Automated Plant Identification with CNN

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 02 | Feb 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 459
Automated Plant Identification with CNN
Akash Yadav1, Smit Viradiya2, Milan Vaghasiya3, Uma Goradiya4
1,2,3BE Student, Department of Computer Science and Engineering, Shree L.R. Tiwari College of Engineering,
Maharashtra, India
4Assistant Professor Department of Computer Science and Engineering, Shree L.R. Tiwari College of Engineering,
Maharashtra, India
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - In both botanical taxonomy and computer vision,
plant image recognition has become a transdisciplinary
priority. We use convolutional deep neural networks to
identify the types of plants photographed in the image and to
explore various factors that affect the performance of these
networks. Determining plant species from field observations
requires a large botanical specialist, whichputsmorethanthe
reach of many nature lovers. Discovery of traditional plants is
virtually impossible for the general public and is a challenge
even for professionals who deal with daily crop problems such
as conservationists, farmers, foresters, andarchitects. Even for
botanists themselves, the speciesidentificationis oftendifficult
task. This change enhances the accuracy of the model
significantly. This model is tested in a leaf set of Flavia leaf
sets. The results of the identification phase show that the
proposed CNN model is still effective in leaf recognition with
an optimal accuracy of more than 88.22%.
Key Words: Deep learning, leaf classification,
convolutional neural networks.
1. INTRODUCTION
Plants are vital to the preservation of natural resources.
Plant species identification gives useful knowledge on plant
classification and traits. Because it requires an individual's
visual perception, manual interpretation is not precise. It is
simple to sample and capture digital leaf photos, which
include textural elements that aid in determining a certain
pattern. The venation and form of a leaf are the most crucial
features to distinguish across plant species. As information
technology advances,toolssuchasimageprocessing,pattern
recognition, and others are used to identify plants based on
the description of leaf shape and venation, which is a critical
concept in the identification process. Itistoughtokeeptrack
of changing leaf properties throughout time. As a result, a
dataset must be created as a reference for a comparable
study. Because of their appealing characteristics and year-
round availability, leaves are included in most plant
identificationprocedures.Automatic plantimagerecognition
is the most promising option for closing the botanical
taxonomic gap, and it has gotten a lot of interestfrom botany
and the computing community. As machine learning
technology progresses, more sophisticated models for
automatic plant identification havebeendeveloped.Millions
of plant photographs have been obtained thanks to the
popularity of cellphones and the launch of PlantNet mobile
apps. In real-world social-based ecological surveillance,
invasive exotic plant monitor, ecological science
popularization, and other applications, mobile-based
automatic plant identification is critical. Improving the
efficacy of mobile-based plant identification algorithms [2]
has piqued the interest of researchers.
In computer vision, despite many attempts [3 - 8] (with
computer vision algorithms), crop identification is still
considered a challenging and unresolved issue. Leaf snap an
automated plant identification system that identified using
integral measure to compute functions of the curvature at
the boundary, plants based on curvature-based shape
aspects of the leaf. The plant recognition process is largely
divided into the following steps shown in Figure 1.
Fig -1 [1]: Steps of an Image-Based Plant Classification
Process
Image Acquisition: The purpose of this step is to obtain an
image of plant species to perform pre-processing and
classification operations.
Pre-processing: The pre-processing process detects the
image as inserted and produces a modified image as output,
which is suitable for the action to remove the feature.
Previous processing usually includes tasks such as sound
removal, image enhancement, and separation. It improves
the availability of plant species.
Feature Extraction: Feature removal is an important step in
identifying plant species. Refers to extracting image
information from the control groups.Featurereleasescan be
divided into global and local features. Earth features mean
representation of elementsthatdefineall imageinformation,

such as shape, texture, and color elements. Local features
refer to the part of the image within a particular region.
Classification: In the separation step, all outputelementsare
connected to the feature vector, which is then separated.
Advanced methods use classification algorithms, such as k-
Nearest Neighbor (k-NN), Decision Tree (DT), Vector
Support Machine (SVM) and Artificial Neural Networks
(ANN) .The efficiency of all methods. Much depends on the
factors described earlier. Graphical engineering itself is a
complex process that requires changes and recalculation of
each problem or set of related data.
With the development of neural networks, neural network
architecture has been used as an effective solutiontoextract
high-quality features from data. Deep Convolutional Neural
Network structures can accurately visualize structures that
cannot be summarized while conserving modern featuresof
raw data. This is useful for dividing or guessing. In more
recent times, the Convolutional Neural Network (CNN) has
emerged as a functional framework fordefiningfeaturesand
identities in image processing. CNN can learn the basics
automatically and integrate them systematically to define
basic concepts to identify patterns. While using CNN one
does not have to do differently the time-consuming
engineering features. The customization of the method
makes it a practical and useful method for the various
problems of the application of isolation and recognition.
2. LITERATURE SURVEY
Recently, a few modern methods have used depth learning-
based methods for identifying plant species. Deep learning-
based approaches use in-depth learning algorithms such as
convolutional neural networks (CNN) to extract features as
well separate leaf pictures.
J Wei Tan, S.W. Chang, S. Abdul-Kareem, H. J. Yap, and K.T.
Yong [9] proposed a CNN-based approach called DLeaf.
Pictures of the leaf of the plant are pre-processed and
extracted features using three CNN models namely pre-
trained AlexNet, well configured AlexNet, and D-Leaf. Five
machines Classes are used for classification as SVM, ANN, k-
NN, NB and CNN. In the process of processing, the captured
images are converted to the marked file format (TIFF) and
sounds are removed using Adobe Photoshop. The image is
redesigned into a square dimension and so on resized to
250x250 adjustment. Sobel edge detection methodisused in
the area of interest (ROI). Because all these images are
converted from RGB to grayscale, background image
classification was processed again skeletonized to find a
pure vein architecture. After preliminary processing,feature
removal was performed using CNN models as mentioned
earlier and the feature is extracted using a common
morphological method called vein features extracted from
the image separated by measuring morphological vein
features. Testing shows CNN's method is better than the
usual morphological method. Proposed D-Leaf feature
issuance with ANN as a classification gives a test of 94.88%
accuracy.
Fig -2: Deep Learning for Plant Species Classification
Truong Quoc Bao, Nguyen Thanh Tan Kiet, Truong Quoc
Dinh and Huynh Xuan Hiep [10] proposed two solutions for
classification leaf by using image processing combined with
a shallow architecture and a deep architecturetoclassifythe
image leaf. We must find parameters by hand or by hand-
designed feature extraction when recognizing shallow
architecture. We use HOG features for recognition of the
result in Table 3. With recognition, deep architecture
combined with the effect of horizontal reflection and
rotation augmentation of datasets further improves the
results. Shallow architecture'saccuracyisdeterminedonthe
image's input size as well as the length of the feature vector.
Deep architecture with the same input image excitation, has
the feature vector is shorter and they arefound bythemodel
itself for greater accuracy. The results show that the
proposed architecture for CNN-based leaf classification
closely competes with the latest extensive approaches on
devising leaf features and classifiers. Comparisons of the
results were compiled in Wäldchen and Mäder [11] with the
method of leaves recognition on the SwedishandFlavia data
sets. From experiment results with Swedish and Flavia data
sets, we can confirm that the CNN-based neural network
depth model, which we propose, works very well on
classification problems of leaves basedontheshapeofveins.
This finding verifies the CNN depth geometry model's
usefulness and simplicity for real-world issues with
enormous data. Building the model and finding the right
parameters is all that is required for the recognition
procedure. The effectiveness oftheclassificationprocesshas
improved, and recognition is no longer as reliant on
discovering and recognizing visual features, which is a time-
consuming and labor-intensive procedure.

Fig -2: Scheme Implementation.
In this work, the grayed stages were compared with a deep
convolutional network.
S. A. Riaz, S. Naz, and I. Razzak [12] introduced the plant
diagnostic method using a multipath deep convolutional
network. In the training phase, the images are pre-
processed, extracted features from the pre-processedimage
and classified. In the testing phase, the images not used for
training are processed and tested by feeding them into a
neural network. The proposed architecture is displayed in
Fig. 4 containing multiple CNN blocks, max pooling layers,
flat and soft layer - maximum size separation of images of
input plants. Each block contains three convolutions, one
batch normalization, one max-pooling layer and a single
thick layer. Features released in one block are combined
with features from the second block. The soft-max layer
separates plant species. This study uses the two data sets
LeafSnap and MalayaKew and achieved 99.38% accuracy
and 99.22% respectively.
Fig -2: INFOPLANT: Plant Recognition Using
Convolutional Neural Networks
3. PROPOSED METHOD
The prediction loss minimization method is commonly used
to train deep neural networks.
Convolutional Neural Network
Convolutional neural network is an effective detection
method, which has been developed recently. In past years,
which garnered widespread attention. Now, CNN has
become one of the most efficient methods in the field of
pattern classification and, more recently, have been used
more widely the field of image processing (Krzyzewskiet al.,
2012; Lecan [13], Bengio, & Hinton, 2015)[14], and It can
reach better performance than traditional methods
(Chatfield, Lempitsky, Vedaldi, and Zisserman, 2011).
Through extensive verification. A CNN consists of one or
more pairs of convolutional and maximum pooling layers. A
convolutional layer applies a set of filters that process
smaller Local parts of the input where these filters are
repeated along the entire input space. Max poolingproduces
a low-resolution versionofactivatingtheconvolutional layer
activation of the top filter from different locationswithinthe
specified window. It addstranslationchangesandtolerances
for minor differences in position of objects parts. Higher
layers use more extensive filters that operate on lower
resolution inputs Process the more complex parts of the
input. The top fully connected layers are finally combined
Input from all positions to classify the overall input. This
hierarchical organization produces good results in image
processing tasks.
Convolutional neural network structure
● Convolution layer: The convolution procedure
extracts several information from the input. The first
layer of convolution produces low-level features such as
edges, lines, and corners.
● Nonlinear layers: A nonlinear "trigger" function is
used by neural networks in general and CNNs in
particular, to signify different identification of likely
features on each hidden layer. CNNs may use a variety of
specific functions such as rectified linear units (ReLUs)
and continuous trigger (nonlinear) functions to
efficiently implement this nonlinear triggering. ReLU
uses the function y = max (x, 0), so the input and output
size of this layer are the same. It increases the nonlinear
properties of the decision function and of the overall
network without electing the receptive fields of the
convolution layer. In comparison to the other nonlinear
functions used in CNNs (e.g., hyperbolictangent,absolute
or hyperbolic tangent, and sigmoid), the advantage of a
ReLU is that the network trains many times faster. ReLU
functionality is illustrated in Fig. 6.
● Factor adjustment reduces the mixing / sample
layer. Strengthens the resistanceof elementstonoiseand
distortion.
Fully Connected Layers : These layers add up the weighted
sum of the previous layer's characteristics, indicating the
exact mix of "ingredients" needed to get a given intended
output result. In the case of a fully connected layer, all the
elements of all the features of the previous layer get used in
the calculation of each element of each output feature.

Table -1: The Proposed CNN Architecture
Attribute L1 L2 L3 L4 L5
Type Conv MaxPool Conv MaxPool Conv Conv
Filtersize 8 × 8 7 × 7 5 × 5 6 × 6 2 × 2 3 × 3
SoftMax
Stride 2 2 2 2 1 1
No. of
filter
100 100 250 250 100 100
Output
size
61 × 61 28 × 28 13 × 13 4 × 4 3 × 3 1 × 1
Training is done using a set of ‘labeled’ input data in a wide
variety of input patterns that represent tags with their
intended output response. Training uses general purpose
methods to repeatedly determine medium and final weights
features of neurons. Figure 5 shows the training process at
block level
The proposed CNN architecture
CNN formats vary in the type of images and especially when
the size of the input images different. In this paper, the size
of the input images is estimated to be 128 × 128 pixels. After
each Conv layer, ReLU activation function is used and for
each compilation layer,the MaxPoolingmethodisused. Fully
connected layers are defined as flexible layers with a 1 × 1
filter size as is common in MatConvnet (Vedaldi & Lenc,
2015) The last layer has n units corresponding to n. and a
leaf set of data sets. After all the layers, the SoftMax losses
are fixed. There are 5 layers including: [Conv - ReLu – Max
pool] → [Conv - ReLu - Max pool] → [Conv - ReLu] → [Conv→
FC] → Softmax.
Fig -5 [15]: Training of Neural Networks.
Fig -6: Pictorial Representation of Relu Functionality
The Final Layer has n units corresponding to n and category
of leaf data sets. After all layers, a SoftMax layer is placed.
4. EXPERIMENTS AND RESULTS
Experiment data sets
In order to test the performance of the partition system, we
have selected two standard sets:
● Swedish leaf data set: Swedish leaf data set taken as
part of joining a leaf-splitting project between
Linkoping University and the Swedish Museum Of
Natural History (Soderkvist, 2001). The data set
contains images of a single leaf and examines the
empty base of 15 Swedish tree species, with 75
leaves per species (1125 photos in total). This data
set is considered to be very challenging due to its
height similarity of different species.
● Flavia Data Set: This data set contains 1907 leaf
images of 32 different types and 50-77pictures of
each type. Pictures of leaves found by scanners or
digital cameras on an empty back.Theindividual leaf
images contain only the blades, without petioles.
Augmentation data
Due to the limitations of most models due to insufficient
data. Augmentation records is a powerful answer. Increase
the number of training photos by creating three copies of
each image after careful thought and rotation. So,eachofthe
original images creates three image enhancements.Thedata
divided for testing is shown in Table 2.
In this study, model input data is scanned or taken from tree
leaves, and makes a training model. The screening processis
as follows:
● Standardized image size: image resizes 128 × 128px
to fit input network. To resize images to the desired
size of 128 × 128 pixels, the first images are size so
that their size is equal to 128, and then smaller
combined with 0 pixels.
● . Image Partition: Each category is categorized as
shown in Table 2 for testing objectives.

Initialization parameter: Reading level: set to 0.00007
(greater than the fastest combination network but not very
good error rate, which is smaller than the network of slow
connections).The above constants are selected primarily
based on experimental effects for the proposed model with
the aid of trial and error method. The amount of timeittakes
to train is determined by the computer's GPU or CPU
resources.
Table -2: Partition Data
Data
set
Number
layers
Number
Samples
Train Val Test
60%
Augmentat
ion data 20% 20%
Swedis
h
15 1125 900 3600 225 225
Flavia 32 1907 1145 4580 381 381
Flavia
+ Swed
ish
47 5825 2045 8180 606 606
Experiment results
The experiment results are aggregatedinto eachtestdata set
and detail in Table 2.
With the CNN model, data sets are divided into 80%, 20%
and 20% as training, validation and sequential test sets,
followed by extension training. Input data is128×128,layer
L4 (Table 1) is a vector of 1 × 100 aspect ratio with Softmax
function.
Table -3: Experiment results of CNN model.
CNN Model
Rate
Train/Val/Test
60-20-20
Augmentation
data
Accuracy test set
CNN
Swedish (15
layers)
No 96.15
Yes 98 . 22
Flavia (32
layers)
No 93.8
Yes 95 . 5
Swedish + Flavia No 94.17
Yes 95.6
The experiment outcomes are summarized in Table 3 for
each test data set.
5. CONCLUSION
Especially in this field, the effective use of in-depth learning
in the agricultural sector reported, mainly to plant
identification from the leaf. The mainresultisthatwegained
improved accuracy usinga standardin-depth readingmodel.
The proposed method will replace the first image channel
and remove the leaf arterial data and augmentation to
reduce overload by resizing and rotating images. It is often
argued that neural networks do not provide any new
perspective on a learning problem, as they fall into the
category of black box models. However, due to the simple
viewing method, it can be shown that images of patterns in
the veins can be helpful in thought-provoking work.
REFERENCES
[1] Manisha M. Amlekar, Ashok T. Gaikwad “Plant
Classification Using Image Processing and Neural
Network.”,10.1007/978-981-13-1274-8 (Chapter 28) ,
375-384. doi:10.1007/978-981-13-1274-8_29
[2] V. E. Balas et al. (eds.), Data Management, Analytics and
Innovation, Advances in Intelligent Systems and
Computing 839, https://doi.org/10.1007/978-981-13-
1274-8_29
[3] Neeraj Kumar, Peter N Belhumeur, Arijit Biswas, David
W Jacobs, W John Kress, Ida C Lopez,and JoãoVBSoares,
“Leafsnap: A computervisionsystemforautomaticplant
species identification,” in ECCV, pp. 502–516. Springer,
2012.
[4] Cem Kalyoncu and Önsen Toygar, “Geometric leaf
classification,” Computer Vision and Image
Understanding, vol. 133, pp. 102–109, 2014
[5] Abdul Kadir, Lukito Edi Nugroho, Adhi Susanto, and
Paulus Insap Santosa, “Leaf classification using shape,
color, and texture features,” arXiv preprint
arXiv:1401.4447, 2013.
[6] Thibaut Beghin, James S Cope, Paolo Remagnino, and
Sarah Barman, “Shape and texture based plant leaf
classification,” in Advanced Concepts for Intelligent
Vision Systems, 2010, pp. 345–353.
[7] James Charters, Zhiyong Wang, Zheru Chi, Ah Chung
Tsoi, and David Dagan Feng, “Eagle: A novel descriptor
for identifying plant species using leaf lamina vascular
features,” in ICME-Workshop, 2014, pp. 1–6.

[8] James S Cope, Paolo Remagnino,SarahBarman,and Paul
Wilkin, “The extraction of venation from leaf images by
evolved vein classifiers and ant colony algorithms,” in
Advanced Concepts for Intelligent VisionSystems,2010,
pp. 135–144.
[9] J. wei Tan, S.W. Chang, S. Abdul-Kareem, H. J. Yap, and
K.T. Yong. Deep learning for plant species classification
using leaf vein morphometric. IEEE/ACM Transactions
on Computational BiologyandBioinformatics,17(1):82–
90, 2018.
[10] Quoc Bao, Truong; Tan Kiet, Nguyen Thanh; Quoc Dinh,
Truong; Hiep, Huynh Xuan (2019). Plant species
identification from leaf patterns using histogram of
oriented gradients feature spaceandconvolutionneural
networks. Journal of Information and
Telecommunication, (), 1–11.
doi:10.1080/24751839.2019.1666625
[11] Wäldchen, J., & Mäder, P. (2017). Plant species
identification using computer vision techniques: A
systematic literature review, in archives of
computational methods in engineering. ISSN: 1134-
3060 (Print) 1886-1784 (Online). Retrieved from
https://www.researchgate.net/publication/312147459
[12] S. A. Riaz, S. Naz, and I. Razzak. Multipath deep shallow
convolutional networks for large scale plant species
identification in wild image. In 2020 International Joint
Conference on Neural Networks (IJCNN), pages 1–7.
IEEE, 2020.
[13] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012).
Imagenet classification with deep convolutional neural
networks. Advances in neural information processing
systems, 1097–1105.
[14] Jassmann, T. J., Tashakkori, R., &Parry,R.M.(2015).Leaf
classification utilizing a convolutional neural network.
Southeastcon 2015, IEEE, 1–3.
[15] Wu, Y. H., Shang, L., Huang, Z. K., Wang, G., & Zhang, X. P.
(2016). Convolutional neural network application on
leaf classification. In D. S. Huang, V. Bevilacqua, & P.
Premaratne (Eds.), Intelligent computing theories and
application. ICIC 2016, Lecture Notes in Computer
Science (Vol. 9771, pp. 12– 17). Springer.

Automated Plant Identification with CNN

More Related Content

Similar to Automated Plant Identification with CNN (20)

More from IRJET Journal (20)

Recently uploaded (20)

Automated Plant Identification with CNN