Mango Leaf Diseases Identification Using Convolutional Neural Network

Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

International Journal of Pure and Applied Mathematics

Volume 120 No. 6 2018, 11067-11079


ISSN: 1314-3395 (on-line version)
url: http://www.acadpubl.eu/hub/
Special Issue
http://www.acadpubl.eu/hub/

Mango Leaf Diseases Identification Using


Convolutional Neural Network
S. Arivazhagan1 , S.Vineth Ligi2
1,2
Center for Image Processing and Pattern
Recognition, Department of ECE,
1,2
Mepco Schlenk Engineering College,
Sivakasi.
August 4, 2018

Abstract
The identification of plant diseases plays a vital role in
taking disease control measures in order to improve the qual-
ity and quantity of crop yield. Automation of plant diseases
is very much beneficial as it reduces the monitoring work in
large farms. Leaves being the food source for plants, the
early and accurate detection of leaf diseases is important.
This work includes a deep learning based approach that au-
tomates the identification of leaf diseases in Mango plant
species. Five different leaf diseases such as Anthracnose,
Alternaria leaf spots, Leaf Gall, Leaf webber, Leaf burn of
Mango has been identified in a dataset consisting of 1200
images of diseased and healthy mango leaves. The proposed
CNN model achieves an accuracy of 96.67% for identifying
the leaf diseases in mango plant thereby showing the feasi-
bility of its usage in real time applications.
Keywords:Deep learning, Convolutional Neural Network,
Mango leaf diseases.

1 Introduction
Mango also called as The King of Fruits is one of the important
fruit crops cultivated in different countries around the world. India

11067
International Journal of Pure and Applied Mathematics Special Issue

produces about 40% of the global mango production and ranks first
among the worlds mango producing countries. It is estimated that,
pests and diseases destroy approximately 3040% of the crop yield
[1]. The common diseases of mango plant includes gall infestation,
webbers attack, mango malformation, stem miner, anthracnose, al-
ternaria leaf spots etc. Such diseases are caused by pathogens like
bacteria, virus, fungi, parasites etc, and even unfavourable environ-
mental conditions. Disease in leaf affects the photosynthesis process
thereby leading to plant death. The symptoms and the affected leaf
area determine the type of disease. In earlier days, identification
of plant diseases was usually carried out by frequent monitoring of
plants by farming experts. In case of small farms it was possible to
identify the diseases easily and take immediate preventive on con-
trol measures. But in the case of large farms, it is time consuming
and expensive. Therefore looking for an automatic, accurate, fast
and less expensive technology for plant disease identification is of
great importance. Image processing and machine learning are most
popular and widely used techniques adopted for plant leaf disease
detection and classification. Deep learning using Neural Networks
is a part of the wider family of machine learning. It has spread
its arms in various fields providing a huge variety of applications.
The development of such computer technology can help farmers to
monitor and control diseases in plants.

2 RELATED WORKS
Leaf disease detection has been a long research topic for the past
few decades. In order to improve the recognition rate of disease
diagnosis, researchers have studied many techniques using machine
learning and pattern recognition. The techniques include machine
learning techniques such as Convolutional Neural Network [2], Ar-
tificial Neural Network [3], Back Propagation Neural Network [4],
Support Vector Machine [5] and other image processing methods
[6,7]. In the above techniques, Convolutional Neural Network per-
forms both feature extraction and classification itself. Others meth-
ods use Color Co-occurrence matrix [8], Angle Code Histogram [4],
Zooming algorithm [9], Canny edge detector [6] and various other
algorithms for feature extraction. Research works have been carried

11068
International Journal of Pure and Applied Mathematics Special Issue

out to classify a single disease in multiple plants varieties or mul-


tiple diseases in a single plant species. These advanced techniques
are applied to many crops such as rice [2], wheat [10], maize [11],
cotton [12]. Also CNN requires less or no preprocessing of images
when compared to other techniques. Recently, several studies on
automated diagnosis of plant diseases have been conducted using
deep learning techniques. Yang Lu et al. proposed a method for
detection and classification of unhealthy leaf images of rice plant
using Deep Convolutional Neural Network. A dataset of 500 natu-
ral images of diseased and healthy rice leaves and stems captured
from rice experimental field was used. CNNs are trained to identify
10 common rice diseases under the 10-fold cross-validation strategy.
Also, since color images are used, the stationary property does not
hold across color channels, therefore the images are rescaled in [0,
1] and Principal Component Analysis and Whitening are applied to
get training features and testing features [2]. Alvaro Fuentes et al.
[13] proposed a deep learning based detection mechanism for real
time leaf disease and pest recognition in tomato plant. Three dif-
ferent detectors such as Faster Region-based Convolutional Neural
Network (Faster R-CNN), Region-based Fully Convolutional Net-
work (R-FCN), and Single Shot Multibox Detector (SSD) are used
to recognize 9 different tomato diseases using a dataset of 5000
images. Additionally data annotation and data augmentation are
performed to reduce false positives and also increase the accuracy.
Mohanty [14] et al. proposed a leaf disease detection model based
on deep Convolutional Neural Network. This work was able to clas-
sify 38 classes consisting of 14 crop species and 26 disease varieties
using a dataset of 54,306 images form Plant Village dataset. In this
work, an automated mango leaf disease identification method based
on deep Convolutional Neural Network is proposed to achieve fast
and accurate recognition results by using ReLU activation function
and batch normalization. This work aims at identifying five differ-
ent Mango leaf diseases such as Anthracnose, Alternaria leaf spots,
leaf gall, leaf burn and leaf webber. The CNN performs automated
feature extraction from the raw inputs in an analytical way. The
classification is based on selecting the features with highest proba-
bility values. The remaining of this paper is organized as follows.
Section 3 describes the architecture of the CNN model. Section
4 describes the implementation of mango leaf disease identification

11069
International Journal of Pure and Applied Mathematics Special Issue

using the proposed CNN model, followed by the results of this work
in Section 5. Section 6 deals with the conclusion.

3 ARCHITECTURE OF DEEP CON-


VOLUTIONAL NEURAL NETWORK
A CNN consists of an input layer and an output layer, as well as
multiple hidden layers between them. The hidden layers basically
consists of the convolution layer, pooling layer, Rectified Linear
Unit, dropout Layer and normalization layers. In the case of image
classification, the input is an image and the output is the class name
also called label. Inspired by various pre-trained Convolutional
Neural Network architectures such as VGG-16, VGG-19, Alexnet, a
deep CNN architecture with three hidden layers has been proposed
for this work. There is no precise rule in organizing the structure
of the individual layer.

3.1 Input Layer


The image can be given as input directly to the CNN model. The
size of the image is denoted as [height width number of color chan-
nels]. For color image the number of color channels corresponds to
3 and for the grayscale image it is 1. Data augmentation can be
done before passing the images to the CNN model. Since neural
networks and deep learning models requires large amount of data,
the dataset is increased by generation of artificial data through ex-
pansion of original dataset. The images are augmented by applying
different transformations that include rotations, zooming, cropping,
transpose and skewing while preserving the label of the image.

3.2 Convolutional Layer


The prime operation of convolutional layer is convolution operation.
The first layer in a CNN is always a convolutional layer for which
the input is an image, in case of first network layer or feature map
from the previous layer. The input is convolved with the filters,
called kernels in order to produce the output feature maps. The

11070
International Journal of Pure and Applied Mathematics Special Issue

convolution output can be denoted as,


 
X
xlj = f  xl−1
i ∗ kijl + blj 
i∈Mj

where, xj represents the set of output feature maps, Mj represents


the set of input maps, Kij represents the kernel for convolution,
bj represents the bias term. The size of the output feature map is
given by,

where, O is the output height/length, W is the input height/length,


K is the filter size, P is the padding, and S is the stride. In order
to preserve the size of the output, padding of zeros can be em-
ployed on the edges of the input. The amount of padding, P can
be determined as follows,

where, K is the filter size.

3.3 ReLU Layer


The Rectified Linear Unit layer also called as the activation layer
is used to introduce some non-linearity to the system, since it per-
forms linear operations during the convolution process. So, it is
introduced after every convolution layer. This layer simply changes
all the negative activation values to 0. The ReLU layer performs a
thresholding operation to each element given by,
f(x) = max(0,x)
This layer plays a significant role in alleviating the vanishing gra-
dient problem and helps to train the system faster. The ReLU
suits well for multiclass classification. For binary classification, the
sigmoid function can be used. instead of ReLU.

11071
International Journal of Pure and Applied Mathematics Special Issue

Fig 1: CNN Architecture

3.4 Max-Pooling Layer


In this layer, the input is divided into multiple non-overlapping
blocks and outputs the maximum among the elements in each block
to form an output of reduced size while preserving the important
information in the input. It is also capable of controlling the over
fitting problem. There is no learning process in this layer.

3.5 Dropout Layer


The basic idea of the dropout layer is that, the input elements
with a certain probability are deactivated or dropped out such that
the individual neurons are able to learn the features that are less
dependent on its surroundings. This process takes place only during
the training phase.

3.6 Batch Normalization Layer


The Batch Normalization layer is usually present between the con-
volution layer and the ReLU layer. It increases the training speed
and reduces the sensitivity of network initialization. In this layer
the activations of each channel are normalized by subtracting the
mini-batch mean and dividing by the mini-batch standard devia-
tion. This is followed by shifting the input by an offset β and then
scaling it by a factor γ. These two parameters are updated during

11072
International Journal of Pure and Applied Mathematics Special Issue

the training phase. The batch normalized output, yi is given by,

yi = BNγ,β (xi ) ≡ γ x̂i + B

where,x̂i is the normalization of activation xi which is given by


equation (3.6).
x i + µB
x̂i = p 2
σB + ε
where, ε is a constant, µB is the mini-batch mean and σB2 is the
mini-batch variance given by,
m
1 X
µB = xi
m i=1

m
1X
σB2 = (xi − µB )2
m i=1
where, m is the mini-batch size. In this way network training can
be made faster by making the optimization problem easier.

3.7 Fully Connected Layer


In the fully connected layer all the neurons of this layer are con-
nected to all the neurons in the previous layer, thereby combining
all the features learned by the previous layer to facilitate classifica-
tion. This layer produces an N-dimensional vector at the output,
where N is the number of classes.

3.8 Output Layer


The output layer consists of the softmax layer followed by the classi-
fication layer. The softmax layer outputs a probability distribution
based on which, the network model classifies an instance as a class
that has the maximum probability value. The Softmax function
also called Normalized Exponential is given by equation (3.9).

P (x, θ|cr )P (cr )


P (cr |x, θ) =
P
k
P (x, θ|cj )P (cj )
j=1

11073
International Journal of Pure and Applied Mathematics Special Issue

P
k
where, ≤ P (cr |x, θ) ≤ 1 and P (cj |x, θ) = 1 P (x, θ|cr ) is the
j=1
conditional probability of an instance given class r and is the class
priori probability. The equation (3.9) can also be written as follows,

exp(ar (x, θ))


P (cr |x, θ) =
P
k
exp(aj (x, θ))
j=1

where, ar = ln(P (x, θ|cr )P (cr ))

4 LEAF DISEASE IDENTIFICATION


USING CNN
Mango leaf diseases image database is created by acquiring images
under challenging conditions such as illumination, size, pose and
orientation, using a digital camera of resolution 4608 x 3456. It
consists of 1200 images of both diseased and healthy leaves. The
diseases include Anthracnose, Alternaria leaf spot, leaf gall, leaf
webber, leaf burn of mango plant. In order to reduce the compu-
tational time complexity, the images are resized from the size 4608
x 3456 to 256 x 256. The proposed CNN architecture consists of
an image input layer followed by three hidden layers and then the
output layer. The layer implementation is represented in Table 1.

Table 1: Layer Implementation of the CNN model

The leaf images of size 256x256x3 are given as input to the input
layer. Data augmentation is performed in order to increase the
dataset by generating artificial data. The images are then passed

11074
International Journal of Pure and Applied Mathematics Special Issue

through the hidden layers. Each hidden layer consists of a con-


volutional layer, batch normalization layer, Rectified Linear Unit
followed by the max pooling layer. Feature extraction is performed
using convolutional and pooling layers, whereas classification is per-
formed by the fully connected layer. Each convolutional layer and
pooling layer consists of different number of filters, of varying size.
The three convolution layers consists of 32, 64, 128 filters of size
11x11, 7x7, 5x5 respectively with stride 2 and padding based on the
formula (3.3). The batch normalization layer and the ReLU layer
increase the training process and network performance. The three
max pooling layers consists of 5x5, 3x3 and 3x3 filters respectively
with stride 1 and padding, P=1 for maxpooling layer 1 and P=0
for maxpooling layers 2 and 3. Then 50% dropout is employed to
deactivate the least learned features. The features learnt by the
convolutional and pooling layers are then classified by using two
fully connected layers of size 64 and 6 respectively. The size of the
second fully connected layer is equal to the number of classes. It
specifies the probability distribution for each class. Steepest Gra-
dient Descent algorithm is used to train the proposed CNN model
with 50 % of the images in each class. The remaining 50% images
are tested. Since CNN is a supervised learning network, the labels
are also trained along with the images. The plot for the network
training progress is shown in Figure 2.

Fig 2: Plot for Training Progress


The Training progress plot shows the increase in training accuracy

11075
International Journal of Pure and Applied Mathematics Special Issue

and simultaneous decrease in loss as the number of iterations in-


creases during the training and validation processes. The loss is
the summation of error made for each sample in the training and
validation sets. Lower the loss, better is the model and recognition
results.

5 RESULTS
The proposed CNN model was trained with 100 images per class
totally accounting for 600 training images. The remaining 600 im-
ages constituting of 100 images per class was tested. The testing
accuracy resulted in 96.67%. The confusion matrix is shown in
Table 2

Table 2: Confusion Matrix

The classes C1 to C6 is listed in the Table 3.


Table 3: List of Classes

Among the 6 classes, classes 3 to 6 have produced 100 % classifi-


cation results since those diseases have distinctive appearance and
features when compared to other classes.

10

11076
International Journal of Pure and Applied Mathematics Special Issue

6 CONCLUSION
The proposed CNN based leaf disease identification model is ca-
pable of classifying 5 different diseases in mango leaves from the
healthy one. Since CNN does not require any tedious preprocess-
ing of input images and hand crafted features, faster convergence
rate and good training performance, it is preferred for many appli-
cations rather than the conventional algorithms. The classification
accuracy can be further increased by providing more images in the
dataset and tuning the parameters of the CNN model. But ob-
taining the optimal parameters for a CNN model is still a research
challenge.

References
[1] http://www.sciencedirect.com/topics/agricultural-and-
biological-sciences/plant-diseases.

[2] Yang Lu, Shujuan, Nianyin Zeng, Yurong Liu, Yong Zhang,
(2017), Identification of Rice diseases using Deep Convolu-
tional Neural Networks , Elsevier Journal on Neurocomputing.

[3] Elham Omrani, Benyamin Khoshnevisan, Shahaboddin


Shamshirband, HadiShaboohi, Nor Badrul Anuar, Mohd
Hairul nizam md nasir, (2014), Potential of redial basis
function- based support vector regression for apple disease de-
tection, Measurement, vol.55,pp:. 512-519.

[4] Kshitij Fulsoundar , Tushar Kadlag , Sanman Bhadale , Pratik


Bharvirkar, (2014), Detection and classification of plant leaf
diseases, International Journal of Engineering Research and
General Science Volume 2, Issue 6.

[5] Shivaputra S.Panchal, Rutuja Sonar, (2016). Pomegranate


Leaf Disease Detection Using Support Vector Machine, Inter-
national Journal Of Engineering And Computer Science ISSN:
2319-7242 ,Volume 5 Issues 6.

[6] R.Preethi, S.Priyanka, U.Priyanka , A.Sheela, (2015), Efficient


knowledge based system for leaf disease detection and classifi-

11

11077
International Journal of Pure and Applied Mathematics Special Issue

cation, International Journal of Advance Research In Science


And Engineering, Vol. No.4, Special Issue (01).

[7] P.Revathi , M.Hema Latha, (2012), Classification Of Cotton


Leaf Spot Diseases Using Image Processing Edge Detection
Techniques International Journal of Applied Research ISBN,
169-173.

[8] S. Arivazhagan, R. Newlin Shebiah , S. Ananthi, S. Vishnu


Varthini, (2013), Detection of unhealthy region of plant leaves
and classification of plant leaf diseases using texture features,
CIGR Journal, Vol. 15, No.1 211.

[9] Santanu Phadikar, Jaya Sil, (2008), Rice Disease Identifica-


tion Using Pattern Recognition Techniques, Proceedings Of
11th International Conference On Computer And Information
Technology, 25-27.

[10] K. Khairnar, R. Dagade, (2014), Disease detection and diagno-


sis on plant using image processing - a review, Int. J. Comput.
Appl. 108 (13), 3639.

[11] L.N. Zhang , B. Yang , (2014), Research on recognition of


maize disease based on mo- bile internet and support vector
machine technique, Trans. Tech. Publ. 108 (13) ,659662.

[12] Z. Shicha , M. Hanping , H. Bo , Z. Yancheng , (2007), Mor-


phological feature extraction for cotton disease recognition by
machine vision, Microcomput. Inf. 23 (4) ,290292 .

[13] Alvaro Fuentes, Sook Yoon, Sang Cheol Kim, Dong Sun Park,
(2017), A Robust Deep-learning-Based Detector for Real-Time
Tomato Plant Diseases and Pests Detection. Sensors, 17(9):
2022.

[14] Mohanty S P, Hughes D, Salathe M, (2016), Using Deep-


Learning for Image Base Plant Disease Detection., Front. Plant
Sci.: 7:1419., 10.3389/fpls.2016.01419.

12

11078
11079
11080

You might also like