Disease Detection in Tomato Leaves using

Machine Learning and Statistical Feature Fusion

S.Arun Kumar,S.Sasikala
Department of Electronics and Communication Engineering
Kumaraguru College of Technology
Coimbatore, India
arunkumar.s.ece@kct.ac.in

Abstract— Tomato is a domestic plant that is pre- Tomato - Lycopersicon esculentum is an essential
dominantly used worldwide and less tolerant to diseases. Among vegetable in all cuisines. The nutritional value present in
the diseases that affect tomato plants, often the type and cause the tomato makes it a protective food [3]. Tomato is grown
of the disease is unknown to the farmer. Accurate plant disease worldwide under 46.16 lakh hectares with a global
diagnosis is critical for increasing crop yield, monitoring plant production of about 1279.93 lakh tonnes. Tomato is an
health, and providing appropriate treatment at an earlier stage. important food crop and is easily susceptible to viral,
Therefore, this work aims at developing a system to identify the bacterial and fungal disease. As a result, crop productivity
type of disease that affects the tomato leaves at an early stage. is gradually falling across the globe over the years. Crop
In the proposed system, Gray Level Co-occurrence Matrix
scouting and conducting a manual crop inspection to
(GLCM) and Local Binary Pattern (LBP) texture-based
features are used to identify the type of disease by means of a
identify diseases is time consuming and inconsistent for
Support Vector Machine (SVM) classifier. To improve the farmers and plant breeders. In addition, naked-eye
diagnostic accuracy of disease detection system, fusion of two observation for the identification of diseases becomes
different texture features is performed using Principal cumbersome for large farmland. This requires techniques
Component Analysis (PCA) and Canonical Correlation Analysis to automate the diagnosis of plant disease so as to save
(CCA). A publicly available tomato leaves dataset is used to time and produce more accurate and reliable results. Image
evaluate the proposed system. A considerable improvement in processing in combination with machine learning is used
relevant performance metrics is observed in CCA based feature to automate the detection of plant disease owing to the
fusion. Henceforth the proposed system could be used as an aid advancements in imaging sensors and computational
to farmers and plant breeders to identify the type of disease and intelligence [4].
provide the necessary treatment.
The proposed work aims to develop a system for the
Keywords— Tomato, plant disease detection, texture features, detection of leaf disease in tomato plants using GLCM and
plant health, Canonical correlation analysis, feature fusion LBP texture based features. Feature extraction is followed
by a back end SVM classifier with linear and non-linear
I. INTRODUCTION kernel functions. To improve the multi class performance,
Agriculture is the primary occupation that provides a system is developed by combining GLCM and LBF
food and helps us to live a healthy life. The plant should features. Furthermore, fusion strategy based on PCA and
be healthy in order to provide people with nutritious food. CCA is employed between GLCM and LBP features. The
In modern days, the agricultural activity along with performance of the leaf disease detection system is
climatic conditions affects the plant causing multitude of assessed in terms of relevant metrics .
diseases. The plant is therefore deprived of nutrition and
II. LITERATURE REVIEW
the yield becomes less as well. In order to increase crop
yield, effective plant health monitoring systems at earlier Considering the importance of automated diagnosis of
stages of plant growth are therefore necessary. Early plant disease, several literature studies have focused on
diagnosis of plant disease is a first step towards the developing a plant disease detection system using image
monitoring and management of plant health [1]. analysis and machine learning techniques [5-15].
Precision farming, the modern, efficient farming In [5], the authors have developed an image processing-
practice combines a variety of sensors with real-time data based solution to detect plant diseases. In the proposed system,
analytics to increase crop yields. Satellite imagery, tools image preprocessing is followed by K-means segmentation
for data processing, smart sensors, Internet of Things and texture-based feature extraction. Classification is
(IOT), autonomous tractors and geo spatial tools helps achieved using Neural network.
precision farming to transform current agricultural An image processing-based support system is developed
practices. Drones equipped with imaging sensors capture in [6], to help farmers for early plant disease diagnostic
plant images in real time and monitor plant health and system. Correlation based feature is classified by Support
stress. Depending on the feedback received from the vector Machine classifier.
support system, farmers take necessary steps to manage
plant health. All critical decisions in precision farming are A survey on different image processing techniques used
made using sensor data. Precision farming increases the for diseases is presented in [7]. The general steps involved in
accuracy of all agricultural activities resulting in reduced an end-to-end image processing framework for disease
operational costs and high crop yields. Automatic disease detection is explained in detail.
diagnosis of plants is also one of the key features of
Authors in [8], reviewed characteristics of various fungal,
precision agriculture. Machine learning techniques are
viral, and bacterial diseases that occur in tomato plant.
used to develop decision support system for early
Correlation and color based study have been used as an
diagnosis of plant disease [2].
attempt for disease classification.

978-1-6654-2829-3/21/$31.00 ©2021 IEEE

 In [9], a study on use of image processing and machine diseases early stages of plant growth is depicted in Fig 1.
learning based framework for automatic plant disease
classification in tomato plant is proposed. A significant
improvement in accuracy was obtained using decision tree
classifier.
An exhaustive study is conducted on various segmentation
methods to discriminate the lesion clearly. Among various
methods for segmentation, histogram-based segmentation
with automatic thresholding outperforms traditional Otsu
method and iterative methods [10].
A detailed review on Neural Network (NN) classifier in
the use of plant disease detection is done in [11]. Mechanism
of NNs, different types of NNs, NN models and classifiers are
discussed exhaustively. A special emphasis is made on the use
of hyperspectral data for plant disease diagnostics. Combining
neural network with hyperspectral data helps in non-
destructive early plant disease diagnostics.
Fig. 1. (i) Bacterial spot (ii) Late blight (iii) Septorial leaf
Authors in [12], discussed the steps involved in plant spot (iv) Mosaic disease (v) Yellow curved disease (vi)Healthy leaf
disease detection and disease severity assessment. B. Preprocessing and Image Segmentation
Approaches used in automatic plant disease detection and
disease severity evaluation have been reviewed. The potential After image acquisition, contrast enhancement is
benefits and shortcomings of various steps in disease detection used as preprocessing step to enhance image quality.
and severity assessment such as segmentation techniques, Image segmentation using HSI (Hue Saturation
classification algorithms, and performance evaluation criteria Intensity) space [17] is used to segment the diseased portion
are explained in a detailed manner. of the leaf. RGB colour space is converted to HSI space.
Saturation is used for masking the image to extract specific
In [13], a disease detection system is developed for region. Green pixels are masked by setting a threshold value
cassava plant using spectral data. Spectroscopy is used to
which is 10% of largest value in the image and as a result
understand how the healthy and diseased plant responds to
diseased area mask is extracted. Then the diseased area
light. Colour and SIFT based features are extracted and
classified using K-NN, SVM and decision trees. Spectral data masked image is multiplied with hue image. Since the mask
resulted in good recognition accuracy than image data only includes 1's and 0's, the area affected by the disease in
the leaves has the value of 1's. Only the diseased region exists
A leaf disease diagnosis system for different crop species when the diseased area mask is multiplied with the hue
is proposed in [14]. Image segmentation is achieved using image.
Grab-cut algorithm. Local binary pattern is used for feature To get rid of further unwanted regions,
extraction. One class SVM is used for identifying the type of morphological analysis and blob-based segmentation is done
disease. The algorithm is generalized to detect diseases on to retrieve the desired Region of Interest (ROI). Further the
different crop species.
desired ROI is converted back to RGB space for further
In [15], a multi sensor network is used to collect real-time analysis.
data about the pests, soil, environmental conditions, climate,
and other metrological factors. From the collected high
dimension data, an early warning model was proposed to
classify the level of pest damage into mild, moderate,
moderately severe, and severe.
III. MATERIALS AND METHODS

System architecture of tomato leaf disease detection

system is depicted in Fig 2.
A. Dataset
A publicly available database [16], with healthy and
diseased tomato leaves is used for the proposed study. The
database consists of images of diseases such as bacterial spot,
late blight, septorial leaf spot, mosaic disease and yellow
curved disease. Bacterial spot and late blight are caused by
bacteria and fungi respectively. Septorial leaf spot, mosaic
disease and yellow curved disease are caused by viral Fig. 2. Architecture of proposed Leaf disease detection system
infections. A total of 1000 images (200 from each disease)
are used in this study. The tomato leaves affected by different
C. Feature Extraction with largest variance. Majority of information is present in
GLCM and LBP based texture features are used in feature first few principal components and the remaining components
extraction phase. Texture is a good descriptor for identifying are removed as they do not carry much information.
the diseased leaf from healthy leaf. Texture also varies from In feature fusion based on PCA, dimensionality
disease to disease. Therefore, texture-based features are reduction is done on both GLCM and LBP feature sets
widely used in image processing-based plant disease separately and the reduced features are combined to form the
diagnostic systems. fused feature set.

1) Gray Level Co-Occurrence Matrix: 2) Canonical Correlation analysis:

Gray Level Co-occurrence Matrix (GLCM) considers spatial CCA is used to examine the mutual relations between
relationships between pixels to compute the texture through two feature sets [20]. In feature fusion using CCA, cross
statistics. An image with varying gray levels can be correlation is used as a metric between feature sets a and b to
represented by GLCM . GLCM specifies the information in generate new sets A=waTa and B = wbTb. CCA transforms in
the gray level variation that co-occur in an image. Texture a way that maximises discriminant information between two
features that provide intensity variation in an image are features. Maximization is achieved by simultaneously
calculated using the co-occurrence matrix. From GLCM maximising cross correlation and minimising auto correlation
matrix, thirteen features are extracted using relevant between the feature vectors.
mathematical equations. Contrast, Correlation, Energy, In feature fusion using CCA, before combining two
Homogeneity are the texture features of GLCM. Statistical feature sets (GLCM and LBP) correlation features are
features include Mean, Standard Deviation, Root Mean extracted between them. Correlation features produce
Square (RMS), Entropy, Variance, Smoothness, Kurtosis, discriminant information from feature vector by eliminating
Skewness and Image difference-measure (IDM) [18]. redundant information.

2) Local binary Pattern: E. Disease Classification

Local binary Pattern (LBP) is a commonly used texture feature
Support Vector Machine (SVM), a supervised ML
in image analysis [19]. LBP is invariant to illumination
algorithm is used for disease classification. After performing
changes, used in real time implementations and
feature fusion, the fused features are classified using SVM
computationally efficient when compared to other features. In classifier. SVM, a maximum margin classifier increases the
LBP, the intensity value of center pixel (3 x 3 mask) is distance between two classes using training data. Once SVM
compared with its 8 neighbors. In thresholding, binary value
model is built, test data is used for validating the classifier
1 is assigned if the pixel of interest is greater or equal to the
performance. SVM uses linear or nonlinear kernels to classify
center pixel. Else 0 is assigned. The output binary value of
the data. Linear SVM is used for classifying linearly separable
LBP is computed by adding the weights corresponding to the
data. On contrary, highly nonlinear data is classified based on
threshold values. Similarly, the LBP codes for all of the pixels kernel transformation. In this work, linear SVM and nonlinear
in the image are generated. Henceforth, the resultant matrix is SVM (Polynomial and Radial-Basis Function (RBF) kernels)
extracts texture features. The calculation of LBP is illustrated
with 5-fold cross validation are used for classification.
with help of an example pixel value in Fig 3.
Confusion matrix obtained from the classifier is used to
calculate relevant performance metrics as specified in Table I.
IV. RESULTS AND DISCUSSIONS
A. Results
The pre-processed image segmented based on hue-based
segmentation is shown in Fig 4. The segmented image is
further classified using SVM classifier and the performance
metrics such as accuracy, sensitivity, specificity, precision, F1
score, Mathews Correlation coefficient and Kappa coefficient
Fig. 3. Illustration of LBP are determined from the confusion matrix.
D. Feature Fusion: The performance metric of baseline system obtained using
GLCM and LBP features provide distinct information. Fusing GLCM and LBP features separately are tabulated in Table II.
two features provide accurate feature representation than a The results of combining GLCM and LBP features without
any feature transformation technique are listed in Table III.
single feature. Hence, feature fusion is employed in the
Table IV shows the performance metrics obtained through
proposed work. In Feature fusion, two feature vectors are
fusion of GLCM and LBP texture features using PCA and
concatenated to form a fused feature vector. In this work, CCA.
features from GLCM and LBP are combined using PCA and
CCA for feature fusion. B. Discussions
From the performance metrics of Table II, it is evident that
1) Principal Component analysis: the performance of the two-class disease detection system is
Principal Component analysis is a linear mean square- significantly better when compared to multi class for the
based dimensionality reduction technique The principal baseline system. An accuracy of 95.65% and 96.68% is
components in PCA are determined by retaining components achieved using GLCM and LBP features respectively for two
classes (Bacterial spot and late blight). More than one TABLE II. METRICS OF BASELINE SYSTEM USING GLCM
performance metric is required to quantify the performance of
a system. Henceforth other metrics listed in Table I are also
calculated. Binary Class
SVM
ACC SEN SPC F1 MC PRE
Type
TABLE I. PERFORMANCE METRICS EVALUATED FROM CONFUSION
MATRIX Linear 96.00 95.29 0.95 0.91 0.95 0.91
95.65
Performance Poly 94.37 94.50 94.24 0.94 0.88 0.94 0.88
Equation
metric RBF 95.40 96.00 94.76 0.95 0.90 0.95 0.90

+ Multi-Class
Accuracy =
+ + + Linear 76.29 76.41 94.07 0.76 0.70 0.76 0.25
Poly 80.42 80.49 95.10 0.80 0.75 0.81 0.38
RBF 73.26 73.36 93.31 0.73 0.66 0.73 0.16
Sensitivity = =
+

The combined feature set resulted in performance

TN
Specificity SPC = = improvement when compared to baseline system for multi
+
class as seen in Table IV. LBP and GLCM predict well the
onset of disease when used separately. Therefore, combining
Precision = them would further increase the diagnostic accuracy of the
+ disease detection. Feature combination resulted in an accuracy
of 96.93% and 85.27% for two class and multiclass systems,
× respectively. A percentage increase of 6.03% is achieved over
F1score = 2 ×
+ the baseline GLCM system.
Matthews Combining two different features using fusion techniques
is proposed in [21]. LBP features are highly discriminative,
Correlation . − .
= computationally simpler, and robust to illumination changes.
Coefficient ( + )( + )( + )( + ) GLCM features respond to inter pixel correlation. Therefore,
Kappa combining LBP and GLCM using efficient fusion strategy
( )
= will result in efficient feature set. Fusion technique based on
coefficient ( )
PCA and CCA is experimented in this work . Fusion based on
† TP - True positive, TN – True Negative, FN – False PCA and CCA reduces the feature dimension.
Negative, FP – False Positive TABLE III. METRICS OF BASELINE SYSTEM USING LBP

Binary Class
SVM
ACC SEN SPC F1 MC PRE
Type
Linear 96.68 96.50 96.86 0.96 0.93 0.96 0.93
Poly 96.16 95.00 97.38 0.96 0.92 0.97 0.92
RBF 95.40 96.00 94.76 0.95 0.9 0.95 0.90

Multi-Class
Linear 72.35 72.53 0.72 0.65 0.73 0.13
93.09
Poly 77.30 77.42 94.32 0.77 0.71 0.77 0.29
RBF 71.24 71.41 92.81 0.71 0.64 0.72 0.10

TABLE IV. METRICS OF COMBINED SYSTEM USING LBP AND GLCM

Binary Class
SVM
ACC SEN SPC F1 MC PRE
Type
Fig. 4. Diseased image after segmentation
Linear 96.93 96.00 97.91 0.96 0.93 0.97 0.93
Poly 95.40 94.50 96.34 0.95 0.90 0.96 0.90
The metrics as observed from Table II and III, clearly
RBF 94.88 94.50 95.29 0.94 0.89 0.95 0.89
illustrates that performance of multiclass system (5 class) falls
behind the two-class system. An accuracy of 80.42 and 77.30 Multi-Class
is obtained using GLCM and LBP features, respectively.
Linear 81.53 81.61 95.38 0.81 0.77 0.82 0.42
An improvement in accuracy was obtained in image
classification by combining multiple texture features. To Poly 85.27 85.32 96.31 0.85 0.81 0.53
0.85
improve the performance of proposed system combining RBF 77.09 77.14 94.27 0.77 0.71 0.77 0.28
texture features is experimented in this work.
 In PCA based feature fusion, resultant fused feature set is the proposed fusion based on CCA could be used as an
of fewer dimensions when compared to combined feature set effective strategy for diagnosing diseases in tomato leaves.
before fusion. In CCA based feature fusion, before combining
two feature sets (GLCM and LBP) correlation features are
extracted between them. The correlation features in CCA TABLE VI. METRICS OF COMBINED SYSTEM FUSED USING CCA
reduces the redundancy at the cost of increased performance
[22].
Binary Class
It is observed from Table V and VI, that feature fusion SVM
ACC SEN SPC F1 MC PRE
outperforms the baseline and combined system performance Type
metrics for both binary and multiclass disease detection. The Linear 96.93 96.00 97.91 0.96 0.93 0.97 0.93
accuracy score of CCA fused system is 87.08%. CCA based Poly 96.42 96.00 96.86 0.96 0.92 0.96 0.92
feature fusion resulted in an increase of 8.28% when RBF 95.65 97.00 94.24 0.95 0.91 0.94 0.91
compared to baseline system.
Multi-Class
Specificity is an important measure for a disease
Linear 83.45 83.53 95.86 0.83 0.79 0.83 0.48
diagnostic system as it reduces the number of false positives.
It is observed from the results, that specificity is almost above Poly 87.08 87.15 96.77 0.87 0.83 0.87 0.59
92% in all the cases irrespective of the class. RBF 83.85 83.92 95.96 0.83 0.79 0.83 0.49

Sensitivity corresponds to true positive rate, and it is

moderate for all cases. Sensitivity is high for CCA based V. CONCLUSION
fusion when compared to other systems. In this work, GLCM and LBP texture-based features are
Precision indicates positive predictive value. High value used for diagnosing diseases in tomato leaves. Feature
of precision corresponds to low false positive. Precision is combination and fusion based on PCA and CCA is proposed
high in case of CCA feature fusion, and its value lies above for improving the system performance. CCA based fusion
0.9 for two class and 0.83 for multiclass systems. F1 score is strategy employed between GLCM and LBP texture features
related to accuracy of the classifier. F1 score is computed by improved the diagnostic accuracy and other performance
striking a balance between false positive and false negative. metrics for multiclass disease detection in tomato leaf. A
F1 score is high in CCA based disease detection system. significant improvement in metrics justifies the use of
proposed system for disease diagnosis. Henceforth the system
TABLE V. METRICS OF COMBINED SYSTEM FUSED USING PCA can be used as an aid for farmers and plant breeders for early
disease diagnosis and increasing the yield.
Binary Class The proposed work focuses only on machine learning.
SVM However deep learning-based approaches offer improved
ACC SEN SPC F1 MCC PREC
Type
accuracy than ML. Despite improved performance, deep
Linear 96.93 97.00 96.86 0.97 0.93 0.97 0.93
learning approaches are computationally expensive, time
Poly 96.68 97.00 96.34 0.96 0.93 0.96 0.93
consuming, and the results are incomprehensible due to the
RBF 95.91 98.00 93.72 0.96 0.91 0.94 0.91
number of hidden layers.
Multi-Class To further improve the accuracy in multiclass, multiple
Linear 82.34 82.41 95.58 0.82 0.78 0.82 0.44 features with a feature selection may be employed. Our future
Poly 86.00 85.97 96.49 0.86 0.82 0.86 0.56 work focuses on using other segmentation methods and
RBF 81.23 81.29 95.30 0.81 0.76 0.81 0.41 evolutionary algorithms for improving the detection
performance. In addition, the system can be implemented in
mobile phone for remote plant disease diagnosis by combining
MCC is a measure of prediction performance. A MCC hardware and software modules.
value of 1 indicates exact prediction, while a value of 0
indicates opposite prediction. From the results MCC value is
above 0.79 in the fused system. REFERENCES
Kappa coefficient, a key metric for multiclass problems,
measures the degree of agreement between the actual and [1] Kacira, M. and Ling, P.P., “Design and development of an automated
predicted classes. Kappa coefficient is also above 0.9 for two and Non–contact sensing system for continuous monitoring of plant
class system and comparatively higher for multiclass in CCA health and growth”, Transactions of the ASAE, 2001, 44(4), p.989.
based fusion. [2] Geoffrey Ling and Blake Bextine., (2017). Precision Farming Increases
Crop Yields https://www.scientificamerican.com/article/precision-
Furthermore, it is observed that linear SVM exhibits farming/ (Accessed on 8/4/2019).
improved results for two class system as the data is linearly [3] TNAU Agritech (2015)
separable. However, for multi class disease detection SVM http://agritech.tnau.ac.in/banking/PDF/Tomato.pdf (Accessed on
with polynomial kernel resulted in improvement in 7/4/2021).
performance metrics. This implies that the data is highly [4] Ghosal, S., Blystone, D., Singh, A.K., Ganapathysubramanian, B.,
Singh, A. and Sarkar, S., “An explainable deep machine vision
nonlinear. Polynomial kernel also requires less computation framework for plant stress phenotyping”. Proceedings of the National
time when compared to RBF. Academy of Sciences, 115(18),2018, pp.4613-4618.
A significant improvement in all the seven-performance [5] Al Bashish, D., Braik, M. and Bani-Ahmad, S., “Detection and
classification of leaf diseases using K-means-based segmentation and
metrics is observed in CCA based feature fusion. Therefore,
 classification:, Information Technology Journal, 10(2), 2011, pp.267- analysis and One Class Classifiers”, Computers and electronics in
275. agriculture, 156, 2019, pp.96-104.
[6] Dandawate, Y. and Kokare, R., “ An automated approach for [15] Cai, J., Xiao, D., Lv, L. and Ye, Y., “An early warning model for
classification of plant diseases towards development of futuristic vegetable pests based on multidimensional data”. Computers and
Decision Support System in Indian perspective”, In 2015 International Electronics in Agriculture, 2019, 156, pp.217-226.
conference on advances in computing, communications and [16] Prajwala TM (2016). Tomato leaf disease detection
informatics (ICACCI),2015, (pp. 794-799). IEEE. https://github.com/PrajwalaTM (Accessed on 3/1/2019).
[7] Pujari, J.D., Yakkundimath, R. and Byadgi, A.S., “ Image processing [17] Hazlyna, H.N., Mashor, M.Y., Mokhtar, N.R., Salihah, A.A., Hassan,
based detection of fungal diseases in plants”, Procedia Computer R., Raof, R.A.A. and Osman, M.K., “ Comparison of acute leukemia
Science, 46,2015, pp.1802-1808. Image segmentation using HSI and RGB color space”, In 10th
[8] Fuentes, Alvaro & Lee, Yujeong & Hong, Youngki & Yoon, Sook & International Conference on Information Science, Signal Processing
Sun Park, Dong. (2016), “Characteristics of Tomato Plant Diseases - A and their Applications, (ISSPA 2010), (pp. 749-752). IEEE.
study for tomato plant disease identification”, In Proc. Int. Symp. Inf. [18] Reza, Z.N., Nuzhat, F., Mahsa, N.A. and Ali, M.H., “Detecting jute
Technol. Converg, 2016,vol. 1, pp. 226-231 plant disease using image processing and machine learning”, In 2016
[9] Sabrol, H. and Satish, K., “ Tomato plant disease classification in 3rd International Conference on Electrical Engineering and
digital images using classification tree”, In 2016 International Information Communication Technology (ICEEICT) ,2016, (pp. 1-6).
Conference on Communication and Signal Processing (ICCSP) (pp. IEEE.
1242-1246),2016, IEEE. [19] Ojala, T, Pietikäinen, M & Harwood, D, “A comparative study of
[10] Sun, G., Jia, X. and Geng, T., “Plant diseases recognition based on texture measures with classification based on featured distributions”,
image processing technology” Journal of Electrical and Computer Pattern Recognition, 1996, vol. 29, no. 1, pp. 51-59.
Engineering, 2018. [20] Sun, Q.S., Zeng, S.G., Heng, P.A. and Xia, D.S., “Feature fusion
[11] Golhani, K., Balasundram, S.K., Vadamalai, G. and Pradhan, B., “A method based on canonical correlation analysis and handwritten
review of neural networks in plant disease detection using character recognition”, In ICARCV 8th Control, Automation, Robotics
hyperspectral data”, Information Processing in Agriculture, 2018,5(3), and Vision Conference, 2004. (Vol. 2, pp. 1547-1552). IEEE.
pp.354-371. [21] Sasikala, S., M. Bharathi, M. Ezhilarasi, M. Ramasubba Reddy, and S.
[12] Pukkela, P. and Borra, S., “Machine Learning Based Plant Leaf Arunkumar. “Fusion of MLO and CC view binary patterns to improve
Disease Detection and Severity Assessment Techniques: State-of-the- the performance of breast cancer diagnosis ” ,Current Medical Imaging
Art”, In Classification in BioApps,2018, (pp. 199-226). Springer, 14, no. 4 (2018): 651-658.
Cham. [22] Haghighat, M., Abdel-Mottaleb, M. and Alhalabi, W., (2016), “Fully
[13] Owomugisha, G., Melchert, F., Mwebaze, E., Quinn, J.A. and Biehl, automatic face normalization and single sample face recognition in
M., “Machine Learning for diagnosis of disease in plants using spectral unconstrained environments”, Expert Systems with Applications, 47,
data.” In Proceedings on the International Conference on Artificial pp.23-34.
Intelligence (ICAI) ,2018, (pp. 9-15).
[14] Pantazi, X.E., Moshou, D. and Tamouridou, A.A., “Automated leaf
disease detection in different crop species through image features