IRJET - Grape Leaf Diseases Classification using Transfer Learning

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 07 Issue: 03 | Mar 2020 www.irjet.net p-ISSN: 2395-0072
© 2020, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 3171
Grape Leaf Diseases Classification using Transfer Learning
Nitish Gangwar1, Divyansh Tiwari2, Abhishek Sharma3, Mritunjay Ashish4, Ankush Mittal5
1,2,3,4Student, Gurukula Kangri Vishwavidyalaya, Haridwar, India (249404).
5Guide, Raman Classes
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - India is the largest producer of fruits in the world.
Various species of fruits like apple, mango, grapes, banana etc.
are exported from India every year. According to the National
Horticulture Board (NHB), the area under grapes is 1.2% of
the total area of the fruit crops (2.8%) in India. Grapes are the
rich source of Fibre, Vitamins (especially C & K) andNutrients.
Grapes contain several antioxidants like resveratrol, which
helps in fighting against multiple diseases like cancer,
diabetes, blood pressure etc. Production and quality of the
grapes are highly influenced by various diseases and it is
difficult to identify these diseases on a grape plant byafarmer
as it requires an expert to discern it. Automatic Detection and
on-time treatment of these diseases can helpinsavingmillions
every year which will also foster the farmer’s yield. Over a
period of time, various implementations of deep learning
stratagem have been used for detecting severalplantdiseases.
Our motive in this research is to detect and classify the grape
diseases using the concept of transfer learning accomplished
with the help of Inceptionv3 followed by variousclassifiers like
logistic regression, SVM, Neural Network etc. using which we
achieved the state-of-the-art solution giving the highest
accuracy of 99.4% over the test dataset.
Key Words: CNN (Convolutional Neural Network), SVM
(Support Vector Machine), ANN (Artificial Neural
Network), LogisticRegression,InceptionV3,fine-tuning.
1. INTRODUCTION
Grape is a commercially important fruit crop of India and
currently ranks seventh-largest producer of grapes in the
world. According to APEDA (Agricultural and Processed
Food Products ExportDevelopmentAuthority), Maharashtra
ranks first in terms of production of grapes as it accountsfor
more than 81.22% of the total production of India. India is
the major exporter of fresh grapes in the world. In India,
more than 20 varieties of grapes are cultivated, covering
more than 123 thousand hectares (2.01%) of the total area.
Fresh Grapes are widely consumed and are also used for
making raisins in India. Fragmented Grapes are widely used
for the production of drinks like wine and brandy.
Diseases present in plants are the genesis of crop losses all
over the world. Grape plants are highly vulnerable towards
diseases like black rot, Esca, Leaf Blight etc, insect pests like
flea beetle, thrips, wasps etc, and disorders like berry drop,
berry cracking. Pest can be controlled using sprays. But in
case of diseases, timely detection and treatment efforts are
needed to be taken so that appropriatecontrol measurescan
be taken to have a healthy grape yield. So, an automated
system is required to clinch this requirement.
For a long time, disease identificationrequiresa well-trained
and an experienced expert but then also it’s a very error-
prone and time-worthy process. So, to combat it we have
made an automated system capable of detecting the disease
present inside the grape plant using the concept of transfer
learning [15] followed by some classifiers among whom
logistic regression outperformed others witha classification
accuracy of 99.4%.
2. Related Works
Sharada P. Mohanty et al [6] has used the concept of transfer
learning with GoogleNet and AlexNet over the PlantVillage
dataset with 38 classes, which resulted in test accuracy of
99.35%. Harvey Wu et al [7] has used 2 stage CNN pipelines
along with heat maps on UAV images in which he
successfully achieved test accuracy of 97.76%. Li et al [5]
have used SVM with different kernels like linear, RBF etc
among which linear kernel resulted in the optimum test
accuracy over the grape downy mildew and grape powdery
mildew classes with 90% and 93.33% respectively.Zhanget
al [8] has fine-tuned his model using GoogleLeNet and
detected podosphaera pannosa usingCNN,SVMandKNN out
of which he obtained a test accuracy of 99.6% using CNN
over the cherry leaves. Ch. Usha Kumari et al [9] has used K-
Means clustering algorithm for feature extraction over the
cotton and tomato leaves dataset and thenANN wasusedfor
Classification resulting in average test accuracy of 92.5%.
Athanikar et al [10] have used segmentation with K-Means
clustering and the features were extracted on the basis of
colour, texture and area in which they obtained a
classification accuracy of 92% over potato leaf samples.
Saraansh Baranwal et al [11] has worked over the apple leaf
dataset from PlantVillage Dataset [3] in which they have
used GoogleLeNet for fine-tuning their model over the four
classes: - Apple Black Rot, Apple Cedar Apple Rust, Healthy
Apple, and Apple Scab and obtained an average accuracy of
98.42%.
3. Dataset Description
The dataset used in this paper is PlantVillage Dataset [3]
which is available on Kaggle and is open source. It has
approximately 55,000 well-labelledimagesofhealthyleaves
and infected leaves of various fruits like apple, blueberry,
cherry, grapes, peach, pepper, orange etc. For each fruit,
more than one type of leaf disease is present and we

consider each type of disease as a separate class for our
classification task. Each image containsa pictureofa leafand
on a broad level for each class, we have two types of dataset
one is segmented which comprises a leaf without
background and other images have a leafwitha background.
Fig -1: Shows the sample image from each class of Dataset
The number of images in a particular class is not uniform, it
varies from 423 images to 1383 images. For our problem
statement, we haveusedonlyGrapeimageswhichcomprises
of four classes i.e. black rot, Esca(Black Measles), Leaf Blight
and healthy leaf images where the train-test-split data is
shown in table 1.
Table -1: Showing the number of samples in training and
testing set of our model
Label Category Number Training
samples
Testing
Samples
0 Black rot 1180 942 238
1 Esca 1383 1099 284
2 Leaf
Blight
1076 834 242
3 Healthy 423 334 89
Total 4062 3209 853
4. Experimental Setup
4.1Feature Extraction
Features of the images were extractedusing theInceptionV3
model [15]. InceptionV3 is based on convolution neural
network which comprises 42 layers as shown in figure 4.
InceptionV3 is trained over the subset of ImageNet dataset
consisting of 1000 images from each of the 1000 categories.
In the ILSVRC [15], the datasetcomprises1.2milliontraining
images,50000 validation images and 100000 testing images
in which inceptionV3 outperformed other models.
InceptionV3 uses a lot of optimizing techniques like
factorizing convolutions, efficient grid size reduction, the
utility of auxiliary classifiers etc. these techniques are
explained below:
4.1.1 Factorizing convolutions
InceptionV3 made several changes regarding the dimension
of convolutions since larger dimension convolution was
resulting into large computation so the surrogate approach
used in the paper [15] is that a 5*5 dimension is multicasted
into two 3*3 dimension .where 5*5 convolution has 25
parameters and updated version has only( 3*3 + 3*3 )=18
resulting into 28% reduction in the parameters
Fig -2: showing the change in the dimension of the feature
map using factoring convolution

This reduction into the calculation results in the very less
number of parameters presented inside the model as
compared to others like AlexNet [16], VGGNet [1] etc.
4.1.2 The utility of Auxiliary Classifier
The auxiliary classifier works as a regularizer. Lee et al [20]
in his work has stated that results obtained using auxiliary
classifier are very converging and accurate. but in case of
InceptionV3 results at the beginning of training phases are
not affected by the auxiliary classifier but results obtained
during the end period of training are quite influential so the
same approach is used in InceptionV3.
4.1.3 Efficient Grid Size Reduction
Usually, max-pooling and average pooling are used to
downsize the feature maps but an efficient method has been
used in the InceptionV3 model in order to downsize the
feature maps usingtheconceptofparallellyperformingmax-
pooling or average pooling andconvolution themethodused
is shown below in the figure.
Fig -3: represents the efficient grid size reduction used in
InceptionV3 model.
Original feature maps were operated parallelly by
convolution and pooling i.e. either average or max pooling
and then result obtainedusingconvolutionandpooling were
concatenated to give the resultantfeaturemap.Through this
way, InceptionV3 has been implemented in a less expensive
and in an efficient way.
4.2 Model
Over the period of time, Deep Convolutional Neural
Networks (DCNNs) [4] have achieved marvellous Results
over multiple places like in [11,13]. There are two ways to
train the model either from the scratch or using the concept
of transfer learning i.e. using the pre-
Fig -4: shows the model image finetuned using InceptionV3 and classified using Logistic Regression

trained models like VGGNet, InceptionV3, AlexNet etcwhich
could provide us with important features of the image. we
have used the InceptionV3 model because it comprises of a
smaller number of parameters as compared to othermodels
like AlexNet, VGGNet and also due to the state-of-the-art
results obtained using InceptionV3 over the plant village
dataset, which can be seen from the table 3
below.InceptionV3 model acts as the feature extractor for
our model, InceptionV3 model includes multiple
optimization techniques like factorizing convolutions,
Auxiliary Classifiers, Efficient grid size reduction etc all of
these techniques are elucidated above in section 4.1. Here
the dataset image is fed as input to the inceptionV3 model
which acts as the feature extractor and provides us with the
appropriate features of the image. After the feature
extraction phase, multiple classifiers were used but Logistic
regression outperformed others in terms of classification
accuracy as shown in table 3 below.
4.3 Classification
Regarding classification, we have used logistic regression
and in the field of statistical mathematics,logistic modelsare
used when dependent variables (target variables) are
categorical like win/lose, pass/fail etc. For using Logistic
Regression, first of all, we have to calculate the hypothesis
function. The output of the function is evaluated in terms of
probability. After which, we have to pass the evaluated
value of Hypothesis function to Cost Function. Cost Function
converts the probabilistic results to categorical results [2].
The flow of the model goes like the following:
H ' (M)=Prob(N=1|M;"∅") _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _1
The probability that given an event M, other event N=1
which is parameterized by variable ‘Φ’ as shown in equation
1.
Prob(N=1│M; ∅)+Prob(N=0│M;∅)=1
Prob(N=0│M; ∅)=1-Prob(N=1│M;∅)
Now cost function shown below in equation 2 is used to
convert probabilistic results to categorical results.
(H’(M), N) = -N*log(H’(M)) - (1-N) * log(1-H’(M)) _ _2
If N=1, (1-N) term will become 0, therefore – log (H ' (M))
will be present.
If N=0, (N) term will become 0, therefore - log (1-H ' (M))
will be present.
5. Result and Discussion
Images were fine-tuned using some well-known modelslike
InceptionV3(in our case), VGG16, VGG19 after the feature
extraction using these models. We demonstrated the results
in terms of evaluation metrics like AUC (Area Under the
Curve), CA (Classification Accuracy), Precision, Recall, F1-
Score using multipleclassifierslikeSVM,LogisticRegression,
Neural Network, K-Nearest Neighbor (KNN). Among all of
these classifiers, Logistic Regression outperforms others in
terms of classification accuracy as shown in table 3 below.
Whereas with this approach we have obtained the state-of-
the-art solution giving us 99.4% accuracy over the test
dataset as compared to others, result comparison report is
shown in table 2 below.
Table -2: Shows the comparison report with other models
Model Proposed Classification
Accuracy
Back Propagation Neural
Network (BPNN) [12]
91%
United Model [13] 98.57%
InceptionV3+LogisticRegression 99.4%

where ROC shows the AUC (Area Under the Curve) with
respect to four classes present in our model shown in
figure 5,6,7,8.

6. Limitations
Although the results obtained using our model were very
influential but using our approach, unlike segmentation we
cannot locate the part containing disease over the grape
leaves. Since we are only performing the classification task
over the present dataset.
7. Conclusion
An expert is a supreme requirement for early detection of
the disease in the grape plants. We designed an automated
system to ameliorate the classification results. We have
provided multiple results by fine-tuning our model with a
number of pre-trained models like VGG16, VGG19,
InceptionV3 etc. and achieved the state-of-the-art solution
over the plant village dataset shown in table 1. Our model
achieved an exceptional classification accuracy of 99.4%.
which can widely meliorate the early detection of diseases
like black rot, Esca, Leaf Blight present overthegrape plants.
The development of an algorithm can help people in
detecting the disease more early so that they can cure it well
on time and can earn a good profit. A similar approach can
also be used with other plants leaves like apple, blueberry,
cherry, grapes, peach, pepper, orange etc.
References
[1] Simonyan, Karen & Zisserman, Andrew. (2014). Very
Deep Convolutional Networks for Large-Scale Image
Recognition. arXiv 1409.1556.
[2] Peng, Joanne & Lee, Kuk & Ingersoll, Gary. (2002). An
Introduction to Logistic Regression Analysis and Reporting.
Journal of Educational Research - J EDUC RES. 96. 3-14.
10.1080/00220670209598786.
[3] Abdallah Ali (2019, September) PlantVillage Dataset,
Version 1, Retrieved 22 Ferbruary
2020,https://www.kaggle.com/xabdallahali/plantvillage-
dataset.
[4] Lee, Chen-Yu et al. “Deeply-Supervised Nets.” ArXiv
abs/1409.5185 (2014): n. pag.
[5] Li, Guanlin & Ma, Zhanhong & Wang, Haiguang. (2011).
Image Recognition of Grape Downy Mildew and Grape
Powdery Mildew Based on Support Vector Machine. IFIP
Advances in Information and Communication Technology.
370. 151-162. 10.1007/978-3-642-27275-2_17.
[6] Mohanty, Sharada & Hughes, David & Salathe, Marcel.
(2016). Using Deep Learning for Image-Based Plant Disease
Detection. Frontiers in Plant Science. 7.
10.3389/fpls.2016.01419.
[7] Wu, Harvey & Wiesner-Hanks, Tyr & Stewart, Ethan &
DeChant, Chad & Kaczmar, Nicholas & Gore, Michael &
Nelson, Rebecca & Lipson, Hod. (2019). Autonomous
Detection of Plant Disease Symptoms Directly from Aerial
Imagery. tppj. 2. 10.2135/tppj2019.03.0006.
[8] Zhang, Keke & Zhang, Lei & Wu, Qiufeng. (2019).
Identification of Cherry Leaf Disease Infected by
Podosphaera Pannosa via Convolutional Neural Network.
International Journal of Agricultural and Environmental
Information Systems (IJAEIS). 10. 98-110.
10.4018/ijaeis.2019040105.
[9] Kumari, Usha & Prasad, S. & Mounika, G.. (2019). Leaf
Disease Detection: Feature Extraction with K-means
clustering and Classification with ANN. 1095-1098.
10.1109/ICCMC.2019.8819750.
[10] Athanikar, ..G., & Badar, M.P. (2016). Potato Leaf
Diseases Detection and Classification System Mr.
[11] Baranwal, Saraansh & Khandelwal, Siddhant & Arora,
Anuja. (2019). Deep LearningConvolutional Neural Network

for Apple Leaves Disease Detection. SSRN Electronic Journal.
10.2139/ssrn.3351641.
[12] Zhu, Juanhua & Wu, Ang & Wang, Xiushan & Zhang,Hao.
(2019). Identification of grape diseases usingimageanalysis
and BP neural networks. MultimediaToolsandApplications.
10.1007/s11042-018-7092-0.
[13] Miaomiao Ji, Lei Zhang, Qiufeng Wu, Automatic grape
leaf diseases identification via UnitedModel based on
multiple convolutional neural networks, Information
Processing in Agriculture,2019, ISSN 2214-
3173,https://doi.org/10.1016/j.inpa.2019.10.003.
[14]https://ai.googleblog.com/2016/03/train-your-own-
image-classifier-with.html
[15] Szegedy, Christian & Vanhoucke, Vincent&Ioffe,Sergey
& Shlens, Jon & Wojna, ZB. (2016). Rethinking the Inception
Architecture for Computer Vision. 10.1109/CVPR.2016.308.
[16] Krizhevsky, Alex & Sutskever, Ilya & Hinton, Geoffrey.
(2012). ImageNet Classification with Deep Convolutional
Neural Networks. Neural Information Processing Systems.
25. 10.1145/3065386.

IRJET - Grape Leaf Diseases Classification using Transfer Learning

More Related Content

What's hot (20)

Similar to IRJET - Grape Leaf Diseases Classification using Transfer Learning (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET - Grape Leaf Diseases Classification using Transfer Learning