Leaf Diseases Detection of Medicinal Plants Based on Support Vector Machine Classification Algorithm

On earth, plants play the most important part. Every organ of a plant plays a vital role in the ecological field as well as the medicinal field. But on the whole earth there are several species of plants are available. The different species of plants have different diseases. Therefore, it is required to identify the plants as well as their diseases correctly. It is difficult and also time consuming to identify the plants and their diseases manually. In this research an automatic disease detection system of plant is proposed. High-quality leaf images are used for training and testing. For detecting the healthy area and diseased area in a leaf, region-based and color-based region thresholding techniques are used. For feature selection Histogram Oriented Gradient (HOG) and Local Binary Pattern (LBP) method were applied. Finally, for classification two-class and multi-class Support Vector Machine (SVM) were used. It is found that both feature selection processes with SVM give 99% accuracy. An user oriented graphical user interface is created for understanding the automated system.


INTRODUCTION
Among all the living things in this world, plants are one of the most important things. Horticulture, a part of agriculture, is used to deal with gardening of useful plants. Horticulture plants like flowers, fruits, vegetables are used to fulfil daily life necessities. So it is necessary to have complete knowledge about these plants as well as the diseases affecting these plants. Detection of plant diseases in the early stages can help to prevent leaf losses and poor production. Detecting of plant diseases manually is a time-consuming process. With the developing of technology, today is quite easy to monitor and detect the plant diseases. Computer Vision technology is much more suitable and gives more accurate results in these cases.
In this paper, an automatic plant diseases detection system is created. This plant disease detection system is based on image segmentation using region-based thresholding and color based thresholding techniques. These techniques are used to detect the areas of disease-affected and healthy leaf region. Finally, a classification algorithm is used to classify all the diseased and healthy leaf categories.
Section II describes the literatures survey. Section III presented proposed methodology for detecting diseases. Experimental results and discussion are in Section IV. The conclusions of the research is given in Section V. Section VI has the reference works that are consulted.

LITERATURES SURVEY
Different researchers have done researches to detect various plant diseases based on various techniques. The researchers used spectroscopic method and imaging techniques to detect the diseases. With the development of artificial intelligence, machine learning and deep learning based on computer vision study gives more satisfying results. These techniques are widely used to detect plant and leaf diseases automatically.
In a research paper, the authors proposed a realtime grape disease detection process based on an improved convolution neural network [1]. First, they used digital image processing techniques to create a grape leaf disease database. Second, they apply a faster-RCNN and a deep learningbased faster DR-IACNN model to detect the disease. They have also shown that the faster DR-IACNN model achieved an 81.1% precision rate and detection speed reaches up to 15.01 FPS.
In another research paper, the authors used a convolution neural network with an image processing technique to detect plant diseases [2]. They mostly used common plants that are available all over the world specifically in Iraq. They classified 15 classes, among them 12 classes were diseased and 3 were healthy. They obtained more than 98% accuracy for both the training and testing dataset.
For all nations, economic growth mostly depends on agricultural products. But due to the various plants' diseases,the growth and productivity are decreased from time to time. In the early stage of diagnosis of these diseases can help to prevent the spreading and helps to better productivity. The authors survey various kinds of classification models based on image processing and machine learning techniques to detectand recognized the diseases of agricultural fields. They also show the proposed method and its accuracy for detailed analysis [3].
In another research paper, the authors also survey a detailed analysis of different classification techniques to detect the diseases of various agricultural plants [4].
To detect the apple leaf disease authors proposed a deep neural base improved convolution neural network model. They create an apple leaf diseases dataset based on the laboratory and real-life complex images. Then to detect the diseases they introduce GoogleNet interception and Rainbow concatenation with a deep neural network. They used this model to train up to five apple leaf diseases. The detection performance is 78.80% with high detection speed is 23.13 FPS [5].
To identify the plant leaf diseases the author proposed a model based on image preprocessing and machine learning approaches. They use different image processing techniques to pre-process the image and the GLCM feature extraction method to extract the features of the leaf. Finally, apply the K Nearest Neighbour classifier to detect the diseases of the leaf. The proposed implementation predicts the 98% accuracy for disease detection [6].
To identify automatic crop diseases the authors survey 19 studies based on Convolution Neural Network (CNN). In this study, the authors describe diseases profile, implementation techniques of CNN, and analysis of the performance of the techniques. Finally, they provide the guidelines for an improved CNN for future research [7]. In another paper, the authors' overview different plant leaf disease detection based on different machine learning classification techniques. They describe different algorithms and their performance [8].

PROPOSED METHODOLOGY
In this experiment to detect the leaf diseases from an input image, the following steps are involved: 1) Pre-Processing, 2) Segmentation, 3) Feature Selection and Extraction, and 4) Classification.

Database Details
The complete set of images consists of 12 plants they arenamed as "Mango", "Arjun", "Saptaporni (AlstoniaScholaris)", "Guava", "Bael", "Jamun", "Jatropha", "Pongamia Pinnata", "Basil", "Pomegranate", "Lemon", and "Chinar". All the plants have their own economical and environmental values. The entire database is divided between two main classes "Healthy" and "Diseased" and 22 subclasses. Figure 1 and 2 shows a detailed view of the database images. This database has a total of 4503 images, among them, 2278 images are for healthy leaf images and 2225 are for diseased images. All the images are collected from the Shri Mata Vaishno Devi University, Katra [9][10]. Table 1 shows the details about the experimental database.

Image pre-processing
Image pre-processing [11] means processing the images before the computation processes. This process commonly involves eliminating the background from the objects, reduces the background noises, resize the input images, enhance the brightness and contrast level. In this experiment, the images in the leaf database are very high in quality [dimension 6000 × 4000] [dimensionɘ×Ɛ All the images are in RGB colour and taken with a high-quality digital camera. Therefore, for experimental usage, here the images reduce to dimensions 600 × 400ɘ×Ɛ for display purposes and dimensions 150 × 100 for feature extraction purposes. Fig. 3 shows the sample leaf example.

Segmentation
Image segmentation [12][13] is one of the most important partsof image processing. These criteria used to divide the input image into the same type of area and extract the region of interest. Threshold [14] segmentation is a common and simple image segmentation technique. It is a region-based segmentation technique. This method can be divided into two categories 1) Global thresholding [15] and 2) Local thresholding [16]. In the Global thresholding technique, the input image is divided into two main regions 1) background and 2) the region of interest. The global thresholding method uses a single thresholding level to segment the input image. The Local thresholding technique uses multiple thresholding levels to divide the input region and background.
In this experiment, the global thresholding technique is used. For segmentation, a global optimum threshold value is selected. In the leaf image database, the background of the input images is much darker therefore the segmentation effect is more effective. This procedure is faster and the calculation is much simple. Figure 4 shows the segmentation result of the input image.

FEATURE SELECTION AND EXTRACTION
Feature selection is a technique in image processing. This technique helps to reduce a large set of features by selecting the best features from the original dataset. It abandons the redundant data from the original dataset. In computer vision object classification or object recognition is one of the most popular subjects. The aim of the object classification is to extract features from the input image and use a proper classifier to classify the feature class label Histogram Oriented Gradient (HOG) and Linear Binary Pattern (LBP) are the two efficient gradient-based feature selection techniques [17]. Their performance is much better than any other feature selection technique.

Histogram Oriented Gradient (HOG)
It is a gradient-based feature descriptor [18]. It calculates the existing gradient and orientation of the input image. Basically, this method broke the input image into a small piece of regions, then calculate the vertical and horizontal gradients of those blocks locally. Finally calculate the magnitude and orientations of those gradients.

Linear Binary Pattern (LBP)
It is a gradient-based effective texture descriptor [18]. This method divided the input image pixels into3 × 3 matrices. Then it considers the central value of 3 × 3 matrices as a threshold value and set a binary value based on the threshold value. This process is continued for all pixel matrix and finally, it converts all the binary numbers into decimal and represents the input image in a better way.

CLASSIFICATION
In machine learning classification [19][20] is a process, used to understanding, recognizing, and grouping the same type of data based on their categories. The classification algorithms are used a pre-trained training dataset to predict the category of an unknown sample whose data fall into the predetermined categories.
For classification of an image support vector machine is one of the most popular and best classification methods. It is a supervised learning method. This algorithm is divided the whole data space with a maximum margin to predict the class of the unknown data sample. Fig. 6 shows the classification model of this experiment.

EXPERIMENTAL RESULT AND DISCUSSIONS
The experimental database has 10 leaf classes and two main classes 'Healthy' and 'Diseased'. Figs. 7a and 7b show the leaf disease detection and classification process. Figs. 8a and 8b show the performance details of each main class and the subclasses using the HOG feature selection method. Figs. 9a and 9 b show the performance details of each main class and the subclasses using the LBP feature selection method.

CONCLUSIONS
In this paper, an automated leaf detection system is proposed. Here, region-based thresholding technique is used to detect the healthy and diseased leaf image. Again, to detect the particular diseased area of a diseased leaf the color-based region thresholding method is used. For feature selection from the input images both HOG and LBP feature selection technique is used. To classify the category healthy and diseased leaf with subclasses, leaf name, and disease name the two-class, and multi-class SVM classifier is used. A detailed performance analysis was done between main classes and subclasses by using different classifiers. Finally, a graphical user interface is created for all users.

CONSENT
It is not applicable.

ETHICAL APPROVAL
It is not applicable.