Local Directional Patterns for Plant Leaf Disease

2023 International Conference on Advances in Electronics, Control and Communication Systems (ICAECCS) | 978-1-6654-6310-2/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICAECCS56710.2023.10104754

Detection
Amine Mezenner, Hassiba Nemmour and Youcef Chibani Adel Hafiane
LISIC, Faculty of Electrical Engineering PRISME Laboratory, INSA CVL
USTHB Univ Orléans
Algiers, Algeria Bourges, France
{amezenner,hnemmour,ychibani}@usthb.dz [email protected]

Abstract—Plant leaf disease detection is an attractive research art reports the use of all famous features of pattern recognition
issue for artificial intelligence and computer vision community, and computer vision. Specifically, the first works focused on
who aims to develop intelligent systems that make an automatic the use of color, texture, and shape features . As Color features
detection of plant leaf diseases. In the recent past years, deep
learning techniques were extensively used since they allow one can note the use of the color histograms generated from the
developing end-to-end systems. However, for several case studies RGB or the HSB (Hue, Saturation, Brightness) representations,
detection scores still need improvement. Presently, we propose as well as the CieLAB and YCbCr features [3]. Commonly,
a new descriptor that is based on local Directional Patterns to pertinent texture and shape features are generated by using
perform feature generation from plant leaves. This descriptor local binary patterns and the histogram of oriented gradients.
is associated with SVM classifier to develop the full detection
system. Experiments are conducted by considering three crop Since they achieve satisfactory results on several datasets [4],
species that are Tomato, Potato, and Bell pepper diseases. [5], [6]. Furthermore, to achieve the classification step, ma-
The proposed LDP features are evaluated comparatively chine learning techniques, especially, artificial neural networks
to convolutional neural networks features as well as to the and Support Vector Machines (SVM) are widely used [5],
histogram of oriented gradients. The results obtained highlight [7]. Recently, the success of artificial intelligence, promoted
the effectiveness of the proposed system which outperforms the
LeNet-5 convolutional neural network by 3% in the over- all the use of Convolutional Neural Networks (CNN), which
accuracy. associate feature generation and classification in the same
process. Precisely, in a CNN, convolutional layers ensure the
Keywords: Convolutional Neural Networks. Local Directional feature verification, while the fully connected layers achieve
Patterns, Plant Disease Detection, Support Vector Machines . the classification task. In this respect, CNN can be used
either as a feature generator or as an end-to-end classification
I. I NTRODUCTION
system. The state of the art shows deep learning models
With the increase of population and climatic disturbances, as the best systems for solving plant leaf disease detection
improving agricultural crops is a major sector for national and classification [8]. Nevertheless, performance obtained for
economies. In this respect, various strategies are adopted to various public datasets, reveal that handcrafted features still
predict favorable conditions for plant disease, such as the give promising performance.
weather effects on developing infections, as well as the mon- Presently, we investigate the usefulness of Local Directional
itoring of plant leaves for early detection of viruses. In fact, Patterns (LDP) as a new image leaf descriptor. This is a
the plant health can be inspected in the leaves color, edges and LBP likewise descriptor that is calculated on image edges
textures. Therefore, automatic plant disease detection systems extracted by using Kirsch detector. LDP is associated with
are based on leaf image analysis. Previously, this task was SVM classifier to develop the plant leaf disease detection.
carried out by agronomists who decide if a plant is infected The performance assessment is carried out comparatively to
or healthy. Currently, Thanks to advances in computer vision CNN that is used as an end-to-end system in a first time,
and artificial intelligence powerful systems were developed to and then utilized as feature generator in associa-tion with
make an automatic disease detection by analyzing plant leaf SVM. Besides, the Histogram of oriented gradients is used for
images [1]. As most of computer vision systems, Plant leaf comparison aim. The results obtained on a public dataset high-
disease detection systems are composed of three main steps, light the effectiveness of the proposed features. The remaining
which are preprocessing, feature generation, and classification of this paper is organized as follow: Section 2 presents the
[2]. Preprocessing operations aim to enhance the visual quality proposed System. Section 3 gives experimental results while
of images and facilitate the extraction of useful information, the last section gives the main conclusions of this work.
such as the background removing, resizing, and image filter-
ing. The feature generation is the most important step within II. P ROPOSED S YSTEM
the detection system since it has to extract information that dis- As already mentioned, a plant leaf disease detection system
tinguishes healthy leaves from infected leaves. The state of the is composed of the steps reported in Fig.1. Presently, the

978-1-6654-6310-2/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: Chaitanya Bharathi Institute of Tech - HYDERABAD. Downloaded on October 16,2024 at 04:02:10 UTC from IEEE Xplore. Restrictions apply.
 preprocessing are limited to dimension reduction, where the values to calculate the histogram of occurrences that is the
image size was reduced to 40×40 pixels, in order to minimize LDP descriptor. Fig 3 illustrates the LDP calculation for
the time consumption of the system development. For feature various values of K, which is a user-defined parameter.
generation, we propose the Local Directional Patterns, as
well as features extracted from a customized LeNeT-5 model.
The classification is carried out by SVM.

Fig. 3: LDP calculation for a given pixel.

2) Histogram of Oriented Gradients (HOG) : HOG is a

Fig. 1: Steps composing a plant leaf disease detection system. famous computer vision descriptor that was introduced for
human detection which showed outstanding performance in
A. Feature generation several pattern recognition and classification tasks [10], [11]. It
is calculated on a dense grid of evenly spaced cells by counting
In this section, we present the Local Directional Patterns occurrences of gradient orientations in localized portions of an
(LDP) that is proposed to generate plant leaf features. This image. For each cell, HOG features are calculated according
descriptor is compared to features extracted from a customized to the following steps:
Convolutional neural network, as well as to the Histogram of
oriented gradients. 1) Compute the horizontal and vertical gradients for each
1) Local Directional Patterns (LDP) : LDP is an edge pixel such that:
descriptor that considers edge responses in different directions
around each pixel [8]. The first step in the LDP calculation Gx (x, y) = I(x, y + 1) − I(x, y − 1) (1)
consists of applying Kirsch edge detector, in order to highlight
the image edges according to 8 directions. As shown in Fig.2,
there are eight filtering masks that characterize all possible
Gy (x, y) = I(x − 1, y) − I(x + 1, y) (2)
edge directions within images.

2) Evaluate the gradient magnitude and phase as follows:

q
M agnitude(x, y) = G2x + G2y (3)

Gy
Angle(x, y) = |tan−1 )| (4)
Gx

3) Accumulate magnitudes according to their orientations

to get the oriented gradient histogram of the cell.
4) Finally, concatenate histograms obtained for all cells to
form the image HOG descriptor.
Fig. 2: Filters utilized in the Kirsch edge detector. 3) Convolutional Neural Network-based features: To ex-
tract CNN features, we propose a customized LeNet-5 model
Note that for each direction the presence of a corner or an that is composed of 3 convolutional layers, associated with
edge corresponds to high values. The eight resulting images maxpooling and bach normalization steps. A flattening layer
are grouped into a single maxima image that takes the maximal is used to adapt convolved maps to the prediction bloc that
value among the eight values of each pixel. In the second is composed of three fully connected layers. The network
step, the LDP histogram is assessed by considering the K architecture was experimentally tuned by considering the CNN
most prominent directions [9]. Precisely, considering the pixel as an end-to-end plant leaf disease detection system. After
neighborhood in the maxima image, the highest K values take the training stage, the CNN is used as a feature generator
1 in the LDP code while remaining neighbors take 0. Then, by selecting outputs of the flatten layer or outputs of a fully
similarly to LBP, binary codes are converted into decimal connected layer as indicated in Fig.4.

Authorized licensed use limited to: Chaitanya Bharathi Institute of Tech - HYDERABAD. Downloaded on October 16,2024 at 04:02:10 UTC from IEEE Xplore. Restrictions apply.
 Fig. 4: Proposed CNN for deep features generation.

B. Classification
Fig. 5: Data distribution in the adopted dataset.
The detection task is achieved by a SVM that aims to
separate plant leaf images into healthy or infected classes.
This classifier is considered as the best off-the-shelf classifier
for various applications. Regarding the plant disease detection
and classification, SVM showed an outstanding performance
for soybeans and bean disease detection by 96.25% and 100%
in the overall accuracy [12], [13]. The training process consists
of finding an optimal hyper plane that maximizes the margin
between two classes [14]. Then, data are assigned to classes Fig. 6: Samples from the adopted dataset.
by using the following decision function:
From this dataset, 2/3 of samples were used in the training
stage, while the remaining samples were used for performance
Sν
X evaluation. The experimental design focuses on evaluating
F (p) = sign( αi K(pi , p) + b) (5) the effectiveness of the proposed features that are LDP and
i=1
CNN-based features. For comparison aims, experiments are
conducted on the end-to-end CNN and the HOG+SVM system
αi : is Lagrange multiplier of the sample Sν, pi is the number as well. First, experiments are carried out according to disease-
of support vectors, and b is a bias. The adopted kernel specific detection which aims to detect a single disease. Then,
function K is the radial basis function that is calculated as: a global detection test is performed in which, all infected
leaves compose the disease class.
In both experiments the detection task is considered as a
−(pi −p)2
binary classification problem where healthy class is confronted
RBF (pi , p) = e 2σ 2 (6)
to all disease classes. In the first step, several run passes
were executed to find the best CNN architecture. Experiments
σ: user defined parameter. based on the end-to-end CNN accuracy conducted to the
configuration reported in table 1:
TABLE I: LeNet5 customized CNN architecture
Layer # Filters /Nodes Filter size Padding Activation function
III. E XPERIMENTAL R ESULTS Conv1 32 3x3 Same ReLu
Batch Norm
Max Pooling
Conv2 64 3x3 Same ReLu
Performance assessment of the proposed system is carried Batch Norm
Max Pooling
out on the PlantVillage dataset which contains 54303 healthy Conv3 128 3x3 Same ReLu
and unhealthy leaf images divided into 38 categories by Batch Norm
Max Pooling
species and diseases. Presently, three species namely, Tomato, Flatten
Potato and Pepper that correspond to strategic agriculture in Dense 1 120 ReLu
Dense 2 120 ReLu
Algeria, are studied. The selected corpus is composed of 15 Dense 3 84 ReLu
classes including healthy classes and various infections of Dense 4 2 Softmax
the considered species. Fig. 5 reports the number of samples
per class, while Fig. 6 shows a sample from each class of Results obtained by using this CNN are reported in table
interest. 2 with those derived by the HOG-SVM system. Note that
HOG is calculated by dividing images into 4×4 cells, while

Authorized licensed use limited to: Chaitanya Bharathi Institute of Tech - HYDERABAD. Downloaded on October 16,2024 at 04:02:10 UTC from IEEE Xplore. Restrictions apply.
 TABLE III: Disease detection Accuracy (%) obtained by using CNN
respective histograms are concatenated to form the full image
and LDP features
descriptor. As can be seen, the CNN presents higher accuracies
Detection task CNN F+SVM FC1+SVM FC2+SVM FC3+SVM LDP+SVM
than HOG-SVM system. Precisely, the detection score exceeds Potato
healthy/ Late blight 97.92 92,19 97,14 97,14 86,72 100
97% in most cases. Therefore, the CNN is used as a baseline healthy/ Early blight 97.65 97,91 99,73 99,48 86,72 100
Pepper bell
to evaluated LDP features. healthy/ Bacterial spot 99,27 98,42 99,27 98,79 98,79 99,84
Tomato
TABLE II: Detection Accuracy (%) obtained for HOG-SVM and healthy/ Target Spot
healthy/ Mosaic Virus
88,88
97,4
99,7
99,85
99,9
99,69
99,5
99,7
99,6
99,7
100
100
CNN systems healthy/ Yellow leaf curl virus 99,87 100 100 100 100 100
healthy/ Bacterial Spot 100 99,92 99,92 100 100 100
healthy/ Early Blight 82,75 99,65 99,54 99,65 99,31 100
Detection Task HOG+SVM CNN healthy/ Late Blight 99,83 100 99,91 99,83 99,83 100
Potato healthy /Potato Early blight 85,71 97.65 healthy/ Leaf Mold 98,94 99,76 99,76 99,76 98,94 100
Potato healthy/Potato Late blight 85.71 97.92 healthy/ Spetoria leaf Spot 99,91 99,82 100 99,91 99,91 100
healthy/ Two spotted spider mite 88,88 96,42 99,27 99,36 99,63 100
Pepper bell healthy/Pepper bell Bacterial spot 72,31 99,27 healthy/ All diseases tomato 95.15 99,46 99,78 99,75 99,66 100
Tomato healthy/Target Spot 58,56 88,88
Tomato healthy/Mosaic Virus 80,65 97,4
Tomato healthy/Yellow leaf curl virus 86,3 99,87 IV. C ONCLUSION
Tomato healthy/Bacterial Spot 82,8 100
Tomato healthy/Early Blight 91,44 82,75 This paper proposed the use of Local Directional Patterns
Tomato healthy/Late Blight 92,5 99,83 as new leaf image descriptor for plant disease detection.
Tomato healthy/Leaf Mold 81,5 98,94
Tomato healthy/Spetoria leaf Spot 76 99,91 LDP is a robust descriptor that combines texture and edge
Tomato healthy/Two spotted spider mite 85,3 99,72 information obtained by applying Kirsch detector on images.
This combination highlights the textural information in various
In the second experiment, LDP features were associated directions. Experiments were conducted on data of three
with SVM to perform the detection task. For comparison, agricultural species extracted from the PlantVillage dataset.
we used the CNN as an end-to-end system, and as a fea- For comparison aim, we used HOG and CNN features. The ob-
ture generator. Specifically, CNN features are extracted by tained results reveal that HOG provides medium performance
considering respectively, outputs of the Flatten layer (F), the compared to other features. The end-to-end CNN as well as the
first Fully Connected (FC1) layer, the second Fully Connected system associating CNN features with SVM provide similar
layer (FC2), and the third Fully Connected layer (FC3). The accuracies that are much better than HOG results. Neverthe-
results in terms of overall detection accuracy are reported in less, the association of SVM with LDP features outperforms
table 3. For each feature, the best SVM parameters that allow all other systems, since it allows an optimal accuracy, when
an optimal training accuracy were selected. Also, for LDP the suitable K value is used. From these outcomes we infer
features the K value was experimentally fixed at 3. As can that LDP can be a typical plant disease descriptor. Further tests
be seen, for all detection tasks, the LDP provides an optimal on larger kinds of species are necessary to confirm again the
precision that reaches 100%. On the contrary, both CNN and effectiveness of this descriptor.
CNN features-SVM systems provide lower accuracies, which
highlight the effectiveness of LDP features. More precisely, R EFERENCES
with CNN features, the accuracy varies between 99.46% [1] Y. M. Oo and N. C. Htun, “Plant leaf disease detection and classifi-
cation using image processing,” International Journal of Research and
and 99.78% which is obtained when using the first Fully Engineering, vol. 5, no. 9, pp. 516–523, 2018.
connected layer that contains 120 nodes. Furthermore, it is [2] U. Shruthi, V. Nagaveni, and B. Raghavendra, “A review on machine
important to note that LDP performance is mainly controlled learning classification techniques for plant disease detection,” in 2019
5th International conference on advanced computing & communication
by the K values. As illustrated in Fig.7 which depicts the systems (ICACCS). IEEE, 2019, pp. 281–284.
Tomato disease detection precision for various K values, the [3] M. El Sghair, R. Jovanovic, and M. Tuba, “An algorithm for plant
performance increases from 94.4% for K=1 to 100 for a K=3. diseases detection based on color features,” Int J Agric Sci, vol. 2, pp.
1–6, 2017.
This finding can be explained by the fact that there is a [4] V. K. Vishnoi, K. Kumar, and B. Kumar, “Plant disease detection using
need of at least 2 orientations to detect curved edges since computational intelligence and image processing,” Journal of Plant
a single orientation cannot highlight linear information in a Diseases and Protection, vol. 128, no. 1, pp. 19–53, 2021.
[5] L. Tan, J. Lu, and H. Jiang, “Tomato leaf diseases classification based
given shape. on leaf images: a comparison between classical machine learning and
deep learning methods,” AgriEngineering, vol. 3, no. 3, pp. 542–558,
2021.
[6] P. Patil, N. Yaligar, and S. Meena, “Comparision of performance of
classifiers-svm, rf and ann in potato blight disease detection using
leaf images,” in 2017 IEEE international conference on computational
intelligence and computing research (ICCIC). IEEE, 2017, pp. 1–5.
[7] A. Cruz, Y. Ampatzidis, R. Pierro, A. Materazzi, A. Panattoni, L. De Bel-
lis, and A. Luvisi, “Detection of grapevine yellows symptoms in vitis
vinifera l. with artificial intelligence,” Computers and electronics in
agriculture, vol. 157, pp. 63–76, 2019.
[8] M. H. Saleem, J. Potgieter, and K. M. Arif, “Plant disease detection and
classification by deep learning,” Plants, vol. 8, no. 11, p. 468, 2019.
[9] T. Jabid, M. H. Kabir, and O. Chae, “Local directional pattern (ldp)–a
robust image descriptor for object recognition,” in 2010 7th IEEE inter-
national conference on advanced video and signal based surveillance.
Fig. 7: Detection results for various LDP K values. IEEE, 2010, pp. 482–487.

Authorized licensed use limited to: Chaitanya Bharathi Institute of Tech - HYDERABAD. Downloaded on October 16,2024 at 04:02:10 UTC from IEEE Xplore. Restrictions apply.
 [10] N. Dalal and B. Triggs, “Histograms of oriented gradients for human
detection,” in 2005 IEEE computer society conference on computer
vision and pattern recognition (CVPR’05), vol. 1. Ieee, 2005, pp.
886–893.
[11] B. S. Kusumo, A. Heryana, O. Mahendra, and H. F. Pardede, “Machine
learning-based for automatic detection of corn-plant diseases using im-
age processing,” in 2018 International conference on computer, control,
informatics and its applications (IC3INA). IEEE, 2018, pp. 93–97.
[12] Q. Yao, Z. Guan, Y. Zhou, J. Tang, Y. Hu, and B. Yang, “Application
of support vector machine for detecting rice diseases using shape and
color texture features,” in 2009 international conference on engineering
computation. IEEE, 2009, pp. 79–83.
[13] U. Mokhtar, N. E. Bendary, A. E. Hassenian, E. Emary, M. A. Mahmoud,
H. Hefny, and M. F. Tolba, “Svm-based detection of tomato leaves
diseases,” in Intelligent Systems’ 2014. Springer, 2015, pp. 641–652.
[14] R. D. L. Pires, D. N. Gonçalves, J. P. M. Oruê, W. E. S. Kanashiro,
J. F. Rodrigues Jr, B. B. Machado, and W. N. Gonçalves, “Local
descriptors for soybean disease recognition,” Computers and Electronics
in Agriculture, vol. 125, pp. 48–55, 2016.

Authorized licensed use limited to: Chaitanya Bharathi Institute of Tech - HYDERABAD. Downloaded on October 16,2024 at 04:02:10 UTC from IEEE Xplore. Restrictions apply.