Plant Disease Detection Using Deep Learning: Anjaneya Teja Kalvakolanu

1
Plant Disease Detection Using Deep learning

Anjaneya Teja Kalvakolanu
Department of Computer Science, California Polytechnic State University
Abstract- Plant disease detection is a huge problem and often

require professional help to detect the disease. This research
focuses on creating a deep learning model that detects the type of
disease that affected the plant from the images of the leaves of
the plants. The deep learning is done with the help of
Convolutional Neural Network by performing transfer learning.
The model is created using transfer learning and is experimented
with both resnet 34 and resnet 50 to demonstrate that
discriminative learning gives better results. This method
achieved state of art results for the dataset used. The main goal is
to lower the professional help to detect the plant diseases and
make this model accessible to as many people as possible.
Index Terms- Deep Learning, Plant disease detection,

Discriminative Learning, Resnet 34, Resnet 50, Plant village.
I. INTRODUCTION
Implementation of RGB and Gray scale images in plant disease
T his Research is focused on Detection of diseases that plant is
affected by using deep learning. The images of affected plant
leaves are taken as input and are sent through a network trained
detection in leaves a study by Padmavathi and Thangadurai have
given the comparative results of RGB and Gray scale images in
using transfer learning from resent 34 and resent 50. We disease finding process from leaves. Color is an important feature
considered three plants for this research Potato, Tomato, Bell in detecting the infected leaves. Color also identifies the intensity
pepper. In particular the diseases with plants considered are early They have considered Grayscale and RGB images and used
blight, mosaic virus, late blight, leaf mold, bacterial Spot, Leaf median filter for image enhancement and segmentation for
curl, Target spot ,spider mite. In addition we also considered extraction of the diseased portion which are used to identify the
Healthy plants as a class. The classes are divided based on plant disease intensity. They identified 13 different diseases from the
and disease type since each plant has a different leaves shape and healthy leaves with the capability to differentiate from the
many diseases appear similar on same plant. We used transfer surroundings from the leaves.
learning with both resent 34 and resnet 50 as our base model. We
trained the model with around 4000 images with discriminative
learning . we achieved an accuracy of 99.44% which is current
state of art for this dataset.
II. BACKGROUND
III. EXPERIMENT
There has been fast paced research for Plant disease detection
using images in the recent years. Penn state has created their own The transfer learning is done using both resent 34 and resnet 50.
Plant village has developed their own dataset to create a deep The data set obtained has over 4000 images that are divide in to
learning model for detection of diseases in plants. There have classes based on the plant type and disease type. The pre
been many models developed and tested on the plant village processing include data augmentation. The images are flipped,
dataset. Methods like segmentation are used to detect the affected warped, transformed to make the dataset more generic.
part of the leaves and they are classified using the deep learning
methods. Vgd is used to classify the images and they achieved an
accuracy of 97.2. In order to develop image classifiers for the
purposes of plant disease detection , there is a requirement of
large and verified dataset of healthy and diseased plants. Until
plant village such dataset is not available.
2
This initial training worked well and we achieved an accuracy of

98.7 percent after training for 4 epochs with a batch size of 64.
Once the training is done to perform discriminative learning all
the other layers are unfrozen and the training is performed with
higher learning rate for the initial layers and the lower learning
rate for he end layers. The end layers are already good at
detecting the features of the images sent in so it makes sense to
train them with low learning rate. The initial layers have only the
pre trained weights so they should be trained to be able to detect
the features of the images so high learning rate is used. This
technique worked in improving the accuracy further more. Hyper
parameter tuning is performed again.
The dataset is turned in to a batch of 64 images, each image is of

size 224 * 224. The normalization is performed with the image
net stats since the imported resnet model is trained on the image
net .The loss function used is categorical cross entropy.
IV. DISCRIMINATIVE LEARNING

The CNN model created from the resnet is trained by using the
discriminative learning. The initial training is done only for the
newly added layers after performing the hyper parameter tuning
for learning rate.
The next training is done with learning rate between 1e-05 and
1e-04 with high learning rate for initial layers and then
distributed the learning rate with the multiples for the next layers
between start and end learning rate. This method gave us an
accuracy of 94.4 percent which is the current state of art.
V. RESULTS
The model is trained with 3e-03 learning rate only for the newly
added layers. We tried working with both resnet 34 and resnet 50 even
though resnet 50 achieved better accuracy than resnet 34 (99.1%)
3
took very less time to enter achieve 99 accuracy. Since the

validation loss and training loss both are lowering and the
accuracy is not falling plotted the training and validation losses
which makes sure that the model is not over fitted. The top most
losses are plotted in the format actual/predicted/loss/probability
and it is clear that most of them happened due to errors in dataset
and even human can go wrong with them. The confusion matrix
is also plotted.
VI. CONCLUSION
Our proposed model with Discriminative learning rate
achieved an accuracy of 99.44 percent which is the current state
of art for this dataset. New data can be added to train the model
and this will not lower the accuracy because of the regularization
and deeper layers of resnet 50. This work can be further extended
by adding more data and training the network.
REFERENCES
Top most losses : Predicted/Actual/Loss/Probability
[1] Mohanty, Sharada P., David P. Hughes, and Marcel Salathé. "Using deep
learning for image-based plant disease detection." Frontiers in plant
science 7 (2016): 1419.
[2] Martinelli, Federico, et al. "Advanced methods of plant disease detection. A
review." Agronomy for Sustainable Development 35.1 (2015): 1-25.
[3] Emmanuel, Plant village dataset, “Kaggle”, 2019.
[4] Jeremy howard, “Fast AI”, fastai.com
[5] Ferentinos, Konstantinos P. "Deep learning models for plant disease
detection and diagnosis." Computers and Electronics in Agriculture 145
(2018): 311-318.
.
4

Plant Disease Detection Using Deep Learning: Anjaneya Teja Kalvakolanu

Uploaded by

Copyright:

Available Formats

Plant Disease Detection Using Deep Learning: Anjaneya Teja Kalvakolanu

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Plant Disease Detection Using Deep Learning: Anjaneya Teja Kalvakolanu

Uploaded by

Copyright:

Available Formats

1

Plant Disease Detection Using Deep learning

Abstract- Plant disease detection is a huge problem and often

Index Terms- Deep Learning, Plant disease detection,

This initial training worked well and we achieved an accuracy of

The dataset is turned in to a batch of 64 images, each image is of

IV. DISCRIMINATIVE LEARNING

took very less time to enter achieve 99 accuracy. Since the

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.