Diagnosis of Leaf Surface Disease Using Two Datasets of Tomato and Rice Obtained from Image Processing Techniques

A method based on image processing techniques and pattern identifying methods to diagnose apple leaf disease was presented in a way that, in the preprocessing stage, first the color change structure was designed for RGB input (red, green and blue) and then the RGB model was converted to HSI (color, saturation and intensity), YUV and gray models [10]. In the segmentation stage, the background is removed based on a certain threshold value and then the image was segmented by the disease point with the region growing algorithm. In feature extraction stage, 38 texture, shape and color features was extracted, which a combination of genetic algorithm and feature selection based on correlation to reduce their dimensions were applied. The classification accuracy for this method has been evaluated to be 0.90.

Sunflower leaf disease was diagnosed and classified. In this method, after preprocessing operation, the leaf segmentation was done using the Particle swarm optimization (PSO) algorithm in order to optimize the clustering process [11]. The average accuracy considered for the classification was 0.98. The reduction of speed and the need for the faster segmentation algorithm were the major problems related to this method.

In another study [12] Artificial bee colony (ABC) has been used to identify and classify diseases in grape leaves using the colony characteristics of artificial bees. After pre-processing and removing the noise of the leaves, the characteristics of texture, shape and color were extracted. Finally, the features reduced by ABC are given to SVM in the classification stage. The accuracy of the proposed method was predicted to be 0.93. Lack of proper segmentation process was found to be one of the problems of this method.

In another study[13], rice plant diseases were diagnosed and classified using convolutional neural network. In this method, a large-scale architecture such as VGG16 and Incep-tionV3 was set up to identify diseases and pests. There was also a small two-step architecture for working with mobile devices. The accuracy of the proposed method was predicted to be 0.93.

Rice disease was specifically addressed. Pictures of disease symptoms on leaves and stems are taken from rice fields [14]. A total of 120 images of sick rice plants have been collected from real farm conditions, according to the decision tree classification. The classification accuracy for the decision tree was 0.57 and Random Forest 0.76.

The improved method of randomized forest classification provided a multi-class method for the problem of rice disease classification [15]. This method was a combination of random forest machine learning algorithm and feature evaluation method and a sample filter. This approach aimed to improve the performance of the Random Forest algorithm with the accuracy of 0.97.

Another method [16] was to classify five types of tomato diseases that was done using the characteristics of shape, texture and color. The classification accuracy for this method was 0.97.

The study [17] aimed to address the issue of fake positive rates and class imbalances by implementing a filtering framework for plant tomato diseases and pest control. This was done using the CNN classifier. The classification accuracy for this algorithm was 96%.

In addition, the authors [18] proposed a method for detecting tomato leaves using the VGG16 and transfer learning architecture, which has been found to have 89% accuracy.

In the proposed method, after pre-processing, improved k-means segmentation is performed. In the feature extraction stage, complete features such as geometric, textural and demarcation features are extracted from the images. Then, to avoid feature saturation, an optimized genetic algorithm will be used to select the optimal features. Finally, the KNN classifier classifies the features.

4.1 Pre-processing

In the preprocessing step, the color space of RGB (Red Green blue) is converted to HIS (Hue, Saturation Intensity) space. The goal of achieving color space is to find an algorithm that amplifies disease symptoms in high-resolution images. The ultimate accuracy of the algorithm in diagnosing and classifying the disease largely depends on the selected color space. In this method, in addition to RGB color space, HSI color components were also examined to determine the best color component that can highlight the diseased parts in the best way. Components H, S, and I indicate color difference, saturation, and color, respectively.

4.2 Improved K-means segmentation

This method, despite simplicity, is a basic method for many other clustering methods (such as fuzzy clustering). This method is an exclusive and flat method. Different styles and methods have been expressed for this algorithm.

In this section, the improved k-means algorithm is used for segmentation. This method is known as Kmeans segmentation with Neutrosophic logic used in the study [21]. The clustering process is described as follows:

• At this point, a pixel called P is selected with the least unknown value so that it is the closest to the local average, which has not been assigned to any cluster so far.
• In this step, the same cluster is assigned to a pixel (P) and to its neighbor (N) whose correct difference is less than or equal to the predefined threshold value or the same as | T (P) - T (N) | ≤ Th.
• Steps one and two are repeated until each pixel is assigned to the C cluster.

The pseudo-code for Neutrosophic Cluster Formation is shown in Figure 3.

3.png

Figure 3. Pseudo-code for cluster filtration [21]

After the refining step, each pair of clusters are merged into a single cluster based on the difference in size and correct values. This makes smaller clusters to be merged into larger clusters, but not vice versa. The pseudo-analysis code of this clustering is shown in Figure 4.

4.png

Figure 4. Pseudo code for segmentation [21]

4.3 Extract features

4.3.1 Geometric features

Geometric features extracted from the image after improved K-means segmentation include the following. It should be noted that in all relations M and N, the dimensions of the image, x and y are also related to the coordinates of each pixel and also I, is the segmented image [22].

Area $=\sum_{x=1}^{M} \sum_{y=1}^{N} \mathrm{I}_{x, y}$        (1)

Perimeter $=2 \times($ length $+$ width $)$      (2)

Majoraxis $=\mathrm{x}+\mathrm{y}$        (3)

In above equation axis = (x, 0) − (−x, 0) and d length of major axis = 2x.

Minoraxislength $=\left((x+y)^{2}+F^{2}\right)^{\frac{1}{2}}$     (4)

Semiaxislength $=c=\sqrt{x-y}$       (5)

Circularity $=2 \times \frac{22}{7}\left(\frac{\sum_{x=1}^{\mathrm{M}} \sum_{y=1}^{\mathrm{N}} \mathrm{I}_{\mathrm{x}, \mathrm{y}}}{(\text { length }+\text { width })^{2}}\right)$                  (6)

Solidity $=\frac{\text { Area }}{\xi}$         (7)

In Eq. (8), ξ denotes the convex area.

mean $=\sum_{(x, y)=1}^{M N} I_{x, y}$     (8)

harmonic mean $=\frac{n}{\sum_{i=1}^{M N}\left(\frac{1}{X_{i}}\right)}$       (9)

Smooth $=1-\frac{1}{1+\sum_{\mathrm{x}, \mathrm{y}} \mathrm{I}_{\mathrm{x}, \mathrm{y}}}$     (10)

Energy $=\sum_{(x, y)=1}^{M N}\left(E_{x, y}\right)^{2}, E_{x, y}=\frac{P_{x, y}}{M N}$        (11)

Px,y which is an element of the GLCM matrix is also gray level co-occurrence matrices (GLCM).

Kurtosis $=\sum_{i=1}^{M N} \frac{\frac{\left(X_{i}-\hat{X}\right)}{M N}}{\sigma^{4}}$           (12)

$\mathrm{Xi} \in(\mathrm{I}(\mathrm{x}, \mathrm{y}))$ and $\hat{X}$ is $\mathrm{I}(\mathrm{x}, \mathrm{y})$ mean.        (13)

$\sigma=\sqrt{\frac{1}{M N}\left(\frac{\left(X_{i}-\hat{X}\right)^{2}}{M N}\right)}$       (14)

Full names of authors are required. The middle name can be abbreviated.

4.3.2 Harlick GLCM (HGLCM) features

In This section, after segmentation, the method of extracting of texture-based statistical features is applied in the proposed method. HGLCM properties are calculated from a statistical distribution which is a combination of the appropriate intensity of each image at specified positions. The classification of images depends on the intensity of the pixels and these statistics. HGLCM is typically used to analyze the second-order probability status of an image, of which 14 features are used in the proposed method [23].

4.3.3 Maximally stable extremal regions feature (MSER)

The MSER descriptor being applied for extracting point properties, is used in the proposed method to distinguish boundary areas from post-segmentation images [24].

In MSER, the threshold is considered as relation (15):

$f_{\theta}(x, y)=f(z)=\left\{\begin{array}{c}1, I_{z} \geq 0 \\ 0 \text { otherwise }\end{array}\right.$       (15)

Thus f (z) ∈ (x, y) defines the value of the threshold, which represents values greater than the white threshold and smaller than the black threshold.

Distinguishing of the local extrema edges from the gray-level highly contrasted images; the mapping of the upper layers set would be determined as:

$\chi_{\lambda}(V)=\left\{u \in \mathrm{R}^{2}, \mathrm{v}(u) \geq \lambda\right.$        (16)

$V(\mathrm{u})=\sup \lambda \in \mathrm{R}, \mathrm{u} \in \chi_{\lambda}(V)$        (17)

This is while in above relationships V (u) = R²⟶R at λ level. extracts spots and local features in the image. Eq. (18) is defined for the stability of these properties.

$\varphi=q(i+b)-q(i-d) / q(i)$      (18)

Φ is also defined as follows:

$\varphi=R_{i}^{a}=\frac{\left(\left|R_{i}^{a-\Delta}\right|-\left|R_{i}^{a+\Delta}\right|\right)}{R_{i}^{a}}$       (19)

In the above relation, $R_{i}^{a}$ a represents the interconnected region and R_i represents the local minimum. $R_{i}^{a-\Delta}$ and $R_{i}^{a+\Delta}$ are the local area up and down, respectively. ∆ and R are the extremum regions used to control the MSER parameters. Moreover, change in the amount of intensity in different areas is defined by Eq. (20):

$\frac{|R(+\Delta)-R|}{|R|}$    (20)

Finally, the property vector obtained from this part is N * 64.

$\chi_{\lambda}(V)=\left\{u \in \mathrm{R}^{2}, \mathrm{v}(u) \geq \lambda\right.$    (21)

4.3.4 LBP features

The original LBP operator has been introduced as a powerful descriptor for image texture. This operator generates a binary number for each pixel according to the 3*3 labels of vicinity pixel. Labels are obtained from thresholding the amount of neighboring pixels with the center pixel value in a way that, for pixels with a value greater than or equal to the value of the center pixel, label 1 is placed, and for pixels with values smaller than the value of the center pixel, label 0 is placed. then these labels are placed next to each other in rotation to form an 8-bit number as can be seen in Figure 5.

5.png

Figure 5. LBP method

4.3.5 Gabor features

The methods presented so far differ from each other in how the features are extracted. They also vary in the type of classification system used, and none of the methods has performed a complete classification. Finally, the use of combined orientation methods can increase accuracy. Therefore, the use of a combination of texture Gabor features can be efficient [25].

The Gabor filter used in the proposed method is created according to the size of the image. Eq. (22) describes how to calculate the Gabor filter.

$G(\mathrm{x}, \mathrm{y})=\frac{\mathrm{f}^{2}}{\pi \gamma \mu} \exp \left(-\frac{\mathrm{x}^{\prime 2}+\mathrm{y}^{\prime 2}}{2 \delta^{2}}\right) \exp \left(\mathrm{j} 2 \pi \mathrm{fx}^{\prime}+\varphi\right)$

$\mathrm{x}^{\prime}=x \cos \theta+y \sin \theta$

$\mathrm{y}^{\prime}=-x \sin \theta+y \cos \theta$                 (22)

In Eq. (20), the Gabor filter is normalized in the frequency domain, which in the above equations f is the sine factor frequency. θ also shows the direction of the normal corrugated lines of the Gabor function to the parallel corrugated lines of the Gabor function. Φ is the phase offset, and δ is equal to the Gaussian cap standard deviation. γ is the spatial visibility ratio that determines the elliptical support of the Gabor function [25].

4.4 Features select

Feature selection is a process that improves the performance of the data analyzing process by selecting effective attributes in a large data set. some of the most important goals of feature selection are the followings:

• Reduce the size and volume of data, increase the speed of operations

• Improve the accuracy of data analyzing and machine learning algorithms and better understanding of the results. Genetic Algorithm (GA) is a search technique in computer science suitable for finding the approximate solutions for optimizing and search problems. Genetic algorithms is a special type of evolutionary algorithms that use concepts of biology science such as inheritance and mutation [7].

Genetic algorithms are usually implemented as a computer simulator in which the population of an abstract sample (chromosomes) of the alternative solutions of optimizing a problem leads to a better solution. Solutions have traditionally been in the form of strings of 0 and 1, but today they are implemented in other ways as well. The hypothesis begins with a completely unique random population and continues through generations. In each generation, the capacity of the entire population is assessed. Several individuals are randomly selected from the current generation (based on competencies) and modified to form a new generation (deducted or recombined), and in the next iteration of the algorithm Becomes the current generation [7].

Solutions are generally shown in pairs 0 and 1, but there are other display methods. Evolution starts from a completely random set of beings and is repeated in the next generations. In every generation, the most suitable ones are selected, but not the best ones.

A solution for a certain problem is represented by a list of parameters called chromosomes or genomes. Chromosomes are generally represented as a simple string of data, although other types of data structures can also be applied. Initially, several attributes are randomly generated to create the first generation. In each generation, every trait is evaluated and the fitness value is measured by the fitness function [7].

fitness $=\alpha E R(K)+(1-\alpha) \frac{|S|}{|C|}$      (23)

In the aforementioned relation, αER (K) is the classification error rate calculated by the classifier (in this method KNN is considered) for K of the property, | C | In fact, is defined a all the properties in the data set, | S | represents the selected properties and α are in fact the same parameter used to control the classification error, which is 0.99 here.

The next step is to create the second generation of society which is based on selection processes and the production is according to the selected traits with genetic operators: connecting chromosomes to each other and changing [7]. For each person, a pair of parent is selected. The choices are made so that the most appropriate elements are selected so that even the weakest elements have a chance to be selected so as the local answer is avoided. An example of a crossover operation is illustrated in Figure 6.

6.png

Figure 6. Crossover operation in the proposed method [7]

The algorithm applied in the proposed method is shown in Figure 7.

7.png

Figure 7. Crossover operation in the proposed method [3]

However, in order to improve the classification accuracy, feature reduction method which was described above and known as the modified genetic algorithm, is presented in the study [3] and has been applied in the proposed method.

4.5 Features normalization

Given that the large amounts in the feature extraction stage may affect the accuracy of the classification, data normalization is applied:

$\hat{x}=\frac{x_{i}-x_{\min }}{x_{\max }\,\,-x_{\min }}$       (24)

In the above relation, x represents the new value in the attribute. x_i describes the initial value of the property. xmin and xmax are considered the smallest and largest values.