Scribd
Upload
9 views6 pages

Tomato

research paper

Uploaded by

Naineni Shiny
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
9 views6 pages

Tomato

research paper

Uploaded by

Naineni Shiny
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 6

ToLeD : Tomato Leaf Disease Detection using Convolution Neural

Network
Abstract:
 Tomato is the most popular crop in the world. India is ranked 2
in production of tomato
 Convolution Neural Network - In this model, there are 3
convolution and 3 max pooling layers followed by 2 fully
connected layer.
 The experimental results shows the efficacy of the proposed
model over pre-trained model i.e. VGG16, InceptionV3 and
Mobile Net. classification accuracy varies from 76% - 100% with
respect to classes and average accuracy of the proposed model
is 91.2% for the 9 disease and 1 healthy class.
Introduction:
Disease identification :
 brown or black patches and holes on the plant
 Tomato need to planted only once a year and remaining time
we need to plant grass family like wheat, corn, sugarcane..
 Diseases are identified into two categories:
1. bacteria or fungi or poor cultivation habits causing - 16
diseases
2. insects causing - 5 other type of diseases
Ralstonia solanacearum bacteria:
 This bacterium can survive in soil for long time period and
enter roots through natural wounds made during secondary
roots emergence or man made during cultivating or
transplanting or even insects.
 High moisture and high temperature favours disease
development
 This results in affecting the vascular system of plant, while the
leaves may stay green. On a cross section view of an infected
plant stem, it appears brown with yellowish material coming
out of it.
Literature Survey:
Deep neural network- Avoids the problem of feature engineering
Mohanty et al: used AlexNet to train classify plant diseases, However the
model accuracy has been decreased as the testing image conditions are
different from the training images

Rangarajan et. al: trained both AlexNet and VGG16net


Brahimi et al [1] introduced saliency map method for localization of infected
regions. This type of visualization improves classification accuracy

Dataset:
 Images of Tomato disease have been taken from Plant Village dataset .
The dataset includes over 50,000 images of 14 crops, such as tomatoes,
potatoes, grapes, apples, corn, blueberry, raspberry, soybeans, squash
and strawberry. We selected Tomato as our target crop.

 There are mainly nine types of diseases in tomato: 1) Target Spot, 2)


Mosaic virus, 3) Bacterial spot, 4) Late blight, 5) Leaf Mold, 6) Yellow Leaf
Curl Virus, 7) Spider mites: Two-spotted spider mite, 8) Early blight and
9) Septoria leaf spot
 Size of all the images is 256 × 256 and format is jpeg
 training dataset – 10000 , 1000 images are healthy and 9000 images are tomato disease
catageory ( belongs to 9 catageories)
validation dataset -7000, each class has 700 images 1 healthy class and 9 unhealthy classes
testing dataset – 500 , 50 images from each class.
 For testing, we randomly picked 50 images from each class from training set and removed
them from those folders, as each class contains 1000 images will decrease as we removed
images from each class we use data augmentation technique to generate some new images.
 Augmentation was done using Augmentor package of python and it helps to build similar
new images by rotating, flipping, cropping and resizing the existing images.

Experimental result and Discussion


In proposed CNN architecture there are three convolution and max pooling layer are used. In each
layer various number of filters has been applied. We can see it in fig2

To get the better accuracy, Panda approach has been applied and the hyper parameters for the
model is described in table1
(a) Image passed in the CNN model (b) extracted features at first convolution layer (c) extracted
features at second hidden layer (d) extracted feature at third hidden layer and (e) extracted
features at fourth layer hidden layer

 For the calculation of loss, the categorical cross entropy method has
been applied. The formula for calculating it represented in following
equation
loss = − M
sigma log(po,c)
c=1
where M - number of classes, y - binary indicator (0 or 1) if class label c is
the correct classification for observation o and predicted probability
observation o is of class c

For the testing purpose, total 500 sample has been used and for the various classes testing accuracy
is different and it is ranging from 76% to 100% and the average accuracy of the proposed model is
91.2%

we ran the proposed model for 1000 epochs and validation and training accuracy
we also ran some pre-trained model for the performance comparison purpose

 we also ran some pre-trained model for the performance comparison


purpose
(1) VGG16 - 77.2%
(2) Inception V3 - 63.4 % when we trained for 1000 epochs on NVIDIA
(3) Mobilenet - 63.7% accuracy
(4) developed CNN based model - 91.2%

Conclusion:
The average testing accuracy of the model is 91.2%. The storage space needed by proposed model is
of order of 1.5 MB whereas pretrained models have storage space needs of around 100 MB thus
showing the benefit of the proposed model over pretrained models.

You might also like