Classification of Fruits and Detection of Disease

using CNN

Dr. Roshani Raut Ms Sakshi Pawar Mr Devanshu Mr Shubham Bhosale Mr Rishabh Nabar
Dagonkar

Professor, Department of Department of Department of Department of

Department of Information Information Information Information Information
Technology, Technology, Technology, Technology, Technology,
Pimpri Chinchwad College Pimpri Chinchwad Pimpri Chinchwad Pimpri Chinchwad Pimpri Chinchwad
of Engineering, Savitribai College of College of College of College of
Phule University Pune, Engineering, Savitribai Engineering, Savitribai Engineering, Savitribai Engineering, Savitribai
India Phule University Pune, Phule University Pune, Phule University Pune, Phule University Pune,
India India India India

[email protected] pawarsakshi2703@gm daggaonkard1998@gm [email protected] [email protected]

rg ail.com ail.com om om

ABSTRACT fruit illness detection techniques and

additionally provide an outline of those
This project demonstrates the tactic for techniques.Analysis on detection of illness
detection of fruit illness and and fruit grading is beneficial for
grade(accuracy).Diseases in fruits and agriculture and farmers.By
plants square up as one of the main characterising the sort of illness in fruits
reasons for the agricultural loss. This and additionally grading of fruit on its
work focuses on developing an easy quality. For detection of illness we need
user-friendly tool that acknowledges the completely different options of fruit and
amount of the illness and grades them classifiers classifying these options. For
consequently.This model uses the fruit grading phase the image
Convolution Neural Networks (CNN) for segmentation calculates the infected and
the classification, that is once more healthy portion of fruit and grading is
retrained with transfer learning supported on proportion of infection on
technique.Identification and classification fruit.The classification accuracy for the
of diseases of fruits are done through projected paper is achieved up to 98%.
varied algorithms. The main focus of our
work is getting the analysis of various
KEYWORDS – CNN classifier, Fruit OBJECTIVE:
detection, feature extraction, disease
detection, fruit ● The primary objective of this work is
to develop a fruit quality observance
system exploiting fruit image
analysis by color image processing.
I. INTRODUCTION ● According to the brightness worth of
the image pixels, the fruit
India is an associate agricultural country and characteristic parameters are
Indian national farmers choose a large measured and picked up from the
variety of fruit and vegetable crops. The given images.
cultivation of crops can be improved by
technological support. This Asian giant is an
agricultural land, India produces 90.2
million metric tonnes of fruits and 169.1 II. RELATED WORK OR
million metric tonnes of vegetables. The LITERATURE SURVEY
area under cultivation of fruits stood at 6.3
million hectares while vegetables were [1] “Fruit Recognition and Grade of
cultivated at 10.1 million hectares. India Disease Detection using Inception V3
produces about 10% of the world's fruit Model
needs. Indian farmers grow a range of fruits Author: Nikhitha M, Roopa Sri S, Uma
such as apple, banana, citrus, grape, mango, Maheswari B
guava, papaya, watermelon. Fruit business In India, crop yield has declined due to
contributes heavily for the development of the post recognition of diseases in
this country. Neural Network techniques fruits/vegetables by the farmers.Farmers
square measure very favoured and helpful face great economic loss worldwide.
for image classification and identification of Diseases in fruits and plants are the main
fruits severally. This work is supported on reasons for the agricultural
the basis of the colour, size, and form of the loss.Knowing the health status of
fruit. To notice the infections in fruits, fruits/vegetables helps farmers to
convolution neural networks techniques are improve their productivity .
measured and used as per size and color. For
technology, this model is employing a [2] Identification of Artificially Ripened
convolution neural network used for Fruits using Machine Learning
classifying the images and visual perception Author: AJAYKUMAR YADAV
Project plan . Here fruits are classified Ripening of fruit is a natural process.
nearly as Good and Rotten Fruits Ethylene is responsible for the ripening
additionally disease,illness levels are process which is produced naturally in
determined on Grades A and B. fruit. But dealers and sellers often use
chemicals like CaC2 (Calcium carbide)
to quicken this process so their product common property to measure quality of
will enter the market early and they can any fruit is its appearance which
maximize profit. Fruits are kept in includes colour, shape, size and surface
storage with chemicals. conditions. The analysis of colour is
especially an important consideration
[3] “Detection and Classification of when determining the efficiency of fruit.
Apple Fruit Diseases Using Complete This study used an inexpensive method
Local Binary Patterns” Author: Shiv to predict cherry and strawberry color
Ram Dubey, Anand Singh Jalal, 2012 parameters by combining image
Third International Conference on processing and artificial neural network
Computer and Communication (ANN) techniques. The automatic vision
Technology. Diseases in fruit cause based technology consisted of a CCD
devastating problems in economic losses camera for image acquisition, MATLAB
and production in the agricultural software for image analysis and ANN
industry worldwide. The image for modeling.
processing based proposed approach is
composed of the following main steps, III. FRUIT CLASSIFICATION
in the first step K-Means clustering MODEL
technique is used for the image
segmentation, in the second step some This model works on the principle of
state of the art features are extracted Convolutional Neural Networks(CNN) ; it
from the segmented image, and finally consists of many stages.
images are classified into one of the
classes by using a Multi-class Support
Vector Machine. The classification
accuracy for the proposed solution is
achieved up to 93%.

[4] “Assessment of Fruit using Digital

Image Processing,” Author: K. Raut, V.
Bora, International journal of Science
technology & Engineering, 2016. In Fig. 1.
recent years automatic vision based This model involves stages as follows firstly
technology has become more powerful data collection which will serve as input to
and more efficient to many areas the machine. Then secondly RGB to
including agricultural fields and food Grayscale conversion of the image will take
industry. An automatic vision based place. This step is required as the associate
system for sorting and grading of fruit RGB Image consists of three layers R,G,B.
like cherry and strawberry based on their It’s a three dimensional matrix, wherever the
maturity level discussed here. The most grayscale image is two dimensional, and
also the values vary between 0–255 (8-bit It takes an input image, processes it and
unsigned integers). Therefore, some classifies it. This step conjointly has sub
algorithms will solely be applied on 2-D layers. Convolution Layer this can be
images instead of 3-D, therefore we have a termed the primary layer to extract features
tendency to convert associate RGB images from an input image. Convolution preserves
into a grayscale image, as an example, Black the link between pixels by learning features
and White conversion of a picture, of the image by treating squares of the input
convolution of a picture, etc. So the file. It's computing mathematically that
algorithm will process quicker, and might takes two inputs like an image matrix and a
involve additional complicated operations. filter or kernel. It conjointly involves
Consecutive step is Image Segmentation padding where the image is contoured with
whereby the digital image is divided into zero(0) in order that it fits, Non Linearity
multiple segments conjointly referred to as (ReLU) that has the aim to introduce
Pixels or Image Objects. This helps simplify non-linearity in ConvNet. Since, the
the image into something substantive and algorithm should be acknowledged for
easier to investigate. Following this learning non-negative values. The
consecutive step is Feature Extraction succeeding layer in CNN is the Pooling
wherever the initial set of the informational Layer that helps scale back the quantity of
data is split and reduced to additional parameters once the images are overlarge.
manageable teams. It helps to induce the Spatial pooling is additionally referred to as
simplest feature from those massive subsampling or downsampling that reduces
information sets by choosing and mixing the spatial property of every map however
variables into options, thus, effectively retains necessary information. There are
reducing the quantity of informational data. varied forms of pooling Max pooling that
This helps to cut back the quantity of takes the most important component from
redundant informational data from the data the corrected feature map. There can be also
set with extracting relevant data. The next taken as Average pooling total of all parts
step is the Convolutional Neural Network within the feature map. Then the total of all
Classifier. Those steps are given as follows the elements in the map can be termed as
shown in Fig. 2. Sum pooling. The next layer in CNN is
Fully Connected Layer or Flatten Layer
whereby the matrix is flattened into a vector
and fed to the neural network as shown in
fig. 3.

Fig. 2.
 Fig. 3. Fig. 4.

In the above Fig. 3. The feature map matrix From the dataset, we have picked pictures
is regenerated as a vector (x1, x2, x3, …). from totally different classes. Among them
With the Fully connected layers, we have a to form the training set we have used (80%)
tendency to combine these features to form a pictures and also the rest (20%) pictures
model. And therefore the last step of this used for testing the model. The network
model is the final output that is extracted model is trained for ten epochs.
when following all above stages.
V. RESULTS
IV. DATASET
In this paper, we've applied a convolutional
For training and testing, all the pictures were neural network on the fruits dataset so as to
designated from Kaggle. This dataset search out the higher classification
contains 9747 pictures. These images performance of the network. For
contain totally different fruits. This dataset determinative classification accuracies,
contains the following: we've taken cases where we have applied
∙ Training set size: 7840 pictures. totally different combinations of hidden
∙ Testing set size: 1907 pictures layers (convolution and pool) for 10 epochs
The fruit pictures were reaped by recording and calculated the accuracies on the
the fruits whereas they're turned by a motor Training and Testing set. The Tensorflow
so manufacturing frames. A study is placed library in python was used to attain these
behind the fruits as a background. Because accuracies.The accuracy obtained which was
of the inequality within the lighting a flood highest was to be found at 98% on test
fill kind formula was developed that images (Conv1, pool1, Conv2, pool2 with
extracted the fruit from the background. dropout). The best performance of the
After removing the background all the fruits network on the test set consecutively
were scaled right down to 100×100 pixels of occurred at 7 up to 10 epochs. The highest
ordinary RGB fruit images. Fig. 4. shows training and test accuracy was attained on
samples of fruit pictures from the dataset. CNN by doing various combinations of
hidden layers and increasing the number of
epochs.
There were losses for various epochs for classification performance gets better For
various combinations of convolution and fruit image detection, CNN additionally
pool. Totally different responses were found showed considerably higher accuracy than a
in several combinations. The loss curve standard methodology did. Besides, this
responded during a downward direction approach is additionally a lot faster to
within the beginning. The loss was the deploy for new fruits.
foremost for the first epochs and step by step
slowly decreased up to 10 epochs. The loss
soothed from about 0.0793 from 8 epochs up
to 10 epochs.
Following are the output figures for Training VI. COMPARISONS
and Testing which classify Good Fruit from
Rotten Fruit and Grade the fruits for Fruit Feature Classifier Accuracy References
diseases.
Orange Color, ANN 88% G.Capizzi,et
Texture al 2016

Tomato Color, PNN 84.4% O.O.Arjenki

Shape , 2013
Fig. 5. Good Fruit
Apple Color, Naive 91% M.Ronald
Size Bayes et. al, 2016

Mango Maturity Thermal 90% S. Naik et.

, Size Imaging al, 2017

Strawbe Color, Cluster >68% K

Fig. 6. Rotten Fruit rry Shape, Analysis, Yamamoto
Size DA et. al, 2015

Apple, Color, CNN 98% This paper

Banana Shape

As shown in the table above the

comparisons are given out. There are
Fig. 7. Grade A various algorithms that can be used for
classification of fruits and disease detection
but CNN Algorithm comes out to achieve
most accuracy.

Fig. 8 Grade B

It is shown that CNN behaves otherwise for

various combinations of layers. However, if
the convolution and pool are used in
repeated order with dropout the
 VII. ADVANTAGES AND time and cut back losses of fruit
DISADVANTAGES because of diseases.
● The leading objective of our
The mathematical model is NP-Complete.
paper is to boost the worth of
In the existing system there is no fruit sickness detection.
computerized system to detect artificial or
●
natural. As everyone knows, chemicals and
damaged fruit may affect human health so VIII. CONCLUSION
usually, machines are used for this purpose.
So we are trying to overcome this problem. This project provides an associate degree
We are image data to detect the fruit economical and accessible technique to
disease. observe fruit unwellness employing an
internet web application. This method uses
Disadvantages: the normal internet app with a dataset to
• Time consuming process. capture fruit and observe. Fruit unwellness
detection strategy would likewise elevate
• Not user friendly system
Indian Farmers to try and do sensible
• CNN does not encode the position cultivating that sets aside chance to time
and orientation of objects. decisions that to boot spare time and reduce
loss of fruit attributable to diseases. This
• Lack of ability to be spatially analysis may be extended to increase the
invariant to the input data. accuracy of techniques and validate with
additional samples. This feature is extremely
• Lots of training data is required.
helpful for the farmers and helpful for
Advantages: various functions.

● Very High accuracy in image

REFERENCES
recognition problems. 1) R. Karthika, K. V. M. Ragadevi and
● Automatically detects the N. Asvini “Detection of Artificially
Ripened Fruit using Image
important features without any Processing,” International journal of
human supervision. Advanced Science and Engineering
research, vol. 2, no. 1, pp. 576-582,
● Weight sharing.
2017.
● It would promote Indian Farmers 2) K. Raut, V. Bora “Assessment of
to try to do good farming that Fruit using Digital Image
Processing,” International journal of
helps to require time to time
Science technology & Engineering,
choices that additionally save vol. 3, no. 1, pp. 273-279, 2016.
3) “Identification of Artificially 10.1007/s11947-010-0411-8, Food
Ripened Fruits Using Smart Phones” Bioprocess Technol 4:487–504,
by Dr.S.Maheswaran in 2017 2011.
11) Ghulam Muhammad, “Date fruits
International Conference on classification using texture
Intelligent Computing and Control descriptors and shape-size features”,
(I2C2), IEEE2017. Engineering Applications of
4) Downey, G. () "Review and Artificial Intelligence, 37, 361–367,
assessment of food and nutrition 2015.
12) Sapan Naik and Bankim Patel,
policies." Food, Health and the “Thermal imaging with fuzzy
Consumer pp 121-212 classifier for maturity and size based
5) Wasserman, Anthony I. (2016) non-destructive Mango (Mangifera
"Software engineering issues for Indica L.) grading”, 2017
mobile application development." International Conference on
Emerging Trends & Innovation in
Proceedings of the FSE/SDP
ICT (ICEI), 978-1- 5090-3404-8/17,
workshop on Future of software 2017.
engineering research. ACM. 13) C. Hung, J. Nieto, Z. Taylor, J.
6) “Assessment of Fruit using Digital Underwood, and S. Sukkarieh,
Image Processing,” Author: K. Raut, "Orchard fruit segmentation using
V. Bora, International journal of multi-spectral feature learning," in
Science technology & Engineering, 2013 IEEE/RSJ International
2016. Conference on Intelligent Robots and
7) Jana, Susovan, and Ranjan Parekh. Systems, 2013, pp. 5314-5320.
"Intra-class Recognition of Fruits 14) H. Cheng, L. Damerow, Y. Sun, and
using Color and Texture Features
M. Blanke, "Early yield prediction
with Neural Classifiers."
International Journal of Computer using image analysis of apple fruit
Applications 148.11, 2016. and tree canopy features with neural
8) Yudong Zhang et al., “Fruit networks," Journal of Imaging, vol.
classification using computer vision 3, p. 6, 2017.
and feedforward neural network”,
15) J. Hemming, J. Ruizendaal, J.
Journal of Food Engineering Volume
143, Pages 167–177, December Hofstee, and E. van Henten, "Fruit
2014. detectability analysis for different
9) Haiwei Dong and Nikolaos Mavridis camera positions in sweet pepper,"
Abdulhamid Haidar, "Image-Based Sensors, vol. 14, pp. 6032-6044,
Date Fruit Classification," 2014. [14] T. F. Chan and L. A. Vese,
International Congress on Ultra
Modern Telecommunications and "Active contours without edges,"
Control Systems, vol. IV, 2012. IEEE Transactions on image
10) Sergio Cubero et al., “Advances in processing, vol. 10, pp. 266-277,
Machine Vision Applications for 2001.
Automatic Inspection and Quality
Evaluation of Fruits and
Vegetables”, doi