Leaf Disease Detection Using Deep Learning

Techniques
Ankit Kumar Khushi Bajaj M. Baskar
Department of Computing Department of Computing Associate Professor
Technologies, School of Computing, Technologies, School of Computing, Department of Computing
College of Engineering and College of Engineering and Technologies, School of Computing,
Technology, SRM Institute of Science Technology, SRM Institute of Science College of Engineering and
and Technology, Kattankulathur, and Technology, Kattankulathur, Technology, SRM Institute of Science
Chengalpattu, Tamil Nadu, Chengalpattu, Tamil Nadu, and Technology, Kattankulathur,
India,603203. India,603203. Chengalpattu, Tamil Nadu,
ak7350@srmist.edu.in kb5122@srmist.edu.in India,603203.
baashkarcse@gmail.com

Abstract — In the agricultural sector, plant diseases are control, playing a crucial role in agricultural management
increasingly prevalent, prompting a growing interest in their and decision-making processes. In recent times, the
detection, particularly in expansive crop fields. Farmers face identification of plant diseases has garnered substantial
challenges in effectively managing various diseases, requiring
attention. Infected plants typically exhibit discernible marks
frequent shifts between control methods. Traditional methods
involve manual observation by surveillance experts to identify
or lesions on their leaves, stems, flowers, or fruits, each
tomato leaf diseases. Failure to implement proper control presenting distinct visible patterns for diagnosis. Leaves
measures can severely impact plant health and productivity. commonly serve as the primary indicators of plant diseases,
Employing mechanized techniques for disease detection is with symptoms often manifesting first in this part of the
advantageous, as it reduces labor-intensive tasks associated plant. Traditionally, agricultural and forestry experts or
with large-scale cultivation. Early detection of symptoms on farmers rely on subjective visual inspection to identify
plant leaves is crucial for effective disease management. diseases on-site, a method prone to errors and inefficiencies.
Inexperienced farmers may make erroneous judgments and
Keywords—deep learning, agriculture, plant disease, mobilenet,
misuse pesticides, leading to environmental pollution and
densenet, CNN
economic losses. To address these challenges, there is a
I. INTRODUCTION growing focus on research into utilizing image processing
techniques for plant disease recognition.
An important development in agriculture that will increase
crop output and quality is the automated detection of plant
Efficient plant disease management is intricately linked to
diseases using leaf analysis. Even seasoned farmers and
sustainable agriculture and climate change challenges.
pathologists may find it difficult to spot illnesses on plant
Studies suggest that climate variations can impact pathogen
leaves visually, despite the wide variety of crops that are
development stages and rates, as well as alter host
grown. This problem is especially severe in rural developing
resistance, resulting in shifts in host-pathogen interactions.
countries because the major means of identifying diseases is
Furthermore, the ease of global disease transmission
still eye inspection, which frequently need professional
exacerbates this complexity, introducing new diseases to
supervision. It may be logistically difficult for farmers in
previously unaffected areas lacking local expertise.
rural areas to consult experts, which might result in
Improper pesticide application by inexperienced individuals
inefficiencies in terms of both time and money.
can foster long-term pathogen resistance, significantly
hindering control efforts. Timely and precise disease
Researchers have investigated several approaches to
diagnosis is essential for precision agriculture, aiming to
overcome these issues, with automated computational
minimize resource wastage and promote healthier
systems showing promise. These technologies use high
production. Addressing the issue of long-term pathogen
throughput and precision to help agronomists and farmers
resistance development and mitigating the adverse effects of
diagnose and identify diseases. Using several feature sets in
climate change are critical steps in achieving this goal.
machine learning is one of the key strategies; deep learning
(DL)-based features and conventional handmade features are
popular choices. Effective feature extraction requires pre- II. LITERATURE REVIEW
processing techniques including segmentation, colour
modification, and picture enhancement. After that, several "Adapted Approach for Fruit Disease Identification using
classifiers are used to properly categorise plant diseases. Images" by Anand Singh Jalal and Shiv Ram Dubey [1].
. Fruit blemishes are indicative of diseases that, if neglected,
can cause significant losses. When pesticides are used
The emergence of plant diseases significantly hampers excessively to treat fruit illnesses, there is a chance that
agricultural productivity, potentially leading to increased harmful residues could accumulate on agricultural goods,
food insecurity if not promptly addressed. Early detection which will greatly contaminate groundwater. Furthermore, a
forms the cornerstone of effective disease prevention and large amount of production expenses is associated with
pesticides, which means that their use must be minimised. literature in every experimental situation when it comes to
Therefore, the goal of our suggested method is to quickly illness categorization.
detect fruit illnesses as soon as symptoms appear, allowing
for rapid treatment actions. Apple scab, apple rot, and apple Song, C., Zhang, D., and Zhang, Y. [4] With a focus on four
blotch are common fruit diseases that have unique visual diseases—powdery mildew, blight, leaf mould fungus, and
characteristics. We suggest and evaluate experimentally an ToMV—this research suggests a new method that uses
adaptive image-based technique for diagnosing fruit Faster RCNN to improve accuracy in identifying crop
diseases. Using K-Means clustering, we first segment fruit disease leaves and detecting infected leaves. In order to
photos. Then, we extract features from the segmented extract deeper illness characteristics from images, we first
images and use a Multi-class Support Vector Machine to substitute VGG16 with a depth residual network. Second,
classify diseases. We emphasise the usefulness of Support we cluster the bounding boxes using the k-means clustering
Vector Machine classification and the significance of technique, which enables better anchoring depending on the
clustering for illness segmentation. Three apple illnesses— clustering outcomes. Crop disease identification is essential
apple blotch, apple rot, and apple scab—are used to validate for crop disease prevention to guarantee crop quality.
our technique and show how effective it is in automatically Experimental findings show that the improved approach for
detecting and identifying fruit diseases. crop leaf disease detection showed a 2.71% improvement in
recognition accuracy and demonstrated quicker detection
Jashinsky, M., Leonowicz, Z., Hassan, S. M., Maji, A. K., & speed compared to the original quicker RCNN. Low
Jasińska, E. [2]: Improving agricultural output requires detection reliability and efficiency are the outcome of
quick and effective detection and prevention of crop traditional detection techniques that rely on manual
diseases. Taking advantage of CNNs' efficiency in machine observation. Farmers frequently lose out on possibilities for
vision tasks, this study focuses on using deep convolutional prevention within the best time period due to a lack of
neural network (CNN) models for the detection and professional knowledge on their part and the impossibility of
diagnosis of plant diseases based on leaf pictures. Our agricultural professionals to be present in the field at all
implemented models—using InceptionV3, times.
InceptionResNetV2, MobileNetV2, and EfficientNetB0—
A. PROBLEM STATEMENT
achieved impressive disease-classification accuracy rates of
98.42%, 99.11%, 97.02%, and 99.56%, respectively, The core problem this project addresses is the difficulty
surpassing traditional handcrafted-feature-based approaches. in accurately and timely identifying plant diseases,
Traditional CNN models typically require many parameters particularly in areas lacking expert resources. Current
and involve high computational costs. Interestingly, not only methods depend heavily on visual inspection by trained
did our models perform better than previous deep-learning professionals, which is not always feasible or efficient,
equivalents, but they also needed less training time. Because especially given the diversity of plant species and diseases.
of its optimised settings, the MobileNetV2 architecture also This limitation can lead to delayed or incorrect diagnoses,
provides interoperability with mobile devices. The disease adversely affecting crop yield and quality. The challenge is
identification accuracy results demonstrate the deep CNN to develop a reliable, automated system that can accurately
models' remarkable ability to greatly improve disease detect and classify plant diseases from leaf images, making
identification efficiency, with possible applications in real- the process more accessible and less reliant on human
time agricultural systems. expertise.
B. MOTIVATION
Muhammad E.H. Chowdhury, Tawsifur Rahman, Amith
The motivation behind the "Leaf Disease Detection using
Khandakar, Nabil Ibtehaz, et al. [3]: Although plants are
Deep Learning Techniques" project stems from the critical
essential for feeding the world's population, their
role that healthy plants play in agriculture and food security.
vulnerability to illness can result in large losses in
Traditional methods of disease detection, heavily reliant on
productivity. To solve this issue, ongoing monitoring is
human expertise, are often inadequate due to the vast variety
necessary, but manual disease monitoring is time-consuming
of plant diseases and the subtle nuances in their visual
and error-prone. A way to lessen these difficulties is to use
symptoms. This challenge is exacerbated in remote and
artificial intelligence (AI) and computer vision for early
under-resourced agricultural areas where expert analysis is
illness detection. In this study, we used 18,162 tomato leaf
scarce. The project is driven by the need to enhance the
pictures for disease detection to extensively assess the
accuracy, efficiency, and accessibility of plant disease
performance of several cutting-edge convolutional neural
diagnosis, leveraging the advancements in machine learning
network (CNN) architectures, including ResNet18, Mobile
(ML) and deep learning (DL) to bridge the gap between
Net, DenseNet201, and InceptionV3. We looked at how well
expert knowledge and practical, on-field applications.
they performed in three different classification tasks: six-
class (healthy vs. diverse groupings of diseased leaves), ten- III. METHODOLOGY
class (healthy vs. numerous categories of unhealthy leaves),
and binary (healthy vs. unhealthy leaves). DenseNet201 The initial stage in our strategy involves performing
outperformed InceptionV3 with 97.99% accuracy in six- convolution operations. Here, we will explore feature
class classification, whereas InceptionV3 showed detectors, which act as filters within the neural network.
remarkable accuracy of 99.2% in binary classification using We'll delve into understanding feature maps, the process of
simple leaf pictures. Furthermore, DenseNet201's accuracy learning parameters for these maps, how patterns are
in classifying ten classes was 98.05%. Our results show that identified, the layers involved in detection, and the
deep architectures perform better than those found in the visualization of the results.
A. SYSTEM ARCHITECTURE
The system architecture for plant disease detection as
depicted in Fig 1 employing MobileNet and DenseNet
involves several key components. Initially, a substantial
dataset comprising images of both healthy and diseased
plant leaves is collected and preprocessed as shown in
Equation 1, encompassing resizing, normalization, and
augmentation techniques to enhance model generalization.
Subsequently, pre-trained convolutional neural networks
(CNNs), notably MobileNet and DenseNet, are employed
for feature extraction due to their efficiency and
effectiveness in analyzing images.

Equation 1 Preprocessing of images

Fig 1. System Architecture
img_height,img_width=256,256
batch_size=20 1) Mobilenet
In computer vision, convolutional neural networks, or
train_datagen = ImageDataGenerator(rescale=1./255) CNNs, have become more popular. However, contemporary
train_generator = train_datagen.flow_from_directory(data_train,
target_size=(img_height,img_width),
CNN architectures are getting deeper and more complicated
batch_size=batch_size, in order to reach better levels of accuracy. We have used
class_mode='categorical', these Dense layers in our project for training model as
) shown in Equation 2. This intricacy presents difficulties for
test_generator = train_datagen.flow_from_directory(data_test,
target_size=(img_height,img_width), practical uses like robots and self-driving automobiles.
batch_size=batch_size,
class_mode='categorical',) MobileNet, depicted in Fig 2. Mobilenet architectureFig 2,
presents a versatile and efficient convolutional neural
These networks offer pretrained weights that have learned
rich hierarchical representations from large-scale datasets, base_model =
making them suitable for transfer learning. The pre-trained tf.keras.applications.MobileNet(input_shape=(img_height,img_width,
MobileNet and DenseNet models are integrated into the 3), include_top=False, weights='imagenet')
system and fine-tuned using the collected dataset, either model8 = Sequential()
model8.add(base_model)
through direct adaptation or via transfer learning, where model8.add(GlobalAveragePooling2D())
only the latter layers are trained on the specific plant disease model8.add(Dense(1024, activation='relu'))
dataset. Optionally, the outputs of the MobileNet and model8.add(Dense(512, activation='relu'))
DenseNet models can be fused to leverage the strengths of model8.add(Dense(256, activation='relu'))
model8.add(Dense(128, activation='relu'))
both approaches and enhance overall performance. model8.add(Dense(64, activation='relu'))
Following integration, the model undergoes rigorous model8.add(BatchNormalization())
training and evaluation, employing techniques like cross- model8.add(Dropout(0.2))
Equation 2 Mobilenet Dense layers
validation to ensure robustness and accuracy. Upon model8.add(Dense(38, activation='sigmoid'))
achieving satisfactory performance metrics, the model is network (CNN) architecture suitable for real-world
deployed in a production environment, accessible through a applications. It introduces a novel approach to achieving
user-friendly interface such as a web or mobile application, lighter models by employing depth-wise separable
catering to farmers or agricultural experts. Continuous convolutions instead of conventional convolutions utilized
monitoring and periodic retraining with new data are in prior designs. Moreover, MobileNet introduces two novel
essential for the model's adaptation to evolving plant global hyperparameters: the width multiplier and resolution
diseases and environmental conditions, ensuring its multiplier. These parameters empower model developers to
effectiveness and reliability in real-world scenarios. fine-tune the balance between accuracy and latency, thereby
trading off speed and reduced size to align with their
specific requirements..
 Fig 4 Densenet Channel

Moreover, a 2x2 average pooling layer and a 1x1

convolution layer follow several dense blocks in the design.
Assuming that the feature map sizes stay constant, joining
the transition layers is simple. Finally, a global average
pooling operation is carried out at the end of the dense block
sequence, and then the result is attached to a softmax
classifier.

3) Convolutional Neural Network

Three layers make up a Convolutional Neural Network

input, hidden layers, and output. The middle layers of a
feed-forward neural network are sometimes called hidden
layers because the activation function and final convolution
hide their inputs and outputs. These layers in CNNs,
however, are made expressly to extract characteristics from
the incoming data. We have used these convolution layers in
Fig 2. Mobilenet architecture our project for training model as shown in Equation 3.
Convolution is the first phase of a CNN's functioning,
2) Densenet during which feature detectors function as filters inside the
DenseNet has been extensively utilized across diverse neural network. The purpose of these feature detectors is to
datasets, with different types of dense blocks employed find patterns in the
based on the dimensionality of the input.
input data so that feature maps may be made. These feature
The Basic DenseNet Composition Layer, shown in Fig. 3, maps' parameters are learnt during training by using
This version of the dense block has a pre-activated batch optimisation strategies like gradient descent and
normalisation layer, a ReLU activation function, and a 3x3 backpropagation. A CNN's layers of detection are in charge
convolution operation after each layer in turn. The dense of gradually removing more complex patterns and structures
block's essential structure is formed by this mixture. from the input data by extracting higher-level features. After
. the detection phase, the learnt characteristics are mapped out
to provide predictions or classifications. Because CNNs can
automatically learn hierarchical representations of data, they
are generally quite good at tasks like object detection,
pattern analysis, and picture recognition.

Equation 3 CNN Layers

model.add(Conv2D(filters = 32, kernel_size = (5,5),padding = 'Same',

activation ='relu', input_shape=(img_height,img_width,3)))
model.add(Conv2D(filters = 32, kernel_size = (5,5),padding = 'Same',
activation ='relu'))
model.add(MaxPool2D(pool_size=(2,2)))
model.add(Conv2D(filters = 64, kernel_size = (3,3),padding = 'Same',
activation ='relu'))
Fig. 3 Densenet composition model.add(Conv2D(filters = 64, kernel_size = (3,3),padding = 'Same',
activation ='relu'))
model.add(MaxPool2D(pool_size=(2,2), strides=(2,2)))
BottleNeck DenseNet (DenseNet-B), as depicted in Fig 4, to model.add(Flatten())
address the computational complexity that arises with each model.add(Dense(256, activation = "relu"))
layer producing k output feature maps. In this structure, a model.add(Dense(38, activation = "softmax"))
bottleneck architecture is employed, incorporating 1x1
convolutions before the 3x3 convolution layer. This design
helps alleviate computational burden while maintaining
effective feature extraction capabilities. B. PROPOSED SYSTEM
The proposed system aims to revolutionize plant disease
detection in agriculture by leveraging state-of-the-art
technology and innovative methodologies. By integrating
MobileNet and DenseNet, two powerful pre-trained achieved by a Convolutional Neural Network (CNN) model
convolutional neural networks renowned for their efficiency across multiple epochs during its training phase on a large-
and effectiveness in image analysis, the system offers a scale image dataset. This table encapsulates the fluctuating
robust framework for early and accurate identification of performance of the CNN model Fig. 6 throughout the
plant diseases. Through a comprehensive dataset collection training process, showcasing the evolution of its accuracy
and preprocessing pipeline, coupled with advanced feature over successive epochs. By examining the average accuracy
extraction techniques, the system empowers farmers and over epochs, this table provides valuable insights into the
agricultural experts with a reliable tool for timely learning trajectory of the CNN model, elucidating its ability
intervention and mitigation strategies. to discern patterns and features within the complex image
data.
The user-friendly interface, accessible through web or
mobile applications, ensures ease of access and utilization, Table 2 CNN Accuracy graph over epoches
even in remote or rural areas. Moreover, continuous
monitoring and periodic retraining mechanisms enable the Epoches Training accuracy(avg) Validation accuracy(avg)
system to adapt to changing environmental conditions and 0-10 0.64 0.50
10-20 0.80 0.73
evolving disease patterns, ensuring its sustainability and 20-30 0.96 0.75
efficacy in addressing the challenges posed by plant diseases 30-40 1.00 0.75
in agriculture. Overall, the proposed system represents a 40-50 1.00 0.76
significant advancement in agricultural technology, poised
to enhance crop management practices and promote
agricultural sustainability worldwide.

IV. RESULT AND ANALYSIS

The results of implementing the plant disease detection

system using CNN, MobileNet and DenseNet have shown
promising outcomes. Through rigorous training and Fig. 6 CNN accuracy graph
evaluation, the model achieved high accuracy and precision
in identifying various diseases affecting plant leaves. By
leveraging the capabilities of MobileNet as depicted in Fig. Table 3 Densnet accuracy graph over Epoches
5, CNN and DenseNet for feature extraction, the model
demonstrated robustness in distinguishing between healthy Epoches Training accuracy(avg) Validation accuracy(avg)
and diseased leaves across different crop species. Table 1 0-10 0.990 0.925
Mobilenet accuracy graph over Epoches presents the 10-20 0.985 0.965
20-30 0.998 0.985
average accuracy of the MobileNet model across various
epochs during its training on a substantial dataset of images.
The performance of the MobileNet model as it progresses
through training iterations. The integration of multiple pre-
trained models and the fusion of their outputs contributed to
enhanced performance and minimized biases.

Table 1 Mobilenet accuracy graph over Epoches

Epoches Training accuracy(avg) Validation accuracy(avg)

0-10 0.90 0.76
10-20 0.95 0.78
20-30 0.96 0.81
30-40 0.96 0.75
40-50 1.00 0.95

Fig. 7 densenet accuracy graph

DenseNet, a type of Convolutional Neural Network (CNN),

distinguishes itself with its densely connected layers (as
depicted in Fig. 7,which promote feature reuse and
propagation throughout the network. This characteristic
allows for a more comprehensive exploration of image
features, resulting in enhanced accuracy by capturing finer
Fig. 5 Mobilenet accuracy graph
details.
Analysis of the system's performance highlighted its
illustrates the mean accuracy attained by a Dense
effectiveness in early and timely disease detection, thereby
Convolutional Network (DenseNet) model across multiple
facilitating prompt intervention and mitigation strategies to
epochs during its training on a comprehensive image
safeguard crop yield and quality Table 2 the mean accuracy
dataset. This table captures the dynamic performance of the [11] Abraham Gastélum-Barrios, Rafael A. BorquezLópez, Enrique Rico-
García, Manuel ToledanoAyala and Genaro M. Soto-Zarazúa*
DenseNet model throughout the training journey, “Tomato Quality Evaluation with Image processing: A review”
showcasing how its accuracy evolves over successive African Journal of Agricultural Research Vol. 6(14), pp. 3333-3339,
epochs. By scrutinizing the average accuracy over epochs, 18 July, 2011
this table offers crucial insights into the learning trajectory [12] P. Vimala Devi and K. Vijayarekha “Machine Vision Application to
of the DenseNet model, shedding light on its capacity to Locate Fruits, Detect Defects and Remove Noise: A Review” Vol.7 |
No.1 | 104-113| January – March | 2014.
recognize intricate patterns and features within the intricate
[13] Shiv Ram Dubey, A. S. Jalal “Detection and Classification of Apple
image data.
Fruit Diseases using Complete Local Binary Patterns”.

V. CONCLUSION
In our setup, the inclusion of both CNN and MobileNet
models highlights a strategic approach to addressing the
trade-off between accuracy and efficiency in image
processing. While traditional CNNs offer higher accuracy,
their reliance on conventional convolution layers can
impede processing speed. This bottleneck is mitigated by the
integration of MobileNet, renowned for its speed and
efficiency, especially in resource-constrained environments.
MobileNet's pretrained features, acquired from ImageNet,
bolster its performance and adaptability across various
image recognition tasks. By harnessing MobileNet's
lightweight architecture and pretrained capabilities, we
significantly accelerate inference speed without sacrificing
accuracy. As a result, MobileNet emerges as a pivotal
component of our image processing framework, enabling
rapid and reliable performance in real-time applications
while optimizing resource utilization and enhancing overall
system efficiency.

VI. REFERENCES

[1] Shiv Ram Dubey, Anand Singh Jalal “Adapted Approach for Fruit
Disease Identification using Images”
[2] Hassan, S.M.; Maji, A.K.; Jasiński, M.; Leonowicz, Z.; Jasińska, E.
Identification of Plant-Leaf Diseases Using CNN and Transfer-
Learning Approach. Electronics 2021, 10, 1388.
[3] M. E.H. Chowdhury et al., ‘Tomato Leaf Diseases Detection Using
Deep Learning Technique’, Technology in Agriculture. IntechOpen,
Oct. 13, 2021. doi: 10.5772/intechopen.97319.
[4] Yang Zhang, Chenglong Song and Dongwen Zhang, Deep Learning-
Based Object Detection Improvement for Tomato Disease, IEEE,
2020.
[5] Jun Sun, Yu Yang, Xiaofei He and Xiaohong Wu, Northern Maize
Leaf Blight Detection under Complex Field Environment Based on
Deep Learning, IEEE, 2020.
[6] Peng Jiang, Yeuhan Chen, Bin Liu, Dongiian He and Chunquan
Liang, Real-Time Detection of Apple Leaf Diseases Using Deep
Learning Approach Based on Improved Convolutional Neural
Networks, IEEE, 2019.
[7] Monica Jhuria, Ashwini Kumar, Rushikesh Borse “Image Processing
for Smart Farming: Detection of Disease and Fruit Grading”
Proceeding of the 2013 IEEE Second International Conference on
Image Processing.
[8] Sudhir Rao Rupanagudi, Ranjani B.S., Prathik Nagaraj, Varsha G.
Bhat “A Cost Effective Tomato Maturity Grading System using
Image Processing for 974 2015 International Conference on Green
Computing and Internet of Things (ICGCIoT) Farmers” International
Conference on Contemporary Computing and Information, 2014.
[9] Manisha A. Bhange, Prof. H. A. Hingoliwala “A Review of Image
Processing for Pomegranate Disease Detection” International Journal
of Computer Science and Information Technologies, Vol. 6 (1), 2015,
92-94
[10] Hetal N. Patel, Dr. M. V. Joshi “Fruit Detection using Improved
Multiple Features based Algorithm” International Journal of
Computer Applications (0975 – 8887), Volume 13– No.2, January
2011