Fine-tuning Deep Learning Models to Identify Pepper Leaf Diseases

Abstract

Deep Learning (DL) has become very popular, especially thanks to the towering achievements in classification, segmentation, and identification in many image types by the use of transfer learning and detailed feature learning stages. Although DL has been criticized by some researchers due to downsampling procedure, it has created a new enlightenment to image processing approaches with its capabilities in generating feature activation maps. In this study, the pre-trained CNN architectures were evaluated for bell pepper leaf identification. The VGG19 architecture with the best identification performance was fine-tuned by iterating various dropout factorization rates and fully-connected layers on different feature activation maps. The fine-tuned VGG-19 CNN architecture separated the healthy and diseased pepper leaves with an accuracy of 97.84%.

Full text

ISBN: 978-605-73802-0-3

ICAII4.0 - International Conference on Artificial Intelligence towards Industry 4.0
11-12 November 2021, İskenderun / Hatay / TURKEY
2
Conference Proceedings Book
Editors :
Assoc. Prof. Dr. Yakup Kutlu
Assoc. Prof. Dr. Sertan Alkan
Assist. Prof. Dr. İpek Abasıkeleş Turgut
Assist. Prof. Dr. Ahmet Gökçen
Assist. Prof. Dr. Gökhan Altan
Assist. Prof. Dr. Mehmet Sarıgül
Assist. Prof. Dr. Levent Karacan
Technical Editors:
Merve Nilay Aydın
Handan Gürsoy Demir
Kadir Tohma
Halil Okur
ISBN: 978-605-73802-0-3 Publication No:1

ICAII4.0 - International Conference on Artificial Intelligence towards Industry 4.0
11-12 November 2021, İskenderun / Hatay / TURKEY
3
International Conference on Artificial Intelligence towards Industry 4.0 (ICAII4.0) has been
organized in Iskenderun, Hatay/Turkey on 11-12 November 2021.
The main objective of ICAII4.0 is to present the latest research and results of scientists related
to Artificial Intelligence, Industry 4.0 and all sub-disciplines of Computer Engineering. This conference
provides opportunities for the different areas delegates to exchange new ideas and application
experiences face to face, to establish business or research relations and to find global partners for future
collaboration.
We hope that the conference results provide significant contribution to the knowledge in this
scientific field. The organizing committee of conference is pleased to invite prospective authors to
submit their original manuscripts to ICAII4.0.
All paper submissions will be double-blind and peer-reviewed and evaluated based on
originality, technical and/or research content/depth, correctness, relevance to conference, contributions,
and readability. Selected papers presented in the conference that match with the topics of the journals
will be published in the following journals:
 Natural and Engineering Sciences
 Journal of Intelligent Systems with Applications
In particular we would like to thank Prof. Dr. Tolga Depci, Rector of Iskenderun Technical
University; Natural and Engineering Sciences, Academic Publisher; Journal of Intelligent Systems with
Applications. They have made a crucial contribution towards the success of this conference. Our thanks
also go to the colleagues in our conference office.
Looking forward to see you in next Conference.
Conference Organizing Committee

ICAII4.0 - International Conference on Artificial Intelligence towards Industry 4.0
11-12 November 2021, İskenderun / Hatay / TURKEY
5
COMMITTEES
HONOUR COMMITTEE
Prof. Dr. Tolga DEPCİ
Rector
Iskenderun Technical University
ORGANISATION COMMITTEE
Assoc. Prof. Dr. Yakup KUTLU
Iskenderun Technical University
Assoc. Prof. Dr. Sertan ALKAN
Iskenderun Technical University
Dr. Ahmet GÖKÇEN
Iskenderun Technical University
Dr. İpek ABASIKELEŞ TURGUT
Iskenderun Technical University
Dr. Gökhan ALTAN
Iskenderun Technical University
Dr. Mehmet SARIGÜL
Iskenderun Technical University
Dr. Levent KARACAN
Iskenderun Technical University
SCIENTIFIC COMMITTEE
Dr. Ahmet GÖKÇEN
Iskenderun Technical University
Dr. Gökhan ALTAN
Iskenderun Technical University
Dr. İpek ABASIKELEŞ TURGUT
Iskenderun Technical University
Dr. Levent KARACAN
Iskenderun Technical University
Dr. Mehmet SARIGÜL
Iskenderun Technical University
Dr. Mustafa YENİAD
Ankara Yildirim Beyazit University
Dr. Sertan ALKAN
Iskenderun Technical University
Dr. Yakup KUTLU
Iskenderun Technical University
Dr. Yalçın İŞLER
Izmir Katip Celebi University

ICAII4.0 - International Conference on Artificial Intelligence towards Industry 4.0
11-12 November 2021, İskenderun / Hatay / TURKEY
7
CONTENT
Detectıng Tagged People in Camera Images ............................................................................. 8
Diagnosis of Ovarian Cancer Using Conventional Machine Learning Methods ...................... 9
EEG Based Spatial Attention Shifts Detection Using Time-Frequency Features on Empirical
Wavelet Transform .......................................................................................................... 10
High Performance Detection and Locking Application for Combat Unmanned Aerial Vehicle
(UAV) .............................................................................................................................. 11
Investigation of Classifier Performances with Respect to The Difference of Flickering
Frequencies of User Commands in Brain-Computer Interfaces .................................... 12
Modification of Posterior Probability Variable with Frequency Factor According to Bayes
Theorem .......................................................................................................................... 13
Prediction of Alzheimer's Disease from MRI Images Using Deep Learning Techniques ....... 14
Thermally Induced Vıbratıon Suppressıon in a Thermoelastıc Beam Structure ...................... 15
Water Quality Measurement System in Iskenderun Technical University Pond ..................... 16
A Review of TDMA Based MAC Protocols with Disjoint Time Slot ........................................ 17
Building a Hybrid Recommendation System for E-Commerce ................................................ 22
Cyberlog: A Generalized IOT Platform Training to Protect Cyber Defence Agents and
Algorithms ....................................................................................................................... 29
Extraction of Lead Oxide from Car Batteries .......................................................................... 34
Fine-Tuning Deep Learning Models to Identify Pepper Leaf Diseases ................................... 41
Interpolation-Based Smart Video Stabilization ........................................................................ 49
Learnıng-Based Image Renderıng............................................................................................ 55
Low Cost Multifunctional Security System That Does Not Require Hardware Design ........... 59
Mobile EEG Measurement System Desıgn ............................................................................... 69
Prediction Diabetes Using Machine Learning Techniques .................................................... 77
Regıonal Sıgnal Recognıtıon of Body Sounds .......................................................................... 82
Scene Construction from Depth Map Using Image-To-Image Translation Model .................. 87
Keywords .................................................................................................................................. 93

ICAII4.0 - International Conference on Artificial Intelligence towards Industry 4.0
11-12 November 2021, İskenderun / Hatay / TURKEY
41
Fine-tuning Deep Learning Models to Identify Pepper Leaf Diseases
Gokhan Altan1
1 Department of Computer Engineering, Iskenderun Technical University, Turkey
Email: gokhan.altan@iste.edu.tr, ORCID: 0000-0001-7883-3131
Abstract
Deep Learning (DL) has become very popular, especially thanks to the towering achievements
in classification, segmentation and identification in many image types by the use of transfer
learning and detailed feature learning stages. Although DL has been criticized by some
researchers due to down sampling procedure, it has created a new enlightenment to image
processing approaches with its capabilities in generating feature activation maps. In this study,
the pre-trained CNN architectures were evaluated for bell pepper leaf identification. The VGG-
19 architecture with the best identification performance was fine-tuned by iterating various
dropout factorization rates and fully-connected layers on different feature activation maps. The
fine-tuned VGG-19 CNN architecture separated the healthy and diseased pepper leaves with an
accuracy of 97.84%.
Keyword(s): Bell Pepper, Deep Learningdepht, VGG-19, Plant Leaf Diseases, PlantVillage
1. Introduction
Food production is a sector that is important in many areas, directly related to social life
and economy. Planting, seed dressing, and harvesting are the main inception processes in the
food industry. Especially in recent years, due to plant diseases and pandemics, the disruptions
in production and the inability to make sufficient supply both impressed the national economy
and prevented sufficient natural products from reaching the consumers. In this case, the
necessity of preventing plant diseases and correct management of the necessary production
policies has come to the forefront for the agricultural sector all over the world. Although it is
difficult to predict some of these situations, plant diseases can contribute to the harvesting
bodyborne leaf diseases, stem diseases, fruit diseases, and root diseases (Singh & Misra, 2017).
Harvest loss caused by late detection of plant diseases and overdue in taking precautionary
measures related to this loss is an application with conventional methods even in developed
countries. Moreover, spraying for identified plant disease is applied to the entire harvest area.
This causes unnecessary expenses for agricultural producers and even spoils the genetics of the
plant due to chemical spraying. In this respect, locally and plant-based diagnosis of plant
diseases is of great necessity to improve the earning in productivity. Therefore, computerized
techniques for early detection of plant diseases and applying reliable preventives are the main
focus of agriculture technologies.
The advancements on computer vision and the increasing computational capacities with
the advantages of GPU have highlighted the artificial intelligence based applications.
Moreover, common use of camera technologies even for mobile devices has contributed these
standpoints on novel machine learning techniques. Use of computer vision in agriculture
increases day by day with the prevalence of drone technologies. It enables the opportunity to
configure plant disease identification models using real-time leaf, bacterial, fruit diseases using

ICAII4.0 - International Conference on Artificial Intelligence towards Industry 4.0
11-12 November 2021, İskenderun / Hatay / TURKEY
42
hybrid techniques and robotic. in the light of these developments, computer-aided techniques
with automated and hand-crafting have frequently been implemented to identify different plant
leaf diseases. Jagan and Mohan handled the paddy spot disease using SIFT technique and
conventional machine learning algorithms, including k-nearest neighbor and support vector
machine (SVM) (Jagan et al., 2016). Phadikar focused on the rice leaf diseases feeding
morphologic and histogram features into SVM with non-linear kernels (Phadikar, 2012). Islam
et al extracted RGB-based features on spot areas to identify the rice leaf disease using Naive
Bayes (Islam et al., 2018). Kumari et al. utilized contrast and statistical features of spots for
tomato and cotton leaf diseases for unfolding to an artificial neural network (ANN) (Kumari et
al., 2019). Chouhan et al. classified the similarity based correlation features for various plant
species using ANN (Chouhan et al., 2018).
Deep Learning (DL) algorithms are popular hybrid techniques with feature learning and
classification stages. Moreover, it is an easy-adaptable technique for many image types with
high classification and segmentation performances on large-scale datasets using transfer
learning on DL with pre-trained architectures. Hence, convolutional neural networks (CNN)
algorithm is the first choice for many researchers without pre-processing and advanced image
processing knowledge. Sladojevic et al. utilized CNN for various plant leaf diseases on their
own large-scale database (Sladojevic et al., 2016). The studies applied transfer learning on pre-
trained ImageNet weights using AlexNet (Amara et al., 2017; Brahimi et al., 2017; Chouhan et
al., 2018; Kumar et al., 2020; Lee et al., 2017; Liu et al., 2018; Sladojevic et al., 2016), LeNet
(Kumar et al., 2020), VGGNet (Amara et al., 2017; Lee et al., 2017), GoogleNet (Amara et al.,
2017; Lee et al., 2017; Sladojevic et al., 2016), ResNet (Lee et al., 2017), and more. Lee et al.
utilized feature learning on AlexNet using fine-tuning and presented high identification
performances of 40 leaves (Lee et al., 2017). Amara et al. applied CNN with LeNet architecture
to classify banana leaf disease on PlantVillage dataset with high accuracy rates (Amara et al.,
2017). The researchers evaluated the performance of various pre-trained architectures on 9
tomato leaf diseases from PlantVillage dataset and apple leaf diseases on their own dataset (Lee
et al., 2017). Ferentinos et al. utilized many pre-trained CNN architectures for separation of
many leaf diseases on PlantVillage database (Ferentinos, 2018). Geetharamani and Pandian
fine-tuned a novel CNN architecture with variations of classification parameters including,
dropout rate, learning rate, and dense layers (Geetharamani & J., 2019). Kurup et al. fed the
Capsule Networks and CNN architectures with plant leaves from PlantVillage to compare the
identification performances (Kurup et al., 2020). Altan also analyzed the efficiency of Capsule
Networks over plant leaf diseases (Altan, 2020). This paper aims at classifying bell pepper leaf
diseases using pre-trained CNN architectures (VGG-16, VGG-19, and MobileNet), comparing
the performance of fine-tuning on the best architecture for transferring dominant pixel
information among convolutional blocks.
The remaining of the paper is organized as a general description of focused plant species
in the PlantVillage database and architectural specifications of experimented CNN models in
materials and methods. The experimental setup and performance evaluation metrics for the
experimented architectures in experimental results. A complete comparison with state-of-the-
art in terms of plant leaf disease identification performances and superiorities of the CNN
models in discussion.

ICAII4.0 - International Conference on Artificial Intelligence towards Industry 4.0
11-12 November 2021, İskenderun / Hatay / TURKEY
43
2. Materials and Methods
2.1. PlantVillage Database
PlantVillage Database is a large-scale plant leaf disease dataset. It enables enhancing the
harvesting techniques with computer-assisted techniques to provide the productivity in
harvesting as early diagnosis. Whereas the food industry is a main problem to supply the
nutrition necessity for people all over the world, it is possible to qualify profitability in time
dressing and watering. Therefore, it was presented as a challenge dataset by Land Grant
University, USA (Hughes & Salathe, 2015). PlantVillage consists of a total number of 54K leaf
images from thirteen species. Most of the plant species have disease spots for various leaf
diseases and healthy leaves to develop diagnosis models for both binary and multi-classes.
Figure 1. Sample bell pepper leaves for healthy (left) and bacteria spot disease (right)
In this study, the bell pepper leaf bacteria disease is focused on with a binary
classification. The PlantVillage database has a total number of 2.4K bell pepper leaf images
including, 1K bacteria diseased and 1.4K healthy leaves. Fig. 1 indicates the randomly selected
pepper leaf images for healthy and diseased. Each leaf image has a standardized square
dimension with a standard background, but uncertain rotations with different angles. Hence, no
cropping and background removal were applied to the inputs before feeding the CNN
architectures.
Although several images have a shadow effect; none of the bell pepper images was
excluded in the analysis. We augmented each leaf image with horizontal-, vertical-, and both-
flipping before splitting training and testing folds.
2.2. Convolutional Neural Networks (CNN)
Deep Learning is an advancing hybrid technique with outstanding classification
performances on various types of inputs. The main philosophy of Deep Learning is combining
unsupervised algorithms in the feature learning and detailed supervised algorithms with novel
factorization and regularization techniques. On the other hand, it has advantages of transferring
the knowledge between related problems for reducing the training time. Transfer learning
enables optimizing pre-trained architectures with a detailed convolutional integration on the
adjacent convolution blocks. Moreover, transfer learning with popular pre-trained CNN

ICAII4.0 - International Conference on Artificial Intelligence towards Industry 4.0
11-12 November 2021, İskenderun / Hatay / TURKEY
44
architectures provides a feature map extraction on a detailed convolutional layer. Hence, pre-
trained CNNs enable reaching better globalization using regularization and factorization on
large-scale datasets.
Figure 2. VGG-19 architecture and fine-tuning process in dense layers
The main benefits of CNN are feature learning, transfer learning, fine-tuning capability
on pre-trained architectures, mobile architectures, and optimizing deep neural networks with
high generalizations. The feature activation maps as a result of many convolution blocks with
a pooling layer are fed into the fully-connected layers (FC) just as a multilayer perceptron.
Whereas the primary convolution layers define low-level features, the subsequent layers
generate high-level features by transferring the dominant pooling results (Altan, 2019). I
evaluated the performance of transfer learning using popular pre-trained architectures,
including VGG-16, VGG-19 (Simonyan & Zisserman, 2014), and MobileNet (Howard et al.,
2017) on the identification of plant leaf diseases.
3. Experimental Results
The CNN has the ability to compose various combinations during the modeling process
due to the sequence of convolutional layers, max pooling, and fully-connected layers. It is a
challenging procedure for proposing novel CNN architectures. However, the popular pre-
trained architectures have proven their usability and efficacy on different problems. Hence,
transfer learning was chosen for the analysis of plant leaf diseases. VGG-16 and VGG-19
architectures were named according to their layer depth by Visual Geometry Group. Although
VGG architectures have a small number of feature learning layers, it is not a mobile network
due to the depth of the fully-connected layers. MobileNet is one of the most lightweight CNN
architectures with the capabilities of adaptability to many data types in artificial intelligence.
Even though very deep neural networks with complicated feature learning stages have many
advantages on extracting detailed features, it is more important to reach approximate
classification performances using simplistic architectures with fine-tuning on CNN
architectures. That is why this study focuses on fine-tuning and pruning on simple CNN
architectures.
Many plant leaf images have uncertain rotations, shadow views and sun shining.
Nonetheless no image enhancing techniques such as background removal and histogram
equalization were performed before CNN architectures.
The diseased leaf images have commonly small bacteria spots. Therefore, it is possible to
diverging the loss of significance for disease separation on complex CNN architectures. It is an
important point using shallow feature learning stages with low dimension of feature activations
maps. Moreover, this point enables evaluating the success of CNN architectures for use as an

ICAII4.0 - International Conference on Artificial Intelligence towards Industry 4.0
11-12 November 2021, İskenderun / Hatay / TURKEY
45
early diagnostic. Therefore, the best CNN architecture (among VGG-16, VGG-19, and
MobileNet) was fine-tuned using various depths of fully-connected layers and neuron ranges in
supervised learning stages. Thus it is aimed to reach the optimum model with highest
generalization performance on leaf disease classification. Hereby, the best achievements were
presented for each CNN architecture and compared with the state-of-art.
Each plant leaf image was converted to gray-scale image to obtain a quick analysis on
one channel instead of three channels (Red, Green, and Blue). Whereas it is a standardized input
size for VGG-16 and VGG-19, MobileNet supports all input sizes bigger than 32x32. The leaf
images were resized to 224×224 for both CNN architectures. Each leaf image was augmented
to enhance the dataset for supporting the pre-trained model with different angles and to avoid
overfitting. Each image was increased by 4 times with vertical-, horizontal, and both-flipping.
Hence, the training of the pre-trained CNN architectures for identification of plant leaf diseases
were performed on a total number of 9900 plant leaf images.
The proposals were trained using 80% and 20% folds for train and test, respectively. The
separate folds of the plant leaf images were utilized for training and testing the pre-trained CNN
architectures. The performance of the proposals was evaluated using statistical test
characteristics, including overall accuracy, sensitivity, specificity, and precision (Altan, 2019).
In the fully-connected layers of the CNN architectures were iterated for various dropout
factorization rates (0.25, 0.5, 0.75) in the fine-tuning, a range of fully-connected layers (2 and
3), and neuron depth (256, 512, 1024, 2048, and 4096 neurons). The output functions of the last
fully-connected layer is softmax.
Table 1. The classification performances (%) for the pre-trained CNN architectures
CNN architectures
Accuracy
Sensitivity
Specificity
VGG-16
82.25
77.17
85.72
VGG-19
91.12
90.35
91.64
MobileNet
90.27
89.51
90.80
The highest classification performances in terms of accuracy, sensitivity, and specificity
are presented in Table 1. The experimental results indicate that the VGG-19 has the highest
identification capability on the plant leaf diseases among the analyzed pre-trained architectures.
Whereas MobileNet reached well enough classification performances, VGG-16 performed a
miscarriage generalization on the re-training of the architectures on ImageNet weights.
Therefore, VGG-19 architecture was fine-tuned using various fully-connected layer
specifications. Table 2 presents the best five classification performances on fine-tuning of
VGG-19 architecture.

ICAII4.0 - International Conference on Artificial Intelligence towards Industry 4.0
11-12 November 2021, İskenderun / Hatay / TURKEY
46
Table 2. The best five classification performances (%) for fine-tuning of VGG-19 architecture
with various fully-connected layers
Fully-connected layers
Dropout
Accuracy
Sensitivity
Specificity
Precision
FC1(512)
FC2(512)
0.50
94.93
94.04
95.53
93.45
FC1(512)
FC2(1024)
FC3(2048)
0.25
94.95
94.11
95.52
93.44
FC1(1024)
FC2(1024)
0.50
95.19
93.41
96.40
94.62
FC1(512)
FC2(512)
FC3(2048)
0.75
96.05
94.89
96.84
95.31
FC1(1024)
FC2(512)
FC3(2048)
0.50
97.84
97.48
98.09
97.19
The enhanced neuron architects in each fully-connected layer provide high identification
accuracy using deeper models. It is experimented that using a big number of neurons for both
fully-connected layers (4096 neurons) performed inefficacious generalization under the average
classification performances for iterated range of dropout and training parameters. Using 1024
neurons at 1st dense layer, 512 neurons at 2nd dense layer, and 2048 neurons at 3rd dense layer
with the proposed fine-tuning approach on VGG-19 architecture reached the best generalization
using a dropout factorization rate of 0.5. The highest bell pepper leaf disease diagnosis
performance was achieved with the rates of 97.84%, 97.48%, 97.49%, and 97.19% for
accuracy, sensitivity, specificity, and precision, respectively.
4. Discussion
The pre-trained architectures are commonly utilized to analyze leaf, fruit, and root
diseases for various plants. The easy-adaptability and denomination of CNN architectures are
supported by transfer learning and pre-trained weights. Moreover, it enables fast convergence
and robust generalizations for many related problems. Although novel architectures using
capsule networks reported the advantages of using spatial information on direct images by
excluding pooling layers, CNN architectures are still in the first position for consecutive
convolutions to extract significant feature maps.
To the best of my knowledge, whereas a limited number of researchers focused directly
on bell pepper leaf disease identification, most of them analyzed it among various plant species.
However, these papers reported an overall identification accuracy rather than plant-based
classification performances. The achievements in the related works on the PlantVillage
database are presented in Table 3.
The achievements are obtained using the contingency table and number of the plant
species for pepper leaves. Geetharamani and Arun Pandian highlighted a light-weight CNN
model. It has nine layers including fully-connected layers and convolutional layers to identify
the plant leaf diseases. They iterated with various batch size, epoch, and dropout factorization

ICAII4.0 - International Conference on Artificial Intelligence towards Industry 4.0
11-12 November 2021, İskenderun / Hatay / TURKEY
47
index in the training of CNN. They reported the identification capacity of their proposal over
AlexNet, VGGNet, Inception-v3 and ResNet. Their proposal achieved the highest pepper leaf
disease classification rates of 93.00%, 92.00%, and 93.00% for accuracy, sensitivity, and
specificity, respectively (Geetharamani & J., 2019). Kurup et al. proposed a capsule network
on the PlantVillage dataset. Their proposal achieved pepper leaf classification accuracy rates
among 91% and 96% (Kurup et al., 2020). Altan also applied a capsule network architecture on
the augmented pepper leaf images. He reached classification performance rates of 95.76%,
96.37%, and 97.49% for accuracy, sensitivity, and specificity, respectively (Altan, 2020).
Although the literature reported well-enough classification performances with popular
CNN architectures and novel capsule network architectures, nevertheless it is possible to reach
better generalization capabilities by fine-tuning the architectures according to the problems.
This study explores the effect of fine-tuning on well-known CNN architectures on pepper leaf
disease identification using various fully-connected layer sequences and dropout factorization
rates.
Table 3. The state-of-art for pepper leaf disease identification on PlantVillage
Related works
Classification
Accuracy
Sensitivity
Specificity
Geetharamani et al.
(2019)
Own CNN architecture
93.00
92.00
93.00
Kurup et al.(2020)
Capsule Network
-
89.17
-
Altan (2020)
Capsule Network
95.76
96.37
97.49
This study
VGG-19 + Fine-tuning
97.84
97.48
98.09
The fine-tuned VGG-19 architecture overcomes the state-of-the-art using three fully-
connected layers and 0.5 of dropout factorization. VGG-19 is the most responsible CNN
architecture for the experimented ranges of classification parameters under the architecture-
based evaluation metrics. Using a small number of neurons at the first fully-connected layer
and detailing the previous fully-connected layers with a big number of neurons provided a better
optimization in both training and testing of the proposal. The fine-tuned VGG-19 architecture
achieved identification rates of 97.84%, 97.48%, and 98.09% for accuracy, sensitivity, and
specificity, respectively.
References
1. Altan, G. (2019). DeepGraphNet: Grafiklerin Sınıflandırılmasında Derin Öğrenme Modelleri. European
Journal of Science and Technology, Special Issue, 319–327. https://doi.org/10.31590/ejosat.638256
2. Altan, G. (2020). Performance Evaluation of Capsule Networks for Classification of Plant Leaf Diseases.
International Journal of Applied Mathematics Electronics and Computers.
https://doi.org/10.18100/ijamec.797392
3. Amara, J., Bouaziz, B., & Algergawy, A. (2017). A deep learning-based approach for banana leaf diseases
classification. Lecture Notes in Informatics (LNI), Proceedings - Series of the Gesellschaft Fur Informatik (GI).
4. Brahimi, M., Boukhalfa, K., & Moussaoui, A. (2017). Deep Learning for Tomato Diseases: Classification and
Symptoms Visualization. Applied Artificial Intelligence. https://doi.org/10.1080/08839514.2017.1315516
5. Chouhan, S. S., Kaul, A., Singh, U. P., & Jain, S. (2018). Bacterial foraging optimization based radial basis
function neural network (BRBFNN) for identification and classification of plant leaf diseases: An automatic
approach towards plant pathology. IEEE Access. https://doi.org/10.1109/ACCESS.2018.2800685
6. Ferentinos, K. P. (2018). Deep learning models for plant disease detection and diagnosis. Computers and
Electronics in Agriculture. https://doi.org/10.1016/j.compag.2018.01.009
7. Geetharamani, G., & J., A. P. (2019). Identification of plant leaf diseases using a nine-layer deep convolutional

ICAII4.0 - International Conference on Artificial Intelligence towards Industry 4.0
11-12 November 2021, İskenderun / Hatay / TURKEY
48
neural network. Computers and Electrical Engineering. https://doi.org/10.1016/j.compeleceng.2019.04.011
8. Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., & Adam, H.
(2017). MobileNets: Efficient convolutional neural networks for mobile vision applications. In arXiv.
9. Hughes, D. P., & Salathe, M. (2015). An open access repository of images on plant health to enable the
development of mobile disease diagnostics.
10. Islam, T., Sah, M., Baral, S., & Roychoudhury, R. (2018). A Faster Technique on Rice Disease Detectionusing
Image Processing of Affected Area in Agro-Field. Proceedings of the International Conference on Inventive
Communication and Computational Technologies, ICICCT 2018.
https://doi.org/10.1109/ICICCT.2018.8473322
11. Jagan, K., Balasubramanian, M., & Palanivel, S. (2016). Detection and Recognition of Diseases from Paddy
Plant Leaf Images. International Journal of Computer Applications. https://doi.org/10.5120/ijca2016910505
12. Kumar, S., Sharma, B., Sharma, V. K., Sharma, H., & Bansal, J. C. (2020). Plant leaf disease identification
using exponential spider monkey optimization. Sustainable Computing: Informatics and Systems.
https://doi.org/10.1016/j.suscom.2018.10.004
13. Kumari, C. U., Jeevan Prasad, S., & Mounika, G. (2019). Leaf Disease Detection: Feature Extraction with K-
means clustering and Classification with ANN. 2019 3rd International Conference on Computing
Methodologies and Communication (ICCMC), 1095–1098. https://doi.org/10.1109/ICCMC.2019.8819750
14. Kurup, R. V., Anupama, M. A., Vinayakumar, R., Sowmya, V., & Soman, K. P. (2020). Capsule network for
plant disease and plant species classification. Advances in Intelligent Systems and Computing.
https://doi.org/10.1007/978-3-030-37218-7_47
15. Lee, S. H., Chan, C. S., Mayo, S. J., & Remagnino, P. (2017). How deep learning extracts and learns leaf
features for plant classification. Pattern Recognition. https://doi.org/10.1016/j.patcog.2017.05.015
16. Liu, B., Zhang, Y., He, D. J., & Li, Y. (2018). Identification of apple leaf diseases based on deep convolutional
neural networks. Symmetry. https://doi.org/10.3390/sym10010011
17. Phadikar, S. (2012). Classification of Rice Leaf Diseases Based on Morphological Changes. International
Journal of Information and Electronics Engineering. https://doi.org/10.7763/ijiee.2012.v2.137
18. Simonyan, K., & Zisserman, A. (2014). Very Deep Convolutional Networks for Large-Scale Image
Recognition. 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track
Proceedings. http://arxiv.org/abs/1409.1556
19. Singh, V., & Misra, A. K. (2017). Detection of plant leaf diseases using image segmentation and soft computing
techniques. Information Processing in Agriculture. https://doi.org/10.1016/j.inpa.2016.10.005
20. Sladojevic, S., Arsenovic, M., Anderla, A., Culibrk, D., & Stefanovic, D. (2016). Deep Neural Networks Based
Recognition of Plant Diseases by Leaf Image Classification. Computational Intelligence and Neuroscience.
https://doi.org/10.1155/2016/3289801