Академический Документы
Профессиональный Документы
Культура Документы
Yogesh Dandawate.
Professor, Dept. of Electronics and Telecommunication Engineering,
Vishwakarma Institute of Information Technology,
Pune, India.
yhdandawate@gmail.com
plant diseases is a forthcoming research area as it proves helpful is made used. The results obtained using Skew divergence and
in monitoring large fields, and automatically detects the plant EPSO technique is 98%.
diseases based on the symptoms that are visible on the plant To detect the diseases automatically and to give the accurate
leaves. Therefore the objective of this paper is to develop image and fast solutions to the farmers, A. Camargo, J.S. Smith [4]
processing technique that can accurately analyse the disease of addressed the problem of developing image processing
the grape plants based on images of the leaves, since leaf is a algorithms that recognizes abnormalities in the crops. Maize
good indicator of the plant health and offer a solution to major plant images are used in the given paper as the data. The author
national issue of increasing agriculture productivity. has considered various diseases like Brown stripe downy
Decision Support Systems (DSS) for agriculture is based on the mildew, Stem borer etc. for classification. The accuracy of the
technology that can be useful for farmers and help to increase classification has been rated up to 95%. In this paper we are
the agricultural productivity. The efficiency of decision making processing leaf images based on texture, shape and colour and
in agriculture sector can be improved by reducing the get back to the farmers with solution with the help of SMS. Two
technological gap between research and farmers using Decision Algorithms are used at the processing side first for extracting
Support System (DSS). But, the current Decision Support HSI values color transformation is used. And for second
Systems (DSS) are call center based which require that the Momentum back propagation Neural Network is used.
farmers should orally give inputs regarding the health of crops Using SVM, Zambre R. [5] have precisely classified cotton
which is not effective. This can be made friendlier for the leaf diseases. Segmentation technique used in the proposed
farmers by integrating image processing with DSS. system involves enhancements of the images first, and then
In this paper, we propose a novel framework that focuses on extraction features such as texture, shape, color and boundary
the disease detection for plants. The farmers can capture the etc., Color and Texture feature extraction techniques are used
images of plant leaves using any mobile camera having to recognize the diseases. The various diseases recognized on
resolution greater than 2 mega pixels. The farmer needs to just the cotton leaf plants are classified as Bacterial blight, Leaf curl,
capture the image of the plant leaf through the mobile camera Grey mildew, Verticillium wilt, Alternaria Macro Spora,
and send the image to a central system through mobile internet. Fusarium wilt and Alternaria Leaf Spot. Efficiently cotton leaf
After acquiring images, through the proposed work, normal and diseases were analysed and classified using SVM accuracy.
abnormal leaves are classified based upon the extracted features In [6], brown spot and the leaf blast diseases of rice plant are
from leaf images. The extracted features, class and the classified. S. Phadikar used Bayes’ formula and Support Vector
percentage area of abnormality will be given as input to the Machine (SVM) for classification. Mathematically Bayes’
DSS. The proposed algorithm will automatically recognize the formula is given as,
plant based on the color, texture and shape of the leaf. The ሺ߱ Τ ݔሻ ൌ ሺ߱ ሻሺ ݔΤ߱ ሻ (1)
purpose of the proposed project and research work is to provide ݎ ሺݔሻ ൌ െ σ ܮ ሺ߱ ሻሺ ݔΤ߱ ሻ (2)
inputs for the DSS, developed for providing advice to the
farmers as when they require over mobile internet. In the first stage, based on the number of peaks in the histogram,
The rest of the paper is organized as follows: Section II diseased and the uninfected leaves are classified The success
describes the literature survey. The proposed approach is rate is 92% for uninfected leaf, 96.4% for leaf with brown spot
illustrated in section III. The pre-processing steps used in the and 84% for leaf with blast, where wrong classification may
algorithm are introduced in section IV. Section V describes the occur due to shadow effect and color distortion of aging leaves.
details regarding classification. The experimentation and In the next level the efficiency of Bayes’ classifier and SVM
results are projected in section VI and finally section VII are computed and compared. The success rate is 79.5% for
concludes the paper. Bayes’ classifier and 68.1% SVM classifier; therefore Bayes’
classifier is proved to be time efficient as compared to SVM.
II. LITERATURE SURVEY To improve agricultural product yields, Arti N. Rathod [7]
proposed various methods to identify leaf disease detection
Extensive research has been conducted to explore various using image processing.For leaf disease detection this paper
methods for automatic detection of plant diseases. The disease provides the survey of different techniques. In order to increase
can manifest in different parts of the plant such as roots, stem, the recognition rate of final classification process various neural
fruit or leaves. As stated before, this work concentrates networks & hybrid algorithms can be developed in the future
particularly on leaves. work.
P. Revathi, M. Hemalatha [3] worked on classification of
diseases in cotton leaves. Authors have considered six types of Revathi [8] proposed a novel approach of identification of
diseases in the cotton plant for classification. Based on Cotton crop diseases from RGB images has been exposed in
advanced computational techniques the significance of this this paper. The author has Proposed Enhanced PSO feature
work design is to reduce the time, cost and complexity. To selection method which adopts user features like variances,
identify the affected region of a leaf the author has used texture, color and edge to extract the features and Skew
Enhanced Particle Swarm Optimization (EPSO) for feature divergence method. Using Back propagation neural network
selection. For calculating the edge, color, texture variance for (BPN), Fuzzy and SVM classifiers the obtained features are
feature analysis of the diseased part Skew divergence technique extracted with the help of and Genetic algorithm (GA) feature
514
2016 3rd International Conference on Signal Processing and Integrated Networks (SPIN)
selection and Edge CYMK color feature. Cross validation of to capture the image of the plant leaf using mobile camera and
three classification models were assessed to test this hypothesis. send it to the DSS, without any additional inputs.
Six types of diseases such as, Root rot, Leaf Blight, Micro
Nurient,Verticilium Wilt, Bacterial Blight, Fusarium wilt have III. PROPOSED TECHNIQUE
been accurately classified to evaluate its efficiency. The Results
of accuracy is as shown in table 1. The proposed approach consists of four main steps: image
acquisition of grape leaves, extraction of grape leaves from
TABLE I . ACCURACY OF DIFFERENT METHODS background, statistical analysis and disease classification.
Sr. Figure 1 shows the block diagram of the overall system.
Methods SVM BPN Fuzzy
No. In the proposed approach, the first step is image acquisition in
1
Edge features extraction
68 70 78 which the images of grape leaves are captured using mobile
method
camera having resolution of 2,5,8,13 mega pixels. A total of
CYMK features extraction
450 images are included in the database which comprises of
2 73 77 78 healthy, diseased and pest-infected images. The image format
method with edge
is jpg. Figure 2 shows sample images from our database.
3
GA features extraction
81 81 82 In the second step, a set of pre-processing steps are applied
method with edge to the input image so that the image becomes suitable for further
GA features extraction processing. Image resizing is one of the important steps so as to
4 83 83 85
method with SVM maintain the uniformity in terms of size of the images without
loss. The next step in pre-processing is background subtraction
CYMK features extraction
5
method with SVM
83 83 85 i.e. segmentation of the region of interest from the image. The
third step is to extract the features from the infected part of the
6
Edge features extraction
91 93 94
leaf based on properties such as color, correlation,
method homogeneity, contrast and energy.
7
Color feature extraction
91 93 94 In the last step, statistical analysis of the features is done and
with SVM final classification is done using Support Vector Machine
Sabine D. Bauer [9] proposed approach based on high- (SVM) which classifies the plant leaves in one of the two
resolution multispectral stereo images, they have described classes as healthy or diseased.
three methods of automatic classification of leaf diseases.
Images of sugar beet leaves are captured by 2 cameras i.e.
multispectral and RGB in a laboratory with proper settings of
lights. The leaves infected with the rust fungus Uroources betae
or Cercospora beticola were considered. The classification rates
achieved in experiments are 86% for Uroources betae and 91%
for Cercospora beticola.
Vinita Tajane [10] analyzed CBIR technique accurately
detecting the diseases in medicinal plants. Edge histogram and
color histogram are the approaches used for identifying leaves
based on its features. For edge histogram canny edge detection
method is used. To check the intensity of each color pixels in
the given sample image color Histogram separates the layers to
plot the blue, green and red layer histogram which also tests for
abnormalities in the leaf. This paper compares testing sample
with the diseased sample, then follow few steps and shows
which type of disease the leaf is suffering from. Fig. 1. General Block diagram for Disease Detection System
Radha Kokare and Yogesh Dandawate [11] proposed an
image processing based approach used for automated plant
iseases classification based on leaf image processing. It is
concerned with the discrimination between diseased and
healthy soybean leaves using SVM classifier. The algorithm is
tested over the database of 120 images taken directly from
different farms using different mobile cameras. The SIFT
algorithm enables to correctly recognize the plant species based
on the leaf shape. The SVM classifier can significantly support
in recognizing normal and affected soybean leaves with an
average accuracy as high as 93.79%. This system will provide
Fig. 2. Sample images from database. (a) Healthy leaf. (b) & (c) Diseased leaf.
help to the farmer with minimal efforts. The farmer only needs
515
2016 3rd International Conference on Signal Processing and Integrated Networks (SPIN)
IV. PRE-PROCESSING
A. Color Transformation
Fig. 3. (a) Original Image (b) Background subtraction results
The next step in pre-processing is color transformation. The
input RGB images are converted into Hue Saturation Value
(HSV) color space. HSV is more close to human perception. C. Analysis of Texture using Opposite Color LBP
Hue component is usually taken into consideration for further
analysis as it the dominant color perceived by an observer. Local Binary Pattern (LBP) is an efficient and uniform
Figure 3 shows the results of RBG to HSV color transformation feature for analyzing local texture structures. Texture formation
is defined on local patterns instead of explaining it on pixel
of the leaves from the database. In the next step, multi
thresholding using Otsu algorithm [12] is applied on the HSV level. The different simple primitives are used by LBP to
images. describe complex structures in an image. LBP consider texture
primitives as well as placement rule property of structure
Multilevel thresholding segment the images into number of analysis [14].
distinct regions which corresponds to one background and The method of opposite color LBP feature extraction is
adapted by Jain and Healey [15]. Here, opposite color
several objects by selecting the threshold value based on the
largest inter-class variance between the foreground and represents the opposite pairs of color perceived by human eyes.
background [13]. In opposite color LBP, LBP operator is applied to each color
component also additionally each pair of color component is
B. Background Subtraction used to form opposite color patterns [14].
Consider a texture ܶ as a joint distribution [14] of the gray
The initial goal in the pre-processing step is to divide the levels ܲ ͳ image pixels normalized as:
input image into two classes as soil (i.e. background) and leaf. ܶ ൎ ݐ൫ݏሺ݃ െ ݃ ሻǡ ǥ Ǥ ǡ ݏሺ݃ିଵ െ ݃ ሻ൯ (3)
Background subtraction techniques are applied to the HSV Where, ݃ is the center pixel value, ݃ ሺ ൌ Ͳǡ ǥ Ǥ ǡ ܲ െ ͳሻ
images to remove the unwanted background from the image and represents gray values of ܲ equally spaced pixels on a circle of
to extract the desired part of the image. radius: ܴሺܴ Ͳሻ
In color based background subtraction, the intensity values
of R, G and B components are used to remove the unwanted
background. Since the green colored pixels contribute to the
healthy part of the leaf, these pixels are preserved and the other
pixels are made black. The pixels having G component greater
than R and B components are preserved and the rest pixels are
made black.
In order to get more accurate background subtraction, color
based background subtraction is used. In this technique, the
onnected components in the image are found out. The largest
connected component (i.e. the leaf) is preserved and the other
unwanted components are removed. Figure 3 shows the results
of background subtraction.
Fig. 4. Results of texture analysis. (a) Original images (b), (c), (d), (e) Results
of texture analysis
516
2016 3rd International Conference on Signal Processing and Integrated Networks (SPIN)
ͳ ݔ Ͳ
And ݏሺݔሻ ൌ ቄ (4)
Ͳ ݔ൏ Ͳ
The local image texture around ܴሺݔ ǡ ݕ ሻ is characterized as
given below:
ܲܤܮǡோሺ௫ǡ௬ ሻ
ିଵ
ൌ ݏ൫݃
ୀ
െ ݃ ൯ʹ ሺͷሻ
Thus, the texture can be approximately represented as:
ܶ ൌ ݐ൫ܲܤܮǡோሺ௫ ǡ௬ ሻ ൯ (6)
517
2016 3rd International Conference on Signal Processing and Integrated Networks (SPIN)
518