International Journal of Applied Mathematics, Electronics and Computers 8(3): 057-063, 2020

INTERNATIONAL JOURNAL OF APPLIED MATHEMATICS International

Open Access
ELECTRONICS AND COMPUTERS
Volume 08
Issue 03
www.dergipark.org.tr/ijamec
e-ISSN: 2147-8228 September, 2020

Research Article

Performance Evaluation of Capsule Networks for Classification of Plant Leaf

Diseases
Gökhan ALTAN a
a
Computer Eng., Iskenderun Technical University, 31200, Hatay, Turkey

ARTICLE INFO ABSTRACT

Article history: Deep Learning (DL) is a high capable machine learning algorithm which composed the advanced
Received 18 September 2020 image processing as feature learning and supervised learning with detailed models with many
Accepted 28 September 2020 hidden layers and neurons. DL demonstrated its efficiency and robustness in many big data
Keywords: problems, computer vision, and more. Whereas it has an increasing popularity day by day, it has
Bell Pepper still some deficiencies to construe the relationship between learned feature maps and spatial
CapsNET
Capsule Network
information. Capsule network (CapsNET) is proposed to overcome the shortcoming by excluding
Deep Learning the pooling layer from the architecture and transferring spatial information between layers by
Plant Leaf Diseases capsule. In this paper, CapsNET architecture was proposed to evaluate the performance of the
Plantvillage model on classification of plant leaf diseases using simple reduced capsules on leaf images. Plant
leaf diseases are common and prevalent diseases that disrupt harvesting and yielding for
agriculture. CapsNET has capability of detailed analysis for even small stains that may lead seed
dressing time and duration. The proposed CapsNET model aimed at assessing the applicability of
various feature learning models and enhancing the learning capacity of the DL models for bell
pepper plants. The healthy and diseased leaf images were fed into the CapsNET. The proposed
CapsNET model reached high classification performance rates of 95.76%, 96.37%, and 97.49%
for accuracy, sensitivity, and specificity, respectively.

This is an open access article under the CC BY-SA 4.0 license.

(https://creativecommons.org/licenses/by-sa/4.0/)

1. Introduction traditional methods even in developed countries.

Moreover, the control of harvest areas is done zone-by-
Food security is at the forefront of many issues covered
zone. According to the locally diagnosed plant diseases,
in healthy life. Food security is a field of study that has a
spraying applied to the entire cultivation area decreases the
wide scope such as planting and growing of seed,
yield and creates additional burden for the producer.
supporting it with the right seed dressing, seed dressing,
Therefore, correct detection of plant diseases is of great
and methods in its harvest [1]. Especially in recent years,
importance and the realization of this structure with
many diseases that cause pathology in normal tissues of
automatic systems will speed up the processes to stop the
cancer type also come to the fore as nutrition with
progression of the disease and improve the quality of the
genetically modified foods and readymade foods. In
harvesting. Therefore, the quest for accelerator, reliable
addition to its impact on human health, proper nutrition of
methods are of great practical importance.
food is also an important issue. Bacterial, insect-borne
In recent years, the widespread use of cameras
leaf, stem, fruit, root, and flower leaf plant diseases are the
especially on mobile devices and the active use of
most influential factors on the harvesting process,
computer vision techniques have enabled the development
harvesting quality, safety and efficiency of food
of image processing and machine learning approaches. It
production [2]. Identification of plant diseases and
provides the opportunity to recommend models supporting
determination of their species is still carried out by
food production and modeling disease identification

* Corresponding author. E-mail address: gokhan.altan@iste.edu.tr

DOI: 10.18100/ijamec.797392
 Gökhan Altan., International Journal of Applied Mathematics Electronics and Computers 08(03): 057-063, 2020

systems in agriculture, which is based on image processing dataset using fine-tuning on AlexNet architecture and
after the use of hybrid analysis methods. Among these reached well enough identification rates for identification
models, traditional machine learning approaches and of forty leaves [11]. Amara et al. used feature learning on
hand-crafted feature extraction commonly performed for LeNet architecture on CNN to identify banana leaf
classification of various plant leaf diseases using images diseases. They experimented on PlantVillage dataset and
of diseased spots on plants. Jagan and Mohan focused on reported accurate achievements [12]. Brahimi et al.
paddy diseases using Scale Invariant Feature Transform compared the efficiency of pre-trained CNN architectures,
(SIFT) on leaf images. They experimented on k-Nearest including GoogleNet and AlexNet, for identification of
Neighbor (k-NN) and Support Vector Machine (SVM) on nine types of tomatoes leaf diseases. They fine-tuned the
SIFT features [3]. Phadikar analyzed rice leaf diseases architectures on PlantVillage dataset [13]. Liu et al.
using morphological features, radial distribution of compared the performance of popular CNN architectures
pathology on leaves and histogram equalization features. including AlexNet, GoogleNet, VggNet, and ResNet for
They used Bayes classifier and SVM for classification and identification apple leaf diseases. They fine-tuned the
optimized classification parameters for non-linear SVM AlexNet architecture using own database [14]. Ferentinos
kernels [4]. Islam et al used percentage of RGB value for et al. experimented on various CNN architectures
spot areas of disease form the leaf images. They classified including AlexNet, VggNet, Overfeat, and GoogleNet for
rice diseases using Naive Bayes classifier using a fast identification of fifty seven leaf classes including diseased
method [5]. Usha Kumari et al. evaluated the efficiency of and healthy in PlantVillage. They iterated on variations of
contrast, energy, homogeneity, and statistical features classification parameters on CNN and reported the
from segmented spots from the leaves for identification of VggNet as the best architecture for leaf classification [15].
tomato and cotton leaf diseases. They fed the hand-crafted Mohanty et al. utilized GoogleNet and AlexNet on
features of spot areas on leaves into artificial neural identification of forty plant leaf diseases on PlantVillage
network (ANN) and reported high classification dataset. They applied low-level image processing
performances [6]. Arivazhagan et al. applied color co- techniques and segmentation for leaves before fine-tuning
occurrence method to extract shape, color and texture the architectures [16]. Zhang et al. analyzed peach leaf
statistical features of spot areas to identify beans, lemon, images to identify the diseased plants using AlexNet
guava, potato, and tomato leaf diseases on a limited variety architecture and compared its efficiency with k-NN, SVM,
of images. They used SVM classifier with non-linear and ANN. They reported the superiority of CNN
kernel to classify the texture features [7]. Chouhan et al. architecture against conventional machine learning
used Region Growing Algorithm, which is based on algorithms [17]. Geetharamani and Pandian proposed their
extracting similarity based correlation of intensity level, own CNN architecture to identify the leaf diseases for
color, or scalar features, to identify various plant leaf thirteen plants in PlantVillage dataset. They iterated their
disease. They classified the diseases using ANN with model using different dropout factorization, learning rate,
radial basis function kernel and reached high plan leaf and fully-connected layers. Their most light-weight
disease identification performances [8]. Kumar et al. architecture for plant leaf disease classification was
proposed exponential spider monkey optimization to established a CNN with nine layers [18].
extract significant features and experimented with SVM, Capsule Network (CapsNet) is a specified Deep
k-NN, and ANN classifiers. They reported the proposal as Learning model proposed by Hinton and his colleagues to
a successful feature extraction [9]. overcome the deficiencies of CNN [19]. Whereas the CNN
Especially, popular Deep Learning (DL) algorithms can has high analysis capability on images with proven
achieve high generalization capacity performances by achievements; pooling layer, which is a down-sampling
detailed analysis with many layers, feature learning stages, approach, causes data losses which may lead the training
and excluding feature extraction and image processing process prone to low generalization performance.
from the classification progresses. DL algorithms were Moreover, CNN cannot transfer spatial information and
commonly fed with raw images without pre-processing. instantiation parameters such as pose of low-level features
DL has own feature extraction stages using convolution to each other, texture and deformation information. These
progress or autoencoder models. Therefore, implementing cases give rise to error rates for classification of whole
DL by transfer learning and modeling novel architectures parts together on an object. The dynamic routing approach
is the most common technique for classification of images. between capsules, which represents likelihood and spatial
Sladojevic et al. used convolutional neural networks information between low-level features, provide
(CNN) for identification of thirteen different types of plant transferring pose parameters and part-whole hierarchy
leaf diseases including Pear, cherry, apple, grapevine, and [19].
peach on a large private database [10]. Lee et al. utilized CapsNet are commonly used for the researching areas
feature learning using the pre-trained AlexNet weights on that CNN achieved well enough classification and
CNN architecture. They analyzed Malayakew plant leaf segmentation performances in a few years. Verma et al.

- 58 -
 Gökhan Altan., International Journal of Applied Mathematics Electronics and Computers 08(03): 057-063, 2020

Figure 1. Sample bell pepper leaf images with bacteria spot (left) and healthy (right)

Figure 2. The CapsNet architecture for generating spatial capsules

utilized CapsNet algorithm to identify potato leaf diseases performance and ability transferring spatial information
in PlantVillage dataset. They also experimented on various between capsules for diseased spots with many CapsNet
pre-trained CNN architectures including ResNet, VggNet, models, and comparing the classification performances
and GoogLeNet to compare the performance of CapsNet with state-of-art.
and highlighted the superiority of CapsNet over CNN The remaining of the paper is organized to detail
architectures in accuracy [20]. Dong et al. modified the PlantVillage dataset and CapsNet algorithm in Section 2.
CapsNet model by stacking three convolutional layers in The experimental setup and statistical test characteristics
addition the conventional CapsNet architecture for to evaluate the CapsNet architectures are shared in Section
identification of peanut leaf diseases in their own dataset. 3. The comparison on related works according to system
They evaluated the performance of CapsNet with SVM performances, advantages and superiority of the models
and CNN and reported the CapsNet as machine learning are discussed in Section 4.
algorithm with the best generalization performance for
peanut leaf diseases [21]. Kurup et al. analyzed plant 2. Materials and Methods
leaves from fourteen species from PlantVillage to identify 2.1. PlantVillage Database
the leaf diseases using CapsNet. They compared the
PlantVillage Database is a challenge dataset that aims at
efficacy of CapsNet and CNN for multiple diseases [22].
changing the traditional harvesting processes with novel
To the best of our knowledge, there is no research which
computer-aided developments for identification of plant
focuses on directly identification of bell pepper leaf
leaf diseases. It is collected by Land Grant University,
disease. However, in the papers on multi-class plant leaf
USA [23]. PlantVillage is comprised a total number of
classification through PlantVillage dataset incorporated
54305 leaf images with healthy and diseased spots from
bell pepper analysis. Majority of them shared an average
thirteen plants. Additionally, background images were
classification performance for all plant leaf diseases in
shared in the dataset for segmentation researches.
PlantVillage with none of class-specific performances.
The existing literature focused on one plant with binary
Therefore, the achievements present overall performance
disease classification (healthy-diseased), but also analyzed
instead of assessing the generalization capacity of models
multi-class plant leaf diseases. In this study, we studied on
for each plant. This paper aims at exploring CapsNet
identification of bell pepper leaf disease using a total
architecture for identification of bell pepper leaf disease in
number of 2475 images (997 bacteria diseased, 1478
PlantVillage dataset, evaluating the generalization
- 59 -
 Gökhan Altan., International Journal of Applied Mathematics Electronics and Computers 08(03): 057-063, 2020

healthy leaves). Sample images for health and diseased many variations in the depth of CONV, number and size
bell pepper plant are depicted in Fig. 1. of filters, and finally classification parameters such as
We analyzed plant leaf images without cropping and depth of FCs, number of neurons at each FC, learning rate,
segmentation of diseased spots. Each leaf image has posed dropout index, and more [25].
with different angle and has a background view. The The squashing function is:
several images have shadow effect; however, none of
∥ 𝑠𝑗 ∥2 𝑠𝑗 (1)
images was excluded from the dataset to include actual 𝑓𝑠𝑞𝑢𝑎𝑠ℎ𝑖𝑛𝑔 (𝑗) =
1+ ∥ 𝑠𝑗 ∥2 ∥ 𝑠𝑗 ∥
cases including background and shadow. We augmented
the leaf images by 4x using horizontal-, vertical-, and both- where 𝑠𝑗 represents for 𝑗𝑡ℎ individual primary capsule
flipping. The plant leaf images were resized to 64×64 to predictions.
obtain a standard input size for the CapsNet. Primary capsule is the lowest layer capsule which
2.2. Capsule Networks (CapsNet) extracts the existence and spatial information of feature.
The next capsules (routing capsules) trace upper level
Transfer learning provides collecting the randomization
features and instantiation parameters.
into a pre-defined space. One of the main reasons for
efficacy of pre-trained CNN architectures is detailed
3. Experimental Results
convolutional analysis stages (CONV layer) for feature
learning. This circumstance enables performing faster The plant leaf images have shadows and sun shining
optimization for pre-trained weights using regularization effects. Due to increase the heterogeneity of the dataset
and factorization techniques. On the other hand, using with actual cases of monitoring plant leaves, each image
pooling layer (down-sampling) may cause significant data was controlled in detail for the analysis. About a half of
loss among feature maps. Capsule network (CapsNET) is the diseased leaf images has small bacteria spots. This case
a novel DL algorithm to overcome the shortcoming of provides assessing the CapsNet for early diagnosis of plant
CNN by excluding the pooling layer from the architecture leaf diseases. Therefore, we experimented on a standard
and transferring spatial information between layers by capsule layer in CapsNet architectures; however, various
capsule [19]. depth and range of neurons and FCs in supervised learning
The input of a capsule is output of CONV layer at a stage was iterated for defining the optimum model with
specified number of filter sizes. The output of a capsule highest classification performance for identification of bell
consists of the likelihood for encoded by capsules between pepper leaf disease. In this study, we shared the best
feature maps and instantiation parameters including pose, achievements for the CapsNet architectures and compared
texture, rotating, and deformation information [24]. The with the state-of-art on PlantVillage dataset.
spatial information enables transferring part-whole The leaf images were resized to 64×64 to obtain a
hierarchy for learned feature map between low-level standard input size for CapsNet. The RGB images
capsules [19]. transformed to gray-scale images before the analysis. Data
The main benefits of CapsNet are dynamic routing augmentation was performed to increase the dataset for
between capsules, spatial information, and the squashing enabling the CapsNet model to learn various
function for defining the output at [0-1] as likelihood. The representations of plant leaf images to and to avoid
capsules with activity vector fed into the fully-connected overfitting. The number of analyzed PlantVillage images
layers (FC) just as CNN [19], [24]. database was increased by 4 times by applying vertical flip,
The activity vectors and spatial information makes horizontal flip, and both. The pre-processing stage of the
CapsNet robustness to overfitting even for small-scale leaf images is indicated in in Fig. 3. Using data
databases, dependent learning for part-whole hierarchy for augmentation procedure, we got 3988 and 5912 plant leaf
rotated and scaled images. Furthermore, the dropout images for diseased and healthy bell peppers, sequentially.
factorization speeds up the training of CapsNet by
excluding the neurons in FCs by a similarity index at each
FC [25].
The CONV in CapsNet represents generating different
representations of input data according to feature maps and
rectified-linear unit (ReLU). The structure of CapsNet for
proposed leaf disease is indicated in Fig. 2. There is no
pooling layer. The number of the CONVs depends on the
level of the features to extract [26]. Whereas the first
CONVs identify low-level features, primary capsules Figure 3. Pre-processing of plant leaf images
transfer the spatial information for the low-level maps.
The experiments were iterated on various CapsNet
Therefore, composing a CapsNet depends on deciding
models using the adaptability of FCs. The proposals were

- 60 -
 Gökhan Altan., International Journal of Applied Mathematics Electronics and Computers 08(03): 057-063, 2020

trained using 80% of the dataset stratified by plant leaves. identification performance, using big number of neurons
None of the data augmentation was included for both for both FC1 and FC2 failed to separate the healthy and
training and testing the CapsNet. The remaining of the diseased leaf images. Using 960 neurons at FC1 and 768
dataset was utilized to validate the performance of trained neurons at FC2 with the proposed Capsule architecture
CapsNet. The test results were evaluated using reached the best achievements. The best bell pepper leaf
independent test characteristics. Accuracy, sensitivity, and disease identification performance was achieved with the
specificity were calculated using confusion matrix of the rates of 95.76%, 96.37%, and 97.49% for accuracy,
trained CapsNet model [27]. sensitivity, and specificity, respectively. The activation
In the CONV layer of the CapsNet, convolution kernel functions are RELU, RELU, and sigmoid for FC1, FC2,
size is 9×9 with 256 filters. Sequentially, 32 primary and FC3, respectively.
capsule layers were generated using a convolution kernel
of 9×9 and stride of 2. Output of the each layer has 24×24 4. Discussion
capsules (24×24×32). Each capsule is an 8-dimensional Most of the novel papers in last decade focused on the
vector that is spatial information. The class capsule layer pre-trained CNN architectures to identify plant leaf
outputs 16 dimensional vector for per capsule. The 8- diseases. Especially, transfer learning approach is the main
dimensional vector is converted into 16-dimensional leaf reason for this choice with popular architectures. The
capsules using an encoder procedure by the weight matrix adaptability and applicability of pre-trained CNN
𝑊𝑖𝑗 . The class capsule layer, leaf capsules, generates a architectures provide a steady optimization for the issues.
matrix with a dimension of 2×16 (healthy-diseased × However, the disadvantage of CNN in data loss with
spatial information). pooling layer makes use of CNN a handicap for
researchers. The achievements extracted from the papers
for bell pepper leaf disease identification on PlantVillage
are presented in Table 2.
The performance of the proposed CapsNet architecture
is comparable to a very limited study, since most studies
presented overall accuracy rather than class-based
classification performance. To the best of our knowledge,
there is no paper that directly focused on identification of
bell pepper leaf diseases. The achievements of related
Figure 4. Decoding stage of the proposed CapsNet architecture
works were calculated using confusion matrix and
The decoder stage of CapsNet consists of FCs (see Fig. weighted average of class-based performances for healthy
4). The number of the neurons at the last FC is the same as and diseased pepper leaves. Geetharamani and Arun
pixels of input leaf image (64×64×1=4096) to establish a Pandian proposed light-weight CNN architecture with
decoder model. The depth of FCs was fixed at 3. The nine layers for identification of plant leaf diseases. They
number of neurons at the first FC (FC1) and second FC experimented on various batch size, epoch, validation
(FC2) was iterated at a range of 128×1024 (increasing by folds, and dropout factorization index on CNN. They
64 neurons). The output functions of the FCs are ReLU, compared the efficiency of their proposal with popular
ReLU, and softmax, respectively. CNN architectures including AlexNet, VggNet, Inception-
Table 1. The best five classification performances (%) for v3 and ResNet and reported the superiority of their
CapsNet architectures with various FC layers. proposal with rates of 93.00%, 92.00%, and 93.00% for
pepper leaf disease identification accuracy, sensitivity, and
CapsNet Architectures ACC SEN SPE
specificity, respectively [18]. Kurup et al. applied
FC1(512), FC2(832), FC3(4096) 92.68 92.86 95.05 conventional CapsNet architecture to the PlantVillage
FC1(1024), FC2(576), FC3(4096) 94.14 94.61 96.27 dataset for leaf classification and leaf disease identification
FC1(384), FC2(896), FC3(4096) 94.85 97.99 98.56 tasks. They achieved pepper leaf diseased-healthy
FC1(960), FC2(704), FC3(4096) 95.30 98.12 98.66 identification performance rates of 91.00-96.00%, 85.00-
FC1(960), FC2(768), FC3(4096) 95.76 96.37 97.49 92.00%, and 88.00-94.10% for precision, sensitivity, and
F1 score, respectively [22].
The best classification performances for identification
of plant leaf diseases for bell pepper are presented in Table
1. The achievements prove that with the CapsNet models
has ability to perform accurate classification performances
for identification of bell pepper leaf diseases using simple
FCs with capsule and spatial information. Whereas
increasing number of the neurons in each FC enhanced the
- 61 -
 Gökhan Altan., International Journal of Applied Mathematics Electronics and Computers 08(03): 057-063, 2020

Table 2. The state-of-art for pepper leaf disease identification using image segmentation and soft computing techniques,” Inf.
on PlantVillage. Process. Agric., 2017, doi: 10.1016/j.inpa.2016.10.005.
[3] K. Jagan, M. Balasubramanian, and S. Palanivel, “Detection
Studies Classifier ACC SEN SPE
and Recognition of Diseases from Paddy Plant Leaf Images,”
[18] CNN 93.00 92.00 93.00 Int. J. Comput. Appl., 2016, doi: 10.5120/ijca2016910505.
[22] CapsNet - 89.17 - [4] S. Phadikar, “Classification of Rice Leaf Diseases Based on
This study CapsNet 95.76 96.37 97.49
Morphological Changes,” Int. J. Inf. Electron. Eng., 2012, doi:
10.7763/ijiee.2012.v2.137.
[5] T. Islam, M. Sah, S. Baral, and R. Roychoudhury, “A Faster
Whereas CNN with various feature learning Technique on Rice Disease Detectionusing Image Processing
of Affected Area in Agro-Field,” in Proceedings of the
architectures is so popular, the deficiencies of CNN such International Conference on Inventive Communication and
as pooling layer and no connection between low-level Computational Technologies, ICICCT 2018, 2018, doi:
features are the main topic of machine learning researchers. 10.1109/ICICCT.2018.8473322.
[6] C. U. Kumari, S. Jeevan Prasad, and G. Mounika, “Leaf
This paper explored CapsNet on pepper leaf disease Disease Detection: Feature Extraction with K-means clustering
classification using no pooling layer, transferring spatial and Classification with ANN,” in 2019 3rd International
information between layers by capsules. The proposed Conference on Computing Methodologies and Communication
(ICCMC), 2019, pp. 1095–1098, doi:
CapsNet architecture defined the most responsible FC 10.1109/ICCMC.2019.8819750.
models using experimental assessments and model-based [7] S. Arivazhagan, R. N. Shebiah, S. Ananthi, and S. Vishnu
evaluation metrics. Using big number of neurons at FC1 Varthini, “Detection of unhealthy region of plant leaves and
classification of plant leaf diseases using texture features,”
and smaller number of neurons at FC2 enabled learning Agric. Eng. Int. CIGR J., 2013.
more responsible and optimized models for identification [8] S. S. Chouhan, A. Kaul, U. P. Singh, and S. Jain, “Bacterial
of bell pepper leaf disease using simplistic supervised foraging optimization based radial basis function neural
network (BRBFNN) for identification and classification of
learning procedures. The proposed CapsNet architecture plant leaf diseases: An automatic approach towards plant
with the highest classification performance reached the pathology,” IEEE Access, 2018, doi:
10.1109/ACCESS.2018.2800685.
rates of 95.76%, 96.37%, and 97.49% for accuracy,
[9] S. Kumar, B. Sharma, V. K. Sharma, H. Sharma, and J. C.
sensitivity, and specificity, respectively. The proposal Bansal, “Plant leaf disease identification using exponential
outperformed the literature for pepper leaf diseases. spider monkey optimization,” Sustain. Comput. Informatics
Syst., 2018, doi: 10.1016/j.suscom.2018.10.004.
Moreover, the proposal presents pepper-specific leaf [10] S. Sladojevic, M. Arsenovic, A. Anderla, D. Culibrk, and D.
disease identification on PlantVillage dataset. Stefanovic, “Deep Neural Networks Based Recognition of
Plant Diseases by Leaf Image Classification,” Comput. Intell.
Neurosci., 2016, doi: 10.1155/2016/3289801.
5. Conclusion [11] S. H. Lee, C. S. Chan, S. J. Mayo, and P. Remagnino, “How
We explored the effect of FC models for CapsNet deep learning extracts and learns leaf features for plant
classification,” Pattern Recognit., 2017, doi:
architecture with various numbers of neurons at each FC. 10.1016/j.patcog.2017.05.015.
Although CNN have been the most popular machine [12] J. Amara, B. Bouaziz, and A. Algergawy, “A deep learning-
learning algorithm in DL for a couple of years, the capsule based approach for banana leaf diseases classification,” in
Lecture Notes in Informatics (LNI), Proceedings - Series of the
architecture is a formidable opponent for image analysis Gesellschaft fur Informatik (GI), 2017.
due to capability of transferring spatial information, pose [13] M. Brahimi, K. Boukhalfa, and A. Moussaoui, “Deep Learning
parameters, activity vectors, and part-whole hierarchy. for Tomato Diseases: Classification and Symptoms
Visualization,” Appl. Artif. Intell., 2017, doi:
The establishing pre-trained architectures on CapsNet 10.1080/08839514.2017.1315516.
have a bigger potential than CNN to be used for transfer [14] B. Liu, Y. Zhang, D. J. He, and Y. Li, “Identification of apple
leaf diseases based on deep convolutional neural networks,”
learning. However, training big number of datasets is still
Symmetry (Basel)., 2018, doi: 10.3390/sym10010011.
time-wasting for CapsNet in DL. The optimization [15] K. P. Ferentinos, “Deep learning models for plant disease
techniques in FCs may reduce the training time. The idea detection and diagnosis,” Comput. Electron. Agric., 2018, doi:
10.1016/j.compag.2018.01.009.
of CapsNet for spatial information must be supported with [16] S. P. Mohanty, D. P. Hughes, and M. Salathé, “Using deep
optimization algorithms on FCs. learning for image-based plant disease detection,” Front. Plant
Sci., 2016, doi: 10.3389/fpls.2016.01419.
Author's Note [17] K. Zhang, Z. Xu, S. Dong, C. Cen, and Q. Wu, “Identification
of peach leaf disease infected by Xanthomonas campestris with
Abstract version of this paper was presented at 9th deep learning,” Eng. Agric. Environ. Food, 2019, doi:
International Conference on Advanced Technologies 10.1016/j.eaef.2019.05.001.
(ICAT'20), 10-12 August 2020, Istanbul, Turkey with the [18] G. Geetharamani and A. P. J., “Identification of plant leaf
diseases using a nine-layer deep convolutional neural network,”
title of “Performance Evaluation of Capsule Networks for Comput. Electr. Eng., 2019, doi:
Classification of Plant Leaf Diseases”. 10.1016/j.compeleceng.2019.04.011.
[19] S. Sabour, N. Frosst, and G. E. Hinton, “Dynamic routing
between capsules,” in Advances in Neural Information
References Processing Systems, 2017.
[1] M. W. Rosegrant and S. A. Cline, “Global Food Security: [20] S. Verma, A. Chug, and A. P. Singh, “Exploring capsule
Challenges and Policies,” Science. 2003, doi: networks for disease classification in plants,” J. Stat. Manag.
10.1126/science.1092958. Syst., 2020, doi: 10.1080/09720510.2020.1724628.
[2] V. Singh and A. K. Misra, “Detection of plant leaf diseases [21] M. Dong, S. Mu, T. Su, and W. Sun, “Image Recognition of

- 62 -
 Gökhan Altan., International Journal of Applied Mathematics Electronics and Computers 08(03): 057-063, 2020

Peanut Leaf Diseases Based on Capsule Networks,” 2019, pp.

43–52.
[22] R. V. Kurup, M. A. Anupama, R. Vinayakumar, V. Sowmya,
and K. P. Soman, “Capsule network for plant disease and plant
species classification,” in Advances in Intelligent Systems and
Computing, 2020, doi: 10.1007/978-3-030-37218-7_47.
[23] D. P. Hughes and M. Salathe, “An open access repository of
images on plant health to enable the development of mobile
disease diagnostics,” Nov. 2015.
[24] M. Kwabena Patrick, A. Felix Adekoya, A. Abra Mighty, and
B. Y. Edward, “Capsule Networks – A survey,” Journal of King
Saud University - Computer and Information Sciences. 2019,
doi: 10.1016/j.jksuci.2019.09.014.
[25] G. Altan, “DeepGraphNet: Grafiklerin Sınıflandırılmasında
Derin Öğrenme Modelleri,” Eur. J. Sci. Technol., pp. 319–327,
Oct. 2019, doi: 10.31590/ejosat.638256.
[26] D. C. Cireşan, U. Meier, J. Masci, L. M. Gambardella, and J.
Schmidhuber, “Flexible, high performance convolutional
neural networks for image classification,” in IJCAI
International Joint Conference on Artificial Intelligence, 2011,
doi: 10.5591/978-1-57735-516-8/IJCAI11-210.
[27] G. Altan, Y. Kutlu, and A. Gökçen, “Chronic obstructive
pulmonary disease severity analysis using deep learning on
multi-channel lung sounds,” TURKISH J. Electr. Eng. Comput.
Sci., 2020, doi: 10.3906/elk-2004-68.

- 63 -