Canadian Medicinal Plant Classification using

Convolutional Neural Network and Transfer

Learning
Shaon Bhatta Shuvo Akram Vasighizaker Akshay Shah Maryam MahdavyRad
School of Computer Science School of Computer Science School of Computer Science School of Computer Science
University of Windsor University of Windsor University of Windsor University of Windsor
Windsor, Canada Windsor, Canada Windsor, Canada Windsor, Canada
[email protected] [email protected] [email protected] [email protected]

Malita Dodti
School of Computer Science
University of Windsor
Windsor, Canada
[email protected]

Abstract—Nowadays, computerized plant species classification most commonly used Neutral Networks for image processing
systems are used to help the people in the detection of the various is Convolutional Neural Network (CNN).
species. However, the automated analysis of plant species is The rest of the paper are distributed as follows: In section
challenging as compared to human interpretation. This research
as been provided in this field for the better classification of plant 2, we gave a brief description of the motivation of our
species. Even now, these methodologies lack an exact classification work. In section 3, we discussed some of the remarkable
of the plant species. The challenge is due to the inappropriate researches related to our work. Data collection procedure
classification algorithm. In Particular, when we consider the has been discussed in section 4. In section 5, we described
medicinal plant species recognition, the accuracy will be the the background study including the CNN, transfer learning,
main criteria. In this research, the suggested system implements
the deep learning technique to obtain high accuracy in the and detailed implementation of our model including the data
classification process using computer prediction methods.The prepossessing. Results are discussed discussed in sections 6.
Convolutional Neural Network (CNN) is employed beside transfer Finally, conclusion and future works are discussed in section
learning for deep learning of medicinal plant images. This 7.
research work has been carried out on the flower images dataset
of four Canadian medical plants; namely, Clubmoss, Dandelion, II. M OTIVATION
Lobelia, and Bloodroot, which is fed as the training dataset for
the CNN and machine learning-based proposed system. Finally, Medicinal plants have been utilized in curing human dis-
an accuracy of 96% has been achieved in classification of the eases for thousands of centuries and are the source of a
medicinal plant species. significant percentage of medicines. Canadian medicinal plants
Index Terms—Image Classification, Medicinal Plant classifica- have a prolonged history of consumption with hundreds of
tion, Convolutional Neural Network, Transfer Learning
classes being employed by First Nations Canadians in their
old-style medicine. Medicinal plant species classification is
I. I NTRODUCTION critical for medicine production and conservation. Local peo-
ples are not enough knowledgeable of their urban medicinal
A medicinal plant is delineated as a plant that is collected plants and their usages. Therefore, classifying the Canadian
from the wild or planted plant for its medicinal value. Plants medicinal plant image using Convolutional Neural Network
have been utilized in curing human diseases for thousands by high accuracy image classification model could be useful
of centuries and are the source of a significant percentage of to identify different types of species.
medicines. Canadian medicinal plants have a prolonged history
of consumption with hundreds of classes being employed III. L ITERATURE R EVIEW
by First Nations Canadians in their old-style medicine [1]. There is no particular research has been done to detect
In recent years, some computational approaches have been medicinal plants of Canada using image classification tech-
introduced, particularly in image processing domain, for plant nique with the best of our knowledge. However, there are
classification. In this regard, Neural Networks represent novel different proposed techniques to classify objects or flowers,
techniques for image processing, with large potentials. The and some of them employed deep learning approaches. In most
researches, classification challenges are deal with using CNN- and 50 layers are considered. Amongst the four sets of layers
based algorithms. The Convolutional Neural Networks (CNN) considered, ResNet26 out-performed the other three models.
shows considerable accomplishment in different research and For experimental training, the learning rate is set to 0.001.
real-world projects. One of the earliest applications of CNN Flavia dataset accuracy (99.65%) result is compared with other
was handwritten recognition [2]. By developing of CNN approaches like Radial Basis Probabilistic Neural Network
technique, innovative models, which include new layers, are (RBPNN), Deep Belief Network with dropout (DBN), Support
presented. In ImageNet challenges, CNN has been utilized Vector Machine (SVM), andResNet26. ResNet26 architecture
mostly with different datasets mixtures. Some researchers produced an accuracy of 91.78% recognition rate for the
compared human detection abilities and a trained network on BJFU100 dataset. Barre et al. [11] developed a LeafNet, a
image recognition. The output of these comparisons depicts CNN-based plant identification system.The leaflet consisted
that humans can detect an image with accuracy about 73.1% of five sets of 2 convolutional layers and one max-pooling
while the trained network shows up 64% accuracy. Also, after layer followed by one convolution, one max-pooling layer,
employing Convolutional Neural Networks to a similar dataset, and three fully connected layers. The leaflet is tested over
it revealed the 74.9% accuracy, so it depicts better results than Leafsnap, Foliage, and Flavia datasets.
human recognition [3]. In recent flower classification researches, various neural
Then, in a study, a deep CNN-based framework of a network classification models are compared; however, their
hierarchical structure was presented that applies a transfer primary focus is on LeNet and AlexNet. Their best result
learning method to modify a deep CNN model. His con- is shown on the AlexNet model, which is implemented with
clusions indicate that this approach can efficiently improve Sigmoid for assigning initial weights [12]. Furthermore, on
classification accuracy [4]. On the other research, a CNN another very recent research for flower classification, a hybrid
method for the flower classification problem was presented. method is utilized together with Convolutional Neural Net-
Their conclusions demonstrate the accuracy of 84.02% [5]. work models and feature selection methods. In this suggested
Additionally, a CNN framework is presented for plant classifi- model Convolutional Neural Network is employed for feature
cation. This framework was presented for the classification of extraction. And then, for selecting between achieved features,
several varieties of plants from the image database collected feature selection methods are utilized. Their classification
from intelligent agriculture stations. In this research, a CNN- model achievement completed by the Support Vector Machine
based architecture is employed for feature extraction of various (SVM) technique was 98.91% [13]. Likewise, Saini and
plant images, which is utilized based on the TARBIL database Khamparia present a plant leaf classification using a deep Con-
and achieved 97.47% accuracy on 16 various plant types. volutional Neural Network method based on a five thousand
This research shows that the CNN-based classification result leaf images of two plant. Their result shows 99.96% accuracy
in more accuracy compares to the SVM-based classification on the training dataset and 99.90% on the test dataset [14].
[6]. In another study, a leaf classification framework has In other recent research, they utilized VGG19, three layers
been proposed by applying the dual-path deep CNN. This CNN and five layers CNN network for classification species
method contains two main functions, firstly the shape and of succulent plant. This method reaches a high accuracy of
texture attributes are analyzed; secondly, the found attributes 99.77%. Their dataset includes 3632 images, which are ten
are optimized for the classification. This method shows high species of succulent plants and non-succulent plants [15].
accuracy in classification of about 99.28% on the Flavia Some researchers also used hybrid models to detect local
dataset [7]. foods [16] and birds [17]. Similarly, in another research,
Ghazi et al. [4] applied transfer learning over the Life a hybrid method is proposed for plant classification. Their
CLEF plant dataset with the help of pre-trained models like recommended method consists of two parts; applying CNN
AlexNet, GoogleNet, and VGGNet. For all these deep convo- for feature extraction and then on the second phase train
lutional neural networks, fine-tuning is performed, and various the random forest model. In this work, PlantCLEF 2019
parameters are analyzed after data augmentation. Parameters dataset was used for the experimental part. Their tested model
like batch size and the number of iterations are analyzed. produces generally higher accuracy than prior strategies [18].
And on a recent research four variety of transfer learning Moreover, in other similar research, a regional convolution
models is compared n four datasets. And the output shows neural network (RCNN) utilized for the detection of plants.
the effectiveness role of transfer learning on increasing the They employed a fast RCNN model, which consists of a
performance of prior plant classification models [8].Also, Convolutional Neural Network for extracting features and
in recent research for classifying natural images by applying support vector machine (SVM) for classification. The plants
the transfer learning approach, they achieved 99.7% overall studied in this research are the medicinal plants that can be
accuracy [9]. Then Sun et al. [10] proposed a 26-layer displayed in various locations like the Himalayas or can be
ResNet (Residual Network) model for plant identification. produced in the local garden [19]. One of another recent
BJFU100 dataset is used, and it consists of 10000 images of main application of transfer learning in classification plants is
100 ornamental plant species found in Beijing Forestry Univer- in smarts farms [20] [21] [22] and plant disease recognition
sity campus. For experimental analysis, BJFU100 and Flavia systems [20] [23] [24].
datasets are utilized. In deep residual networks, 18, 26, 34, Based on the significant number of the literature survey, it is
clear that the reported work on plant classification over Cana- compared to a typical NN is that it automatically detects the
dian plant species is sparse. Also, numerous research works significant features without any supervision. A CNN is consists
are carried out using features such as shape, texture, color, of various layers that transform an input into the output. The
morphological, or physiological features. Reported works on complexity of the learned features increases in every hidden
plant species classification using deep learning architecture are layer. For example, detection simple features are learned in
limited. Hence in this project, an investigation is performed the first hidden layer, like edges, and the detection of more
using Convolutional Neural Network in order to achieve a complex shapes in the last one. A CNN model is composed
higher plant classification rate. of two main components: the feature extraction part and the
classification part (Fig 1).
IV. DATA
Numerous medicinal species of woodlands as slow-growing
perennials are found near Canada’s Waterfall area. In this
study, we utilize dataset contains 1805 images, includes
approximately 400 images for each species. Four group
of species are selected in this study namely Clubmoss,
Dandelion, Lobelia, and Bloodroot. We use texture, shape,
color,physiological or morphological as the features set of
the data. The most important advantageous of clubmoss are
its usage to treat kidney and urinary disorders. Moreover,
Dandelion is a native plant of Canada and commonly used
as a weed, growing at one of the many Canadian camping
grounds. Health values dandelion provides is to treat joint
Fig. 1. The Overview of a CNN and its main components [26]
complaints, liver disorders, skin conditions, and anemia. Also,
Lobelia is local to Prince Edward Island, Nova Scotia, New In order to understand a CNN architecture, we introduce
Brunswick, and parts of Ontario and Quebec. It also widely several concepts. Compared to a typical neural network in
known as Indian tobacco. It is beneficial mostly for relieving which each input layer’s neuron is linked to the hidden layer’s
respiratory ailments such as bronchitis and asthma. smoking neurons. In a CNN we have Local Receptive Fields, a small
the leaves or brewing them into a tea are two main usage of it. number of input layer’s neurons which are connected to the
Furthermore, great Lakes, as well as Nova Scotia and parts of hidden layer’s neurons. The local receptive field use convolu-
New Brunswick are residence of Bloodroot. It alleviate skin tion to translate an image into a feature map. Convolution can
issues and respiratory ailments. Moreover, Bloodroot is toxic perform the mathematical convolution operation by moving
and one should use it carefully under the prescription of a a filter across the image. At every region, an element-wise
qualified physician. matrix multiplication and summation of the result are done.
In this study we follow some steps for data collection. i) This sum create the feature map, the yellow area in Fig 2.
Articulate the problem; Knowing what one wants to predict Non-linearity features make powerful any type of neural
helps in deciding the data valuable to collect. Data Exploration network. A neural network can achieve this using an activation
in the categories of Classification, Clustering, Regression, and function by passing the weighted sum of its inputs to the next
Ranking helps with the decision. ii)Establish Data Collection layer. CNN use the same function and applies the transforma-
Mechanism; Process of collecting the Data which can be tion to the output of each neuron by passing the result of the
Automated or Manual based on the requirement. iii)Format convolution operation through an activation function. ReLU
Data; File format of the images stored need to be same for or Rectified Linear Unit is a popular activation function that
maintaining the consistency. iv) Reduce Size; Data need to be maps the output of a neuron to the highest value in the next
collected based on the target needs to be achieved which is layer (Fig 3).
critical for our Dataset. v) Complete Data Cleansing; Data with A pooling step can use for the dimensionality reduction of
missing, erroneous or fewer representative values is removed the features map by compressing the output of small number
to make prediction more accurate. of neurons into a one output (Fig 4). We call this a kind of
down-sampling the feature map which keeps the important
V. BACKGROUND S TUDY AND M ETHODOLOGY
features on CNN automatically. Meanwhile, this can leads to
A. Background Study reduction of the number of parameters to learn the model.
1) Convolutional Neural Network (CNN): A Convolutional A fully connected layer can make it flat the output of the
Neural Network, CNN, is a deep learning architecture [25]. last pooling layer to a 1 dimension vector of values. As it
Image classification is one of the problems that a CNN could can be seen in Fig 5, between 4 nodes and 5 nodes a fully
do and is a trained network that can classify images into connected layer is just a dot product of the 1x4 input vector,
one of a thousand pre-determined categories. One can employ yellow nodes, with the 4x5 weight matrix W1. The result of
a CNN to do image processing including image detection, this matrix multiplication is a 1x5 vector, shown as the red
segmentation, and classification. The main advantage of CNN nodes. We then multiply this 1x5 vector with a 5x5 matrix
 Fig. 5. A fully connected layer is used to flatten the output of the last pooling
layer

Fig. 6. Dropout regularization technique

is also a regularization technique that is advantageous in some

applications in which achieving new training data is difficult.
Therefore, we can use data augmentation as a way to enriches
Fig. 2. Creation of feature map by local receptive field translation across an or augments by creating new training examples via a random
image transformation of the current dataset. Common transformations
are shifting, rotation, exposure, resizing, contrast change,
adjustment, and etc. (Fig 7). It is worth mentioning that, data
augmentation should only apply to the training data, but not
the validation or test set.

Fig. 3. ReLU activation function maps the output of a neuron to the highest
value

Fig. 7. Data augmentation

2) Transfer Learning: There are some challenges before

building the model; small dataset, low-resolution images, lots
of noise, and the variations inside the same class images. Deep
Fig. 4. Pooling layer reduces the dimensionality of the features map CNN with transfer learning is the solution to overcome these
challenges [27]. The idea is to use the information learned
from one specific problem to apply to a comparable one.
W2, representing in a 1x5 vector, the green nodes. Finally For example, a CNN model trained to differentiate between
using 5x1 matrix W3 we get the output. animals could be used to train a new model that identify trucks
Dropout is the most popular regularization technique, a and cars (Fig 8).
technique which is used in order to prevent overfitting for deep The process that takes a pre-trained network, re-training it
neural networks. The idea is that during training time, a neuron on a new data, and modifying it is called Transfer Learning.
and all the inputs and outputs to this neuron is temporarily Based on the definition: Model A is successfully trained to
dropped or disabled with the dropout-rate p (Fig 6). solve source task A using a large dataset A. However, the
In the case of having a few samples to train, overfitting is dataset B for a target task B is too small, preventing Model
unavoidable. But fortunately, data augmentation is a solution B from training efficiently. Thus, we use part of model A to
to enhance generalization performance on small datasets. This predict results for task B (Fig 9).
 B. Data Preprocessing
1) Data Distribution: In this section, we explain data dis-
tribution, data augmentation, and data rescaling and resizing.
First of all we divide dataset into three subsets including train,
test, and validation set. The initial dataset was included 1805
images. We separate 80% of it for training, 10% for test, and
10% for validation set. The size of each classes is shown in
Fig. 8. The overview of Transfer Learning [26]
Table I. Also, the distribution of each subset is displayed in
Fig 10. As we can see, the proportion of each class is equal
in the three subset.

TABLE I
T HE SIZE OF SAMPLES IN TRAIN , TEST, AND VALIDATION SET

The number of samples

Bloodroot Clubmoss Dandelion Lobelia
Training set 360 524 250 371
Test set 36 52 25 37
Validation set 36 52 25 37

Fig. 9. The process of a pre-trained network in Transfer Learning [28]

There are several different strategies for Transfer learning.

The simplest strategy is to solve a on a different Target task
but in the same Domain by directly using a full or part of an
already pre-trained model from a source task. Also, feature
extractor is another usage of a pre-trained model instead of
using the whole model. This approach allows us to directly
apply new dataset to solve an entirely different problem.
Since the new task employ the pre-trained complex model as
features, we are allowed to train a simpler and faster linear
model to modify the output based on the new dataset. This
strategy is best when the target task dataset is very small.
Finally, we can go one step further, by not only training the
output classifier, but also fine-tune weights in some layers of
the pre-trained model. This strategy is best when the target
task dataset is significantly big.
3) Inception V3: When we have a relatively small dataset,
a super-effective technique is to use Transfer Learning where Fig. 10. The proportion of samples in train, test, and validation set
we use a pre-trained model. This model has been trained on
an extremely huge dataset, and we would be able to transfer 2) Data Augmentation: Since we have a small number
weights which were learned through hundreds of hours of of data, we use data augmentation technique to increase
training on multiple high powered GPUs. Big corporations the size of data by generating different shapes of images
tend to release such models to the public, aiming to enhance from the given data. We use ImageDataGenerator class form
the development of this field. Some pre-trained models used keras.preporssing library for training dataset to obtain aug-
directly include Inception V3, BERT as well as YOLO, GloVe, mentation by creating train data object. Rotation, shifting,
UnsupervisedMT and etc. Many such models are open-sourced shearing, zooming, and flipping are different data augmen-
such as VGG-19 and Inception-v3. They were trained on tation we apply to our dataset. Rotation do randomly rotate
millions of images with extremely high computing power pictures in range 0 to 180. Shifting do randomly translate
which can be very expensive to achieve from scratch. images horizontally or vertically in a range of a fraction of
InceptionV3 is a model that is made up of results of several total height or width. Flipping do at a random manner flip
research [29]. It includes convolutions, pooling layer, dropout half of the images horizontally. In summary, augmentation
techniques, and fully connected layers. A dataset of 1,000 transforms images randomly into different shapes, so that
categories from the ImageNet [12] dataset is employed to train there is impossible to have an image duplicat. This helps the
this model. The ImageNet consists of ten million URLs of model generalize better and eventually reduces the chance of
labeled images. overfitting.
 3) Data Rescaling and Resizing: All the images are consist
of RGB values with pixel range from 0 to 255. However,
we need to rescale these values between 0 and 1, because
values in range 0 to 255 would be too high for our model
to process. Therefore, we use the same ImageDataGenerator
class to update train data and create two new object validation
data and test data from validation and test set respectively.
Also, since we collect data from different sources, images
have different height and width. Therefore, we need to resize
images to have the same size in terms of width and height in
training, validation, and testing set. In this case we keep height Fig. 12. Reduce of the spatial dimensions of a three dimensional tensors
and width pixel size 150 to generate the final training set. using Global Average Pooling layer (GAP) [28]
Moreover, we apply the same approach to generate validation
and test set using validation data and test data object respec- TABLE II
tively to resize each images into 150X150 pixels. A random T HE SUMMARY OF THE FINAL CONSTRUCTED MODEL
batch of training set of size 25 is shown in Fig 11. The total number Non-trainable Trainable Layers
of layers Layers
314 249 65

In order to compile the method, we use RMSprop with

learning rate 0.0001 as the optimizer and categorical cross-
entropy as loss function. After rescaling the images, and using
Image Augmentation, we flow them in batches of 15 using
Early Stopping regularization function on the train and test
data. Validation data do not augmented, just test and train set.
In order to obtain the best result, we also investigate other
optimizers, such as RMSProp, Adam, and SGD and compare
the results. We calculate precision, recall, and accuracy as the
performance evaluation metrics.

VI. R ESULTS AND D ISCUSSION

Fig. 11. A random batch of training set
The result is one of the most important parts of any project.
In this project, we achieved 95.3% test accuracy on the Incept
C. Model Implementation V3 model. It requires 4 epochs to achieve this accuracy.
Table III displays the detailed result of all epochs. The first
In this section, we explain model construction, compilation, 4 column includes training loss, training precision, training
and validation. To construct the model, we use InceptionV3 recall, and training accuracy while the rest four-column
model as a pre-trained model/base model. The model is includes validation loss, validation precision, validation recall,
available in keras.application library. The model construction and validation accuracy. Finally, we achieved 95.33% test
includes the following steps: i) First, we initialize InceptionV3 accuracy with 0.91 precision and 0.85 recall.
model as our base model. The parameters trained on ImageNet
dataset. ii) Then, we add a Global Average Pooling (GAP)
layer to the model to reduce the spatial dimensions of a TABLE III
three dimensional tensors, which eventually leads to minimize T RAINING AND VALIDATION R ESULT
overfitting. It reduces h * w * d dimensions to 1 *1 * d
Epoch Train Train Train Train Val Val Val Val
dimensions by taking the average of h and w values (Fig 12). Loss Pre- Re- Ac- Loss Pre- Re- Ac-
iii) Then, we add a fully connected layer, FC, of 1024 ci- call cu- ci- call cu-
sion racy sion racy
hidden nodes. iv) Finally, the FC layer is connected to our
1 0.3098 0.8848 0.7687 0.8983 0.2555 0.8931 0.7961 0.9200
output layer to predict output. The FC layer mainly takes the 2 0.3321 0.8952 0.8044 0.8937 0.1318 0.8970 0.8157 0.9067
output from GAP as input and provides output to our final 3 0.2737 0.9000 0.8253 0.9116 0.7863 0.9025 0.8342 0.9133
output layer to predict the best label for each image. We use 4 0.2282 0.9056 0.8418 0.9336 0.2996 0.9090 0.8496 0.9200
‘Softmax’ as the activation function, because the number of
output classes are more than two. We use 65 layers in the Moreover, we achieved 95.33% accuracy with RMSProp. In
model and make 249 layers non-trainable. The summary of the case of SGD and Adam, it requires 6 epochs and accuracy
the constructed model is provided in II. is also quite good but lower than the RMSProp. Table IV
shows the detailed result from different optimizers to compare TABLE VI
them. T HE RESULTS FOR TESTING THE MODEL IN TERMS OF ACCURACY, LOSS ,
PRECISION , AND RECALL

TABLE IV Test Loss 0.02

T EST R ESULT W ITH D IFFERENT O PTIMIZER Test Precision 0.91
Test Recall 0.85
Optimizer Total Test Test Test Test Ac- Test Accuracy 95.33%
Epoch Loss Preci- Recall curacy
sion
RMSProp 4 0.02 0.91 0.85 95.33
SGD 6 0.07 0.95 0.94 92.67 VII. C ONCLUSION
Adam 6 0.04 0.95 0.94 93.33
To conclude, we aim to create an application that can
In order to compare the results, we also implement our classify Canadian medicinal plants. In this study, we use Incep-
algorithm on the dataset without applying image augmentation. tionV3 module with transfer learning and image augmentation.
In this case, we saw a sharp reduction in terms of accuracy In a dense layer, we use ReLU as an activation function
with RMSProp and SGD. Without Image augmentation, we because it is one of the most popular activation function. We
achieved only 75.33% test accuracy on RMSProp. But, in the also used early stopping regularization with patience parameter
case of Adam, we found 90.00% test accuracy which is nearly 3. By conducting our experiment with different optimizers, we
the same as the image augmentation case. Table V displays find that our application RMSProp provides better accuracy.
the detailed result from different optimizers without image At the end, we achieve 95.33% test accuracy using RMSProp.
augmentation. In the future, we plan to increase the accuracy by expanding
the size of dataset. Currently, our application can only classify
TABLE V the four plants. In the Future, we would like that it can classify
T EST R ESULT W ITHOUT I MAGE AUGMENTATION more plants. We also will test our application on the other deep
models like Inception V4 and Inception ResNet.
Optimizer Total Test Test Test Test Ac-
Epoch Loss Preci- Recall curacy
sion R EFERENCES
RMSProp 8 4.05 0.91 0.73 75.33
SGD 5 0.15 0.87 0.82 78.67 [1] Y. Uprety, H. Asselin, A. Dhakal, and N. Julien, “Traditional use of
Adam 5 0.28 0.94 0.88 90.00 medicinal plants in the boreal forest of canada: review and perspectives,”
Journal of ethnobiology and ethnomedicine, vol. 8, no. 1, p. 7, 2012.
[2] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet:
As a result, after performing an operation on different test A large-scale hierarchical image database,” in 2009 IEEE conference on
cases, we found that for the small dataset, image augmentation computer vision and pattern recognition, pp. 248–255, Ieee, 2009.
[3] Y. Yang and T. M. Hospedales, “Deep neural networks for sketch
can increase the accuracy. In our study, the performance of recognition,” arXiv preprint arXiv:1501.07873, vol. 1, no. 2, p. 3, 2015.
RMSprop optimizer is higher than SGD and Adam. With the [4] G.-h. Song, X.-g. Jin, G.-l. Chen, and Y. Nie, “Two-level hierarchical
help of default Inception V3, we achieve 95.33% accuracy. Fig feature learning for image classification,” Frontiers of Information
13 shows the accuracy, loss, precision, and recall for training Technology & Electronic Engineering, vol. 17, no. 9, pp. 897–906, 2016.
[5] Y. Liu, F. Tang, D. Zhou, Y. Meng, and W. Dong, “Flower classification
and validation of the model. The results for testing the model via convolutional neural network,” in 2016 IEEE International Con-
are provided in Table VI. ference on Functional-Structural Plant Growth Modeling, Simulation,
Visualization and Applications (FSPMA), pp. 110–116, IEEE, 2016.
[6] H. Yalcin and S. Razavi, “Plant classification using convolutional
neural networks,” in 2016 Fifth International Conference on Agro-
Geoinformatics (Agro-Geoinformatics), pp. 1–5, IEEE, 2016.
[7] M. P. Shah, S. Singha, and S. P. Awate, “Leaf classification using
marginalized shape context and shape+ texture dual-path deep convolu-
tional neural network,” in 2017 IEEE International Conference on Image
Processing (ICIP), pp. 860–864, IEEE, 2017.
[8] A. Kaya, A. S. Keceli, C. Catal, H. Y. Yalic, H. Temucin, and B. Tekin-
erdogan, “Analysis of transfer learning for deep neural network based
plant classification models,” Computers and electronics in agriculture,
vol. 158, pp. 20–29, 2019.
[9] D. Ezzat, A. E. Hassanien, M. H. N. Taha, S. Bhattacharyya, and
S. Vaclav, “Transfer learning with a fine-tuned cnn model for classifying
augmented natural images,” in International Conference on Innovative
Computing and Communications, pp. 843–856, Springer, 2020.
[10] Y. Sun, Y. Liu, G. Wang, and H. Zhang, “Deep learning for plant
identification in natural environment,” Computational intelligence and
neuroscience, vol. 2017, 2017.
[11] P. Barré, B. C. Stöver, K. F. Müller, and V. Steinhage, “Leafnet:
A computer vision system for automatic plant species identification,”
Ecological Informatics, vol. 40, pp. 50–56, 2017.
Fig. 13. The comparison between train and validation in terms of accuracy, [12] K. Mitrović and D. Milošević, “Flower classification with convolutional
loss, precision, and recall neural networks,” in 2019 23rd International Conference on System
Theory, Control and Computing (ICSTCC), pp. 845–850, IEEE, 2019.
[13] M. TOĞAÇAR, B. ERGEN, and Z. CÖMERT, “Classification of flower
species by using features extracted from the intersection of feature se-
lection methods in convolutional neural network models,” Measurement,
p. 107703, 2020.
[14] G. Saini, A. Khamparia, and A. K. Luhach, “Classification of plants
using convolutional neural network,” in First International Conference
on Sustainable Technologies for Computational Intelligence, pp. 551–
561, Springer, 2020.
[15] A. K. Das, M. A. Iqbal, B. Paul, A. Rakshit, and M. Zahid, “Classifi-
cation of succulent plant using convolutional neural network,”
[16] T. Nishat, I. Md Romyull, and S. Shaon Bhatta, “A convolution neural
network based classification approach for recognizing traditional foods
of bangladesh from food images,” in Proceedings of the Interna-
tional Conference on Intelligent Systems Design and Application(ISDA),
pp. 844–852, Springer, 2018.
[17] I. Md Romyull, T. Nishat, and S. Shaon Bhatta, “Mobilenet model
for classifying local birds of bangladesh from image content using
convolutional neural network,” in Proceedings of the 10th International
Conference on Computing, Communication and Networking Technolo-
gies (ICCCNT), pp. 1–4, IEEE, 2019.
[18] S. Banerjee and R. Pamula, “Random forest boosted cnn: An empirical
technique for plant classification,” in Proceedings of the Global AI
Congress 2019, pp. 251–261, Springer, 2020.
[19] R. N. Lochan, A. S. Tomar, and R. Srinivasan, “Plant detection and
classification using fast region-based convolution neural networks,” in
Artificial Intelligence and Evolutionary Computations in Engineering
Systems, pp. 623–634, Springer, 2020.
[20] B. Espejo-Garcia, N. Mylonas, L. Athanasakos, S. Fountas, and
I. Vasilakoglou, “Towards weeds identification assistance through trans-
fer learning,” Computers and Electronics in Agriculture, vol. 171,
p. 105306, 2020.
[21] P. Bosilj, E. Aptoula, T. Duckett, and G. Cielniak, “Transfer learning
between crop types for semantic segmentation of crops versus weeds in
precision agriculture,” Journal of Field Robotics, vol. 37, no. 1, pp. 7–19,
2020.
[22] R. G. de Luna, E. P. Dadios, A. A. Bandala, and R. R. P. Vicerr, “Tomato
growth stage monitoring for smart farm using deep transfer learning
with machine learning-based maturity grading,” AGRIVITA, Journal of
Agricultural Science, vol. 42, no. 1, 2020.
[23] J. Arunnehru, B. Vidhyasagar, and H. A. Basha, “Plant leaf diseases
recognition using convolutional neural network and transfer learning,”
in International Conference on Communication, Computing and Elec-
tronics Systems, pp. 221–229, Springer, 2020.
[24] K. Rangarajan Aravind and P. Raja, “Automated disease classification
in (selected) agricultural crops using transfer learning,” Automatika,
vol. 61, no. 2, pp. 260–272, 2020.
[25] Y. Li, Z. Hao, and H. Lei, “Survey of convolutional neural network,”
Journal of Computer Applications, vol. 36, no. 9, pp. 2508–2515, 2016.
[26] P. udikandula, “Deep view on transfer learning with iamge classification
pytorch available at: https://medium.com/@purnasaigudikandula/deep-
view-on-transfer-learning-with-iamge-classification-pytorch-
5cf963939575. [accessed: 21- mar- 2019].”
[27] C. Tan, F. Sun, T. Kong, W. Zhang, C. Yang, and C. Liu, “A survey on
deep transfer learning,” in International conference on artificial neural
networks, pp. 270–279, Springer, 2018.
[28] A. Cook, “Global average pooling layers for object localization,
available at: https://alexisbcook.github.io/2017/global-average-pooling-
layers-for-object-localization [accessed: 21- mar- 2019].”
[29] X. Xia, C. Xu, and B. Nan, “Inception-v3 for flower classification,” in
2017 2nd International Conference on Image, Vision and Computing
(ICIVC), pp. 783–787, IEEE, 2017.