Abstract: This study exposes the automatic computation system to analysis the cotton leaf spot diseases. First to initialize the images from the database (Image features) that are highly related to the test image (new image), where test image is given by the user. Three features are used for matching the train image features in database images, namely color feature variance, shape and texture feature variance, these features are extracted by PSO. The feature selection method which helps to identify the injured leaf spot of cotton and at the same time improve the accuracy of the system and reduce the error rate. These features are calculated by different techniques. The proposed Skew divergence color variance feature is calculated by color histogram and color descriptor. The shape Skew divergence feature is calculated by Sobel and canny through the find out edge variance, edge location using Edge detection method. The skew divergence texture feature is calculated by texture descriptor. This study based on six types of diseases has been like Bacterial Blight, Fusarium wilt, Leaf Blight, Root rot, Micro Nutrient, Verticillium wilt. In this study, utilize these three features and combined the classifier of proposed Cross Information gain Deep forward Neural Network (CIGDFNN) which helps to recognize and identify cotton leaf spot diseases. The forceful feature vector set is a combination of three features lead to obtain the higher accuracy rate and sensitivity, specificity when tested with the cotton leaf dataset.
INTRODUCTION
The leaf spot of diseases plays a vital role in the plant growth environment, also diseases can be easily identified with the help of the infected area of the crop (Badnakhe and Deshmukh, 2012) Commonly the leaf will naturally going to show the infected part in a clear way which can be easily identified. So, we can say the transform in the crop colors are the significant aspect of the notification. When the health of the crop is in good stage then the color of the crop is different, but as soon as the crop is going to precious by some harming pathogens, the color changes automatically. Crop diseases have affected in a particular part. It may cause a reduction in productivity. The early days exposed eye observation of experts is the main approach adopted in practice for detection of plant diseases. In this investigation based on feature extraction the proposed skew divergence of edge, color, texture variance features has been used to analyze the affected part of the leaf. Feature extraction stage deals with the color, size and shape of the spot and finally classification are done by the proposed Cross information gain Deep forward Neural Network. This technique has been used to predict the six types of diseases. Finally obtained performance evaluation results were obtained.
This study discussed is based on how can to reduce the computational complexity by enhancements of automatic detection of Crop diseases. As this can give much damage in the agricultural fields of crops and detect the symptoms of diseases. Again we can tell the potential prevention and treatment of the infected plant. In this way want to look for the correctly, robust and less expensive and accurate method to detect and classify the crop diseases.
First, test images of various cotton leaves are captured using a digital mobile phone. Then image processing techniques are applied to acquire the images to extract useful features that are necessary for further analysis. Several statistical techniques are applied to classify the images according to the specific problem at affected leaf spot area. Feature selection PSO has been used to analyze the best matching of affected leaf feature results owing the optimal solution. In the classification phase, Edge, color, texture feature variance corresponding feature values stored in the image domain. Fusion method has been applied and one of feature level fusion do the Joint Feature Vector (JFV) calculate, score level fusion techniques like sum rule and the product rule of edge, color, texture features variance are calculating the affected part of in that block. And testing image feature for the leaves are extracted and compared with the training part features, if correctly pattern matching based on the affected region of cotton leaf disease.
Li et al. (2011) the proposed work based on the Web-Based Intelligent Diagnosis System for Cotton Diseases Control system used the proposed method in a BP neural network as a decision-making system (Hayat et al., 2013). The proposed work discussed the automated system that can identify the pest or disease affecting a cotton leaf, boll or flower by using image analysis. In this proposed method is CMYK based image cleaning technique to remove shadows, hands and other impurities from images.
Bernardes et al. (2013) this proposed method for automatic classification of cotton diseases through feature extraction of leaf symptoms from digital images. Wavelet transform energy has been used for feature extraction while SVM has been used for classification.
Zhang et al. (2007) the proposed study discussed the fuzzy feature selection approach-fuzzy curves and Fuzzy surfaces are to select features of cotton disease leaves. In classifying to get the best information for diagnosing and identifying, a subset of independent significant features is identified exploiting the fuzzy feature selection approach.
Meunkaewjinda et al. (2008) the proposed work organized the cotton leaf disease segmentation process using a self organizing feature map with a genetic algorithm for optimization and support vector machines for classification. Gulhane and Gurjar (2011) the proposed work described in the features that could be extracted intention using the Self organizing feature map with a back-propagation neural network is used to recognize the color of the images. Finally classification of the cotton diseases.
Gurjar and Gulhane (2012) this proposed work addresses the disease analysis is possible for the cotton leaf disease. The analysis of the various diseases present on the cotton leaves can be effectively detected in the early stage before it will injure the whole crops, initially we can be able to detect three types of diseases of the cotton leaves by the methodology of Eigen feature regularization and extraction technique in more extent. If it detects in early stage we can say that, we able to make better manufacture.
He et al. (2013) this proposed work described the three different color models for extracting the injured image from cotton leaf images were developed, address the RGB color model, HIS color model and YCbCr color model. The ratio of damage (γ) was chosen as feature to measure the degree of damage which caused by diseases or pests. In this study shows the comparison of the results obtained by the implementing in different color models, the comparison of outcomes shows good accuracy in both color models and YCbCr color space is considered as the best color model for extracting the injured leaf images.
MATERIALS AND METHODS
This proposed study based on cotton leaf diseases classification has been used for the Proposed CIG-DFNN classifier and proposed skew divergences edge, color, texture variance feature techniques are used to extract the affected spot of leaf features are like the shape, color, texture features methods to investigate. Finally obtained the performance evaluation of six types of diseases accuracy, sensitivity, specificity of this system as shown in Fig. 1.
Proposed CIG-DFNN algorithm
Configuration: The architecture has followed us:
| • | Input: 1<i<n where i = 1-150 |
| • | One hidden layer: 1<j<p where February 3, 2014j = 1-20 |
| • | Output: 1<k<m where k = 1-11 |
The according to DFNN theory weight ranges from[-1/S1 to 1/S1] where S1 is number of hidden layers in the network.
| Fig. 1: | Proposed CIG_DFNN system architecture diagram |
Proposed CIG- DFNN algorithm:
| • | In the first phase greedily train subsets of the parameters of the network using a layer-wise and unsupervised learning criterion, by repeating the following steps for each layer (i∈{1,…, l}). Until a stopping criteria is met, iterate through training by: | |
| • | Mapping input training sample xt to representation |
|
| • | Updating parameters bi-1, bi and Wi of layer is using some unsupervised learning algorithm | |
Also, initialize (e.g., randomly) the output layer parameters bl+1, Wl+1.
| • | In the second and final phase, fine-tune all the parameters q of the network using backpropagation and gradient descent on a global supervised cost function C(xt, yt, ∈), with input xt and label yt that is trying to make steps in the direction: |
In general the deep forward neural network initialization of the parameters randomly to perform the classification becomes poor solution. Standard gradient descent from random initialization is done, so poorly solution with deep neural networks, to overcome this problem in DFNN first observes the influence of the nonlinear activation functions. Find that the logistic sigmoid activation is unsuited for deep networks with random initialization because of its mean value which can drive especially the top hidden layer into saturation.
In this algorithm the first phase that initializes the parameters of the whole network will ultimately have an impact on the solution found in the second phase (the fine-tuning phase). In the first phase, initializations of the parameter are measured by the gradient descent error rate and also extend the work to which configuration of the neural network parameters can be different from the initial configuration. Hence, similarly to using a regularization term on the parameters of the model that constrains them to be close to a particular value (e.g., 0 for weight decay), the first phase here will ensure that the parameter solution for each layer found by fine-tuning will not be far from the solution found by the unsupervised learning algorithm. In addition, the non-convexity of the supervised training criterion means that the choice of initial parameter values can greatly influence the quality of the solution obtained by gradient descent with error gradient function.
In DFNN also, we perform CIG rate to diminish the error rate of the hidden layer and input vector layer. Reduce the error rate calculate the CIG for the hidden layer and the input vector layer (Fig. 2).
Proposed CIG_DFNN algorithm: First, simple edge detection is carried out and blocks with edge pixels inside are judged into the structural category. Then, color variance is calculated in the remaining blocks. Find the variance across the edge with canny edge detector and color splitting methods. Variance in the gray level in a region in the neighborhood of a pixel is a measure of the texture. Here we calculate the feature level fusion to combine the Edge, color and texture feature sets after normalization in order to yield a Joint Feature Vector (JFV). Instead of using GA for selection of the best features in the feature vector. This study has been used for PSO feature selection of the best parameters for both global and local features results.
| Fig. 2: | Proposed CIG-DFNN algorithm flow |
In this algorithm of feature as input to extract the features.next classify the diseases. Finally display the performance of outcomes. Initially select the feature vectors for training, The proposed CIG-DFNN classification algorithm the input file is read by the system. We define the decision function (W) to extract the feature vectors W and based on decision function only, we classify the training samples into -1 or +1. If the objective or decision function f (W) is greater than zero. Then we define the class label as +1 otherwise, It is considered as class label of -1. Finally the performance evaluation based on six types of diseases is obtained.
RESULTS AND DISCUSSION
Discussion about the dataset: Cotton leaf disease datasets collected from south zone Tamil Nadu at Andhiyaur district from 2012 in June month. On cotton the incidence on hybrids and Surabi varites from maximum incidence was recorded during upto one week the diseases images 270 data sets collected from the field. Directly congregate the farmers and get the suggestion from them.
In this investigation dataset collected was done using camera mobile (Nokia) captured that infected cotton leaf images. In this dataset used has an input image and tested in advance version 2010 MATLAB tool environments. In this investigation analysis about six types of diseases like the Bacterial Blight, Fusarium wilt, Leaf Blight, Root Rot, Micronutrient, Verticillium wilt. This process has been reduced the computational complexity and less cost expensive. Identification process has been used for proposed image processing with data mining techniques are combined and analyzed classifying the diseases image of the output. Finally accuracy of the diseases is obtained.
| Table 1: | Actual classes of six types of diseases |
| Table 2: | Performance evaluation of proposed CIG-DFNN algorithm |
| Fig. 3: | Performance evaluation of six types of diseases |
Table 1 shows the actual classes of six types of diseases has been used for CIG-DFNN classifier classify the diseases.
Table 2 described the performance Evaluation of proposed CIG-DFNN algorithm has been used to analyzed the six types of diseases, finally out comes the true positive of diseases, like vise the precision, sensitivity, specificity the overall accuracy is 95%.
Figure 3 Represent the performance evaluation of six types of diseases has been used for proposed CIG-DFNN algorithm by correctly classified the six types of diseases. Lastly out comes the overall accuracy of this system.
CONCLUSION
The proposed Skew divergence color variance feature is calculated by color histogram and color descriptor. The shape feature extracted has been used for Skew divergence feature is calculated by Sobel and canny through the find out edge variance, edge location of Edge detection method. As follow this has been used for the skew divergence texture feature is calculated by Gober filter and texture descriptor. In this investigation classified the six types of diseases like Bacterial Blight, Fusariumwilt, Leaf Blight, Root rot, Micro Nutrient, Verticilium Wilt and we utilize these three features and combined the classifier of proposed Cross Information gain Deep forward Neural Network (CIG-DFNN) were analyzed which are the most appropriate features for training images which one has been recognized the cotton leaf spot diseases. Finally performance evaluation of proposed CIG-DFNN system based on the overall accuracy is 95%. This method helps to correctly classify the diseases. It has been used for accurate classification the diseases by increasing the accuracy rate and diminish the error rate.
ACKNOWLEDGMENTS
I would like to express my special thanks of gratitude to Dr. S. Nakkeeran Savage Perumal, Associate Professor (Pathology), Department of Plant Pathology of Tamil Nadu Agricultural University. And thankful to Our Karpagam University in to do my research work. My heartily thankful them for the overall support.