2023 26th International Conference on Computer and Information Technology (ICCIT)

13-15 December, Cox’s Bazar, Bangladesh

ANN and CNN Based Ensemble Learning for

Recognizing Renowned Medicinal Plants
Nahian Sourov, A.S.M Redowan, Faiad Hossain Chowdhury, Md. Nazmus Sakib, and Raiyan Rahman
Department of Computer Science and Engineering
United International University, Dhaka-1212, Bangladesh
Email: [email protected], [email protected],
[email protected], [email protected],
[email protected]

Abstract—The automatic identification and classification of relies on the condition of the image dataset and the feature
medicinal plants with the help of convolutional and artificial extraction and classification methods used. Artificial neural
neural networks is the basis of this work. This paper focuses network classifiers are functional for classifying medicinal
on the ten most renowned leaves of Bangladesh. These are Bo-
hera, Devilbackbone, Haritoki, Lemongrass, Nayontarat, Neem, plants, but their output relies on the condition of the image
Pathorkuchi, Thankuni, Tulsi, and Zenora. The paper emphasizes dataset . [2] Plants contain various bioactive combinations that
how these technologies could help scientific research, advance give them antimicrobial possessions. These compounds could
environmental preservation, and integrate traditional medicinal be used to develop new antibiotics.
products with modern medicine. Since the beginning, medicinal CNNs are a DL(Deep Learning) algorithm mostly used for
plants have been used to heal illnesses because they contain
the active compounds in medicine, which are derived from processing optical image Data. CNNs have been shown to be
various plant sections. People use natural and organic items effective in classifying medicinal plants. The input layer of
most frequently, and curative plants have a large economic a CNN loads the whole plant image, and the output of this
impact. Bangladesh is a region where a wide variety of medicinal layer is then sent to the convolutional layer. The convolutional
plants are found, and it is assumed that roughly 1,500 of the layer extracts features from the image being processed by
estimated 5,000 plant species there contain medicinal properties.
The system uses three pre-trained models (VGG16, ResNet50, and manipulating the result of neurons connected to particular
ResNet152) as base models and adds custom layers to alter them regions in the input layer. The pooling layer lessens the size of
to get the particular job done. The system also uses a majority the feature map by collaborating with a similar feature points
voting mechanism to combine the predictions from the different into one. This helps to address issues such as dimensionality,
models. The system pulls off an imposing accuracy of 99% on a noise, and receptive field. Shape feature extraction is used
dataset of 5000 images of 10 medicinal plant leaves. This work
offers a prospective method for the automated identification and for identifying leaf edges. This necessitates the use of image
classification of medicinal plants, which may have significant thresholds as well as a Canny edge detection. The process
imputations for medical care, scientific study, and environmental of morphological feature extraction extracts raw, acquired
sustainability. features like perimeter, area, and diameter. The Binary Local
Index Terms—Artificial Neural Network, Convolutional Neural Pattern, intended to represent texture, is used to extract texture
Network, Medicinal plants, Classification
features. [3]
In our work, we have used the leaves of 10 widely used
I. INTRODUCTION
medicinal plants in Bangladesh. Each medicinal plant had 500
Medicinal plants have been utilized to treat ailments since photos in the datasets that we used. After examining each
ancient times. These are the raw elements for herbal medicine, image in the data set, we used ANN and CNN to categorize
which blends traditional and science expertise. The active the medical images. We used pre-trained models VGG16,
substances in medicine are wrenched out from different parts ResNet50, and ResNet152 along with ANN and CNN. Then,
of Plants. People are becoming more aware of the importance we combine these algorithms and use a majority vote to get
of health and are increasingly using natural and organic the best outcome. We discover that the maximum average
products. Medicinal plants are used in many processes and accuracy is 99%.
contribute notably to the economy. In Bangladsh, a tropical Section II introduces the current related works. Section III
country, a varid collction of 5,000 plant spcis includs 1,500 confers about the methodology. Section IV explains the inves-
rcognizd for thir mdicinal attributs, playing a vital rol in both tigation results analysis. Lastly, Section V shows a conclusion
traditional and contmporary halthcar approachs. Bangladesh for the study and puts forward future study instructions.
has extensively researched the antibacterial properties of plants
such as neem, turmeric, ginger, and garlic. [1] II. R ELATED W ORKS
The field of automated image processing for identifying Raghukumar et al. [4] used KNN and SVM (support vector
and classifying medicinal plants is a rapidly growing area of machines) to detect Ayurvedic medical plants. They use the
research. The precision of plant identification and classification image of plant leaves to bring out shape, texture, and color

979-8-3503-5901-5/23/$31.00 ©2023 IEEE

features and then classify them. Borkatulla et al. [5] identify III. M ETHODOLOGY
plants used as raw materials for herbal medicine and apply A. ANN and CNN
the ImageNet pre-trained ResNet50, DenseNet201, VGG16,
ANNs are machine-learning models persuaded by human
and InceptionV3 models to the dataset. Abdullahi et al. [6]
brainiac structures. They consist of input, hidden, and output
studied improving identifying medicinal plants over previously
layers, processing data using weighted connections and ac-
applied ANN methods. Deep Learning (CNN) was used in
tivation functions. Deep learning, a subset of ANNs, excels
this research. Kareru et al. [7] FTIR spectroscopy was used to
in tasks like picture recognition, language translation, and
identify saponins in medicinal plants used by herbalists in the
gaming, revolutionizing various sectors. [19] In our paper we
Embu and Mbeere districts of Kenya’s Eastern province. The
applied ANN with VGG16, ResNet50 and ResNet152. CNNs
FTIR technique was validated by phytochemical screening of
are artificial neural networks for pattern detection and image
saponins in the selected medicinal plants. Mallick et al. [8]
processing. They capture spatial hierarchies, recognize edges
presented studies on a straightforward, rapid, accurate, and
and complex patterns, and improve translation in variance.
well-aimed HPTLC technique constructed and verified using
CNNs have transformed computer vision tasks, including
ICH guidelines for simultaneous resolution and quantification
picture classification and facial recognition. [20] In our paper
of alpha-amyrin, beta-sitosterol, lupeol, and n-triacontane from
we use CNN with VGG16,ResNet50 and ResNet152.
dried whole plant powder of L. reticulata and P. lanceolata.
Travadi et al. [9]aimed to develop a new rbcL and ITS B. Experimental Design
metabarcoding for the identification of medicinal plant species, In Fig 1, We collected datasets and used 80% of them
validate the primer’s specificity as well as effectiveness using for training and 20% for testing. Secondly, we use data
mock controls, and determine whether a multibarcoding step preprocessing, then we use three algorithms with ANN and
could be used for herbal product pharmacovigilance. three algorithms of CNN and compare the results. When we
Yusupova et al. [10] It is a field survey about which kind complete the algorithms, we do majority voting, take the high
of medicinal plants grow in Uzbekistan. Chanyal et al. [11] answer for us, and see our machine prediction answer.
provided the productivity and predictability of algorithms
utilizing deep learning(DL) and machine learning(ML) for
classifying medicinal plants using images of their leaves,
particularly using CNN methods. Rai et al. [12] reviewed
the use cases of laser-induced breakdown spectroscopy, also
known as the LIBS method, and how LIBS has been used to
identify glycemic elements of Indian medicinal plants used for
anti-diabetic drugs. Shailendra et al. [13] reviewed a system
for classifying therapeutic plants using a combination of IoT
and machine learning. An application is developed for users
to submit images of plant leaves for classification.
Simirgiotis et al. [14] examined phnolic and flavonoid
content in infusions of four spcis, offring HPLC fingrprints
for fastr, mor dirct analysis of crud drugs, facilitating fficint
comparison and valuation of thir composition. Hong et al.
[15] present study is aimed at investigating the mutagenic Fig. 1: Experimental Design
potentials of Alchornea cordifolia, Cnestis ferruginia, Lon-
chocarpus sericius, Trema orientalis, and Senna alata medicine
plants that are commonly used in Nigeria. Acharjee et al. [16] C. Dataset Description
develop antibiotics from plants that will help to overcome the In Fig 2, we use a medical leaf picture dataset. Our dataset
side effects of conventional antibiotics. The solvent extraction has ten classes: Bohera, Devilbackbone, Haritoki, Lemongrass,
method was used, the conventional cell culture technique, the Nayontarat, Neem, Pathorkuchi, Thankuni, Tulsi, and Zenora.
minimum inhibitory concentration (MIC) assay, and the disc Each class has 500 images; in our experiment, we arranged
diffusion method. Palash et al. [17] investigated supply chain 400 images from each class for training and 100 for testing.
actors’ performance in Bangladesh’s medicinal plant business.
Quantitative and Qualitative methods are studied here. Steele D. Work Flow
et al. [18]aimed to identify compounds from South American Firstly, the images are collected and undergo preprocessing
plants with potential anti-malarial properties by assessing their steps to prepare them for machine learning analysis. Ma-
ability to bind haemin. The study utilized two screening chine learning algorithms are used for image classification,
methods, demonstrated correlations between binding affinity specifically ANN and CNN. Three well-established models,
and antiplasmodial activity, and highlighted the significance namely VGG16, ResNet50, and ResNet152, are utilized for
of isoquinoline compounds in this context. the CNN architecture. A majority voting mechanism is applied
 knowledge embedded in pre-trained models, offering the po-
tential for effective classification with limited training epochs,
making it valuable for various image recognition tasks. Table
II demonstrates image classification tasks using three different
pre-trained deep learning models (VGG16, ResNet50, and
ResNet152). In each case, the pre-trained model is loaded, and
custom layers are added to adapt it to a specific classification
task. These convention layers include concentrated layers with
dropout regularization to intercept overfitting, followed by a
final dense layer with a sigmoid activation function for multi-
class classification. The models are assembled with the Adam
optimizer, sparse categorical cross-entropy loss, and accuracy
as the evaluation metric. They are then trained on image
datasets for 10 epochs with batch sizes 32 and validated for
their performance. This approach harnesses transfer learning
Fig. 2: Dataset of medicinal plants to fine-tune pre-trained models for specific classification tasks
while ensuring robustness through regularization techniques,
potentially yielding effective classification results.
to combine the predictions from these models. The training
process involves using a dataset of 500 images distributed
across ten classes, with a total dataset size of 5000 images.
IV. R ESULT A NALYSIS
1) Image acquisition and preprocessing: The images were
preprocessed in order to remove noise and improve data A confusion matrix is a fundamental tool in statistics and
quality. Resize the images, convert them to grayscale, and machine learning to assess the effectiveness of classification
normalize the pixel values. . models. It gives a graphical representation of how well a model
2) Machine learning algorithm selection: We made use categorizes instances into various classes or groups. Analyz-
of ANN and CNN. An artificial neural network (ANN) is a ing the model’s performance and pinpointing any possible
type of neural network that can be utilized in tasks such as shortcomings is easier when the matrix compares the model’s
regression and classification. CNN is a deep learning algorithm projected labels with the exemplification labels in a tabular
that is specifically designed for image recognition. processing format. [21]
jobs done. A. Individually Result Analysis
3) Model training: This involved feeding the data to the
In Table III, we see that the average accuracy is 0.9946,
algorithms and adjusting the parameters of the algorithms until
and in Table IV, the average accuracy is 0.9950. ANN with
they accurately classified the images.
ResNet152, we see the highest accuracy. The average loss of
4) Model evaluation: We evaluated them to determine their
tableIII is 0.0163, and the average loss of Table IV is 0.0279.
accuracy. This was done using a hold-out data set that was not
So, the ANN works better because the loss of Table III is
used for training. The accuracy of the models was measured
less than tableIV, so the ANN is better. The precision of both
using metrics such as accuracy, precision, and recall.
tables is the same, and in the Recall, the ANN performs better,
5) Majority voting: This involved first getting the predic- and the f1-score Table III performs better.
tions of each model for each image. The predictions are then In figure 3, VGG16 with CNN the Devilbackbone, Hari-
ranked, and the most common prediction is chosen as the final toki, Nayontarat, Thankuni shows 99% accuracy as well as
prediction. Tulsi and Zenora show 98% accuracy and other shows 100%
accuracy.
E. Model Architecture In figure 4, ResNet50 with CNN only Zenora shows 99%
In Table I, the erecting and training of a CNN using accuracy other all detect by 100%.
the VGG16, ResNet50 and ResNet152 architectures as base In figure 5, ResNet152 with CNN Devilbackbone, Zenora
models for image classification, employing transfer learning. shows 98% accuracy, and Nayontarat shows 99% accuracy
In both cases, the pre-trained models are loaded with their other all detect by 100%.
layers frozen to preserve learned features from the ImageNet In figure 6, VGG16 with ANN the Devilbackbone, Nayon-
dataset. Custom CNN layers, including convolutional, pooling, tarat, Neem, Thankuni cannot detect properly other all detect
and dense layers, are added on top, with a final softmax output by 100%.
for multi-class classification. The models are compiled with In figure 7, ResNet50 with ANN only Tulsi cannot detect
the Adam improver and sparse categorical cross-entropy loss, properly accuracy shows 99% other all shows 100% result.
suitable for multi-class tasks, and trained for ten epochs with a In figure 8, ResNet152 with ANN only Tulsi cannot detect
batch size of 32. This transfer learning strategy harnesses the properly accuracy shows 99% other all shows 100% result.
 TABLE I: VGG16,ResNet50 and ResNet152 base models with custom CNN layers on top
VGG16 ResNet50 ResNet152
Layer Output Shape Param# Layer Output Shape Param# Layer Output Shape Param#
InputLayer (None, 224, 224, 3) 0 InputLayer (None, 224, 224, 3) 0 InputLayer (None, 224, 224, 3) 0
vgg16 (None, 7, 7, 512) 14714688 resnet50 (None, 7, 7, 2048) 23587712 resnet152 (None, 7, 7, 2048) 58370944
(Conv2D) (None, 5, 5, 64) 294976 (Conv2D) (None, 5, 5, 64) 1179712 (Conv2D) (None, 5, 5, 64) 1179712
(MaxPooling2D) (None, 2, 2, 64) 0 (MaxPooling2D) (None, 2, 2, 64) 0 (MaxPooling2D) (None, 2, 2, 64) 0
(Conv2D) (None, 2, 2, 128) 73856 (Conv2D) (None, 2, 2, 128) 73856 (Conv2D) (None, 2, 2, 128) 73856
(MaxPooling2D) (None, 1, 1, 128) 0 (MaxPooling2D) (None, 1, 1, 128) 0 (MaxPooling2D) (None, 1, 1, 128) 0
(GAP2D) (None, 128) 0 (GAP2D) (None, 128) 0 (GAP2D) (None, 128) 0
dense-6 (Dense) (None, 256) 33024 dense-6 (Dense) (None, 256) 33024 dense-6 (Dense) (None, 256) 33024
dense-7 (Dense) (None, 128) 32896 dense-7 (Dense) (None, 128) 32896 dense-7 (Dense) (None, 128) 32896
dense-8 (Dense) (None, 10) 1290 dense-8 (Dense) (None, 10) 1290 dense-8 (Dense) (None, 10) 1290
Total params: 15150730 Total params: 24908490 Total params: 58968010
Trainable params: 436042 Trainable params: 1320778 Trainable params: 597066
Non-trainable 14714688 Non-trainable params: 23587712 Non-trainable params: 58370944
params:

TABLE II: Pre-train VGG16,ResNet50 and ResNet152 with custom ANN

VGG16 ResNet50 ResNet152
Layer Output Shape Param# Layer Output Shape Param# Layer Output Shape Param#
vgg16 (Functional) (None, 7, 7, 512) 14714688 resnet50 (Functional) (None, 7, 7, 2048) 23587712 resnet152 (Functional) (None, 7, 7, 2048) 58370944
dense (Dense) (None, 7, 7, 256) 131328 dense 8 (Dense) (None, 7, 7, 256) 524544 dense 12 (Dense) (None, 7, 7, 256) 524544
dropout (Dropout) (None, 7, 7, 256) 0 dropout 6 (Dropout) (None, 7, 7, 256) 0 dropout 9 (Dropout) (None, 7, 7, 256) 0
dense 1 (Dense) (None, 7, 7, 128) 32896 dense 9 (Dense) (None, 7, 7, 128) 32896 dense 13 (Dense) (None, 7, 7, 128) 32896
dropout 1 (Dropout) (None, 7, 7, 128) 0 dropout 7 (Dropout) (None, 7, 7, 128) 0 dropout 10 (Dropout) (None, 7, 7, 128) 0
dense 2 (Dense) (None, 7, 7, 64) 8256 dense 10 (Dense) (None, 7, 7, 64) 8256 dense 14 (Dense) (None, 7, 7, 64) 8256
dropout 2 (Dropout) (None, 7, 7, 64) 0 dropout 8 (Dropout) (None, 7, 7, 64) 0 dropout 11 (Dropout) (None, 7, 7, 64) 0
(GAP2D) (None, 64) 0 flatten 1 (Flatten) (None, 3136) 0 flatten 2 (Flatten) (None, 3136) 0
dense 3 (Dense) (None, 10) 650 dense 11 (Dense) (None, 10) 31370 dense 15 (Dense) (None, 10) 31370
Total params: 14887818 Total params: 24184778 Total params: 58968010
Trainable params: 173130 Trainable params: 597066 Trainable params: 597066
Non-trainable params: 14714688 Non-trainable params: 23587712 Non-trainable params: 58370944

TABLE III: Using ANN

Model Accuracy Loss Precision Recall f1-score
VGG16 0.9890 0.0357 0.99 0.99 0.99
ResNet50 0.9970 0.0058 1.00 1.00 1.00
ResNet152 0.9980 0.0075 1.00 1.00 1.00

TABLE IV: Using CNN

Model Accuracy Loss Precision Recall f1-score
VGG16 0.9910 0.0493 0.99 0.99 0.99
ResNet50 0.9990 0.007 1.00 1.00 1.00
ResNet152 0.9950 0.0276 1.00 0.99 0.99

In figure 9, we see that the pre-trained model ResNet50 with

CNN works better for our datasets.

B. Ensemble Result Analysis

Fig. 3: VGG16 with custom CNN
In Table V, The provided classification detail summarizes
the performance of a machine learning model on a dataset with
ten classes. The model achieved an impressive accuracy of
99%, specifying that it accurately predicted the class labels for
most of the 1000 instances. With minor dissimilarities, it gives
an idea of high precision, recall, and F1-score for most classes,
exhibiting its solid predictive capabilities. Classes Devilback- classes with high confidence. The model’s performance is
bone, Lemongrass, Neem, Pathorkuchi, Thankuni, and Tulsi excellent, as indicated by the high weighted average F1-score
achieved perfect scores across accuracy, recall, and F1-score, of 99, suggesting its effectiveness in multi-class classification
specifying the model’s capability to correctly classify these tasks.
 Fig. 4: Resnet50 with custom CNN Fig. 7: Resnet50 with custom ANN

Fig. 5: Resnet152 with custom CNN

Fig. 8: Resnet152 with custom ANN

Fig. 6: VGG16 with custom ANN Fig. 9: Test accuracy of all model

V. C ONCLUSION AND F UTURE W ORK medicinal plants. We use three pre-trained models (VGG16,
In this research, we provide a unique ANN and CNN-based ResNet50, and ResNet152) as base models and add custom
automated system for the recognition and classification of
 TABLE V: Ensemble Classification Report [12] P. K. Rai, A. K. Srivastava, B. Sharma, P. Dhar, A. K. Mishra,
precision recall f1-score G. Watal et al., “Use of laser-induced breakdown spectroscopy for the
detection of glycemic elements in indian medicinal plants,” Evidence-
Bohera 98 100 99 Based Complementary and Alternative Medicine, vol. 2013, 2013.
Devilbackbone 100 100 100 [13] R. Shailendra, A. Jayapalan, S. Velayutham, A. Baladhandapani, A. Sri-
Haritoki 100 98 99 vastava, S. Kumar Gupta, and M. Kumar, “An iot and machine learning
Lemongrass 100 100 100 based intelligent system for the classification of therapeutic plants,”
Nayontara 97 100 98 Neural Processing Letters, vol. 54, no. 5, pp. 4465–4493, 2022.
Neem 100 100 100 [14] M. J. Simirgiotis, M. Silva, J. Becerra, and G. Schmeda-Hirschmann,
Pathorkuchi 100 100 100 “Direct characterisation of phenolic antioxidants in infusions from four
Thankuni 100 100 100 mapuche medicinal plants by liquid chromatography with diode array
Tulsi 100 100 100 detection (hplc-dad) and electrospray ionisation tandem mass spectrom-
Zenora 100 97 98 etry (hplc-esi–ms),” Food chemistry, vol. 131, no. 1, pp. 318–327, 2012.
[15] C.-E. Hong and S.-Y. Lyu, “Genotoxicity detection of five medicinal
accuracy 99 plants in nigeria,” The Journal of Toxicological Sciences, vol. 36, no. 1,
macro avg 99 99 99 pp. 87–93, 2011.
weighted avg 100 99 99 [16] “In-vitro anti-bacterial activity of medicinal plants against urinary
tract infection (uti) causing bacteria along with their synergistic
effects with commercially available antibiotics,” New Microbes and
New Infections, vol. 51, p. 101076, 2023. [Online]. Available:
layers to adapt them to the specific task and a majority voting https://www.sciencedirect.com/science/article/pii/S2052297522001287
mechanism to combine the predictions from the different [17] “Medicinal plant business in bangladesh: Exploring the performance
of supply chain actors,” Journal of Agriculture and Food
models. The system achieves an impressive accuracy of 99% Research, vol. 6, p. 100230, 2021. [Online]. Available:
on a dataset of 5000 images of 10 medicinal plant leaves. https://www.sciencedirect.com/science/article/pii/S2666154321001320
In the discussion of future work, we improve our dataset, [18] J. Steele, R. Phelps, M. Simmonds, D. Warhurst, and D. Meyer,
“Two novel assays for the detection of haemin-binding properties of
add many classes to collect many types of medical plants, antimalarials evaluated with compounds isolated from medicinal plants,”
collect many rare medical plant pictures using drones, and Journal of Antimicrobial Chemotherapy, vol. 50, no. 1, pp. 25–31, 2002.
use different methods to classify data in many ways. [19] A. K. Jain, J. Mao, and K. M. Mohiuddin, “Artificial neural networks:
A tutorial,” Computer, vol. 29, no. 3, pp. 31–44, 1996.
[20] S. Albawi, T. A. Mohammed, and S. Al-Zawi, “Understanding of a
R EFERENCES convolutional neural network,” in 2017 international conference on
[1] A. Ghani et al., Medicinal plants of Bangladesh: chemical constituents engineering and technology (ICET). Ieee, 2017, pp. 1–6.
and uses. Asiatic society of Bangladesh, 1998. [21] J. T. Townsend, “Theoretical analysis of an alphabetic confusion matrix,”
[2] A. Gopal, S. P. Reddy, and V. Gayatri, “Classification of selected Perception & Psychophysics, vol. 9, pp. 40–50, 1971.
medicinal plants leaf using image processing,” in 2012 International
Conference on Machine Vision and Image Processing (MVIP). IEEE,
2012, pp. 5–8.
[3] T. N. Quoc and V. T. Hoang, “Medicinal plant identification in the wild
by using cnn,” in 2020 International Conference on Information and
Communication Technology Convergence (ICTC). IEEE, 2020, pp. 25–
29.
[4] A. M. Raghukumar and G. Narayanan, “Comparison of machine learning
algorithms for detection of medicinal plants,” in 2020 Fourth Interna-
tional Conference on Computing Methodologies and Communication
(ICCMC). IEEE, 2020, pp. 56–60.
[5] B. Borkatulla, J. Ferdous, A. H. Uddin, and P. Mahmud, “Bangladeshi
medicinal plant dataset,” Data in Brief, vol. 48, p. 109211, 2023.
[6] J. Abdollahi, “Identification of medicinal plants in ardabil using deep
learning : Identification of medicinal plants using deep learning,” in
2022 27th International Computer Conference, Computer Society of Iran
(CSICC), 2022, pp. 1–6.
[7] P. Kareru, J. Keriko, A. Gachanja, and G. Kenji, “Direct detection of
triterpenoid saponins in medicinal plants,” African Journal of Tradi-
tional, Complementary and Alternative Medicines, vol. 5, no. 1, pp.
56–60, 2008.
[8] S. S. Mallick, V. V. Dighe et al., “Detection and estimation of alpha-
amyrin, beta-sitosterol, lupeol, and n-triacontane in two medicinal plants
by high performance thin layer chromatography,” Adv. Chem, vol. 2014,
no. 7, 2014.
[9] T. Travadi, A. P. Shah, R. Pandit, S. Sharma, C. Joshi, and M. Joshi,
“A combined approach of dna metabarcoding collectively enhances
the detection efficiency of medicinal plants in single and polyherbal
formulations,” Frontiers in Plant Science, vol. 14, p. 1169984, 2023.
[10] Z. Yusupova, S. F. Baratjon ogli, and M. Z. Abduqunduzovna, “Medic-
inal plants growing in our republic medicinal properties,” Periodica
Journal of Modern Philosophy, Social Sciences and Humanities, vol. 15,
pp. 5–7, 2023.
[11] H. Chanyal, R. K. Yadav, and D. K. J. Saini, “Classification of
medicinal plants leaves using deep learning technique: A review,”
International Journal of Intelligent Systems and Applications in
Engineering, vol. 10, no. 4, p. 78–87, Dec. 2022. [Online]. Available:
https://ijisae.org/index.php/IJISAE/article/view/2199