IEEE Consumer Electronics Magazine - May/June 2022 - 46

Special Section on Consumer Technologies for Smart Agriculture
Table 1. Dataset details of the proposed system.
Crop
Disease
name
Apple
Grape
Tomato
Mango
Alstonia
Guava
Figure 1. Sample images of diseased leaf images.
phase. However, such a collection of millions of
data samples is formidable in practice and also
time-consuming. In contrast, a transfer learning
mechanism is adopted to reduce the training
time of the model. Furthermore, it allows the convolution
weights to reuse, where the model has
trained on massive large image datasets instead
of training from scratch by random weight initialization.
In this work, we leverage the pretrained
weights from a massive COCO image dataset.
In this work, we consider two datasets to
carry out our experiments. i.e., to classify the
diseased images and locate the infected portions
of the disease leaves based on symptoms. The
two datasets, namely, PlantVillage Dataset10 and
Leaves Images Dataset11 contain a collection of
leaf images of healthy and diseased plants with
distinct crop species. Here, we utilize only eight
crop species with a total of 3,038 images are collected
from these two dataset repositories,
which are labeled into eight classes. A few samples
of plant disease combinations and (including nondiseased)
anddetails of the selectedimage dataset
are furnished in Figure 1 and Table 1, respectively.
For deep neural networks, it is necessary to collect
sufficient data tomake the training process stable.
After the dataset split, the data augmentation technique
applied to the training set. This technique
helps to boost detection performance and also
prevent overfitting.
Manual Annotation
Image annotation aims to mark the infected
portion in the diseased leafy image and assign a
46
Pomegranate
Jamun
Healthy
Total
Label
name
Apple-Scab AAS
Black-rot GBR
Target-spot TTS
Anthrocnose AN
FoilarGalls
FG
Fungal
disease
Cercospora
spot
Fungal
disease
FD
CS
FD
Nondiseased Healthy
9
9
#
Original
images
510
400
350
265
254
142
272
345
500
3,038
label to it. There are many image annotation
tools available for this purpose. The " labelImg "
annotation tool is an open-source tool for performing
annotations manually. This tool uses a
bounding box as a representation for marking
the diseased portion in the leaf. These annotated
leafy images are saved as XML files. Each annotated
image contains the coordinate values
xmin, ymin, width, height of bounding boxes,
and the associated labels. Finally, all annotated
images are converted into COCO json format
given as ground truth to the model to extract
deep feature representations on the diseased
leaves images within a stipulated time.
METHODOLOLGY
The proposed framework used in this work is
inspired based on the EfficientDe12 deep learning
model is depicted in Figure 2. It is a composite
network and relies on three main components:
Backbone Network, BiFPN Network, and Class/
Box Prediction Network.
The Backbone Network part of the framework
used to extract the in-depth features from
images. In this work, EfficientNet-B213 was chosen
from the EfficientNet family to extract the indepth
image features for the CNN architecture to
balance the computational cost and detection
IEEE Consumer Electronics Magazine
IEEE Consumer Electronics Magazine - May/June 2022

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - May/June 2022
