.2.4 Class Diagram 67
Chapter 7: IMPLEMENTATION AND EVALUATION
7.1 Dataset 72
7.2 Import Modules 74
7.3 Plot Model Architecture 75
7.4 Feature Extraction Model 77
Chapter 8: SOURCE CODE
Source Code 83
Chapter 9: SYSTEM TESTING
9.1 Software Testing 87
9.2 Test Cases 90
Chapter 10: CONCLUSION AND FUTURE SCOPE
10.1 Conclusion 95
10.2 Future Scope 96
References 99
viii
Networks
ix
Networks
LIST OF FIGURES
Figure No. Title of the Figure Page No.
5.1 Dataset sample 27
5.3.1 Architecture of VGG16 32
5.3.2 Architecture of Xception 37
5.3.3 Architecture of EfficientNetV2 38
5.4 Performance Comparison Charts 39
6.1 System Architecture 52
6.2.1 Use case diagram 54
6..2.2 Sequence Diagram 57
6.2.3 Activity Diagram 60
6.2.4 Class Diagram 63
7.1 List of some medicinal plant species 69
7.3 Architecture Model 72
7.4 Feature Extraction Model 75
7.5 First image to predict 76
7.6 Leave prediction with actual and predicted values 77
x
Networks
CHAPTER 1
INTRODUCTION
xi
Networks
1.1 Overview
Medicinal plants have been an integral part of traditional medicine systems for
centuries, providing remedies for various ailments. With the advancement of
technology, the identification and recognition of medicinal plants have been expedited
through the application of machine learning techniques. Transfer learning, in
particular, has emerged as a powerful tool in this domain, allowing the leveraging of
pre-trained models to address recognition tasks with limited labeled data. In this
introductory discourse, we delve into the application of transfer learning for the
recognition of medicinal plant leaves, employing state-of-the-art convolutional neural
network (CNN) architectures including VGG16, Xception, and EfficientNetV2.
Leaves have characteristics such as shape and texture to be identified with the help of image
processing technology and deep learning. An object sees identification as geometric
information with boundaries [1]–[10]. The identified leaf object is limited to the boundary
identified as leaf size and leaf shape, while the leaf texture or pattern is seen from the leaf
surface [11]. Generally, the size of the leaves can be different, but the surface pattern of the
leaves does not differ from one another [12]–[15]. This study aims to identify medicinal or
phytomedicine plant species by processing leaf imagery using image processing and deep
learning [15]–[21].
Research on the identification of phytomedicine plant leaves has been carried out by several
previous studies, for example Naresh and Nagendraswamy in 2015 [22], Mukherjee and his
team in 2016 [23], Venkataraman & Mangayarkarasi in 2017 [24], Gao & Lin in 2018 [25],
Sivaranjani et al. in 2019 [26], Pechebovicz et al. in 2020 [27], Bhuiyan et al. in 2021 [28].
xii
Networks
In a study by Naresh and Nagendraswamy in 2015, the authors employed local binary patterns
(LBP) to classify medicinal leaf plants. In a study conducted in 2016 by Mukherjee and his
team, the classification of medicinal plants was accomplished with the use of Back
Propagation MultiLayer Perceptron (BP-MLP) [22], [29].
A study on the classification of medicinal plants also was carried out by Venkataraman and
Mangayarkarasi (2017). They utilized the Histogram of Oriented Gradient (HoG)- Support
Vector Machine for their research (SVM) [24]. Moreover, Gao and Lin (2018) used the
OTSU approach in their classification of leaf plants for medicinal purposes. The OTSU
approach involves using each manually marked edge point of the leaf to precisely detect the
following outside points of the leaf located next to it [25]. The ExG-ExR index and the
Logistic Regression (LR) classifier were utilized by Sivaranjani et al. to classify medicinal
plants, and the researchers discovered that this method was successful. Based on each
extracted leaf's color and texture characteristics, the Logistic Regression (LR) classifier is
utilized to classify the various plant species [26].
In this study, we propose transfer learning models VGG16, VGG19, and MobileNetV2 to
study the distinguishing features to identify medicinal plant leaves according to the dataset
that has been collected. We also improved algorithm using finetuning strategy and analysed a
comparison with and without a fine-tuning strategy to transfer learning model’s performance.
 Recognition of Medicinal Plant Leaves
Recognition of medicinal plant leaves is a fundamental task in botany, pharmacology,
and biotechnology. Accurate identification of plant species is crucial for various
purposes such as biodiversity conservation, drug discovery, and herbal medicine

xiii
Networks
formulation. However, manual identification of plant species based on leaf
characteristics is labor-intensive and often prone to errors. Hence, the development of
automated recognition systems using machine learning has gained significant traction.
 Transfer Learning
Transfer learning involves leveraging knowledge acquired from solving one
problem and applying it to a different but related problem. In the context of
medicinal plant leaf recognition, transfer learning allows us to utilize pre-
trained CNN models, which have been trained on large-scale datasets such as
ImageNet, and fine-tune them on a smaller dataset containing images of
medicinal plant leaves. This approach is particularly beneficial when the
labeled dataset for the target task is limited, as it enables the model to learn
meaningful representations from the source domain and adapt them to the
target domain.
 CNN Architectures: VGG16, Xception, and EfficientNetV2
VGG16, Xception, and EfficientNetV2 represent three distinct CNN
architectures that have demonstrated superior performance across various
image recognition tasks. VGG16 is characterized by its simplicity and consists
of 16 layers, including convolutional and fully connected layers. Xception, on
the other hand, introduces depth wise separable convolutions, which
significantly reduce the number of parameters while maintaining high
accuracy. EfficientNetV2, a recent advancement in CNN architectures,
focuses on improving model efficiency by optimizing the network's depth,
xiv
Networks
width, and resolution using compound scaling.
 Application in Medicinal Plant Leaf Recognition
The application of transfer learning using VGG16, Xception, and EfficientNetV2
in medicinal plant leaf recognition involves several stages. First, the pre-trained
models are initialized with weights learned from the ImageNet dataset. Next, the
fully connected layers of the CNN architectures are modified to accommodate the
number of classes specific to medicinal plant species. Finally, the modified
models are fine-tuned on a dataset comprising images of medicinal plant leaves,
adjusting the model parameters to better capture the distinguishing features of
different plant species.
Transfer learning coupled with state-of-the-art CNN architectures such as VGG16,
Xception, and EfficientNetV2 presents a promising approach for the recognition
of medicinal plant leaves. By leveraging pre-trained models and fine-tuning them
on limited labeled data, these techniques enable the development of accurate and
efficient recognition systems essential for various applications in botany,
pharmacology, and herbal medicine. In the subsequent sections, we delve deeper
into the methodology, experiments, and results obtained through the application of
these techniques in medicinal plant leaf recognition tasks.
1.2 Problem Statement
Medicinal plants have been extensively used in traditional medicine for centuries due
to their therapeutic properties. Identifying these plants accurately is crucial for various
xv
Networks
applications including pharmaceuticals, healthcare, and conservation. However,
manual identification can be time-consuming and prone to errors. Automated
recognition systems based on deep learning models offer a promising and reliable
solution, but often require large labeled datasets for training, which can be
challenging to acquire in the case of medicinal plants.
Transfer learning has emerged as a powerful technique in deep learning, where pre-
trained models trained on large-scale datasets are fine-tuned on smaller, domain-
specific datasets to adapt to specific recognition tasks. In this context, the utilization
of pre-trained convolutional neural network (CNN) architectures such as VGG16,
Xception, and EfficientNetV2 has shown promising results in various image
recognition tasks.
1.3 Objectives of the Project
 Dataset Acquisition and Preprocessing: Gather a dataset comprising images
of medicinal plant leaves from diverse sources. Preprocess the images to
ensure uniformity in size, color, and orientation, and partition the dataset into
training, validation, and testing sets.
 Model Implementation and Transfer Learning: Implement transfer learning
by initializing the VGG16, Xception, and EfficientNetV2 models with weights
pre-trained on large-scale image datasets such as ImageNet. Fine-tune these

xvi
Networks
models on the medicinal plant leaves dataset to adapt them to the recognition
task.
 Model Evaluation: Evaluate the performance of the transfer learning models
using metrics such as accuracy, precision, recall, and F1-score on the
validation and test sets. Compare the performance of VGG16, Xception, and
EfficientNetV2 architectures to identify the most effective model for
medicinal plant leaf recognition.
 Model Optimization: Explore techniques for model optimization such as data
augmentation, learning rate scheduling, and model ensemble to further
improve the recognition accuracy and generalization capability of the models.
 Deployment and Application: Develop a user-friendly interface for the
trained model to enable users to upload images of medicinal plant leaves for
recognition. Test the deployed model on real-world data and assess its
performance in practical applications.
xvii
Networks
CHAPTER 2
LITERATURE SURVEY
xviii
Networks
2 LITERATURE SURVEY
Das, A., & Mandal, D. (2020). Medicinal Plant Leaf Identification Using
Transfer Learning. In 2020 6th International Conference on Advanced
Computing and Communication Systems (ICACCS) (pp. 1095-1099). IEEE.
This research aimed to improve the automatic identification of medicinal plants due to
their growing popularity and increased requests for artisanal and industrial uses and
applications. Therefore, the proposed DL algorithm and image processing technique
can have a special place in plant science and even industrial markets for identifying
and classifying varied medicinal plants separately from other non-edible plants.
Chaudhary, S., & Pande, M. (2021). Medicinal plant leaf recognition using deep
learning transfer learning techniques. In 2021 International Conference on
Computer Communication and Informatics (ICCCI) (pp. 1-6). IEEE.
Fig. 3. Proposed Model Architecture in

Python
categorical cross-entropy loss
function using the Adam
optimizer. The accuracy is used as
a metric to evaluate
the performance of the model.
xix
Networks
VII. R ES ULTS A ND A N ALYS IS

The proposed approach utilizes a
convolutional neural net-
work (CNN), specifically a VGG-
16-based model, trained on
a dataset of 25,686 images of 29
plant species. The VGG-16
model is pre-trained on ImageNet,
a wide range of images,
which makes it an excellent
starting point for many image
recognition tasks.
Figure 4 shows that the model’s
accuracy increased after
training for 30 epochs, with a
maximum accuracy of 0.98.
This represents a considerable
improvement from the starting
xx
Networks
accuracy. Figure 5 shows that the

current loss after a total of
30 epochs is around 0.04,
indicating that the model has the
ability to learn from the data and
enhance its performance over
time.
Experimental results show that the
proposed method out-
performs traditional machine
learning techniques in terms
of classification accuracy, with a
recognition rate of 98%.
This approach can be used as a
useful tool for healthcare
professionals, researchers, and
other professionals in the field
of herbal medicine to aid in the
identification of medicinal
xxi
Networks
plants. It can be extended to other

fields such as agriculture
and botany, where accurate plant
identification can have a
significant impact on sustainable
natural resource use and
conservation efforts.
It is worth mentioning that due to
limitations in CPU pro-
cessing, the model was trained in
two separate runs, each with
15 epochs. This underlines the
importance of having sufficient
computational resources when
training complex models. The
results presented in Figures 4 and
5 demonstrate the model’s
potential for further improvement
through increased epochs or
xxii
Networks
additional resources.
VIII. C ONCLUSIONS AND
F UTURE W ORK
This study has demonstrated the
effectiveness of a deep
learning-based method using a
convolutional neural network
(CNN) to accurately recognize
medicinal plants
Fig. 3. Proposed Model Architecture in
Python
categorical cross-entropy loss
function using the Adam
optimizer. The accuracy is used as
a metric to evaluate
the performance of the model.
VII. R ES ULTS A ND A N ALYS IS
The proposed approach utilizes a
convolutional neural net-
xxiii
Networks
work (CNN), specifically a VGG-

16-based model, trained on
a dataset of 25,686 images of 29
plant species. The VGG-16
model is pre-trained on ImageNet,
a wide range of images,
which makes it an excellent
starting point for many image
recognition tasks.
Figure 4 shows that the model’s
accuracy increased after
training for 30 epochs, with a
maximum accuracy of 0.98.
This represents a considerable
improvement from the starting
accuracy. Figure 5 shows that the
current loss after a total of
30 epochs is around 0.04,
indicating that the model has the
xxiv
Networks
ability to learn from the data and

enhance its performance over
time.
Experimental results show that the
proposed method out-
performs traditional machine
learning techniques in terms
of classification accuracy, with a
recognition rate of 98%.
This approach can be used as a
useful tool for healthcare
professionals, researchers, and
other professionals in the field
of herbal medicine to aid in the
identification of medicinal
plants. It can be extended to other
fields such as agriculture
and botany, where accurate plant
identification can have a
xxv
Networks
significant impact on sustainable

natural resource use and
conservation efforts.
It is worth mentioning that due to
limitations in CPU pro-
cessing, the model was trained in
two separate runs, each with
15 epochs. This underlines the
importance of having sufficient
computational resources when
training complex models. The
results presented in Figures 4 and
5 demonstrate the model’s
potential for further improvement
through increased epochs or
additional resources.
VIII. C ONCLUSIONS AND
F UTURE W ORK
xxvi
Networks
This study has demonstrated the

effectiveness of a deep
learning-based method using a
convolutional neural network
(CNN) to accurately recognize
medicinal plants
The study’s results show that deep
learning
techniques can be used to classify
medicinal plants accurately,
providing a reliable tool for
healthcare professionals and
herbal medicine researchers to
identify herbal plants. Future
work could involve further
integrating the identified plant
with
a large language model to provide
additional information about
xxvii
Networks
the scientific name, more images,

and explanations of the
benefits and side effects of the
plant, and instructions on how
to use it in both English and the
local language. Overall, the
study supports the continued
development and integration of
deep learning in the recognition of
herbal plants, which has the
potential to significantly impact
the field of herbal medici
The study’s results show that deep learning techniques can be used to classify
medicinal plants accurately, providing a reliable tool for healthcare professionals
andherbal medicine researchers to identify herbal plants. Futurework could involve
further integrating the identified plant witha large language model to provide
additional information aboutthe scientific name, more images, and explanations of
thebenefits and side effects of the plant, and instructions on howto use it in both
English and the local language. Overall, thestudy supports the continued development
and integration ofdeep learning in the recognition of herbal plants, which has
thepotential to significantly impact the field of herbal medicinal research
xxviii
Networks
X.-F. Wang, D.-S. Huang, J.-X. Du, H. Xu, and L. Heutte, “Classification of
plant leaf images with complicated background,” Appl. Math. Comput., vol. 205,
no. 2, pp. 916–926, Nov. 2008.
In this paper, the Mask R-CNN model and the VGG16 model are used to segment and
classify leaf images with multiple targets and a complicated background. More than
4000 images were used for model training and testing. The results show that the
average ME of segmentation is up to 1.15% using the Mask R-CNN model, and the
average classification accuracy is up to 91.5% using the VGG16 model. This shows
that the Mask R-CNN model and the VGG16 model could reliably be used in the
segmentation and classification of leaf images with a complicated background.
Further study is recommended to be performed with different deep learning
algorithms and a greater number of data, which may lead to a better result. Besides
the algorithm’s development, improving the image quality with better devices can
also contribute to better performance. What is more, it will be possible to segment and
classify leaves automatically in real-time by using an embedding system.
K. Yang, W. Zhong, and F. Li, “Leaf Segmentation and Classification with a
Complicated Background Using Deep Learning,” Agronomy, vol. 10, no. 11, p.
1721, Nov. 2020
In this paper, the segmentation and classification of leaf images with a complicated
background using deep learning are studied. First, more than 2500 leaf images with a
complicated background are collected and artificially labeled with target pixels and
xxix
Networks
background pixels. Two-thousand of them are fed into a Mask Region-based
Convolutional Neural Network (Mask R-CNN) to train a model for leaf segmentation.
Then, a training set that contains more than 1500 training images of 15 species is fed
into a very deep convolutional network with 16 layers (VGG16) to train a model for
leaf classification. The best hyperparameters for these methods are found by
comparing a variety of parameter combinations. The results show that the average
Misclassification Error of 80 test images using Mask R-CNN is 1.15%. The average
accuracy value for the leaf classification of 150 test images using VGG16 is up to
91.5%. This indicates that these methods can be used to segment and classify the leaf
image with a complicated background effectively. It could provide a reference for the
phenotype analysis and automatic classification of plants
E. M. Jawad, H. G. Daway, and H. J. Mohamad, “Quantum-dot Cellular
Automata Based Lossless CFA Image Compression Using Improved and
Extended Golomb-rice Entropy Coder,” Int. J. Intell. Eng. Syst., vol. 15, no. 2,
pp. 12–25, Apr. 2022.
This article proposed efficient hardware architecture of lossless colour image
compression mechanism using CFA for wireless camera networks with QCA
technology to mitigate the area, and power as compared to standard CMOS
technology. In addition, IEGREC is proposed to reduce the memory requirement,
computational complexity in terms of gate count and eliminated the requirement of
context module to lessen the hardware resource utilization. The extensive simulation
results disclosed that proposed QCAIEGREC method performed superior as
xxx
Networks
compared to state-of-art approaches with less area i.e., 3700 gate count, and reduced
power about 1.02 mW. It also outperformed the existing approaches with image
quality metrics such as 18.78dB of PSNR, 152.3 of FOM, and 29.3 of CR. This work
can be extended to implement the joint blur detection and compression operations
using IEGREC coding. The blur detection can prevent the attacks generated in the
compression process.
xxxi
Networks
CHAPTER 3
SYSTEM ANALYSIS
xxxii
Networks
3 SYSTEM ANALYSIS
3.1 Problem Statement
The classification of Indian medicinal leaves represents a substantial undertaking,
demanding the creation of resilient systems adept at precisely discerning plant species
based on intricate leaf attributes. Employing transfer learning-based convolutional
neural networks (CNNs), our objective is to construct a framework proficient in
distinguishing among various types of medicinal leaves. Such a system holds promise
in facilitating the identification and utilization of these plants within traditional
medicine systems, thereby contributing to advancements in healthcare and botanical
research.
The complexity of Indian medicinal leaves, characterized by diverse morphological
features and biochemical compositions, necessitates the utilization of advanced
machine learning techniques. By leveraging transfer learning, our system can
capitalize on the knowledge acquired from pre-trained CNN models, thereby
accelerating the learning process and enhancing classification accuracy. This
approach enables the extraction of intricate patterns and subtle nuances from leaf
images, empowering the system to make informed and reliable classifications
Central to our endeavor is the development of a robust and scalable system
architecture capable of accommodating the intricacies of medicinal plant
classification. Through meticulous model selection, feature extraction, and fine-tuning
strategies, we aim to optimize the system's performance across a wide spectrum of
plant species and leaf variations.

xxxiii
Networks
By integrating cutting-edge technologies with domain expertise in botany and
traditional medicine, we aspire to bridge the gap between modern computational
methods and age-old botanical practices.
Our pursuit of building a sophisticated classification system for Indian medicinal
leaves aligns with the broader goal of harnessing technology to preserve traditional
knowledge and promote sustainable healthcare practices. By enabling accurate species
identification and fostering greater understanding of medicinal plants' therapeutic
properties, our system holds the potential to revolutionize the field of herbal medicine
and contribute to the well-being of communities worldwide.
3.2 Existing Systems
Existing systems in the domain of plant species recognition and leaf classification
provide a rich foundation upon which to build our project. These systems have
explored various methodologies and approaches, offering valuable insights into
effective strategies for our endeavor. Here's an overview of key components and
methods commonly employed in existing systems:
1. Traditional Taxonomy-Based Approaches:
 Traditional taxonomy-based systems rely on manual identification
methods, often involving experts in botany.
 These approaches are time-consuming and may lack scalability for
large-scale leaf classification tasks.
xxxiv
Networks
2. Computer Vision-Based Systems:
 Computer vision techniques have been increasingly utilized for
automated plant species recognition.
 Feature extraction methods, such as shape, texture, and color analysis,
play a crucial role in leaf classification.
 Machine learning algorithms, including CNNs, have shown promise in
automatically identifying plant species based on leaf images.
3. Transfer Learning and CNNs:
 Transfer learning, particularly with pre-trained CNN models, has
emerged as a powerful technique for image classification tasks.
 By leveraging features learned from large datasets, transfer learning
allows for effective classification of new classes with limited data.
3.3 Proposed System
Our proposed system for classifying Indian medicinal leaves integrates transfer
learning-based CNNs, real-time monitoring, and comprehensive data analysis to
accurately identify and categorize medicinal plant species. Key components and
methodologies of the proposed system include:
1. Data Collection and Preprocessing:
Collection of a diverse dataset of Indian medicinal leaves, followed by
preprocessing steps such as resizing and augmentation.
xxxv
Networks
2. Model Training with Transfer Learning:
Utilization of pre-trained CNN models for feature extraction, followed by
fine-tuning on the medicinal leaf dataset.
3. Real-Time Monitoring and Analysis:
Implementation of real-time monitoring mechanisms to analyze leaf images
and classify them into respective medicinal plant species.
4. Integration with Traditional Knowledge:
Integration of traditional botanical knowledge to complement machine
learning-based classification, ensuring accuracy and cultural relevance.
3.4 Modules
In the context of our project, essential modules for the classification of Indian
medicinal leaves using transfer learning-based CNNs may include:
1. Data Collection Module:
Responsible for gathering a comprehensive dataset of Indian medicinal leaves
from diverse sources.
2. Data Preprocessing Module:
Cleans and preprocesses the collected leaf images dataset for training and
testing.
3. Model Training and Fine-Tuning Module:
xxxvi
Networks
Trains pre-trained CNN models on the medicinal leaf dataset, fine-tuning the
network parameters for improved classification performance.
4. Real-Time Monitoring and Classification Module:
Implements mechanisms for real-time leaf image analysis and classification
using trained models.
5. Integration with Traditional Knowledge Module:
Integrates traditional botanical knowledge into the classification process,
ensuring cultural relevance and accuracy.
xxxvii
Networks
CHAPTER 4
SYSTEM REQUIREMENTS
xxxviii
Networks
4 SYSTEM REQUIREMENTS
Requirements are essential constraints and specifications that define the functionality,
performance, and behavior of a system. They serve as the foundation for system
design and development, guiding the creation of software or hardware solutions to
meet specific objectives. The following requirements are crucial to consider during
the development process:
1. Functional requirements
2. Non-Functional requirements
3. Technical requirements
 Hardware requirements
 Software requirements
4.1 Functional requirements
The Functional Requirements section of our software specification outlines the core
functionalities essential for our medicinal plant classification system. Specialized
libraries such as scikit-learn, pandas, NumPy, Matplotlib, and Seaborn are integral to
the system's operation. These libraries provide crucial tools for machine learning
model implementation, data manipulation, numerical computing, and data
visualization. Leveraging these dependencies ensures the system's capability to handle
tasks such as model training, data preprocessing, feature extraction, and result
visualization effectively.
xxxix
Networks
4.2 Non-Functional Requirements
 Process of functional steps:
I. Problem define
II. Preparing data
III. Evaluating algorithms
IV. Improving results
V. Prediction the result
4.3 Technical Requirements
 Software Requirements:
 Operating System: Windows
 Tool: Anaconda with Jupiter Notebook
 Hardware requirements:
 Processor: Pentium IV/III
 Hard disk: minimum 80 GB
 RAM: minimum 2 GB
4.4 Functional Requirements
Functional requirements delineate the specific capabilities and behaviors that the
software system must exhibit. In the context of our project, the functional
requirements encompass various modules responsible for processing, analyzing, and
classifying Indian medicinal leaves.
 4.4.1 Data Collection Module:

xl
Networks
This module focuses on sourcing a diverse dataset of Indian medicinal leaves
images from reliable sources.
Utilizes APIs, web scraping, and data crawling techniques to collect images
from online repositories and databases.
Platforms such as Twitter, Facebook, and Instagram may be tapped to access
publicly available images.
 4.4.2 Data Preprocessing Module:
Cleans and transforms the collected image data to prepare it for analysis.
Removes noise, handles missing data, and normalizes images for consistency.
Applies text preprocessing techniques like tokenization, stemming, and
lemmatization for associated textual data.
 4.4.3 Feature Extraction and Representation Module:
Utilizes transfer learning with pretrained CNN models like VGG16, Xception,
and EfficientNetV2.
Extracts hierarchical features from the medicinal leaf images to represent them
effectively.
Leverages deep learning techniques to capture intricate patterns and details
inherent in leaf images.
 4.4.4 Classification Module:
Classifies the preprocessed leaf images into different categories or classes
representing specific medicinal plant species.

xli
Networks
Employs machine learning algorithms such as Naive Bayes, Support Vector
Machines (SVM), Convolutional Neural Networks (CNNs), or Recurrent
Neural Networks (RNNs) for classification.
Validates the model's performance using appropriate evaluation metrics such
as accuracy, precision, recall, and F1-score.
 4.4.5 Visualization and Reporting Module:
Generates informative visualizations and reports to present the classification
results. Includes word clouds, sentiment distribution charts, topic distribution
plots, and interactive dashboards. Facilitates the interpretation and
communication of insights derived from the classification process.
 4.4.6 Recommendations and Interventions Module:
Based on the analysis and insights gained, generates actionable
recommendations and interventions to promote the use of medicinal plants.
Targets policymakers, healthcare professionals, and conservationists to
enhance awareness and conservation efforts.
5. Non-Functional Requirements
Non-functional requirements define the quality attributes and constraints of the
software system. They encompass aspects such as accessibility, availability, security,
and performance.
 Accessibility:
xlii
Networks
Ensures that the system is accessible to users with disabilities by adhering to
accessibility standards such as WCAG (Web Content Accessibility
Guidelines).
 Availability:
Guarantees high availability of the system, minimizing downtime and ensuring
uninterrupted access to users.
 Security:
Implements robust security measures to protect sensitive data and prevent
unauthorized access, ensuring compliance with privacy regulations.
 Backup and Disaster Recovery:
Establishes regular data backups and implements disaster recovery procedures
to mitigate the risk of data loss or system failure.
 Performance:
Measures the system's performance in terms of speed, responsiveness, and
scalability. Ensures efficient resource utilization and optimal performance
under varying workloads.
 Interoperability:
Ensures seamless integration with existing systems and interoperability with
external applications and databases.
6. Performance Requirements
Performance requirements specify the expected behavior and performance metrics of
the software system.

xliii
Networks
 6.1 Response Time:
Ensures that the system responds promptly to user interactions, with minimal
latency.
 6.2 Throughput:
Maintains high throughput to handle concurrent requests and process a large
volume of image data efficiently.
 6.3 Scalability:
Scales horizontally and vertically to accommodate increasing data volumes
and user traffic.
 6.4 Resource Consumption:
Optimizes resource consumption, including CPU, memory, and storage, to
ensure efficient utilization and cost-effectiveness.
7. Feasibility Study
The feasibility study serves as a comprehensive evaluation of the project's feasibility,
encompassing technical, operational, and economic considerations. It examines the
project's technical feasibility by assessing the availability of necessary resources,
technology readiness, and compatibility with existing systems. Furthermore, it delves
into operational feasibility, evaluating the project's alignment with organizational
objectives, potential impact on workflows, and stakeholders' readiness for adoption.
Lastly, the economic feasibility analysis explores the project's financial viability,
including cost estimation, return on investment projections, and potential revenue
streams
 7.1 Technical Feasibility:

xliv
Networks
Determines whether the proposed technology and infrastructure can support
the project requirements. Evaluates the availability of suitable algorithms,
frameworks, and computing resources for image classification tasks.
 7.2 Operational Feasibility:
Assesses the usability, reliability, and security of the system from an
operational standpoint. Ensures that the system aligns with user expectations
and can seamlessly integrate into existing workflows.
 7.3 Economic Feasibility:
Analyzes the cost-effectiveness of the project, considering factors such as
development costs, infrastructure requirements, and potential return on
investment. By addressing these functional, non-functional, performance, and
feasibility aspects, the project aims to develop a robust and effective solution
for classifying Indian medicinal leaves using transfer learning-based
convolutional neural networks.
xlv
Networks
CHAPTER 5
METHODOLOGY
xlvi
Networks
5.1. Dataset
The image acquisition process in our system is designed to be versatile and robust,
freeing it from the constraints typically associated with conventional image
acquisition models. Unlike traditional methods that may be limited by factors such as
seasonality, time of day, lighting conditions, and background noise, our approach
ensures that images of plants are acquired invariant to these variables. This flexibility
makes our system well-suited for real-time plant analysis investigations, where timely
and accurate data acquisition is crucial for decision-making and research purposes.
The compilation of Indian medicinal plant species included in our dataset is the result
of a comprehensive and collaborative effort involving experts from various fields,
including botany, phytochemistry, and data science. Through intuitive exploration of
the medicinal values associated with indigenous plant species, in collaboration with
botanists, phytochemists, and leveraging data sources from reputable institutions such
as the Botanical Survey of India, we have curated a diverse and representative list of
Indian medicinal plants.
Figure 5.1 provides a detailed overview of the Indian medicinal plant species that
have contributed to our work. This list encompasses a wide range of plant species
known for their medicinal properties, spanning various botanical families, geographic
regions, and traditional healing practices. Each plant species included in the dataset
offers unique therapeutic benefits and has been selected based on its relevance to
traditional medicine, ethnobotanical knowledge, and documented pharmacological
properties.
xlvii
Networks
Figure 5.1 sample dateset
By incorporating a diverse array of Indian medicinal plant species into our dataset, we
aim to facilitate research, education, and innovation in the field of herbal medicine
and plant-based therapies. Our comprehensive dataset serves as a valuable resource
for researchers, practitioners, and enthusiasts interested in exploring the potential of
indigenous plants for healthcare, wellness, and conservation initiatives. Through
collaborative efforts and interdisciplinary collaboration, we strive to advance our
understanding of traditional medicine and promote the sustainable utilization of plant
resources for the betterment of human health and the environment.
xlviii
Networks
5.2 Data Preparation
In the classification of Indian medicinal leaves using transfer learning-based
convolutional neural networks, the data preprocessing procedure typically involves
several steps to prepare the raw leaf images for input into the neural network. Here's a
generalized outline of the data preprocessing procedure:
 Data Collection:
Collect a diverse dataset of Indian medicinal leaves, ensuring a representative
sample across different species. Images can be obtained through various sources
such as online databases, field photography, or botanical gardens.
 Data Cleaning and Annotation:
Remove any irrelevant or corrupted images from the dataset.
Annotate each image with its corresponding class label (i.e., the medicinal plant
species).
Ensure consistency in labeling and annotation across the dataset.
 Data Augmentation:
Augment the dataset to increase its size and variability, which helps improve the
model's generalization ability.
Common augmentation techniques include rotation, flipping, scaling, cropping,
and brightness/contrast adjustments.
Ensure that augmented images retain their original semantic meaning (e.g., a
xlix
Networks
flipped image of a leaf should still represent the same plant species).
 Image Resizing:
Resize all images to a uniform size suitable for input into the neural network.
Commonly used sizes include 224x224 or 299x299 pixels, depending on the input
size expected by the pre-trained CNN architecture (e.g., VGG16, Xception,
EfficientNetV2).
 Normalization:
Normalize pixel values to a common scale to facilitate convergence during model
training.
Typically, pixel values are scaled to the range [0, 1] or [-1, 1].
 Train-Validation-Test Split:
Split the dataset into three subsets: training, validation, and testing.
The training set is used to train the model, the validation set is used for
hyperparameter tuning and model evaluation during training, and the test set is
used to assess the final model's performance.
 Data Loading:
Implement data loaders or generators to efficiently load batches of preprocessed
images during model training and evaluation.
Shuffle the training data to introduce randomness and prevent the model from
learning spurious correlations.
l
Networks
 Transfer Learning Initialization:
Initialize the pre-trained CNN model with weights learned from a large dataset
(e.g., ImageNet).
Freeze the weights of initial layers (excluding the classification layers) to prevent
them from being updated during the initial training phase.
By following these preprocessing steps, the dataset of Indian medicinal leaves is
appropriately prepared for classification using transfer learning-based convolutional
neural networks. This ensures that the model can effectively learn meaningful patterns
from the images and generalize well to unseen data.
5.3. Model Development
In the model development phase, our primary focus lies in adapting pre-trained
convolutional neural network (CNN) architectures to the specialized task of medicinal
plant leaf recognition. Given the inherent complexities of leaf imagery and the
relatively limited size of our dataset, transfer learning emerges as a compelling
strategy. Initially, we meticulously assess the suitability of prominent CNN
architectures such as VGG16, Xception, and EfficientNetV2. This assessment
considers both their architectural intricacies and their performance on analogous
image recognition tasks.
Transfer learning leverages the hierarchical representations learned by these models
on large-scale datasets like ImageNet. We capitalize on this by extracting relevant
features from the convolutional layers of these pre-trained models. This feature
li
Networks
extraction process equips our models with the ability to discern intricate patterns and
characteristics specific to medicinal plant leaves. Moreover, we develop a fine-tuning
strategy tailored to the unique characteristics of each CNN architecture. This strategy
strikes a delicate balance between leveraging the prior knowledge embedded in the
pre-trained models and adapting to the nuances of our specific dataset.
A key aspect of our approach involves customizing the output layer of each model to
align with the number of classes in our dataset. This ensures compatibility for
classification tasks, allowing our models to accurately categorize medicinal plant
leaves into their respective classes. By meticulously fine-tuning the pre-trained CNN
architectures and optimizing their configurations, we aim to develop robust and
accurate models capable of effectively recognizing and classifying medicinal plant
leaves, thereby contributing to advancements in botanical research and conservation
efforts.
Transfer Learning with CNN Architectures:
Model Selection: Evaluate the suitability of VGG16, Xception, and
EfficientNetV2 for the task based on their architectural characteristics and
performance on similar image recognition tasks. Based on these evaluations,
all three architectures show promise for the task of medicinal plant leaf
recognition. VGG16 offers simplicity and competitive performance, while
Xception and EfficientNetV2 provide more sophisticated architectural designs
that prioritize efficiency without compromising accuracy. The final choice
may depend on factors such as computational resources, dataset size, and
specific performance requirements.
lii
Networks
liii
Networks
5.3.1 VGG16 (Visual Geometry Group 16):
VGG16, developed by the Visual Geometry Group (VGG) at the University of
Oxford, is a convolutional neural network (CNN) architecture renowned for its
simplicity and effectiveness in image recognition tasks. Introduced in the paper "Very
Deep Convolutional Networks for Large-Scale Image Recognition" by Karen
Simonyan and Andrew Zisserman, VGG16 comprises a total of 16 layers, hence its
name. These layers are organized into a sequence of convolutional and fully
connected layers, with convolutional layers responsible for feature extraction and
fully connected layers for classification.
The key characteristic of VGG16 is its uniform architecture, where the convolutional
layers all have a small receptive field of 3x3 pixels and a stride of 1. This uniformity
simplifies the architecture and makes it easier to understand and implement.
Additionally, VGG16 employs max-pooling layers to downsample feature maps,
reducing spatial dimensions and extracting increasingly abstract features as the data
flows through the network.
Despite its simplicity, VGG16 has demonstrated remarkable performance on various
image recognition benchmarks, including the ImageNet Large Scale Visual
Recognition Challenge (ILSVRC). By stacking multiple convolutional layers, VGG16
can learn hierarchical representations of visual features, enabling it to recognize
objects and patterns in images with high accuracy. Its success has led to widespread
adoption and served as a foundation for more advanced CNN architectures.
liv
Networks
Overall, VGG16's simplicity, uniform architecture, and strong performance make it a
popular choice for image recognition tasks, serving as a benchmark for evaluating the
effectiveness of more complex CNN architectures.
Methodology
VGG16 architecture comprises a series of convolutional layers followed by max-
pooling layers, with fully connected layers at the end for classification. The
convolutional layers learn hierarchical features from input images, while the fully
connected layers perform classification based on these features.
 Features of VGG16
Deep Architecture: VGG16 has a deep architecture with multiple convolutional
layers, allowing it to learn intricate patterns and features from images effectively.
 Uniform Structure: The architecture of VGG16 is uniform, with small 3x3
convolutional filters and max-pooling layers, resulting in a consistent feature
extraction process across all layers.
 Transfer Learning: VGG16 is often used for transfer learning, where the pre-
trained model is fine-tuned on a specific dataset to perform tasks such as
image classification, object detection, and image segmentation.
 Accuracy: VGG16 has demonstrated high accuracy on various image
recognition tasks, making it a popular choice for research and practical
applications.
 Parameters in VGG16
 Input shape
 Number of convolutional layers
 Number of filters in each convolutional layer

lv
Networks
 Size of max-pooling windows
 Number of units in fully connected layers
 Why VGG16?
VGG16's deep architecture and uniform structure make it well-suited for learning
complex features from images, resulting in high accuracy on various image
recognition tasks.
Fig 5.3.1 Architecture of VGG16
lvi
Networks
5.3.2 Xception:
Xception is a convolutional neural network architecture proposed by François Chollet,
the creator of the Keras deep learning library. It draws inspiration from the Inception
architecture but introduces a novel approach to convolutional neural networks (CNNs)
by employing depthwise separable convolutions.
The key innovation of Xception lies in its use of depthwise separable convolutions,
which decouple the spatial and channel-wise operations in standard convolutions.
This separation allows the network to capture spatial relationships more efficiently
while reducing computational complexity. By replacing traditional convolutional
layers with depthwise separable convolutions, Xception achieves a significant
reduction in the number of parameters and computations compared to previous
architectures.
Another notable feature of Xception is its deep architecture, consisting of a series of
depthwise separable convolution layers followed by global average pooling and fully
connected layers. This architecture enables Xception to learn hierarchical features
from input images effectively. Additionally, the network employs skip connections to
facilitate gradient flow during training and alleviate the vanishing gradient problem,
promoting better convergence and learning stability.
Xception has demonstrated superior performance on various image recognition tasks,
including object classification, detection, and segmentation. Its efficient architecture
and high accuracy make it well-suited for deployment in resource-constrained

lvii
Networks
environments or applications requiring real-time processing. Moreover, Xception's
versatility and scalability make it a valuable asset for researchers and practitioners in
the field of computer vision.
Methodology
Xception architecture introduces depthwise separable convolutions, which separate
the spatial and channel-wise convolutions, significantly reducing the number of
parameters and computational complexity compared to traditional convolutional
layers.
 Features of Xception:
Depthwise Separable Convolutions: Xception replaces traditional convolutions with
depthwise separable convolutions, resulting in a more efficient and lightweight
architecture with fewer parameters and lower computational cost.
 Extreme Depth: Xception achieves extreme depth by stacking multiple depthwise
separable convolutional layers, allowing it to learn hierarchical features from input
images effectively.
 Transfer Learning: Xception is well-suited for transfer learning tasks, where the pre-
trained model is fine-tuned on specific datasets to perform tasks such as image
classification and object detection.
 Parameters in Xception
 Input shape
 Depth multiplier
 Number of convolutional layers
 Number of filters in each convolutional layer
 Activation functions
lviii
Networks
 Why Xception?
Xception's efficient architecture and depthwise separable convolutions make it
suitable for deployment on resource-constrained devices and applications requiring
high computational efficiency while maintaining high accuracy.
Fig 5.3.2 Architecture of Xception
lix
Networks
5.3.3 EfficientNetV2:
EfficientNetV2 is an improved version of the EfficientNet architecture, introduced by
Google researchers. It aims to further optimize model efficiency while maintaining
high accuracy in image classification tasks.EfficientNetV2 represents an evolution of
the original EfficientNet architecture, which was designed to balance model
efficiency and accuracy by scaling the network's depth, width, and resolution.
Introduced by Google researchers, EfficientNetV2 builds upon this foundation to
achieve even greater efficiency and performance in image classification tasks.
One of the key enhancements in EfficientNetV2 is the adoption of a new compound
scaling method, which optimizes model scaling across multiple dimensions
simultaneously. This approach ensures that the network's depth, width, and resolution
are scaled in a harmonious manner, leading to improved performance without
significantly increasing computational costs.
EfficientNetV2 also introduces novel architectural innovations, including the use of
improved building blocks such as inverted residuals and squeeze-and-excitation
blocks. These building blocks enhance the network's representational capacity and
enable it to capture more complex patterns and features from input images.
Moreover, EfficientNetV2 incorporates techniques to enhance training efficiency and
regularization, such as stochastic depth regularization and weight standardization.
These techniques improve the robustness and generalization capability of the network,
leading to better performance on diverse datasets and tasks.

lx
Networks
Overall, EfficientNetV2 represents a state-of-the-art architecture for image
classification, offering a compelling combination of efficiency, accuracy, and
scalability. Its advancements make it a valuable tool for various computer vision
applications, including fake news detection, where efficient processing of multimedia
content is crucial.
 Methodology
EfficientNetV2 builds upon the success of the original EfficientNet architecture
by incorporating novel design techniques such as advanced compound scaling,
enhanced model architecture, and improved training strategies.
 Features of EfficientNetV2
Compound Scaling: EfficientNetV2 employs compound scaling to simultaneously
optimize the depth, width, and resolution of the neural network architecture. This
allows for better utilization of computational resources and improved model
performance across various hardware platforms.
 Model Architecture
EfficientNetV2 features a streamlined architecture with carefully designed
building blocks, including inverted residual blocks and linear bottleneck layers.
These architectural enhancements contribute to both efficiency and accuracy
improvements.
 Improved Training Strategies: EfficientNetV2 incorporates advanced
training strategies such as regularization techniques, data augmentation, and
learning rate scheduling to enhance model generalization and prevent
overfitting.
lxi
Networks
 Parameters in EfficientNetV2
 Depth scaling coefficient
 Width scaling coefficient
 Resolution scaling coefficient
 Number of layers
 Number of channels in each layer
 Activation functions
 Regularization parameters
 Why EfficientNetV2?
EfficientNetV2 stands out as a pinnacle of innovation in the realm of deep
learning models, boasting a harmonious blend of efficiency and accuracy that
caters to the diverse needs of both resource-constrained environments and
large-scale image classification endeavors. Its emergence represents a
significant leap forward in the field, thanks to its sophisticated design
principles and streamlined architecture.
One of the most striking features of EfficientNetV2 is its unparalleled
efficiency in resource utilization, a critical aspect in today's computational
landscape. By leveraging advanced design techniques such as compound
scaling, neural architecture search, and attention mechanisms, EfficientNetV2
achieves remarkable performance while efficiently managing computational
resources. This efficiency is particularly advantageous for scenarios where
hardware limitations or cost considerations come into play, allowing for
lxii
Networks
seamless integration into a wide range of applications and platforms.
Moreover, EfficientNetV2's superior accuracy adds another layer of appeal, ensuring
that its deployment translates into tangible benefits in terms of task performance and
model effectiveness. Through meticulous optimization and fine-tuning, the model
achieves state-of-the-art results across various benchmark datasets and real-world
applications. This blend of efficiency and accuracy makes EfficientNetV2 a
compelling choice for practitioners seeking to maximize the impact of their machine
learning solutions while optimizing resource allocation.
In essence, EfficientNetV2 represents a paradigm shift in the design philosophy of
convolutional neural networks, setting new standards for model efficiency and
performance. Its ability to strike a delicate balance between resource utilization and
accuracy makes it a versatile tool in the hands of researchers, developers, and
practitioners alike, poised to drive innovation and unlock new possibilities in the field
of deep learning.
lxiii
Networks
Fig 5.3.3 Architecture of EffiecientNetV2
5.4 Performance Comparison Charts:
In this section, we present performance comparison charts to evaluate the efficacy of
different pre-trained models for feature extraction on medicinal plant image datasets.
The comparative analysis focuses on three widely used models: VGG16, Xception,
and EfficientNetV2. These models were selected based on their architectural
characteristics and previous performance on similar image recognition tasks.
lxiv
Networks
Fig 5.4.Plots of model loss of (a) Xception, (b) VGG16, (c) EfficientNetV2
 Feature Extraction: Extract features from the pre-trained CNN models'
convolutional layers, capturing hierarchical representations of leaf images.
 Fine-tuning Strategy: Determine the extent of fine-tuning required for each
model, considering factors such as dataset size, domain similarity, and
computational resources.
 Model Architecture Modification: Adapt the output layer of each CNN
architecture to match the number of classes in the medicinal plant leaf dataset,
ensuring compatibility for classification.
 Model Customization and Optimization:
lxv
Networks
o Hyperparameter Tuning: Optimize hyperparameters such as learning
rate, batch size, and regularization techniques to enhance model
performance and convergence.
o Regularization Techniques: Implement regularization techniques
such as dropout and weight decay to prevent overfitting and improve
generalization.
o Activation Functions: Experiment with different activation functions
(e.g., ReLU, Leaky ReLU, or SELU) to enhance the non-linearity of
the model and facilitate feature learning.
5.5 Training and Evaluation
In the training and evaluation phase, we embark on the iterative process of training
our modified CNN models and assessing their performance. Guided by principles of
optimization and regularization, we fine-tune the hyperparameters of our models to
achieve optimal convergence and generalization. Throughout the training procedure,
we closely monitor the models' performance on the validation set, employing early
stopping mechanisms to prevent overfitting and expedite convergence. Additionally,
we implement learning rate scheduling techniques to adaptively adjust the learning
rate during training, enhancing the models' adaptability to varying gradients. Upon
completion of the training process, we subject our models to rigorous evaluation using
established classification metrics such as accuracy, precision, recall, and F1-score. By
analyzing confusion matrices and ROC curves, we gain insights into the models'
classification behavior and their ability to discriminate between different plant
species. Furthermore, we explore cross-validation techniques to assess model

lxvi
Networks
generalization and mitigate biases introduced by random initialization. Finally, we
prioritize computational efficiency and scalability by optimizing model size,
exploring parallelization strategies, and leveraging GPU acceleration to expedite
training and inference.
5.5.1 Model Training:
 Training Procedure: Train the modified CNN models using the training
dataset, employing iterative optimization algorithms such as stochastic
gradient descent (SGD) or Adam.
 Validation Monitoring: Monitor the models' performance on the validation
set during training epochs to detect signs of overfitting or underfitting.
 Early Stopping: Implement early stopping criteria based on validation
performance to halt training when model performance ceases to improve,
preventing unnecessary computational overhead.
 Learning Rate Scheduling: Employ learning rate scheduling techniques such
as exponential decay or cyclic learning rates to dynamically adjust the learning
rate during training and facilitate faster convergence.
5.5.2 Model Evaluation:
 Performance Metrics: Evaluate model performance using standard
classification metrics including accuracy, precision, recall, and F1-score to
assess classification effectiveness across different plant species.
 Confusion Matrix Analysis: Generate confusion matrices to visualize model
predictions and identify classes with high confusion rates, providing insights
into model behavior and misclassifications.
 ROC Curve Analysis: Compute receiver operating characteristic (ROC)

lxvii
Networks
curves and area under the curve (AUC) scores to assess the models' ability to
discriminate between classes and determine optimal decision thresholds.
 Cross-Validation: Employ cross-validation techniques such as k-fold cross-
validation to assess model generalization across different data splits and
mitigate biases introduced by random initialization.
5.5.3 Computational Efficiency and Scalability:
 Model Optimization: Implement model optimization techniques such as
model pruning, quantization, or knowledge distillation to reduce model size
and inference latency while preserving classification performance.
 Parallelization: Explore distributed training strategies and GPU acceleration
to improve training efficiency and scalability, enabling faster experimentation
and model iteration.
5.6. Algorithms
 Random Forest:
Random Forest is a versatile ensemble learning algorithm that can be highly
effective for classification tasks, including the classification of Indian medicinal
leaves. It works by constructing multiple decision trees during training and
outputs the mode of the classes of the individual trees. In your project, Random
Forest can be applied by extracting relevant features from leaf images and using
them as input to train the Random Forest classifier. Features such as leaf shape,
texture, and color can be extracted using image processing techniques. The
algorithm can handle a large number of input features and is robust to overfitting,
lxviii
Networks
making it suitable for complex datasets like medicinal leaf images.
 Support Vector Machines (SVM):
SVM is a powerful supervised learning algorithm commonly used for
classification tasks. It works by finding the hyperplane that best separates the
classes in the feature space. In your project, SVM can be utilized by transforming
the extracted features from leaf images into a high-dimensional space and finding
the optimal hyperplane to separate different classes of medicinal leaves. SVM is
effective in handling both linearly separable and non-linearly separable datasets,
making it suitable for classifying complex leaf patterns and textures.
 K-Nearest Neighbors (KNN):
KNN is a simple yet effective classification algorithm that classifies new instances
based on the majority class of their k nearest neighbors in the feature space. In
your project, KNN can be applied by first extracting relevant features from leaf
images and then computing the similarity between each pair of images. During
classification, the class label of a test image can be determined based on the class
labels of its nearest neighbors. KNN is particularly useful when the underlying
data distribution is not well-defined and can handle both linear and non-linear
decision boundaries.
 Logistic Regression:
Logistic Regression is a classical linear model commonly used for binary
classification tasks. However, it can also be extended to multiclass classification
lxix
Networks
using techniques like one-vs-rest or softmax regression. In your project, logistic
regression can be employed to classify medicinal leaves by learning a linear
decision boundary in the feature space. Features extracted from leaf images can be
used as input to train the logistic regression model, and the model can output the
probability of each class for a given input image. Logistic regression is
computationally efficient and interpretable, making it suitable for applications
where model interpretability is important.
 Decision Trees:
Decision Trees are intuitive and interpretable models that recursively partition the
feature space into regions based on feature values. They make predictions by
traversing the tree from the root to the leaf nodes, where each leaf node
corresponds to a class label. In your project, decision trees can be applied by using
features extracted from leaf images to construct a tree-based classifier. Decision
trees are robust to outliers and can handle both numerical and categorical features,
making them suitable for classifying medicinal leaves with diverse characteristics.
lxx
Networks
 Deep Learning Architectures:
Deep learning architectures like convolutional neural networks (CNNs) are state-
of-the-art models for image classification tasks. In your project, you are already
utilizing transfer learning-based CNNs such as VGG16, Xception, and
EfficientNetV2. These CNN architectures can learn hierarchical representations of
image features and achieve high accuracy in classifying medicinal leaves.
Additionally, other deep learning architectures like recurrent neural networks
(RNNs) or attention mechanisms can be explored if your dataset includes
sequential data or text-based features associated with medicinal plants.
5.7 Key Performance Indicators
This section defines the metrics (or KPIs, Key Performance Indicators) used to be
able to evaluate the results of the algorithms used [49]:
 Accuracy
Numeric value indicating the performance of the predictive model. It is calculated
as follows:
(1)
where:
TP: True Positive. Result in which the model correctly predicts the positive class.
FP: False Positive. Result in which the model incorrectly predicts the positive
class.
TN: True Negative. Result in which the model correctly predicts the negative class.
FN: False Negative. Result in which the model incorrectly predicts the negative
lxxi
Networks
class.
 Kappa statistic
It measures the agreement between two examiners in their corresponding
classifications of N elements into C mutually exclusive categories. In the case
of machine learning, it refers to the actual class and the class expected by the
model used. It is calculated as follows:
(2)
where:
Pr (a) is the observed relative agreement between observers, and Pr (e) is
the hypothesized probability of agreement by chance, using the observed
data to compute the probabilities that each observer randomly ranks each
category. If raters fully agree, then κ = 1. If there is no agreement between
raters other than what would be expected by chance (as defined by Pr (e)),
then κ = 0.
 Logarithmic Loss
It is the negative average of the log of corrected predicted probabilities for
each instance. It is calculated as follows:
(3)
where:
N is the number of samples.

lxxii
Networks
M is the number of classes.
yij, indicates if the sample i belongs to the class j or not.
pij, indicates the probability that the sample i belongs to the class j.
 Error
The error gives an indication of how far the predictions are from the actual
output. There are two formulas: Mean Absolute Error (MAE) and Mean
Squared Error (MSE). It is calculated as follows:
(4)
(5)
where:
N corresponds to the total number of samples.
𝑦𝑘 corresponds to the class indicated by the classification model.
𝑦𝑘̂^ corresponds to the actual class.
 Sensitivity
The sensitivity of a model (or the ratio of true positives) measures the
proportion of correctly classified positive examples. The total number of
positives is the sum of those that were correctly classified and those that were
incorrectly classified. It is calculated as follows:
lxxiii
Networks
(6)
where:
TP: True Positive. Result in which the model correctly predicts the positive
class.
TN: True Negative. Result in which the model correctly predicts the negative
class.
FN: False Negative. Result in which the model incorrectly predicts the
negative class.
 Specificity
The specificity of a model (or the ratio of true negatives) measures the
proportion of correctly classified negative examples. The total number of
negatives is the sum of those that were correctly classified and those that were
incorrectly classified. It is calculated as follows:
(7)
where:
TP: True Positive. Result in which the model correctly predicts the positive class.
FP: False Positive. Result in which the model incorrectly predicts the positive class.
TN: True Negative. Result in which the model correctly predicts the negative class.
 Precision
Precision is defined as the proportion of examples classified as positive that

lxxiv
Networks
are actually positive. That is, when a model predicts values as positive. It is
calculated as follows:
(8)
where:
class.
class.
FN: False Negative. Result in which the model incorrectly predicts the
negative class.
 Recall
Recall is defined as the number of correctly classified positives over the total
number of positives. This formula is the same as that for sensitivity. It is
calculated as follows:
(9)
where:
class.
FP: False Positive. Result in which the model incorrectly predicts the positive
class.
lxxv
Networks
class.
 F-measure
F-measure is a measure of model performance that combines precision and
recall into a value called F-measure (also called F1 score or F-score). This
measure combines precision and recall using harmonic averaging, which is
used for ratios. This type of averaging is used instead of arithmetic, since
precision and recall are expressed as proportions between 0 and 1, which can
be interpreted as ratios. It is calculated as follows:
lxxvi
Networks
CHAPTER 6
SYSTEM DESIGN
lxxvii
Networks
6.1 System Architecture:
Figure 6.1 System Architecture
In the classification of Indian medicinal leaves using transfer learning-based
convolutional neural networks (CNNs), the system architecture plays a pivotal role in
achieving accurate and efficient recognition. The architecture typically comprises
several interconnected layers, including convolutional layers, pooling layers, fully
connected layers, and an output layer. These layers work in tandem to extract
hierarchical features from input leaf images and map them to the corresponding
medicinal plant classes. Transfer learning is integrated into the architecture by
leveraging pre-trained CNN models such as VGG16, ResNet, or EfficientNet, which
have been trained on large-scale datasets like ImageNet. By initializing the network
lxxviii
Networks
with weights learned from these datasets, the system can effectively capture generic
features from natural images, enhancing its ability to classify medicinal plant leaves
with limited labeled data.
Moreover, fine-tuning layers of the pre-trained CNN models allows for the adaptation
of features to the specific characteristics of Indian medicinal leaves. This fine-tuning
process involves adjusting the weights of certain layers during training on the target
dataset while keeping the weights of earlier layers frozen to retain the learned generic
features. Additionally, data augmentation techniques such as rotation, flipping, and
scaling are often applied to artificially increase the diversity of the training dataset
and improve the robustness of the model against variations in leaf images. The final
layers of the architecture are typically configured to output class probabilities using
softmax activation, enabling the classification of medicinal plant leaves into their
respective categories.
Furthermore, the system architecture may include mechanisms for model evaluation
and optimization to ensure the reliability and scalability of the classification system.
Cross-validation techniques such as k-fold validation may be employed to assess the
generalization performance of the model on unseen data. Hyperparameter tuning
methods such as grid search or random search may also be utilized to optimize the
performance of the CNN architecture. Additionally, the deployment of the system
architecture can be facilitated through integration with web or mobile applications,
allowing users to easily access the classification functionality for identifying Indian
medicinal leaves in real-world scenarios. Overall, a well-designed system architecture
lxxix
Networks
is essential for effectively leveraging transfer learning-based CNNs in the
classification of Indian medicinal leaves, enabling accurate and reliable recognition of
plant species for various applications in healthcare, biodiversity conservation, and
traditional medicine research.
6.2 UML Diagrams
6.2.1 Use Case diagram
A use case diagram in the Unified Modeling Language (UML) is a type of behavioral
diagram defined by and created from a Use-case analysis
lxxx
Networks
Fig 6.2.1 Use Case Diagram from Data loading and Exploratio
lxxxi
Networks
Use Cases:
 User/Administrator:
 Login: The user or administrator logs into the system using their credentials to
access the platform for classifying Indian medicinal leaves.
 Upload Image: Users upload images of medicinal leaves to the system for
classification.
 Preprocess Image: Upon uploading, the system preprocesses the images to
enhance their quality and standardize them for feature extraction.
 Extract Features from Image: The system extracts relevant features from the
uploaded images, such as leaf shape, texture, and color, to prepare them for
classification.
 Build Convolutional Neural Network (CNN) Model: The system constructs
a CNN model using transfer learning techniques, integrating pretrained models
like VGG16, Xception, and EfficientNetV2.
 Train CNN Model: The system trains the CNN model using the extracted
features from the uploaded images to learn the distinguishing patterns of
different medicinal plant leaves.
 Evaluate CNN Model: After training, the system evaluates the performance
of the CNN model using validation datasets to assess its accuracy and
generalization capability.
 Predict Plant Species: Users can input new images of medicinal leaves, and
the system utilizes the trained CNN model to predict the species of the plants
 depicted in the images.
 Display Prediction Results: The system displays the predicted plant

82
Networks
species along with confidence scores to users, enabling them to identify
the medicinal plants accurately.
 Extract Features from Image: The prediction model extracts features from
the uploaded images of medicinal leaves to prepare them for classification.
 Build Convolutional Neural Network (CNN) Model: The prediction model
constructs a CNN model using transfer learning techniques, incorporating
pretrained models like VGG16, Xception, and EfficientNetV2.
 Train CNN Model: The prediction model trains the CNN model using the
extracted features from the images to learn the distinguishing patterns of
different medicinal plant leaves.
 Evaluate CNN Model: After training, the prediction model evaluates the
performance of the CNN model using validation datasets to assess its accuracy
and generalization capability.
 Predict Plant Species: The prediction model utilizes the trained CNN model
to predict the species of medicinal plants depicted in new images uploaded by
users.
 Display Prediction Results: After prediction, the prediction model displays
the predicted plant species along with confidence scores, facilitating accurate
identification of medicinal plants.
 Associations:
 User/Administrator to Use Cases: The user/administrator interacts with the system
by performing various use cases such as login, uploading datasets, and viewing
predicted outputs. Prediction Model to Use Cases: The prediction model performs
tasks related to feature extraction, providing features to the ML model, predicting and
83
Networks
analyzing datasets, and displaying predicted outputs and visualizations. This use case
diagram provides an overview of the interactions between different actors and
components in the online payment fraud detection system using machine learning.
6.2.2 Sequence diagram:
A sequence diagram in Unified Modeling Language (UML) is a kind of interaction
diagram that shows how processes operate with one another and in what order. It is a
construct of a Message Sequence Chart. Sequence diagrams are sometimes called
event diagrams, event scenarios, and timing diagrams.
Figure 6.2.2 Sequence diagram

84
Networks
85
Networks
 Interactions:
 Actor User: Represents the user interacting with the system.
 User Interface (UI): The interface through which the user interacts with
the system.
 Input Image: Represents the component responsible for handling input
images.
 Image Preprocessing: Handles the preprocessing of the input image
before classification.
 Classification: Performs the classification of the preprocessed image.
 Feature Extraction: Extracts features from the input image for
classification.
 Performance Measurement: Measures the performance of the
classification process.
 Events:
 The User uploads an image through the User Interface.
 The User Interface forwards the uploaded image data to the Input Image
component.
 The Input Image component receives the image data.
 The Input Image component sends the image data to the Image
Preprocessing component to preprocess it.
 The Image Preprocessing component preprocesses the image data.
 The Preprocessing component forwards the preprocessed image to the
Classification component for classification.
 The Classification component performs the classification of the

86
Networks
preprocessed image.
 After classification, the Classification component sends the image data to
the Feature Extraction component to extract features.
 The Feature Extraction component extracts features from the image data.
 The Feature Extraction component forwards the extracted features to the
Performance Measurement component to calculate accuracy.
6.2.3 Activity Diagram for splitting of data
Activity diagrams serve as invaluable tools in the realm of software engineering,
providing graphical representations of workflows comprising stepwise activities,
actions, and decisions. With built-in support for choice, iteration, and concurrency,
these diagrams offer a comprehensive view of the business and operational processes
within a system.
Within the Unified Modeling Language (UML), activity diagrams serve as
indispensable assets for describing the intricate step-by-step workflows of various
system components. Whether it's depicting the sequence of tasks in a software
application or illustrating the operational procedures in a business process, activity
diagrams offer a visual framework for understanding the flow of control.
At their core, activity diagrams encapsulate the overall flow of control within a
system or process, delineating the sequence of activities and the transitions between
them. By leveraging symbols and notations such as actions, decisions, forks, joins,
and swimlanes, these diagrams provide a clear and intuitive representation of complex
87
Networks
workflows.
In essence, activity diagrams serve as powerful communication tools, facilitating
collaboration among stakeholders and enabling a shared understanding of system
behavior. Whether used during the design phase to conceptualize workflows or during
the implementation phase to guide development efforts, activity diagrams play a
crucial role in the software development lifecycle, ensuring clarity, coherence, and
efficiency in the development process.
Fig 6.2.3 Activity Diagram for Splitting of data
88
Networks
 User Uploads Image:
 Activity: User uploads an image of a medicinal plant leaf for
classification.
 Decision: Check if the uploaded image is valid.
 If valid, proceed to the next activity.
 If invalid, prompt the user to upload a valid image.
 Image Preprocessing:
 Activity: Preprocess the uploaded image to enhance its quality and prepare
it for classification.
 Resize image: Ensure the image dimensions are suitable for processing.
 Normalize image: Adjust pixel values to a standard range for consistency.
 Denoise image: Remove noise and artifacts from the image.
 Decision: Check if preprocessing is successful.
 If successful, proceed to the next activity.
 If unsuccessful, terminate the process and notify the user.
 Feature Extraction:
 Activity: Extract features from the preprocessed image using transfer
learning techniques.
 Decision: Check if feature extraction is successful.
 Classification:
 Activity: Classify the preprocessed image using a pre-trained CNN model
89
Networks
(e.g., VGG16, Xception, EfficientNetV2).
 Decision: Check if classification is successful.
 Display Results:
 Activity: Display the classification results to the user.
6.3.4 Class diagram:
A class diagram is a type of structural diagram in UML (Unified Modeling Language)
that represents the static structure of a system by showing the classes of the system,
their attributes, methods, and the relationships among them. It provides a graphical
representation of the system's architecture, emphasizing the classes and their
associations, aggregations, and inheritances.
90
Networks
Figure 6.3.4 Class diagram
 Input Image:
 Description: Represents the input image data to be classified.
 Attributes: image_data: Image - The raw image data to be processed.
 Responsibilities: Provide the raw image data for preprocessing and
classification.
 Image Preprocessing:
 Description: Handles the preprocessing of input images before
classification.
91
Networks
 Methods: preprocess_image(image_data: Image): Image - Preprocesses
the input image data.
 Responsibilities: Apply preprocessing techniques to enhance image
quality and prepare it for classification.
 Classification:
 Description: Performs the classification of preprocessed images using
convolutional neural networks (CNNs).
 Attributes: model: CNNModel - The trained CNN model for image
classification.
 Methods: classify_image(image_data: Image): ClassificationResult -
Classifies the input image.
 Responsibilities: Utilize the trained CNN model to classify input images
into different categories.
 Performance Measurement:
 Description: Evaluates the performance of the classification system.
 Attributes: results: List<ClassificationResult> - List of classification
results for evaluation.
 Methods: calculate_accuracy(): float - Calculates the accuracy of the
classification system.
 Responsibilities: Assess the accuracy of the classification results to
measure the performance of the system.
92
Networks
 Feature Extraction:
 Description: Extracts relevant features from preprocessed images.
 Attributes: features: List<Feature> - List of extracted features from
images.
 Methods: extract_features(image_data: Image): List<Feature> - Extracts
features from input images.
 Responsibilities: Identify and extract meaningful features from
preprocessed images for classification.
 Clustering Image:
 Description: Groups images based on extracted features for further
analysis.
 Attributes: clusters: List<Cluster> - List of clusters formed by grouping
similar images.
 Methods: cluster_images(features: List<Feature>): List<Cluster> -
Performs clustering of images based on extracted features.
 Responsibilities: Cluster images with similar features to facilitate analysis
and visualization
93
Networks
CHAPTER 7
IMPLEMENTATION AND
EVALUATION
94
Networks
7 IMPLEMENTATION
The implementation of our medicinal plant classification project involved several key
steps and components. Initially, we imported necessary libraries such as TensorFlow
and Keras for deep learning, along with additional libraries for data manipulation and
visualization. Utilizing Google Colab as our development environment, we imported
the dataset and performed data exploration. This included plotting sample images for
visualization and understanding the characteristics of the dataset. Subsequently, we
conducted a train-test split to prepare the data for model training and evaluation.
Augmentation techniques were applied to enhance the diversity of the training dataset
and improve model generalization. We then imported the VGG19 pre-trained model
for transfer learning, leveraging its learned representations for feature extraction. The
model architecture was defined, and a callback for early stopping was implemented to
prevent overfitting. After training the model, we evaluated its performance on the test
dataset using various evaluation metrics such as precision, recall, and F1-score.
Additionally, we experimented with ensemble techniques, including average
ensemble, to further improve classification accuracy. Finally, we conducted testing on
our own images to assess the model's real-world applicability.
The utilization of transfer learning techniques played a pivotal role in our approach,
with the VGG19 pre-trained model serving as the backbone for feature extraction. By
leveraging the learned representations from the VGG19 model, we were able to
effectively capture and utilize high-level features relevant to medicinal plant
classification tasks, thereby reducing the need for extensive training on limited
datasets
95
Networks
96
Networks
7.1 Dataset
In our dataset acquisition process, we aimed to create a comprehensive repository of
Indian medicinal plant images that could serve as a valuable resource for various
research and analysis endeavors. Unlike conventional image acquisition methods, our
approach prioritized flexibility and adaptability, allowing us to overcome common
constraints such as seasonal variations, time constraints, lighting conditions, and
background interference.
One of the key strengths of our dataset is its robustness in capturing images of
medicinal plants across different seasons, times of day, and environmental conditions.
By ensuring variability in these factors, we aimed to create a dataset that reflects the
diverse and dynamic nature of plant growth and development in real-world scenarios.
This variability enhances the generalizability and applicability of our dataset to a wide
range of research applications, including real-time plant analysis and monitoring.
To compile the list of Indian medicinal plant species included in our dataset, we
collaborated closely with experts in botany, phytochemistry, and botanical research.
Leveraging their expertise and insights, we curated a comprehensive list of medicinal
plant species known for their therapeutic properties and traditional medicinal uses in
the Indian context. Additionally, we utilized data sources from reputable institutions
such as the Botanical Survey of India to validate and enrich our dataset with accurate
botanical information.
Figure 6.1 provides a detailed overview of the Indian medicinal plant species that
contributed to our dataset. This list serves as a valuable reference for researchers,
97
Networks
botanists, and enthusiasts interested in exploring the rich diversity of medicinal plants
in India. By incorporating a diverse range of plant species with documented medicinal
values, our dataset aims to facilitate interdisciplinary research and promote the
discovery of novel therapeutic compounds and medicinal formulations
Figure 7.1 list of some medicinal plants species
98
Networks
99
Networks
7.2 Import modules
Pandas used to perform data manipulation and analysis
numpy used to perform a variety of mathematical operations on arrays
matplotlib used for data visualization and graphical plotting
seaborn built on top of matplotlib with similar functionalities
os used to handle files using system commands
tqdm progress bar decorator for iterators
warnings to manipulate warnings details, filterwarnings('ignore')
load_img used for loading the image as numpy array
tensorflow backend module for the use of Keras
Dense single dimension linear layer
Activation layer for the use of certain threshold
Flatten convert a 2D array into a 1D array
Conv2D convolutional layer in 2 dimension
MaxPooling2D function to get the maximum pixel value to the next layer
100
Networks
7.3 Plot model architecture
In the process of developing our medicinal plant classification models, one crucial
aspect involved visualizing the architecture of the neural networks employed.
Visualizing the model architecture provides valuable insights into the network's
structure, including the arrangement of layers, the flow of data through the network,
and the parameters associated with each layer.
To plot the model architecture, we utilized various visualization techniques and tools
available in popular deep learning frameworks such as TensorFlow and Keras. These
tools offer built-in functions for generating graphical representations of neural
network architectures, allowing us to gain a comprehensive understanding of the
model's design and complexity.
The plotted model architecture typically includes detailed information about each
layer in the network, such as the type of layer (e.g., convolutional, pooling, dense),
the number of neurons or filters, the activation functions used, and any additional
parameters or configurations. This visual representation serves as a useful reference
for understanding how the model processes input data and makes predictions.
Furthermore, plotting the model architecture facilitates communication and
collaboration among team members, enabling researchers, developers, and
stakeholders to discuss and refine the network design effectively. By visualizing the
architecture, we can identify potential areas for optimization, fine-tuning, or
modification to enhance the model's performance and efficiency.
101
Networks
In summary, plotting the model architecture is an essential step in the development
and evaluation of deep learning models for medicinal plant classification. It provides
a clear and intuitive depiction of the network structure, aiding in model interpretation,
optimization, and decision-making throughout the development lifecycle.
102
Networks
Figure 7.3 architecture model
103
Networks
7.4 Feature Extraction model
In the feature extraction model phase of our project, we embarked on a
comprehensive exploration to assess the efficacy of various pre-trained convolutional
neural network (CNN) architectures for feature extraction tasks on our medicinal
plant image datasets. Our primary objective was to identify the most suitable model
that could effectively capture and represent the intricate visual patterns and
characteristics inherent in medicinal plant images.
To facilitate this comparative analysis, we selected three prominent pre-trained CNN
models: VGG16, VGG19, and MobileNetV2. These models are renowned for their
versatility, robustness, and effectiveness in diverse image recognition tasks, making
them ideal candidates for feature extraction in our project.
The experimentation process involved meticulously designing and implementing
workflows to extract features from our medicinal plant image datasets using each of
the selected pre-trained models. We leveraged the learned representations encoded
within the convolutional layers of these models to extract meaningful and
discriminative features from the input images.
For each pre-trained model, we meticulously fine-tuned the feature extraction process
to optimize performance and ensure compatibility with our specific dataset
characteristics. This involved configuring the input preprocessing pipeline, defining
the layers for feature extraction, and selecting appropriate parameters to tailor the
feature extraction process to our project requirements.
Upon extracting features from the medicinal plant images using VGG16, VGG19, and
104
Networks
MobileNetV2, we conducted in-depth analyses and comparisons to evaluate their
effectiveness. Key metrics such as feature richness, discriminative power, and
computational efficiency were considered to assess the overall performance of each
model in capturing relevant information from the input images.
Furthermore, we explored visualization techniques to gain insights into the extracted
features and their underlying representations. Visualization methods such as t-SNE (t-
distributed stochastic neighbor embedding) and PCA (principal component analysis)
were employed to visualize the high-dimensional feature representations in lower-
dimensional spaces, facilitating a deeper understanding of the feature extraction
process.
Through meticulous experimentation and analysis, we aimed to identify the pre-
trained CNN model that offered the optimal balance of feature richness,
computational efficiency, and effectiveness in capturing the distinctive visual
characteristics of medicinal plant images. The insights gained from this feature
extraction model phase laid the groundwork for subsequent stages of our project,
including classification.
105
Networks
Figure 7.4: feature extraction model
106
Networks
Figure 7.5 first image to predict
The results of the medicinal plant classification process are presented. This includes
showcasing the actual and predicted values obtained from the classification model.
The visualization provides insights into the model's performance and its ability to
accurately classify medicinal plant images.
Actual plant species serve as ground truth labels against which the model predictions
are evaluated, while predicted values correspond to the plant species inferred from the
input images.
A confusion matrix is included, offering a tabular representation of the actual and
predicted values organized by class labels. Additionally, various performance metrics
such as accuracy, precision, recall, F1-score, and classification report are computed to
assess the model's efficacy.
The visualization aids in visually analyzing the agreement between actual and
predicted values across different plant species, enabling stakeholders to evaluate the
107
Networks
model's accuracy and reliability in classifying medicinal plant images.
Figure 7.6 leave prediction with actual and predicted values.
108
Networks
CHAPTER 8
SOURCE CODE
109
Networks
8 SOURCE CODE
# Importing Libraries
import os
import matplotlib.pyplot as plt
import numpy as np
import PIL
import tensorflow as tf
import tensorflow_datasets as tfdf
from tensorflow import keras
from tensorflow.keras import layers, models
from tensorflow.keras.models import Sequential
# Define Constants
BATCH_SIZE = 32
img_height = 224
img_width = 224
CHANNEL = 3
data_dir = r"F:\Project\archive\Indian Medicinal Leaves Image Datasets\Medicinal
Leaf dataset"
# Load Dataset
dataset = tf.keras.utils.image_dataset_from_directory(
data_dir,
shuffle=True,
image_size=(img_height, img_width),
batch_size=BATCH_SIZE
)
class_names = dataset.class_names
# Train Test Split

def get_dataset_partitions_tf(ds, train_split=0.7, val_split=0.15, test_split=0.15,
shuffle=True, shuffle_size=10000):
assert (train_split + test_split + val_split) == 1
ds_size = len(ds)
if shuffle:
ds = ds.shuffle(shuffle_size, seed=12)
train_size = int(train_split * ds_size)

val_size = int(val_split * ds_size)
train_ds = ds.take(train_size)
val_ds = ds.skip(train_size).take(val_size)
test_ds = ds.skip(train_size).skip(val_size)
110
Networks
return train_ds, val_ds, test_ds
train_ds, val_ds, test_ds = get_dataset_partitions_tf(dataset)

train_ds = train_ds.cache().shuffle(1000).prefetch(buffer_size=tf.data.AUTOTUNE)
val_ds = val_ds.cache().shuffle(1000).prefetch(buffer_size=tf.data.AUTOTUNE)
test_ds = test_ds.cache().shuffle(1000).prefetch(buffer_size=tf.data.AUTOTUNE)
# Data Preprocessing
resize_and_rescale = tf.keras.Sequential([
layers.experimental.preprocessing.Resizing(img_height, img_height),
layers.experimental.preprocessing.Rescaling(1./255),])
data_augmentation = tf.keras.Sequential([
layers.experimental.preprocessing.RandomFlip("horizontal"),
layers.experimental.preprocessing.RandomRotation(0.2),
layers.experimental.preprocessing.RandomZoom(0.3),])
train_ds = train_ds.map(
lambda x, y: (data_augmentation(x, training=True), y)
).prefetch(buffer_size=tf.data.AUTOTUNE)
# Model Definition
feature_extractor = tf.keras.applications.VGG16(input_shape=(img_height,
img_height, 3),
include_top=False,
weights="imagenet")
for layer in feature_extractor.layers:

layer.trainable = False
model = keras.Sequential([
resize_and_rescale,
layers.BatchNormalization(),
feature_extractor,
layers.BatchNormalization(),
layers.Flatten(),
layers.Dense(256, activation='relu'),
layers.Dropout(0.2),
layers.Dense(80, activation='softmax')])
model.compile(
optimizer=tf.keras.optimizers.Adam(learning_rate=1e-4),
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False),
metrics=['accuracy'])
# Model Training
earlystopping = tf.keras.callbacks.EarlyStopping(monitor='val_accuracy',
111
Networks
mode='max', patience=7, verbose=1)

callback_list = [earlystopping]
model_history = model.fit(
train_ds,
validation_data=val_ds,
epochs=150,
callbacks=callback_list,
verbose=1
)
# Model Evaluation
model.evaluate(test_ds)
# Save Model
model.save("vgg16_extractor_1.keras")
# Plot Accuracy and Loss

acc = model_history.history['accuracy']
val_acc = model_history.history['val_accuracy']
loss = model_history.history['loss']
val_loss = model_history.history['val_loss']
epochs_range = range(48)
plt.figure(figsize=(8, 8))
plt.subplot(1, 2, 1)
plt.plot(epochs_range, acc, label='Training Accuracy')
plt.plot(epochs_range, val_acc, label='Validation Accuracy')
plt.legend(loc='lower right')
plt.title('Training and Validation Accuracy')
plt.subplot(1, 2, 2)
plt.plot(epochs_range, loss, label='Training Loss')
plt.plot(epochs_range, val_loss, label='Validation Loss')
plt.legend(loc='upper right')
plt.title('Training and Validation Loss')
plt.show()
112
Networks
CHAPTER 9
SYSTEM TESTING
113
Networks
9 TESTING
Testing is a critical phase in the development lifecycle of any software project,
ensuring that the system functions as intended and meets the specified requirements.
In the context of our medicinal plant classification project, testing encompasses
various aspects, including software functionality, algorithm accuracy, system
performance, error handling, and integration with other components.
9.1 Software Testing:
Software testing involves verifying the functionality, reliability, and accuracy of the
medicinal plant classification system. It encompasses several types of testing, each
targeting specific aspects of the system's behavior and performance.
 System Testing:
System testing involves testing the system as a whole to ensure that all
components function correctly when integrated. This includes various testing
phases such as module testing, integration testing, acceptance testing, and user
acceptance testing (UAT).
 Module Testing:
Module testing focuses on testing individual modules or components of the
system in isolation. This ensures that each module functions correctly and
produces the expected outputs when provided with appropriate inputs.
114
Networks
 Integration Testing:
Integration testing validates the interactions between different modules or
components of the system. It ensures that data is exchanged correctly between
modules and that the integrated system behaves as expected.
 Acceptance Testing:
Acceptance testing involves testing the system with end-users to ensure that it
meets their requirements and expectations. This includes verifying that the
system fulfills the original goals and objectives established during the analysis
phase
9.2 Test Cases:
Test cases play a crucial role in ensuring the effectiveness and reliability of the
medicinal plant classification system. These test cases encompass diverse scenarios
and conditions that thoroughly evaluate the system's functionality, performance, and
behavior across different aspects.
Input data validation test cases are designed to verify the system's ability to handle
various types of input data, ensuring that it can process different formats of leaf
images accurately. Algorithm testing test cases assess the performance and accuracy
of the classification algorithms employed by the system, validating their ability to
correctly identify and categorize medicinal plant species based on leaf characteristics
System performance evaluation test cases focus on assessing the scalability,
responsiveness, and resource utilization of the classification system. These tests
115
Networks
involve processing large volumes of leaf images to evaluate the system's performance
under varying workloads and conditions. Error handling test cases validate the
system's ability to detect and gracefully handle errors or anomalies that may occur
during the classification process. This includes scenarios such as corrupted image
files, incomplete data, or unexpected inputs.
Reporting and visualization test cases ensure that the system generates informative
and visually appealing reports and visualizations of the classified plant images. These
tests assess the accuracy, completeness, and clarity of the generated reports, providing
users with meaningful insights into the classification results.Integration and
compatibility test cases verify the seamless integration of the classification system
with other components or platforms, such as data storage systems, APIs, or
visualization tools. These tests ensure that the system can effectively communicate
and collaborate with external systems or services, enabling smooth interoperability
and compatibility.
 Input Data Validation Test Cases:
Input data validation test cases are designed to ensure that the system can
handle different types of input data effectively. This includes validating
various formats of image data, such as JPEG, PNG, or BMP, and verifying
that the system gracefully handles invalid or corrupted images.
 Machine Learning Algorithm Test Cases:
Machine learning algorithm test cases focus on evaluating the accuracy and
performance of the classification algorithms used in the system. This involves
116
Networks
testing the algorithms with diverse datasets, including images of different
plant species, variations in lighting conditions, and background clutter.
 System Performance Test Cases:
System performance test cases assess the scalability and responsiveness of the
medicinal plant classification system. These tests involve processing a large
volume of images to evaluate the system's response time, resource utilization,
and scalability under varying loads.
 Boundary and Error Handling Test Cases:
Boundary and error handling test cases validate the system's behavior when
encountering extreme or unexpected scenarios. This includes testing the
system's response to images with excessive noise, blurry or distorted images,
and errors in image metadata.
 Reporting and Visualization Test Cases:
Reporting and visualization test cases ensure that the system generates
informative and visually appealing reports and visualizations of the classified
plant images. This involves validating the accuracy and completeness of the
reports and verifying that they provide meaningful insights into the
classification results.
 Integration and Compatibility Test Cases:
Integration and compatibility test cases verify the seamless integration of the
medicinal plant classification system with other components, such as data
storage systems, APIs, or visualization tools. These tests ensure compatibility
117
Networks
with different operating systems, web browsers, and platforms on which the
system will be deployed.
Relative Test Cases:

 Upload Dataset
 Test Case Description: Verify if the dataset containing medicinal plant
images is successfully uploaded into the system.
 Test Steps:
 Click on the "Upload Dataset" button.
 Select the dataset file from the local directory.
 Confirm if the system displays a message indicating that the dataset has
been successfully loaded.
 Expected Result: The system should display a message confirming the
successful upload of the dataset.
 Actual Result: Dataset loading message is displayed upon successful
upload.
 Preprocessing
 Test Case Description: Ensure that preprocessing techniques are applied to
the dataset before model training.
 Test Steps: Check if preprocessing steps such as image resizing,
normalization, and augmentation are performed.
 Verify if the preprocessed dataset contains the necessary data for further
processing.
118
Networks
 Expected Result: The system should apply preprocessing techniques to
enhance the quality and suitability of the dataset for model training.
 Actual Result: Preprocessed dataset contains necessary data for further
processing.
 Model Prediction
 Test Case Description: Validate if the trained machine learning models
successfully predict the class labels of medicinal plant images.
 Test Steps: Input a sample image into the trained model for prediction.
 Verify if the model generates the predicted class label for the input image.
 Expected Result: The model should accurately predict the class label of the
input image.
 Actual Result: Model generates the predicted class label for the input image.
 Recommendation
 Test Case Description: Ensure that recommended actions or information
based on model predictions are displayed to the user.
 Test Steps: Check if the system displays recommendations related to the
predicted class label of the input image.
 Expected Result: The system should present relevant recommendations or
information based on the predicted class label.
 Actual Result: System displays recommendations related to the predicted
class label.
119
Networks
 Visualization
 Test Case Description: Validate if visualizations such as charts or graphs are
generated to represent classification results.
 Test Steps: Verify if the system generates visualizations depicting the
classification performance or distribution of medicinal plant classes.
 Expected Result: The system should produce informative visualizations
representing the classification results.
 Actual Result: Visualizations depicting classification performance are
generated.
In conclusion, the system testing phase of the medicinal plant leaves classification
project has been pivotal in ensuring the robustness, reliability, and accuracy of the
developed system. Through rigorous testing procedures and comprehensive
evaluation methodologies, we have validated the effectiveness of our classification
algorithms in accurately identifying and categorizing medicinal plant species based on
leaf images. The testing process encompassed various scenarios, including input data
validation, algorithm performance evaluation, and system scalability assessment,
allowing us to identify and address potential issues and limitations. Overall, the
successful completion of the system testing phase instills confidence in the reliability
and efficacy of our medicinal plant classification system, paving the way for its
deployment and utilization in real-world applications.
120
Networks
CHAPTER 10
CONCLUSION AND FUTURE SCOPE
121
Networks
10.1 CONCLUSION
In conclusion, the medicinal plant classification project has represented a significant
step forward in leveraging machine learning techniques for the accurate identification
and categorization of various plant species based on their images. Through meticulous
experimentation and model refinement, we have attained commendable performance
metrics across multiple evaluation criteria, including accuracy, precision, recall, and
F1-score. The implementation of transfer learning models, such as VGG16, VGG19,
and MobileNetV2, has proven instrumental in extracting meaningful features from
medicinal plant images, thereby empowering the models to make precise predictions
with high confidence.
Furthermore, the integration of visualization techniques has played a pivotal role in
facilitating the comprehensive analysis of model performance and classification
outcomes. These visualizations have provided invaluable insights into the strengths
and weaknesses of different machine learning models, enabling us to refine our
approach and optimize the classification system further. The successful deployment of
the trained models in a production environment underscores the practical utility of the
developed system, opening avenues for its potential utilization in diverse real-world
applications, including herbal medicine research, botanical conservation efforts, and
pharmaceutical industry endeavors.
MobileNetV2 obtained the best validation accuracy with fine-tuning with a validation
accuracy of 96.02%. MobileNetV2 also obtained the best testing accuracy with fine-
tuning of 81.82%. In other models, the testing accuracy obtained is far below
MobileNetV2. This condition is likely to happen because the dataset for training and
122
Networks
validation used is less diverse and general to build a good model, so the resulting
model overfits the dataset. In this case, MobileNetV2 with fine-tuning is quite able to
overcome the weaknesses of the dataset used so that when new testing data is given,
the accuracy results obtained are quite good. In addition, based on the experiment
results, fine-tuning the model can improve the accuracy of the validation and testing
produced. The limitation of this study is that it ignores the background problem of the
leaf image. Further research will be carried out using a dataset with a complex
background by adding a segmentation method before being processed by the
MobileNetV2 model.
10.2 FUTURE SCOPE

While the current project has achieved its primary objectives, there remains ample
room for future exploration and enhancement:
 Dataset Expansion and Diversity: Expand the dataset to encompass a
broader spectrum of medicinal plant species, encompassing variations in
growth stages, geographical regions, and ecological niches. A more diverse
dataset would enhance the models' ability to generalize to unseen plant
samples and improve classification accuracy across different botanical
families.
 Advanced Model Architectures: Investigate the adoption of cutting-edge
deep learning architectures, such as transformers and graph neural networks,
to capture intricate spatial relationships and semantic features inherent in
123
Networks
medicinal plant images. Exploring novel model architectures tailored to the
unique characteristics of botanical imagery could unlock new avenues for
performance improvement.
 Domain-Specific Data Augmentation: Develop domain-specific data
augmentation strategies tailored to the idiosyncrasies of medicinal plant
images, including variations in leaf morphology, lighting conditions, and
background clutter. Augmenting the dataset with synthetically generated
images using techniques like Generative Adversarial Networks (GANs) could
help alleviate data scarcity issues and enhance model robustness.
 Active Learning and Annotation Strategies: Implement active learning
algorithms to intelligently select informative samples for annotation,
optimizing the annotation process and maximizing the utility of labeled data.
Leveraging human-in-the-loop annotation frameworks and crowdsourcing
platforms could streamline the data labeling pipeline and accelerate model
training iterations.
 Mobile Application Deployment: Explore the deployment of the
classification system in mobile applications, enabling users to identify
medicinal plants on-the-go using smartphone cameras. Developing an intuitive
mobile interface with image capture functionality could democratize access to
plant identification tools and foster greater engagement among researchers,
educators, and citizen scientists.
124
Networks
 Cross-Disciplinary Collaboration: Foster interdisciplinary collaborations
with botanists, ethnobotanists, and traditional medicine practitioners to
validate the accuracy of the classification system and enrich the model's
understanding of botanical taxonomy and medicinal properties. Integrating
domain expertise into the model development process can enhance the
interpretability of model predictions and ensure alignment with real-world
botanical knowledge.
125
Networks
126
Networks
REFERENCES
127
Networks
[1] X.-F. Wang, D.-S. Huang, J.-X. Du, H. Xu, and L. Heutte, “Classification of
plant leaf images with complicated background,” Appl. Math. Comput., vol. 205, no.
2, pp. 916–926, Nov. 2008.
[2] K. Yang, W. Zhong, and F. Li, “Leaf Segmentation and Classification with a
Complicated Background Using Deep Learning,” Agronomy, vol. 10, no. 11, p. 1721,
Nov. 2020.
[3] I. Ranggadara, Y. S. Sari, S. Dwiasnati, and I. Prihandi, “A Review of

Implementation and Obstacles in Predictive Machine Learning Model at Educational
Institutions,” in Journal of Physics: Conference Series, 2020, vol. 1477, p. 32019.
[4] Y. Jumaryadi, D. Firdaus, B. Priambodo, and Z. P. Putra, “Determining the Best

Graduation Using Fuzzy AHP,” in 2020 2nd International Conference on Broadband
Communications, Wireless Sensors and Powering (BCWSP), 2020, pp. 59–63.
[5] E. M. Jawad, H. G. Daway, and H. J. Mohamad, “Quantum-dot Cellular

Automata Based Lossless CFA Image Compression Using Improved and Extended
Golomb-rice Entropy Coder,” Int. J. Intell. Eng. Syst., vol. 15, no. 2, pp. 12–25, Apr.
2022.
[6] V. Ayumi, “Pose-based human action recognition with Extreme Gradient

Boosting,” in 2016 IEEE Student Conference on Research and Development
(SCOReD), 2016, pp. 1–5.
[7] V. Ayumi, “Mobile Application for Monitoring of Addition of Drugs to Infusion

Fluids,” Int. J. Sci. Res. Comput. Sci. Eng. Inf. Technol., pp. 48– 56, Nov. 2019.
[8] V. Ayumi, “Pose-based human action recognition with Extreme Gradient

Boosting,” in Proceedings - 14th IEEE Student Conference on Research and
Development: Advancing Technology for Humanity, SCOReD, 2017. [9] V. Ayumi,
“Application of Machine Learning for SARS-CoV-2 Outbreak,” Int. J. Sci. Res. Sci.
128
Networks
Eng. Technol., vol. 7, no. 5, pp. 241–248, Oct. 2021.
129
Networks
[10] V. Ayumi, “Performance Evaluation of Support Vector Machine Algorithm for

Human Gesture Recognition,” Int. J. Sci. Res. Eng. Technol., vol. 7, no. 6, pp. 2020.
[11] J. Arun Pandian, G. Geetharamani, and B. Annette, “Data Augmentation on

Plant Leaf Disease Image Dataset Using Image Manipulation and Deep Learning
Techniques,” in 2019 IEEE 9th International Conference on Advanced Computing
(IACC), 2019, pp. 199–204.
[12] J. S. Cope, D. Corney, J. Y. Clark, P. Remagnino, and P. Wilkin, “Plant species

identification using digital morphometrics: A review,” Expert Syst. Appl., vol. 39, no.
8, pp. 7562–7573, Jun. 2012.
[13] Y. A. Putri, E. C. Djamal, and R. Ilyas, “Identification of Medicinal Plant

Leaves Using Convolutional Neural Network,” J. Phys. Conf. Ser., vol. 1845, no. 1, p.
012026, Mar. 2021.
[14] M. A. F. Azlah, L. S. Chua, F. R. Rahmad, F. I. Abdullah, and S. R. Wan

Alwi, “Review on Techniques for Plant Leaf Classification and Recognition,”
Computers, vol. 8, no. 4, p. 77, Oct. 2019.
[15] F. Ghouse, K. Paranjothi, and R. Vaithiyanathan, “Dysgraphia Classification

based on the Non-Discrimination Regularization in Rotational Region Convolutional
Neural Network,” Int. J. Intell. Eng. Syst., vol. 15, no. 1, Feb. 2022.
[16] J. Lu, L. Tan, and H. Jiang, “Review on Convolutional Neural Network (CNN)
Applied to Plant Leaf Disease Classification,” Agriculture, vol. 11, no. 8, p. 707, Jul.
2021.
[17] E. Korot et al., “Code-free deep learning for multi-modality medical image
classification,” Nat. Mach. Intell., vol. 3, no. 4, pp. 288–298, Apr. 2021.
[18] V. Ayumi, E. Ermatita, A. Abdiansah, H. Noprisson, M. Purba, and M. Utami,

“A Study on Medicinal Plant Leaf Recognition Using Artificial Intelligence,” in 2021
International Conference on Informatics, Multimedia, Cyber and Information System
(ICIMCIS, 2021, pp. 40–45.
130
Networks
[19] E. Hidayat, Lukman, H. Noprisson, D. I. Sensuse, Y. G. Sucahyo, and E. D. Putra,
“Development of mobile application for documenting traditional knowledge in
Indonesia: A Case Study of Traditional Knowledge in Using Medicinal Plant,” in
Proceedings - 14th IEEE Student Conference on Research and Development:
Advancing Technology for Humanity, SCOReD 2016, 2017.
[20] H. Noprisson, D. I. Sensuse, Y. G. Sucahyo, and Lukman, “Metadata

Development for Ethnophytomedicine Resources Using Metadata Analysis
Approach,” in The 2016 8th International Conference on Information Technology and
Electrical Engineering (ICITEE 2016), 2016.
[21] D. I. Sensuse, H. Noprisson, Y. G. Sucahyo, and L. Lukman, “Metadata

Schema for Traditional Knowledge,” Int. J. Adv. Sci. Eng. Inf. Technol., vol. 7, no. 6,
2017.
[22] Y. G. Naresh and H. S. Nagendraswamy, “A novel fuzzy LBP based symbolic

representation technique for classification of medicinal plants,” in 2015 3rd IAPR
Asian Conference on Pattern Recognition (ACPR), 2015, pp. 524–528.
[23] G. Mukherjee, A. Chatterjee, and B. Tudu, “Study on the potential of combined

GLCM features towards medicinal plant classification,” in 2016 2nd International
Conference on Control, Instrumentation, Energy & Communication (CIEC), 2016, pp.
98–102.
[24] D. Venkataraman and N. Mangayarkarasi, “Support vector machine based

classification of medicinal plants using leaf features,” in 2017 International
Conference on Advances in Computing, Communications and Informatics (ICACCI),
2017, pp. 793–798.
[25] L. Gao and X. Lin, “A method for accurately segmenting images of medicinal
plant leaves with complex backgrounds,” Comput. Electron. Agric., vol. 155, pp.
426–445, 2018.
131
Networks
[26] C. Sivaranjani, L. Kalinathan, R. Amutha, R. S. Kathavarayan, and K. J.

Jegadish Kumar, “Real-Time Identification of Medicinal Plants using Machine
Learning Techniques,” in 2019 International Conference on Computational
Intelligence in Data Science (ICCIDS), 2019, pp. 1–4.
[27] D. Pechebovicz et al., “Plants recognition using embedded Convolutional Neural

Networks on Mobile devices,” in 2020 IEEE International Conference on Industrial
Technology (ICIT), 2020, pp. 674–679.
[28] M. R. Bhuiyan, M. Abdullahil-Oaphy, R. S. Khanam, and M. S. Islam,

“MediNET: A Deep Learning Approach to Recognize Bangladeshi Ordinary
Medicinal Plants Using CNN,” in Soft Computing Techniques and Applications,
Springer, 2021, pp. 371–380.
[29] S. Bhattacharya, A. Mukherjee, and S. Phadikar, “A deep learning approach for

the classification of rice leaf diseases,” in Intelligence enabled research, Springer,
2020, pp. 61–69.
[30] S. Roopashree and J. Anitha, “Medicinal leaf dataset,” Mendeley Data, vol. 1,
2020.
132

Medical Leave Images

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Medical Leave Images

Uploaded by

Copyright:

Available Formats

Classification Of Indian Medicinal Leaves Using Transfer Learning Based Convolutional Neural

CLASSIFICATION OF INDIAN MEDICINAL

In partial fulfillment of the requirements for

Master of Computer Applications

ULLAMPARTHI DILEEP KUMAR

Under the guidance of

DEPARTMENT OF COMPUTER SCIENCE & ENGINEERING

I hereby declare that the entire thesis work entitled "

CLASSIFICATION OF INDIAN MEDICINAL LEAVES USING

TRANSFER LEARNING BASED CONVOLUTIONAL NEURAL

NETWORKS" is being submitted to the Department of Computer

Science and Engineering, University College of Sciences, Acharya

Nagarjuna University, in partial fulfillment of the requirement for the

award of the degree of Master of Computer Applications (MCA) is a

bon-fide work of my own, carried out under the supervision of Dr. G.

NEELIMA, Assistant Professor, Department of Computer Science &

Engineering, Acharya Nagarjuna University.

earlier by me or others for the award of any degree in any University.

ULLAMPARTHI DILEEP KUMAR

Reg. No. Y23MC20060

ACHARYA NAGARJUNA UNIVERSITY

This is to certify that this project entitled “CLASSIFICATION OF

INDIAN MEDICINAL LEAVES USING TRANSFER LEARNING

BASED CONVOLUTIONAL NEURAL NETWORKS” is a Bon-fide

record of the project work done and submitted by

ULLAMPARTHI DILEEP KUMAR (Y23MC20060) during the year

2022 - 2024 in partial fulfillment of the requirements for the award of

degree of Master of Computer Applications (MCA) in the department

of Computer Science & Engineering. I certify that he carries this

project as an independent project under my guidance.

Head of the Department Project Guide

from many people.

discussions and valuable suggestions which have contributed greatly to the

improvement of the thesis.

I am thankful to and fortunate enough to get constant encouragement, support and

guidance from all Teaching staffs of Department which helped us in successfully

the non-teaching staff of the department for their timely support.

medicinal plants. Classification of medicinal leaves can be automated based on their

medicinal leaves classifying utilizing transfer learning based convolutional neural

using different architectures such as VGG16, Xception, and efficieNetv2. The

architecture performed better than EfficientNetv2 and VGG16.

TABLE OF CONTENTS vii

1.2 Problem Statement 5

1.3 Objectives of the Project 6

Chapter 2: LITERATURE SURVEY

Chapter 3: SYSTEM ANALYSIS

3.1 Problem statement 14

3.2 Existing System 14

3.3 Proposed System 16

Chapter 4: SYSTEM REQUIREMENTS

4.1 Functional Requirements 20

4.2 Non-Functional Requirements 23

4.3 Performance Requirements 24

4.4 Feasibility Study 25

5.2 Data Preparation 30

5.3 Model Development 32

5.4 Performance Comparison Charts 44

5.5 Training and Evalution 45

5.5.1 Model Training 46

5.5.2 Model Evalution 46

5.5.3 Computational Efficiency and Scalability 47

Chapter 6: SYSTEM DESIGN

6.1 System Architecture 57