Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

Using Machine Learning to Identify Diseases and

Perform Sorting in Apple Fruit
Arpit Patidar 1, Abir Chakravorty1*
1
Agricultural and Food Engineering Department, IIT Kharagpur
*Orc Id: https://orcid.org/0000-0002-9253-2594

Abstract:- Fruit diseases play a major role in global I. INTRODUCTION

agriculture, leading to substantial crop losses and
influencing food production and economic stability. In In India, while the apple cultivation area has expanded
this age of Industry 4.0 the fruit sorting is an important by 20%, production has only seen amodest increase of 1-2%
part in the food processing wherein this work plays a (National Horticulture Board 2020). Pests and diseases pose
vital role. In this study, a solution for the detection and significant challenges to apple crop production worldwide,
classification of apple fruit diseases is proposed and with fungal diseases particularly impacting fruit quality in
experimentally validated. Deep learning models offer regions like Himachal Pradesh, India's second-largest
promise for automating disease identification using fruit producer of apples. Plant infections are broadly categorized
images, but encounter obstacles such as the requirement as biotic and abiotic, with pathogens such as viruses, fungi,
for extensive training data, computational complexity, and bacteria responsible for biotic diseases (V. K. Vishnoi et
and the risk of overfitting. This study introduces an al. 2021). Biotic diseases, being highly transmissible and
innovative convolutional neural network (CNN) hazardous, pose greater risks compared to abiotic diseases
architecture aimed at addressing these challenges by caused by factors like mineral deficiency or environmental
incorporating a reduced number of layers, thus stressors.
alleviating computational burdens while maintaining
performance. Additionally, augmentation techniques  Introduction of Apple Disease Detection
such as shift, shear, scaling, zoom, and flipping are Apple fruits are frequently affected by biotic diseases
employed to diversify the training set without additional such as scab, cedar rust, blotch, powdery mildew, blight,
image acquisition. Our CNN model is specifically trained mosaic, and black rot. It is important to identify scab,
to identify common apple crop diseases like Scab, Rot, blotch, and black rot. Traditional disease detection methods
and Blotch. Rigorous experimental evaluation rely on visual observation by experts, which can be costly
demonstrates the effectiveness of our model, achieving a and time-consuming, particularly in remote regions. Since
remarkable classification accuracy of 95.37%. diseases can result in significant losses in both yield and
Significantly, our model demonstrates reduced storage quality at harvest, it becomes essential to identify symptoms
requirements and faster execution times compared to early in the fruit development stages through automatic
existing deep CNN architectures, enabling deployment disease detection. Identifying diseases promptly aids in
on handheld devices and resource-limited environments. planning control measures for the following year to mitigate
While other CNN models may offer similar accuracy losses. Additionally, some diseases spread to the tree's
levels, our approach emphasizes efficiency and resource branches, twigs, and fruit. Scratching, rot, and blotch are
optimization, rendering it practical for real-world common diseases of apples. When an apple rot infection
applications in agriculture. Furthermore, our CNN occurs, the affected area becomes circular and brown or
model exhibits resilience to environmental variations black, with a red halo around it. On the other hand, apple
and imaging parameters, enhancing its applicability scabs appear as gray or brown corky spots. A fungus called
across diverse agricultural settings. By leveraging apple blotch causes dark, asymmetrical, or lobed lesions on
advanced machine learning techniques, the approach the fruit's surface.
developed in this experimental work contributes to
modernizing fruits and vegetables sorting operations in While machine vision systems are currently used in the
food processing, crop management practices thus industry to automatically inspect apples for size and color,
promoting agricultural sustainability. The scalability it is still difficult to identify flaws in apples. This challenge
and portability of our model make it suitable for is caused by the inherent differences in skin tone between
deployment in both small-scale farms and large-scale apple varieties, the variety of defect kinds, and the existence
agricultural operations. of calyxes and stems. Many existing methods for apple
defect segmentation rely on simple thresholding approaches.
Keywords: Apple diseases, classification, convolutional For instance, a globally adaptive thresholding method, based
neural network, deep learning, disease detection, image on a modified version of Otsu's algorithm, was proposed for
processing, machine learning, fruit sorting, automation. segmenting fecal contamination defects on apples.
Classification-based techniques aim to categorize pixels into
different classes using various classification methods.

IJISRT24JUL195 www.ijisrt.com 732

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

Bayesian classification is commonly utilized, where pixels LBP has also found applications in face recognition,
are compared against a pre-calculated model and classified dynamic texture recognition, and shape localization.
as either defective or healthy. Unsupervised classification, Convolutional neural networks (CNNs) have been proposed
on the other hand, lacks target values and thus provides no as a comprehensive modeling approach for pattern
guidance during the learning process. Some researchers recognition tasks. I have proposed and experimentally
have employed unsupervised classification for defect validated the effectiveness of using CNNs as classifiers for
segmentation purposes. In another approach, (Ojala et al. the automatic detection and classification of fruit diseases.
2002) utilized the Local Binary Pattern (LBP) histogram for To validate our approach, I focused on three common apple
rotation-invariant texture classification, demonstrating diseases: blotch, rot, and scab.
impressive outcomes on representative texture databases.

Fig 1 Three Common Apple Diseases. a) Scab, (b) Rot and, (c) Blotch

Fig 2 Sample Images from the Data Set of Type (a) Apple Scab (b) Apple Rot (c) Apple Blotch, and (d) Normal Apple

IJISRT24JUL195 www.ijisrt.com 733

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

Machine Learning (ML) techniques such as k-nearest  To develop and optimize machine learning models
neighbor (KNN), artificial neural networks (ANN), and capable of classifying diseased and healthy apple fruits
support vector machine (SVM) are commonly employed for based on image data.
fruit disease detection and identification. However, deep  To evaluate the performance of the developed models
learning methods have emerged as a more effective approach in terms of accuracy, sensitivity, specificity, and
for hierarchical feature learning, capable of representing computational efficiency.
both low-level and high-level features with increased  To explore the potential applications of the proposed
complexity. Convolutional neural networks, or CNNs, have approach in real-world agricultural settings and its
proven effective in a number of applications, such as speech implications for crop management and disease control.
recognition, object detection, image classification, and
natural language processing. CNNs' effectiveness in II. LITERATURE REVIEW
identifying fruit diseases has also been demonstrated.
However, deploying CNN models on portable devices like  Automated Image Capturing System for Deep Learning-
smartphones, which are essential for supporting farmers based Tomato Plant Leaf Disease Detection and
becomes difficult due to the computational demands of Recognition
processing large datasets. By creating a lightweight deep De Luna et al. (2019) introduced an innovative method
CNN model that reduces computational demands and to enhance agricultural productivity, specifically in tomato
effectively diagnoses illnesses, this work attempts to address cultivation, by integrating smart farming systems. Their
this challenge. I present a novel four-convolutional-layer research addresses the challenge of detecting diseases in
deep CNN model (Conv4 DCNN) that is specifically tomato plants by utilizing recent advancements in computer
intended to use fruit images for the diagnosis of three vision, particularly deep learning techniques. They
different apple fruit diseases. Techniques for augmenting developed a motor-controlled image capturing system
data are used to improve model performance, and the capable of photographing every side of tomato plants to
random search technique is used to optimize facilitate disease identification and recognition. The study
hyperparameters and find the best possible values. This focused on the Diamante Max tomato variety and aimed to
work's main contribution is the creation of a low-complexity detect three common diseases: Phroma Rot, Leaf Miner, and
deep neural network model specifically designed to identify Target Spot. To classify tomato diseases, a deep
apple fruit diseases with minimalcomputational burden. convolutional neural network (CNN) was trained using a
dataset containing both diseased and healthy plant leaves.
 Importance of Apple Fruit Disease Detection The Transfer Learning disease recognition model achieved
The early detection and timely management of an accuracy of 95.75%, while the F-RCNN anomaly
diseases in apple orchards are crucial for ensuring detection model scored 80% confidence. Additionally, the
sustainable production and minimizing economic losses. automated image capture system demonstrated a 91.67%
Traditionally, disease detection has relied on visual accuracy rate in identifying diseases on tomato plant leaves.
inspection by agricultural experts, which is labor-intensive,
subjective, and prone to errors. Moreover, the delay in  CNN based Leaf Disease Identification and Remedy
disease diagnosis can lead to ineffective control measures, Recommendation System
resulting in reduced yields and compromisedfruit quality. Rohan et al. (2019) presented a novel approach
named CNN based Leaf Disease Identification and
 Role of Machine Learning Remedy Recommendation System. They underscored the
The advent of machine learning techniques has vital role of agriculture in the econoour and society,
revolutionized disease detection in agriculture by highlighting the challenges faced by farmers in identifying
providing automated and accurate methods for identifying leaf diseases, which often lead to reduced crop yields. The
fruit diseases. Machine learning algorithms, particularly authors stressed the potential of utilizing leaf videos and
those based on deep learning, have shown remarkable images to provide valuable insights for effective disease
performance in analyzing large datasets, including images, management and increased productivity. Early detection and
and extracting meaningful patterns. By leveraging these treatment of diseases are crucial to minimize negative
algorithms, it becomes possible to develop efficient and impacts on harvest yields and ensure adequate crop
reliable systems for detecting apple fruit diseases at an early nutrition. The emergence of smart devices capable of
stage. identifying anddetecting plant diseases has facilitated more
efficient responses to disease outbreaks in the agriculture
 Objectives sector. Leveraging publicly available datasets containing
The primary objective of this research is to develop a 5,000 photos of both healthy and diseased plant leaves, the
machine learning-based approach for the detection of apple study focused on plant disease detection through image
fruit diseases. Specifically, the study aims to: processing techniques. Crop types were classified, and
diseases were categorized into four classes using semi-
 To investigate the feasibility of using machine supervised techniques.
learning algorithms for identifying common diseases
affecting apple fruits.

IJISRT24JUL195 www.ijisrt.com 734

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

 Real-Time Detection of Apple Leaf Diseases Using Deep techniques. The model is trained on a dataset comprising
Learning Approach Based on Improved Convolution images of crop leaves under both healthy and diseased
Neural Networks conditions. Through this approach, the model effectively
Liu et al. (2019) focus on the identification of common classifies leaf images into diseased categories based on the
apple leaf diseases affecting apple crops, such as aria leaf pattern of defects.
spot, brown spot, mosaic, grey spot, and rust. They employ
deep learning methods, particularly enhanced convolutional  Identification of Plant Disease using Image Processing
neural networks (CNNs), to detect these diseases. The study Technique
utilizes the Apple Leaf Disease Dataset (ALDD), which Rathan et al. (2019) address the evolving role of
includes both intricate and lab-created images of apple agriculture, emphasizing its significance in sustaining
leaves. To enhance the dataset, the researchers utilize data populations, particularly in regions where a substantial
augmentation and image annotation technologies, portion of the population relies on agriculture for
facilitating the development of a new deep CNN-based livelihoods. However, crop diseases frequently undermine
model for disease detection. This model incorporates the agricultural productivity, leading to significant losses. Early
GoogleNet Inception structure and Rainbow concatenation. detection of leaf diseases is crucial for mitigating such
Through training and evaluation on a dataset comprising losses. The paper aims to develop software capable of
26,377 disease images, the proposed INAR-model achieves automatically identifying and categorizing plant diseases.
a high detection speed of 23.13 frames per second (FPS) The process involves steps such as image acquisition,
with a detection performance of 78.80%. The INAR-SSD preprocessing, segmentation, feature extraction, and
model demonstrates a promising method for early detection classification. Leaf images are utilized for disease
of apple leaf diseases, offering real-time detection identification, highlighting the importance of employing
capabilities with enhanced accuracy and speed compared to image processing techniques in the identification and
previoustechniques. categorization of agricultural diseases.

 Identification of plant leaf diseases using a nine-layer  GUI based Detection of Unhealthy Leaves using Image
deep convolutional neural network Processing Techniques
Geetharamani and Arun Pandian (2020) propose a Vamsi et al. (2019) emphasize the critical role of
method for identifying plant leaf diseases employing deep enhancing agricultural productivity for the growth of the
convolutional neural networks (CNNs). Their approach Indian econoour. They propose the utilization of image
involves utilizing a dataset consisting of background images processingtechniques to identify unhealthy leaves, aiming to
and more than 39 distinct classes of plant leaf diseases to facilitate the development of an intelligent and efficient
train the CNN model. To enhance the dataset's quality, they farming system. Focusing on ladies finger plant leaves,
employ six data augmentation techniques: gamma the study aims to detect early signs of diseases such as
correction, image flipping, principal component analysis powdery mildew, leaf spot, and yellow mosaic vein. The
(PCA), color augmentation, rotation, noise injection, and process involves capturing leaf images, preprocessing,
scaling. Through experimentation, the authors find that segmenting, extracting features, and classifying them.
integrating data augmentation techniques significantly Additionally, the study accounts for practical challenges
improves the model's performance. They explore various such as noisy image datasets, terrain conditions, and
combinations of training epochs, batch sizes, and dropout climate variations. For classification, Support Vector
rates to optimize the model training process. Comparative Machines (SVM) and Artificial Neural Networks (ANN)
analysis with transfer learning techniques indicates that the are employed, while K-Means clustering is utilized for
proposed CNN model outperforms them, especially when segmentation. Principal Component Analysis (PCA) is
assessed with independent data. The CNN model achieves a applied to reduce the feature set. Experimental results
classification accuracy of 96.46% through simulations, demonstrate that SVM and ANN achieve average disease
surpassing the performance of transfer learning techniques. detection accuracies of85% and 97%, respectively. In noise-
Overall, the study demonstrates the superior accuracy of the free conditions, these accuracies improve to 92% and 98%,
CNN model compared to transfer learning methods in respectively. The study lays the foundation for advancing
identifying plant leaf diseases. automation in the agriculture sector.

 Crop Disease Detection Using Deep Learning  Automatic Disease Symptoms Segmentation Optimized
Kulkarni (2018) addresses the rising prevalence of for Dissimilarity Feature Extraction in Digital
crop diseases attributed to recent climate changes and crops' Photographs of Plant Leaves
susceptibility to various pathogens. This phenomenon leads Abdu et al. (2017) address the critical initial step of
to financial losses for farmers, widespread crop destruction, applying machine learning to plant disease detection, which
and reduced agricultural productivity. Identifying and involves segmenting diseased symptom regions in plant leaf
treating these diseases pose significant challenges due to the images. This process, referred to as Region of Interest (ROI)
rapid proliferation of diverse diseases and limited segmentation, aims to isolate areas exhibiting symptoms
knowledge among farmers. The task is further complicated with color variations from the surrounding green tissue.
by the similarities in texture and appearance among diseased These segmented regions serve as the basis for extracting
and healthy leaves. To tackle this issue, Kulkarni proposes a discriminant features. However, existing approaches often
deep learning-based model leveraging computer vision fall short in capturing the comprehensive evolution of

IJISRT24JUL195 www.ijisrt.com 735

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

disease symptoms from onset to complete manifestation,  Dataset Details

which could provide valuable dissimilarity features for The apple images used in this study were captured
disease characterization. Moreover, challenges arise from from the Machua fruit mandi in Kolkata and some data were
the structure of the disease itself and the circumstances taken from publicly available on Kaggle (Shiv Ram Dubey
during image capture, complicating traditional ROI Et al. 2012). The dataset comprises a total of 5915 RGB
segmentation techniques. For example, disease symptoms images of apple fruits, categorized into four classes. These
may blend into healthy green tissue, blurring the separation classes correspond to three common apple diseases: blotch,
line and potentially leading to inaccuracies. To overcome scab, and apple rot. Additionally, there is a fourth class
these limitations, the study proposes an automated extended representing healthy/uninfected apple images. Sample
Region of Interest (EROI) segmentation technique. This images from each class are illustrated in Figure 3.1. The
method utilizes color homogeneity thresholding to expand naming convention for the disease classes aligns with the
the border region, incorporating portions of healthy tissue specific apple diseases, while the healthy class represents
to provide insights into symptom progression. Both standard images of fruit without any disease. The diversity of images
ROI segmentation and a reduced ROI approach were within the dataset is crucial for effective training, allowing
applied to a well-known Plant Village dataset to validate the the model to learn important variations. The deep CNN
proposed methodology. model's generalizability is improved by this diversity.
Several image transformations, including shift, shear,
III. MATERIALS AND METHODS scaling, zoom, and flipping, are used to expand the dataset
and add more variations (Peng jiang 2019). These
This section covers complete details related to the adjustments produce subtle changes to the images, adding a
development of deep CNN (Conv-4 DCNN) model to variety of examples to the training set. Thus, by enhancing
investigate the diseases in apple crops. The process of the model's ability to generalize and reduce overfitting, this
training and validating the proposed deep CNN model is augmentation technique enhances the model's performance
presented along with the detailed mathematical and tolerance.
formulations.

a. Apple Scab Dataset

b. Apple rot Dataset

IJISRT24JUL195 www.ijisrt.com 736

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

c. Apple Blotch Dataset

d. Normal Apple Dataset

Fig 3 Apple Dataset

 Implementation Work from the input data. Figure 4 illustrates the process of data
A deep Convolutional Neural Network (CNN) is an manipulation.
architecture of a feed-forward Artificial Neural Network
(ANN) commonly utilized in deep learning (Vihaor Kumar
Vishnoi et al. 2023). In CNNs, the term 'deep' refers to the
incorporation of numerous layers, unlike traditional neural
networks. Typically, a deep CNN is built by stacking
various layers, including convolutional layers with non-
linear activation functions, pooling layers, and fully
connected layers. Deep CNNs offer numerous advantages
over traditional machine learning methods, notably in their
ability to automatically extract features without the need for
manual feature engineering. The initial layers of a deep Fig 4 Representation of the Process of Data
CNN are responsible for capturing basic, low-level features Augmentation/Manipulation

IJISRT24JUL195 www.ijisrt.com 737

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

 Convolutional Layer
Convolution is carried out by the convolutional layer
by applying a filter (kernel) to an input image. By locating
regional patterns found in the layers before, it creates feature  Fully Connected Layer
maps. The linear convolution operation and the non-linear In order to create output layers with a predetermined
activation function are the two primary components of the number of outputs, the fully connected (FC) or dense layers
convolutional layer. Equation (1) illustrates how this in a CNN are usually positioned toward the end of the CNN
operation, which convolves the filter over multi-channel architecture. These layers bear similarities to those found in
image volumes like RGBimages, is expressed. traditional neural networks. The data that FC layers work
with is one-dimensional (1-D). The flatten layer comes
before the FC layers and transforms the two-dimensional (2-
D) output from earlier layers into a one-dimensional (1-D)
representation. The transformations carried out by the FC
layers are both linear and non-linear, and they can be
mathematically represented as follows in equations (6) - (7):
In the convolution process, the kernel K

convolves with the image I of similar

channel numbers nC but potentially different sizes,
resulting in the generation of a feature map O
. Here, represent the height and
width of the kernel respectively, while nH and nW denote In the provided equations, X represents the input
the height and width of the input image. Typically, the feature map, W denotes the weight, b signifies the bias
kernel is traditionally regarded as an odd-dimensional square terms, and O represents the output of the fully connected
window, with layer. To enhance prediction accuracy, optimal weights are
essential to minimize the loss function. The gradient descent
method is commonly employed to determine these optimal
weights. Another technique, the Adam algorithm, offers a
smoother and less noisy optimization process by adjusting
In the given equation 2, the symbol p represents the the learning rate based on adaptive estimates of lower-order
padding value, s denotes the stride, and z signifies the moments. This facilitates learning rate annealing during
number of kernels convolved with the input image. gradient optimization.
The Rectified Linear Unit (ReLU) stands as one of the
 Proposed Method
most commonly used activation functions. ReLU operates in
The proposed Conv-4 DCNN model comprises four
a manner where it does not activate all neurons
convolutional layers and two fully connected layers
simultaneously. Neurons are activated only if the output of
following three max-pooling units. Rectified Linear Unit
the convolution unit or other linear transformation is equal to
(ReLU) serves as the non-linear activation function in each
or greater than zero. This activation function is represented
convolutional layer and the initial dense layer. For
as shown in equation (3):
classifying apple fruit diseases, the softmax function is
applied at the output layer. This function facilitates
multiclass classification by assuming that each sample
belongs to precisely one class. Figure 5 illustrates the
 Pooling Layer sequencing of different layers alongside the activation
The feature maps created by the convolutional layers functions integrated into the developed deep CNNmodel.
are downsampled by the pooling layer. Its goal is to reduce Furthermore, the dropout unit functions by randomly
the size of activation maps that might have too many eliminating certain neurons that rely on any one feature,
parameters. As such, it reduces processing overhead, lessens therefore some neurons are ignored to spread out the
overfitting, and shortens training times. Max, min, and weights for better generalization. Initially, at the first
average pooling are three of the main pooling techniques. convolution level 16 filters (3 × 3) with valid padding and
But the most common method is max pooling, which takes stride (1, 1) are selected to convolute over RGB images of
the maximum value out of every input patch. The following size 256 × 256.
equation (3) shows the max pooling operation:

Where R denotes a receptive field containing P pixels. The

dimension of the generatedfeature map is defined in (4).

IJISRT24JUL195 www.ijisrt.com 738

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

Fig 5 Pipelining of Layers in Deep CNN Model

 Model Building this dataset using the hold-out validation technique. That is
The deep CNN model is evaluated and trained using why the training, validation, and test sets consist of 3549,
the dataset that is covered in Section 3.1. In a 60-30-10 1775, and 591 images, respectively, as Fig 6 shows the deep
ratio, the training, validation, and test sets are divided into CNN model.

Fig 6 Deep CNN model

The deep CNN model is built in Python on the In the given equation, p is the probability of correctly
Jupyter Notebook virtual platform, which has an Nvidia predicting that observation i belongs to class m, N is the
P100 TPU and 12 GB of RAM, using the Keras, Scikit-learn number of classes (where N > 2), and y is a binary indicator
libraries. The Matplotlib library is used to display the indicating the correct labeling of class m for observation i.
outcomes for visualization. The model is trained with a The Adam optimization algorithm is used for gradient
minibatch size of 56 and a learning rate of 0.0001 over a optimization, with the first and second moments'
span of 20 epochs. Equation (8) defines categorical cross- exponential decay rates set at β1 = 0.9 and β2 = 0.999,
entropy, which is the loss function of choice because it respectively. To find the best hyperparameters, random
works well for multiclass classification tasks. search techniques are used for hyperparameter tuning.
Moreover, the feature map produced by the second
convolutional layer offers more precise information about
the lesions connected to a specific disease and less general
information about the image.

IJISRT24JUL195 www.ijisrt.com 739

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

Fig 7 Layered architecture of proposed deep CNN model for apple disease detection

 Image Acquisition:  Integration with Flask Backend

Image acquisition in the first load the image in a digital The Flask backend managed image preprocessing,
picture process that consists of capturing the image through model inference, and result rendering. Upon image upload,
a digital camera and stores it in digital media for additional preprocessing steps such as resizing and normalization were
python operations (Vibhor Kumar Vishnoi et al. 2022). applied before passing the image through the CNN model
for diseaseidentification.
 Image Preprocessing:
The main aim of image pre-processing is to enhance  Real Time Feedback and Visualization
the image information containing unwanted distortions or to The Flask application provided real-time feedback on
reinforce some image features for any processing. disease presence in uploaded apple fruit images. Results
Preprocessing techniques use various techniques like from the model inference were visualized and displayed to
dynamic image size and form, filtering of noise, image users, enabling interaction and exploration of detected
conversion, enhancing image and morphological operations. diseases.

 Classification:  Scalability and Deployment

Classification is employed for testing fruit diseases. The Flask application was designed for scalability and
The Random Forest classifier is utilized for classification. deployment flexibility. Deployment options included local
hosting for testing and development, as well as cloud-based
 Model Deployment using Flask: Website Creation with solutions for broader accessibility and scalability across
user Friendly Interface multiple platforms.
The trained CNN model was deployed using the Flask
web framework in conjunction with HTML, CSS, and  Real Time Interface
JavaScript to develop a user-friendly interface for disease "Real-time interface" refers to a user interface (UI) that
detection. Flask's lightweight and scalable platform, coupled provides immediate feedback or response to user input or
with its ability to seamlessly integrate with front-end changes in data. In the context of our CNN model for fruit
technologies, makes it an ideal choice for hosting web disease detection, implementing a real-time interface would
applications that leverage machine learning models. allow users, such as farmers or agricultural practitioners, to
interact with the system seamlessly and receive timely
 Flask Application Setup feedback on disease detection results.
The Flask application was structured following the
Model-View-Controller (MVC) pattern. It was configured  Tools and Technologies
with route definitions, static file directories, and template
rendering settings to ensure smooth operation.  Python
The Python community is large and friendly, and
 User Interface Design with Html and CSS Stack Overflow is a great place to start looking for solutions
HTML and CSS were utilized to design a visually to different problems you may run into when developing. It
appealing interface featuring upload buttons and image is one of the most widely used languages on the website, so
placeholders. CSS styling was applied for improved there's a good chance you can find direct answers to any
usability and aesthetics. questions. Additionally, Python provides a wealth of
powerful tools designed specifically for scientific
 Enhancing Interactivity with Java Script computing. Rapid iteration and significantly less code are
JavaScript was used to enhance user interaction, required to implement specific functionalities thanks to the
implementing event listeners for image uploads and form easily accessible and well-documented packages such as
submissions. Client-side validation mechanisms were NumPy, Pandas, and SciPy. Python's ability to tolerate code
employed to ensure a seamless user experience. that closely resembles pseudo code is one of its noteworthy
features. This is helpful because it makes the process of

IJISRT24JUL195 www.ijisrt.com 740

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

implementing and testing pseudo code described in  Keras

academic papers simpler. But there are some disadvantages A high-level neural network API that runs on top of the
to Python. It can be annoying to deal with the dynamic TensorFlow, CNTK, or Theanoframeworks, Keras (Ganatra,
nature of the language and the well-known "duck typing" N. et al. 2020) is primarily designed in Python. It was
issue within packages. Unexpected return types from designed with the intention of promoting quick
package methods, like an object that looks like an array experimentation, understanding that quick changes from
instead of an actual array, can be problematic. Furthermore, idea to result are essential for carrying out meaningful
a lot of trial-and-error testing may be necessaryto determine research. Keras supports both convolutional and recurrent
method return types due to the standard Python networks, as well as their combinations, to make
documentation's lack of clarity, which makes learning new prototyping simple and quick. It also runs computations on
Python packages and libraries more difficult. both CPU and GPU architectures with ease. A library
comprising multiple implementations of commonly used
 Numpy neural network building blocks, such as layers, objectives,
NumPy is a Python package that provides sophisticated optimizers, and activation functions, is included in Keras,
and scientific mathematical functions inside the Python along with several tools to facilitate the handling of image
environment. Known as the scientific computing core and text data.
library (Geetharamani & Pandian, 2019), it provides a
powerful multidimensional array object and makes  Compiler Option
integration with C and C++ easier. Among other things, Anaconda is a top-tier open-source Python distribution
NumPy is very helpful for linear algebra tasks and has designed for large-scale data processing, predictive
random number generation capabilities. NumPy's array type analytics, and scientific computing applications. Simplifying
is a useful data structure that enhances Python by being the deployment and package management procedures is its
specifically designed for numerical calculations. main goal.
Furthermore, NumPy provides basic arithmetic operations,
such as Eigenvector computation functions.  Jupyter Notebook
The Jupyter Notebook is an open-source web
 Scikit Learn application that makes it easier to create and share
A free machine learning library for Python called documents with interactive code, equations, graphics, and
Scikit-learn (Mohanty et al. 2016) has many algorithms for text. Its broad range of applications is applicable to a variety
tasks like classification, regression, and clustering. Among of fields, such as machine learning, statistical modeling,
these algorithms are k-nearest neighbors, random forests, data visualization, numerical simulation, and data cleaning
and support vector machines. Moreover, Scikit-learn and transformation (Brownlee, J. 2016).
interfaces with Python's scientific and numerical libraries,
including SciPy and NumPy, with ease. Although Scikit- IV. RESULTS AND DISCUSSION
learn is mostly written in Python, Python is also used to
implement some of the core algorithms for best  Results
performance. Achieving such a high accuracy rate holds significant
 Tensorflow implications, particularly in agriculture, where early disease
TensorFlow is an open-source software library detection can greatly influence crop yield and quality. A
intended for numerical computation with data flow graphs CNN model with a 95.37% accuracy rate shows promise for
(Geetharamani & Pandian, 2019). These graphs show practical deployment in real-world scenarios, enabling
nodes for mathematical operations and edges for the timely interventions to mitigate disease spread and minimize
multidimensional data arrays (tensors) that are exchanged yield losses. In this implementation, the proposed CNN
between them. Its adaptable architecture enables model is evaluated for its accuracy. The training accuracy is
computation to be distributed across one or more CPUs or plotted along with model validation accuracy v/s number of
GPUs in desktop, server, or mobile devices using a single epochs in Fig. 9 It can be observed that there is a small
API. TensorFlow was first created for machine learning and difference in training and validation accuracy curves. It
deep neural network research by scientists and engineers in shows that the proposed model fits well for the addressed
Google's Machine Intelligence research group, specifically problem. It can also be observed from the graph that model
the Google Brain Team. However, its adaptability makes it accuracy initially increases sharply with increasing epochs.
applicable to a wide range of fields outside of machine Later it improves slowly.
learning.

IJISRT24JUL195 www.ijisrt.com 741

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

Fig 8 Model Training

 Model Validation Accuracy  Model Loss

Model validation accuracy represents the degree of Model loss (Hossein Azgomi et al. 2023), often
correctness of predictions made by the model concerning the referred to as the objective function or cost function,
true labels or outcomes (Hasan et al., 2022). It is typically quantifies the disparity between the model's predictions and
expressed as a percentage and reflects the proportion of the actual ground truth labels. It represents the error incurred
correct predictions out of the total number of predictions by the model during training and is minimized through
made. Higher accuracy values indicate that the model is optimization techniques such as gradient descent. Model
making more accurate predictions, thereby demonstrating its loss is a crucial metric as it guides the learning process by
proficiency in understanding and capturing patterns within signaling to the model how far off its predictions are from
the data. However, it's crucial to note that accuracy alone the true labels. Lower values of loss indicate that the model
may not provide a comprehensive understanding of the is converging towards better performance, effectively
model's performance, especially in scenarios where class capturing the underlying patterns in the data. Different loss
imbalances exist or when certain classes are more critical functions are employed based on the nature of the task, such
than others. Thus, while accuracy serves as a primary as mean squared error (MSE) for regression tasks and
indicator of model performance, it should be complemented categorical cross-entropy for classification tasks.
with other evaluation metrics for a more holistic assessment.

Fig 9 Model Validation Accuracy v/s No. of Epochs Curve Fig 10 Model Loss v/s No. of Epochs Curve

IJISRT24JUL195 www.ijisrt.com 742

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

 Confusion Matrix
A classification model's performance is graphically
summarized by the confusion matrix (Peng Jiang et al.
2019), which shows accurate predictions on the diagonal
and incorrect classifications off-diagonal. It helps identify
difficult classes and offers insights into classification
accuracy. The performance of the model can only be
improved and maximized with the use of this tool. When
presented with several classes that share a similar shape,
classifiers may become confused. Reduced performance
may also arise from highly complex patterns presented in
the same class due to infected apple images at different
stages or against different backgrounds. A model's
classification accuracy can be visually estimated using a
confusion matrix. The final test results' confusion matrix is
displayed in Figure 11. The model's prediction accuracy in
the relevant class increases with deeper coloration in the
visualization results. Since every accurate prediction is on Fig 11 Confusion Matrix
the diagonal and every inaccurate prediction is off the
diagonal, it is easy to identify which classes have confused  Model Prediction Sample
the detecting system.

Fig 12 Model Prediction

IJISRT24JUL195 www.ijisrt.com 743

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

 Model Deployment using Flask expect predictions for four classes of apple diseases: apple
In the development of our Flask-powered website blotch, apple scab, apple rot, and normal apple.
interface for the CNN image classification model, our main Additionally, I've integrated detailed documentation and
focus was on enhancing usability, functionality, and visual help resources directly within the Flask application, aiming
appeal. The homepage serves as the entry point, providing to assist users in understanding and effectively utilizing the
users with a brief project overview and a clear call-to-action classification tool. The design prioritizes a visually pleasing
(CTA) button leading them to the image classification tool. color scheme, typography, and responsive layout, enhancing
Using Flask, this tool enables users to upload images for accessibility and performance across various devices for a
classification, complete with straightforward instructions seamless user experience. Leveraging Flask's capabilities,
and visual indicators like progress bars and loading icons. the website interface operates in real-time, allowing users to
I've ensured that classification results are transparently interact with the classification tool immediately.
presented, showing predicted classes. Specifically, users can

Fig 13 Website Interface

 Performance Metrics used for Model Evaluation  Recall or Sensitivity:

The model performance is evaluated in respect of Recall, or sensitivity (0.9531), assesses the classifier's
various metrics which can be derived from the confusion capability to identify all positive samples accurately. It is
matrix. determined by the ratio of true positive predictions to the
total number of actual positive samples.
 Accuracy:
Accuracy (0.9537) is a widely used performance  Specificity / True Negative Rate (TNR):
metric in classification tasks. It measures the ratio of correct Specificity, or True Negative Rate (TNR), measures
predictions to the total number of predictions made by the the accuracy of a classifier in identifying negative samples.
classifier. The confusion matrix enables the determination of It is computed by dividing the number of correctly predicted
both overall and individual class accuracies. negative samples by the total number of actual negative
samples.
 Precision:
Precision (0.9537) measures the proportion of positive  F1-Score:
predictions that are correctly identified by the classifier. It F1-Score (0.9525) gives the harmonic mean of the
evaluates the accuracy of positive predictions by calculating precision and recall, and refers to the no. of correctly
the ratio of true positives to the total number of predicted classified instances. It ranges from 0 to 1.
positives.

IJISRT24JUL195 www.ijisrt.com 744

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

V. CONCLUSION AND FUTURE WORK [6]. Awate, A., Deshmankar, D., Amrutkar, G., Bagul,
U., & Sonavane, S. (2015, October). Fruit disease
 Conclusions detection using color, texture analysis and ANN. In
In our study, I delved into the feature-based approach 2015 international conference on green computing
utilized in existing systems, predominantly employing and internet of things (ICGCIoT) (pp. 970-975).
image processing techniques. I thoroughly examined steps IEEE.
such as image acquisition, preprocessing, feature extraction, [7]. Balakrishna, K., & Rao, M. (2019). Tomato plant
and classification. To enhance disease detection in apples, I leaves disease classification using KNN and PNN.
proposed a novel system utilizing Convolutional Neural International Journal of Computer Vision and Image
Networks (CNNs). Through extensive experimentation, our Processing (IJCVIP), 9(1), 51-63.
CNN-based approach achieved a commendable accuracy of [8]. Brownlee, J. (2016). Introduction to the Python Deep
95.37%. This result underscores the effectiveness of CNNs Learning Library Theano. [Online]. Available.
in disease identification within apple fruit, showcasing their [9]. De Luna, R. G., Dadios, E. P., & Bandala, A. A.
potential for advancing agricultural disease management (2018, October). Automated image capturing system
systems. for deep learning-based tomato plant leaf disease
detection and recognition. In TENCON 2018-2018
 Future Work IEEE Region 10 Conference (pp. 1414-1419). IEEE.
[10]. Devaraj, A., Rathan, K., Jaahnavi, S., & Indira, K.
 Incorporation of More Fruit Varieties: Extend disease (2019, April). Identification of plant disease using
detection to various fruits, broadening agricultural image processing technique. In 2019 International
applications. Conferenceon Communication and Signal Processing
 Mobile Application: Enable farmers to access our model (ICCSP) (pp. 0749-0753). IEEE.
easily through a mobile app. [11]. Dutot, M., Nelson, L. M., & Tyson, R. C. (2013).
 Cloud Integration: Utilize cloud services for enhanced Predicting the spread of postharvest disease in stored
scalability and collaboration, facilitating real-time fruit, with application to apples. Postharvest biology
updates and accessibility. and technology, 85, 45-56.
[12]. Ferentinos, K. P. (2018). Deep learning models for
REFERENCES plant disease detection and diagnosis. Computers and
electronics in agriculture, 145, 311-318.
[1]. Abdu, A. M., Mokji, M. M., Sheikh, U. U., & Khalil, [13]. Ganatra, N., & Patel, A. (2020). A multiclass plant
K. (2019, March). Automatic disease symptoms leaf disease detection using image processing and
segmentation optimized for dissimilarity feature machine learning techniques. International Journal on
extraction in digital photographs of plant leaves. In Emerging Technologies, 11(2), 1082-1086.
2019 IEEE 15th International Colloquium on Signal [14]. Geetharamani, G., & Pandian, A. (2019).
Processing & Its Applications (CSPA) (pp. 60-64). Identification of plant leaf diseases using a nine-layer
IEEE. deep convolutional neural network. Computers &
[2]. Adedoja, A., Owolawi, P. A., & Mapayi, T. (2019, Electrical Engineering, 76, 323-338.
August). Deep learning based on nasnet for plant [15]. Han, J., & Moraga, C. (1995, June). The influence of
disease recognition using leave images. In 2019 the sigmoid function parameters on the speed of
international conference on advances in big data, backpropagation learning. In International workshop
computing and data communication systems on artificial neural networks (pp. 195-201). Berlin,
(icABCD) (pp. 1-5). IEEE. Heidelberg: Springer BerlinHeidelberg.
[3]. Al Bashish, D., Braik, M., & Bani-Ahmad, S. (2011). [16]. Ilic, M., Spalevic, P., Veinovic, M., & Ennaas, A. A.
Detection and classification of leaf diseases using K- M. (2015, November). Data mining model for early
means-based segmentation and. Information fruit diseases detection. In 2015 23rd
technology journal, 10(2), 267-275. Telecommunications Forum Telfor (TELFOR) (pp.
[4]. Arivazhagan, S., Shebiah, R. N., Ananthi, S., & 910-913). IEEE.
Varthini, S. V. (2013). Detection of unhealthy region [17]. Jhuria, M., Kumar, A., & Borse, R. (2013,
of plant leaves and classification of plant leaf December). Image processing for smart farming:
diseases using texture features. Agricultural Detection of disease and fruit grading. In 2013 IEEE
Engineering International: CIGR Journal, 15(1), 211- secondinternational conference on image information
217. processing (ICIIP-2013) (pp.521-526). IEEE.
[5]. Arivazhagan, S., Shebiah, R. N., Ananthi, S., & [18]. Jiang, P., Chen, Y., Liu, B., He, D., & Liang, C.
Varthini, S. V. (2013). Detection of unhealthy region (2019). Real-time detection of apple leaf diseases
of plant leaves and classification of plant leaf using deep learning approach based on improved
diseases using texture features. Agricultural convolutional neural networks. IEEE Access, 7,
Engineering International: CIGR Journal, 15(1), 211- 59069-59080.
217. [19]. Kim, M. S., Lefcourt, A. M., Chen, Y. R., & Tao, Y.
(2005). Automated detection of fecal contamination
of apples based on multispectral fluorescence image
fusion. Journal of food engineering, 71(1), 85-91.

IJISRT24JUL195 www.ijisrt.com 745

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

[20]. Kleynen, O., Leemans, V., & Destain, M. F. (2005). [31]. Patel, S., Jaliya, U. K., & Patel, P. (2020). A survey
Development of a multi-spectral vision system for on plant leaf disease detection. Int.J. Mod. Trends
the detection of defects on apples. Journal of food Sci. Technol, 6, 129-134.
engineering, 69(1), 41-49. [32]. Qin, F., Liu, D., Sun, B., Ruan, L., Ma, Z., & Wang,
[21]. Kulkarni, O. (2018, August). Crop disease detection H. (2016). Identification of alfalfa leaf diseases using
using deep learning. In 2018 Fourth international image recognition technology. PloS one, 11(12),
conference on computing communication control and e0168274.
automation (ICCUBEA) (pp. 1-4). IEEE. [33]. Sahithya, V., Saivihari, B., Vamsi, V. K., Reddy, P.
[22]. Kumar, A., & Gill, G. S. (2015, May). S., & Balamurugan, K. (2019, April). GUI based
Automatic fruit grading and classification system detection of unhealthy leaves using image processing
using computer vision: A review. In 2015 techniques. In 2019 International Conference on
Second International Conference on Advances in Communication and Signal Processing (ICCSP) (pp.
Computing and Communication Engineering (pp. 0818-0822). IEEE.
598-603). IEEE. [34]. Soni, P., & Chahar, R. (2016, July). A segmentation
[23]. Lee, S. H., Wu, C. C., & Chen, S. F. (2018). improved robust PNN model for disease
Development of image recognition and classification identification in different leaf images. In 2016 IEEE
algorithm for tea leaf diseases using convolutional 1st International Conference on Power Electronics,
neuralnetwork. In 2018 ASABE Annual International Intelligent Control and Energy Systems (ICPEICES)
Meeting (p. 1). American Society of Agricultural and (pp. 1-5). IEEE.
Biological Engineers. [35]. Suma, V., Shetty, R. A., Tated, R. F., Rohan, S., &
[24]. Liu, B., Zhang, Y., He, D., & Li, Y. (2017). Pujar, T. S. (2019, June). CNN based leaf disease
Identification of apple leaf diseases based on deep identification and remedy recommendation system.
convolutional neural networks. Symmetry, 10(1), 11. In 2019 3rd International conference on Electronics,
[25]. Lu, Y., Yi, S., Zeng, N., Liu, Y., & Zhang, Y. (2017). Communication and Aerospace Technology (ICECA)
Identification of rice diseases using deep (pp. 395-399). IEEE.
convolutional neural networks. Neurocomputing, [36]. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., &
267, 378-384. Wojna, Z. (2016). Rethinking the inception
[26]. Mahlein, A. K., Rumpf, T., Welke, P., Dehne, H. architecture for computer vision. In Proceedings of
W., Plümer, L., Steiner, U., & Oerke, the IEEE conference on computer vision and pattern
[27]. E. C. (2013). Development of spectral indices for recognition (pp. 2818-2826).
detecting and identifying plant diseases. Remote [37]. Tm, P., Pranathi, A., SaiAshritha, K., Chittaragi, N.
Sensing of Environment, 128, 21-30. B., & Koolagudi, S. G. (2018, August). Tomato leaf
[28]. Mehl, P. M., Chao, K., Kim, M., & Chen, Y. R. disease detection using convolutional neural
(2002). Detection of defects on selected apple networks. In 2018 eleventh international conference
cultivars using hyperspectral and multispectral image on contemporary computing (IC3) (pp. 1-5). IEEE.
analysis. Applied engineering in agriculture, 18(2), [38]. Waghmare, H., Kokare, R., & Dandawate, Y. (2016,
219. February). Detection and classification of diseases of
[29]. Mohanty, S. P., Hughes, D. P., & Salathé, M. (2016). grape plant using opposite colour local binary pattern
Using deep learning for image-based plant disease feature and machine learning for automated decision
detection. Frontiers in plant science, 7, 215232. support system. In 2016 3rd international conference
[30]. Ojala, T., Pietikainen, M., & Maenpaa, T. (2002). on signal processing and integrated networks (SPIN)
Multiresolution gray-scale and rotation invariant (pp. 513-518). IEEE.
texture classification with local binary patterns. IEEE [39]. Wu, C., Luo, C., Xiong, N., Zhang, W., & Kim, T. H.
Transactions on pattern analysis and machine (2018). A greedy deep learning method for medical
intelligence, 24(7), 971-987. disease analysis. IEEE Access, 6, 20021-20030.

IJISRT24JUL195 www.ijisrt.com 746

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

APPENDIX

FLASK APP PROGRAMMING CODE

IJISRT24JUL195 www.ijisrt.com 747

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

HTML PROGRAMMING CODE

IJISRT24JUL195 www.ijisrt.com 748

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

CSS PROGRAMMING CODE

IJISRT24JUL195 www.ijisrt.com 749

Volume 9, Issue 7, July – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24JUL195

JAVA SCRIPT PROGRAMMING CODE

IJISRT24JUL195 www.ijisrt.com 750