LEAF DISEASE CLASSIFICATION USING IMAGE RECOGNITION

2 3
1
Senthil Prabhu N Chethan S Jeevanatham S
Department of Computer Science Department of Computer Science Department of Computer Science and
and Engineering, and Engineering, Engineering,
Karpagam Institute of Technology, Karpagam Institute of Technology, Karpagam Institute of Technology,
Coimbatore,India,641021 Coimbatore,India,641021 Coimbatore,India,641021
[Link]@[Link] 21csa13@[Link] 21csa22@[Link]

4 5
Kuganesh S Manikandan N
Department of Computer Science and Department of Computer Science and
Engineering, Engineering,
Karpagam Institute of Technology, Karpagam Institute of Technology,
Coimbatore,India,641021 Coimbatore,India,641021
21csa31@[Link] 21csa34@[Link]

Abstract – The project provides an enhanced automation of diagnostics, as well as greater accuracy in
approach to disease management in tea plantations identifying diseases [3]. Besides, by the developed
using image technology, recommendation system recommendations about the use of fertilizers that are
fertilizer, and easy to use system. These advances in suitable for a specific disease, the project wants to
Roxz algorithms such as transformers for disease enhance the health of the crops, thereby raising the
identification envision accurate diagnosis and yield. Accessibility remains another critical feature
bolstered agricultural productivity. The user interface achieved by Python and helped by Gradio, as such a
for these technologies is pleasant and tea farmers can user interface will enable tea farmers to operate using
conveniently use them since Gradio is in Python. The sophisticated technologies [4]. This is in line with the
problems that the proposed system will solve relate to overall goals of the project in attaining improved
the problems of manual inspection and generalized agricultural results through achieving efficiency,
fertilizer recommendations that are already in place improvements in accessibility, as well as user
and are not very effective in the existing system, thus experience in an effort towards realizing sustainable
providing farmers with specific solutions that help development of tea plantations [5].
improve agricultural results in tea-growing.
Experimental results show that the proposed method 1.2 Scope of the System
is quite efficient as the model EfficientNetB0 attains
striking accuracy in disease differentiation while the
The current study is a part of a large-scale project that
Gradio delivers intuitive interface for ease of dealing
focuses on several key initiatives towards improving
with the system. In summary, this case falls neatly
disease management strategies in tea farming. First, it
into a single and globally accessible solution for
includes implementation of the complex image
enhancing disease management practices on tea
processing technique to assess the diseases existing in
plantations and hence sustainable agriculture.
tea leaves. These diagnostics will have to be integrated
utilizing advanced formulas and machine learning
Keywords models including the likes of Transformers. Moreover,
Gradio, PyTorch, Torchvision, EfficientNetB0, SVM, the project will work on the implementation of
CNNmodels (ConventionalNeuralNetworks), MLP particular recommendation of fertilizers depending upon
classifiers and DL(deep learning) modelling. the various diseases, by using statistical approaches for
improvement of crop quality and yields. One of the key
elements of the initiative is the use of the Gradio for
1. Introduction creating pleasant GUI in Python, so that tea farmers will
easily manage complex technologies. The project will
1.1 Objectives of the Project therefore offer efficiency in diagnostics and precision
in disease detection with a view of getting better results
This aspect of disease management in tea plantations in the production of tea through agriculture. Altogether,
is a major threat to crop productivity, subsequently the specific objectives of the project are associated with
requires an accurate screen and classification of disease diagnosis, access of the user, productivity of the
diseases affecting tea leaves. This initiative agriculture in the field of tea plantation, and improving
appreciates the need to address this challenge sustainability.
holistically [1]. Through the application of
professional image processing technology, 1.3 Existing System
individualized fertilization advice, as well as a clear
and practical menu interface, the project aims to
In the current system of disease control in the tea
disrupt the management of diseases in the tea
producing fields, visual examination is still the most
production industry [2]. By employing transforms
common practice.
such as transformers, the system claims to achieve the
In order to avoid infections, the tea farmers physically because early signs of the diseases may not be visible
assess the leaves in order to identify components such and the diseases progress rapidly and may lead to high
as discoloration, appearance of spots or any abnormal yields loss.
growing pattern on the leaves. Although this approach Secondly, the present system includes generic
is farmer oriented and based on their experience the recommendation of the fertilizer, which cause
process is more subjective and consumes a lot of constraints in the enhancement of the health of the crop
time. Also, human factors in disease diagnosis create and its yield. Such recommendations are usually
problems like misdiagnosis, and wrong disease standard practice or a broad range of practices without
handling or management. Furthermore, some reference to the actual nutrient requirement of diseased
plantations employ simple diagnostic methods such as tea plants. For this reason, fertilization may not correct
hand lenses as the study revealed but these offers little nutrient deficiencies or imbalances that may have been
information on the pathogenesis of the diseases and compounded by disease pressure. This degrades the
often misclassify diseases hence subjectivity will general health and perhaps the resistance of tea crops;
further be enhanced. they become vulnerable to diseases or pests, their output
reduces. Without highly effective nutrient management
The other major feature of the current system is strategies that are specific to the variable nutrient
generic application of fertilizer recommendations. requirements of diseased tea plants in particular, the
These recommendations benchmark on the past current system may not allow for the obtainment of the
practices or standardized principles, which do not better set of agricultural results as well as better
undergo detailing concerning the diseases affecting profitability to the tea growers.
the tea plants. Therefore, care is not taken to see that
these fertilizers reduce disease effects or enhance crop
performances. Additionally, the lack of specific In addition, it shares one disadvantage in that more
nutrient management approaches may lead to the patients lack the advanced technology and resources for
problem of nutrient imbalances that, on the one hand, disease control in the given system. A cross-sectional
worsen disease susceptibility and, on the other hand, study revealed that many tea- growing areas, and
yield potential. especially those in developing or rural environments,
may not possess sufficient physical and human
resources, or funding, to adopt newer technologies
Technology, especially for the processing, to manage including remote sensing, molecular diagnostics, or data
diseases is not easily available to patients especially analysis for disease identification and management.
those living in remote tea growing regions. There is This constraint makes it almost impossible to detect
no infrastructure support and there are hardly any diseases early enough, monitor disease progression as
qualified professionals or financials to support well as prevent and control spread of diseases to reduce
technological adoption such as remote sensing or losses in crops.
molecular diagnostics. This limitation makes effective
disease surveillance as well as timely interventions 1.4 Proposed System
difficult to implement. As a result, coffee farmers
struggle when it comes to the observation of diseases
Therefore, in our proposed system, we used
and in turn, the management of the health of their
Transformers for image classification due to its greater
crops. All these disadvantages in disease management
feature extraction performance compared to CNN
are the major gaps needed to be addressed for
models. More particularly, we employ EfficientNet, a
enhanced agricultural results in tea farming by a one
model that uses Transformer-based feature blocks
stop technological promoted approach earlier and
therein. Such an architecture makes it possible to
targeted recommendations.
decode complex features from the tea images, hence
improving the models capacity for disease detection.
1.3.1 Drawbacks Of Existing System
Through Transformers, we leverage the potential of
taking into consideration long- term signals as a solution
The current disease management system used in the to get a broader view of the monitored data and a more
tea plantations is fraught with several problems which accurate differentiation between diseases.
limits its efficiency [1]. First, reliance on the visual
methods to identify diseases results into
inconsistencies and subjectivity in the entire process. Also, our integration of EfficientNet into the system
IMPRINTS may fail due to human error or guarantees the efficiency and good performing of the
variability, and diagnosis may be either missed or organizations. We especially adjust the parameters and
assigned to the wrong category. This means that structures of the model for images of tea leaves and
vulnerable technical interventions can lead to poor diseases which are more common in tea gardens. While
management practices, as treatments and fine-tuning, hyperparameters, including learning rates
interventions are unlikely to be accurately targeted to and batch sizes, are altered while the overall model
the diseases encountered in tea plants. Furthermore, design is also optimized for highest accuracy in the
manual assessment takes much time especially when extraction of diseases. Moreover, to further improve the
the plants in the plantation are many and an early system and increase efficiency, we apply transfer
application of control measures may not be deployed learning methods for utilizing pre-trained Transformer
models. In the sphere of recommendations for 2. Literature Survey
fertilizers, our system uses data-oriented solutions
related to the geographic climate of India. We use The popularity of deep learning has led to a rising trend
artificial intelligence that helps us enter statistics on of studying image recognition technologies, including
composition of the soil, weather and crop yields to using computerized image classification for plant
develop administering of fertilizer. They have disease diagnosis.
considered nutrient requirement of the tea plant under
different climatic factor to ensure healthy growth and India which is the largest consumer and second largest
yield of crop. We can help tea farmers get the right and producer of tea is at losing end when diseases that harm
targeted advice based on the data analytics and the the health of the tea plant and act as a hinderance to
predictive modeling which meets the plantation needs growth occur. It is known that various diseases are
exactly. aimed at the rate of tea crop growth, so there is a high
requirement for effective disease diagnosis and control
1.4.1 Features of Proposed System measures. This paper seeks to build a powerful deep
CNN known as the LeNet to diagnose the tea plant
A number of novel components are included in the diseases through analyzing a dataset of the images of
proposed system for the purpose of reforming the the leaves. The effectiveness of LeNet, a CNN model,
approaches to the control of diseases and the will be employed within the study to improve the tea
enhancement of the results in tea production. Firstly, diseases diagnostic precision and possibly expand the
our system incorporates the best image classification ability to diagnose diseases of other plants.
capability in existence using Transformers for
identifying diseases in tea leaves. Our system can India being a large exporter of tea, its crops around the
easily compliment Transformer-based models like world suffer significant yield decrease from diseases in
EfficientNet to give a precise classification of tea leaves caused by prolonged pathogen contact. It is
diseases on the visual appearances of tea leaves. This well understood that early disease detection is useful in
particularly enhanced and intricate feature extraction minimizing harm, but the task proves to be a nightmare
network makes it easier and more precise to diagnose when done manually. Based on CNNs, this paper
ailments affecting tea plants hence reducing crop introduces a Deep CNN with multiple hidden layers to
damage and recovery time among tea farmers. assist in classifying the tea leaves to various diseases
with a 96.56% accuracy possible. Based on different
Also, our system provides appropriate fertilizer to disease classes, the proposed model shows the following
plant in tea gardens considering the climate of India specific accuracies of 92%, 92%+0.5%,
and the soil structure of such a field. By using facts 93%+0.5%, 99.10%, 85%, and 90% respectively that are
and AI, we create unique fertilizing schedules considerably higher than the existing methodologies. In
dependent on the concrete disease and other addition, the model has been designed to be
circumstances present. In making these implemented in IoT devices, thereby indicating that the
recommendations, aspects like soil nutrient status, model may not only be used in cases of tea crop
climate and past performance of crops in relation to diseases but also in managing diseases of other crop
diseases are given due consideration for effective plants in future, opening up great opportunities for more
input in growing inputs to feed the tea plants without enhanced management of diseases in agriculture.
compromising disease resistance. In specifically
recommending the right fertilizer to be applied in tea The latest upsurge of image recognition technologies
plants and when to apply it, our system ensures that has led to the extensive application of convolutional
tea farmers can make necessary improvements on neural networks (CNN) for image classification
crop health, realize its productivity potential hence especially to identify plant diseases [3]. In this work, the
improve on the overall sustainability and income goal was to design a new DCNN model called LeafNet
possibility of tea farming. for disease diagnosis of tea plant from its leave images.
In the design of LeafNet, feature extractor filters of
1.5 Software Specification varying sizes were included in order to semi- automate
the extraction of disease features from images. Also,
1. Anaconda (Python environment) DSIFT features were extracted to construct a bag of
visual words model BOVW) that formed the basis of
2. Python 3.8v
disease classification using SVM and MLP classifiers.
3. Chrome Browser We found that LeafNet had a higher overall
classification accuracy than the two algorithms, SVM
1.6 Hardware Specification and Fuzzy MLP algorithms, with the accuracies being
90.16% for LeafNet, 60.62% for SVM and 70.77% for
1. 8 GB RAM 256 ROM MLP. Thus, the obvious high performance of LeafNet in
2. i5 Gen or Ryzen 5 Gen required disease identification confirms the possibility of
improving the effectiveness and reliability of the
1.7 Software Environment disease’s diagnostics in tea plants in future applications.

1. Anaconda – Prompt
This study proposes a method using Convolutional information from soil samples and plant health status to
Neural Network (CNN) with EfficientNet-B0 to identify nutrients that are scarce or in excess concerning
distinguish between organic and non-organic wastes the diseases [1]. Fertilizer recommendations using this
which are relevant in sustainable waste management. analysis may be developed through Machine learning
To this end, it seeks to accurately differentiate algorithms.
between waste types within lens-captured images,
sparing organizers the time and effort of manually
categorizing images. Precise settings of the By making the web application as friendly for the tea
EfficientNet-B0 architecture and CNN parameters are farmers as possible, the best elements of a GUI are
challenging and can be adjusted only with deployed, possibly in the shape of Gradio in Python.
hyperparameters [4]. Hence, the trade- off between Using this interface, farmers can navigate and navigate
efficiency as well as accuracy is fairly impressive in through the system in order to explore the functionality
EfficientNet-B0. The method also proves efficiency of the exports for example disease diagnosis and
by obtaining a very high level of accuracy for test fertilizer advice to improve agricultural results while
cases, 98%. keeping the user in mind.
Third, through automating the it identifies wastes and
thus offers benefits in relation to waste management 3.1 Transformers EfficientNet Bo
mistakes and environmentally sustainable treatment.
EfficientNet-B0 can also be described as an exemplar of
In two empirical away-line officer studies, we tested convolutional neural network (CNN) architecture since
the effectiveness of the most contemporary it is designed as a model that achieves the best balances
EfficientNetB0 model to diagnose malaria parasite in of performance and computational cost[2]. EfficientNet-
blood smears [6]. The used dataset was clean and B0 is by researchers from Google and belongs to the
labelled by professionals from the Lister Hill National family known as EfficientNet, which tries to address the
Centre for Biomedical Communications (LHNCBC). computational requirements of the models. The structure
In contrast to other strategies, the presented approach of EfficientNet-B0 also employs a compound scaling
directly vented on the feature extraction capability of approach to scale up the network’s depth, width, and
the EfficientNetB0-reference model. For the transfer resolution uniformly by a set of fixed scalars.
learning approach, we have increased the initial
feature sets and decreased training time. After this, we
retrained the model to reflect our identified layers and
used them to further train the model on our selected
dataset. We identified the best generalization of the
results with a throughput of 94.70% after fifty epochs,
followed by a close second of 94.68% after only ten
epochs. Furthermore, further visualization and
analysis using the Grad-CAM algorithm with the fine-
tuned EfficientNetB0 showed better infection
detection than other modern DCNN models. The
findings of this research are that the latest
EfficientNetB0 model achieves state-of-art accuracy,
and fine-tuning it provides extremely efficient deep
learning solutions for malaria parasite detection
directly from blood smears, without the need for
Fig 1 Efficient B0 Archiecture
complicated optimization, augmentation, or weighting
functions.
The structure of EfficientNet-B0 begins with several
3. Proposed Methodology sets of operations made of convolutional layers in
addition to batch normalization, activation functions
Procedures that involve the use of advanced image like ReLU, and pooling layers. But, it is important to
technology, the provision of customized fertilizer note that the EfficientNet-B0 uses depth-wise separable
recommendation and a well-developed application convolution, which simplifies the standard convolution
interface. This technology entails using advanced operation into depth and point convolution
image technology, which probably entails using operation. This type of architectural decision greatly
sophisticated cameras or drone like instruments with minimizes the number of parameters and calculations
specific sensors for disease identification and needed, decreased while not affecting the performance.
differentiation on tea leaves. Computer aided Further, EfficientNet-B0 use of squeeze-and-excitation
algorithms evaluate captured images in order to (SE) block adapts channel-wise feature response during
identify diseases symptoms and ease the process of effective training to increase only the importance and
diagnostic. Also, individual nutrient recommendations hence improving the accuracy of the model [13].
particular to diseases are established using
Due to the EfficientNet-B0’s modularity and
scalability, it can be applied to a broad range of
computer vision applications involving image
classification, object detection and semantic
segmentation. It has a high-performance to size ratio,
which makes it both powerful and economical with
resources, so it is a highly useful tool for practical
usage, especially in highly computational and
possibly other computationally intensive cases.

4. Architecture Diagram

Fig 3 Training and validation of the proposed model

model’s training schedule is and how well it can perform

on data that was not used during training. Furthermore, in
the case of Tea Leaf disease classification, the efficient
net b0 model has shown to be the go-to pick with better
predictive ability to classify between various forms of
disease classes with high accuracy. Exploring precision,
recall, and F1 score in the enhancement of the
classification model identified that the model achieved
remarkable accuracy of approximately 99.97%. Such
detailed analysis is brought out to demonstrate the WX
model’s precision in identifying that the right instances
Fig 2 Archiecture diagram are recognized, the ability to capture all instances
bearing relevance to the knowledge base, and how the
The context of most importance in an architectural balance of precision and recall rates is reflected on the
diagram is the arrangement of data and processes F1 Score. Such level of technical competency puts
within an application. First, the users are invited to EfficientNet B0 at a vantage point of being a
enter the data through a Web browser, which triggers a dependable solution for complex classification problems
number of processes in the background. This input especially in the diagnosis of the Tea Leaf disease.
passes through the back end where it faces a Deep
Learning (DL) model which is EfficientNet 6. Gradio
Classification model. This DL model very carefully
demarcates different components depending on the Gradio is a python library that is used for building a
input and identifies different leaf diseases correctly frontend for machine learning models. That is why it is
classifying them. Afterward, the processed data is in machine learning engineers’ interest for it to exist:
forwarded to the validation recommendation system because it makes creating interactive web interfaces
that carefully easier so that users can deploy and share their models to
others with greater facility. In this context, developers
analyze the disease percentage rate and extent of plant are able to easily specify the input and output part of the
invasion comprehensively as well. Besides, based on target model, which could be text, boxes and sliders or
the analysis, the recommendation system provides even image uploaders that enable the users to run the
individual recommendations regarding the application model in live streams. Furthermore, Gradio provides
of fertilizers. This elaborate flow successfully designs tremendous native support to a whole host of
the optimal interaction between abstract user inputs, predesigned widgets and layouts, which can help a
intelligent DL analysis, and rational recommendations developer reduce a great deal of coding that needs to be
on managing plant health issues. done. Few downtimes with simplicity coming in, it
supports easily adaptable API coupled by good
5. Experimental Result documentation that can suit beginners and experts in
creating interactive user interfaces to many machine
During the training and validation of the learning activities and systems.
EfficientNetB0 model, it was realized that the model Furthermore, it has interfaces to popular other deep
achieves its best performance at the 6th epoch as learning toolkits which are used often which include
validation accuracy recorded a very high of 99.96 %. TensorFlow, PyTorch and scikit- learn, thus developers
This milestone also highlights how effective the
who develop their models using such technique can [2] Pothineni, S. Upsetting Tea Development: A
continue using them without a lot of changes. Novel DCNN Approach for High-Precision Leaf
Globally the formats which the library receives input Infection Classification.
data is quite liberal and capable of handling any input
such as image, text input and any other in the various [3] Rahman, H., Ahmad, I., Jon, P. H., Rabbi, M. F.,
classification projects. Moreover, catalogued & Salam, A. (2024). Robotized Discovery of Chosen Tea
functionalities are employed to provide specialized Leaf Illnesses by Computerized Picture Handling
graphical interface characteristics that will enable Utilizing Convolutional Neural Organize (CNN):
developers to design the interface for their use. More Bangladesh Perspective.
so, the additional parameters set down by Gradio
when deploying a model makes the interface highly [4] Talukder, M. A., Layek, M. A., Kazi, M.,
enhanced making AI solutions more approachable and Uddin, M. A., & Aryal, S. (2024). Engaging covid-19
useful in different sectors, sub-disciplines, discovery: Optimizing execution through fine-tuned
occupations, and consumers. efficientnet profound learning architecture. Computers in
Science and Medicine, 168, 107789.
7. Conclusion
[5] Soh, K. S., Moung, E. G., Danker, K. J. J.,
Gradio is used in creating the frontend to the machine Dargham, J. A., & Farzamnia, A. (2024). Cocoa
learning models with the help of the python Infections Classification utilizing Profound Learning
programming language. Built for machine learning Calculation. In ITM Web of Conferences (Vol. 63, p.
engineers, it extends current frameworks to create 01014). EDP Sciences.
interactive web interfaces with less code, which helps
users deploy and share their models with others as well.
In Gradio, developers are able to easily input the input [6] Salian, S. R., & Sawarkar, S. D. (2022).
and output interfaces for the target models, which Melanoma skin injury classification utilizing made
could be text boxes, sliders and image uploaders where strides efficientnetb3. Jordanian Diary of Computers
the users can run the model live stream. Furthermore, and Data Technology, 8(1).
Gradio has auto-complete support for a lot of
predesigned overruling widgets and layouts that can
reduce the quantity of code to be produced by the
programmer. It can be easily implemented, not too
complicated for the APIs and documentation make it
suitable for any first time programmers and expert
programmers to build innovative user interfaces for a
complex multilevel process of machine learning and
for building sophisticated systems.

Also, it has interface to interconnect with more

elaborate deep learning environments that are
commonly used; these includes TensorFlow, PyTorch
and scikit-learn this makes developers who develop
their models using such technique to continue using
them without straining much. The format of the input
data that the library is able to manage is also relatively
versatile and not limited to image input, text input only
or any other relevant input that may be used in the
projects oriented towards classification. Moreover,
catalogued functionalities are employed to employ
graphical interface characteristics in order to provide
developers with more opportunities to create an
interface they need. Even more importantly, the
additional features defined by Gradio as part of model
deployment strongly enhance the interface, which in
turn enhances the application of AI technologies in
various industries and domains.

References

[1] Shinde, S. V., & Lahade, S. (2024). Profound

Learning for Tea Leaf Infection Classification:
Challenges, Think about Holes, and Rising
Technologies. Applied Computer Vision and Delicate
Computing with Interpretable AI, 293- 314.