classification of rice plant leaf diseases using feature matching
Transcription
classification of rice plant leaf diseases using feature matching
International Journal On Engineering Technology and Sciences – IJETS™ ISSN(P): 2349-3968, ISSN (O): 2349-3976 Volume I, Issue VII, November - 2014 CLASSIFICATION OF RICE PLANT LEAF DISEASES USING FEATURE MATCHING Dr.C.Kumar Charliepaul 1 Principal A.S.L Pauls College of Engg & Tech, Coimbatore . [email protected] ABSTRACT—Technological advances have brought about drastic changes in farming like plant disease identification, pest outbreaks and crop management. The proposed system identifies diseases of rice plant leaf by extracting features from the infected regions of the rice plant leaf images. Fermi energy based segmentation method used to segment the infected region from its background region. Symptoms of the diseases are characterized using features like color and shape of the infected portion and extracted feature used for identifying diseases. Color features are determined by calculating mean and standard deviation of the infected and background pixels as well as change of color of the infected region in comparison with the background in three different color planes, Red (R), Green (G) and Blue (B). Shape of the infected region is a major symptom to predict the diseases. When the plant is infected by diseases having the symptoms of shapes like oval, circular and irregular spot. DRSLE (Distance Regularized Level Set Evolution) algorithm used for identifying desired shape of the infected region. Rough Set theory reduces the complexity of the system and minimizes loss of information by selecting core features. Finally using features matching predict diseases of rice plant leaf images and provides superior result compare to traditional method. Keywords: Fermi energy segmentation, Rough set theory, Feature Extraction, Rice plant leaf, Feature Matching. I. is similar with respect to some properties like color, INTRODUCTION intensity, or texture. Interested region provide more useful Image processing is a process of analyzes and information for the model. Feature Extraction reduces the manipulation of digital image in order to improve the quality amount of unnecessary information to describe a model. The of image. Image is a collection of pixels. Pixel is a main major problem in analyzing the complex data involves large element in digital image. Digital image is processed in digital number of variables. The large number of variable analyzes computer. The digital image is composed of a finite number require a large amount of memory and computational time. of elements and each element has a particular value and The color analyzes is the process of extracting the interested location. Image Segmentation is the process of segmenting a color information. The main element of image is color. The digital image into several segments. The main aim of analysis of color feature in image database retrieval is most segmentation in image processing is to simplify or represent important. an image into more meaningful and also easier to analyze. So the color feature is more domain independent Image segmentation mainly used to isolate the objects and compared to other feature. Color having three dimensional boundaries in images. It assigns a label to every pixel in an space Red, Green and Blue but RGB color space is not image, the pixel with same label share certain characteristics. uniform in images. It is used to eliminate the false hit. The Thresholding is the simplest method of image segmentation. Data contains both redundant and irrelevant features. It can be used to isolate the interested region from its Irrelevant feature does not provide more useful information background region. Each of the pixels in an interested region and redundant feature leads to computational process as 290 International Journal On Engineering Technology and Sciences – IJETS™ ISSN(P): 2349-3968, ISSN (O): 2349-3976 Volume I, Issue VII, November - 2014 complex. Feature selection is the process of selecting image of the disease infected plant or leaf, into the H, I3a relevant and core feature for constructing the model. Feature and selection is also called as selection of variable, selection of transformations are developed from a modification of the variable subset and selection of attribute. It is subset of original image in to color transformation to meet the feature extraction. Feature extraction generate new feature requirements of the plant disease data set. The transformed from original feature while feature selection returns a subset image is then segmented by analyzing the distribution of of feature. Rough set theory is a feature selection method. It intensities in a histogram. The threshold cut-off value is is determined by lower and upper boundary of a set. It is a determined according to their position in the histogram. This mathematical concept dealing with uncertainty in data. technique is particularly useful when the target in the image Feature matching is a process of finding similar feature in data set is one with a large distribution of intensities [2]. I3b color transformations. The I3a and I3b different images in image processing. The feature of testing S.L.S. Abdullah et al (2012) proposed improved images is compared with the feature of training images. The thresholding based technique for image segmentation. highest feature matching images are taken as resultant Different illuminations may produce different color intensity images. of the object surface and thus lead to inaccurate segmented images and traditional methods were unable to produce good quality segmented. Therefore, an improved thresholdingbased segmentation integrated with an inverse technique (TsTN) that was able to partition natural images correctly. The analysis results showed that TsTN has the ability to produce good quality segmented images for dark images. Furthermore, this segmentation technique was proven to be more accurate than the traditional thresholding and clustering II. RELATED WORK techniques [1]. A.K Das et al (2012) proposed SVM and Bayes’ Z.Xue et al (2003) proposed Bayesian shape model classifier to classify the diseases of rice plant leaf. An (BSM) to find contour points in the face. A full-face model automated system has been developed to classify the leaf consisting of the contour points is designed to describe the brown spot and the leaf blast diseases of rice plant based on face patch, using which the normalization of the extracted the morphological changes of the plants caused by the face patch can be performed efficiently. In BSM, the diseases. Otsu method is used to isolate the infected region prototype of the face contour can be adjusted adaptively from the background. Radial distribution of the hue from the according to its prior distribution. Moreover, an affine center to the boundary of the spot images has been used as invariant internal energy term is introduced to describe the features to classify the diseases by Bayes’ and SVM local shape deformations between the prototype contour in Classifier [3]. the shape domain and the deformable contour in the image R.Lu et al (2013) proposed Support Vector Machine domain. The face patch is extracted and normalized using the for classification and Otsu method for segmentation. This piece-wise affine triangle warping algorithm.[5] A.Camarago et al (2009) proposed method automatically adjusts the classification hyper plane color calculated by using linear SVM and requires minimum transformation method. This method converting the RGB 291 International Journal On Engineering Technology and Sciences – IJETS™ ISSN(P): 2349-3968, ISSN (O): 2349-3976 Volume I, Issue VII, November - 2014 training and time. It also avoids the problems caused by with variation in color from gray to light brown at centers, variations in the lighting condition and color of the fruit. surrounded by distinct dark reddish brown margins. Fig. 2. Leaf Brown Spot Disadvantage of this method is segmentation is not accurate [6]. 3) Rice Blast III. PROPOSED WORK Rice blast is caused by the fungus Magnaportheoryzae Treatment the rice plant leaf based on diseases and observed in both lowland and upland. Initially white to saves the products from quantitative and qualitative loss and grayish green circular lesions or spots with dark green plays significant role in country’s economic growth. The borders are found on the leaves. proposed system aims at developing a predicting system to Fig. 3. Rice Blast predict the diseases of rice plant leaf by performing the B. Fermi Energy Based Segmentation steps: Identification of the Infected Region, Extraction of Features, Selection of Features, Feature matching and System quality depends on the segmented result of Identification of Diseases. infected leaf images. Thresholding is a widely used segmentation technique that determines threshold value and A. Description of Rice Plant Leaf Diseases based on that value segment the infected leaf images. The The Rice plant leaf disease classified by the energy-based segmentation method consists in finding the proposed system is described below. optimal segmentation .This method is robust because the 1) Leaf Brown Spot Brown spot symptoms are observed at tillering stage. The shapes of the infected region vary from circular to oval with light brown color to gray at the center and reddish brown color at margin. segmentation criteria are objectively defined in the energy 2) Sheath Rot function and the optimization process is global and automatic. Fermi energy based segmentation method is used to predict the infected regions using RGB color components of the images. The Fermi energy or Fermi level is expressed in Eq. (1). 2 h2 2 3N 3 EF 2ML2 Fig. 1. Sheath Rot (1) Sheath rot caused by the pathogen Sarocladiumoryzae. Where N is the number of particles, M is the mass Rotting occurs on the leaf sheath that encloses the young of the particles, L is the length of the cube and h is the panicles. The lesions start as oblong or some irregular spots Planck constant. When an image is acquired using a physical source, the information content in the image is 292 International Journal On Engineering Technology and Sciences – IJETS™ ISSN(P): 2349-3968, ISSN (O): 2349-3976 Volume I, Issue VII, November - 2014 proportional to the energy radiated by the source. Fermi with the background in three different color planes, Red (R), EF of an image is computed using Eq. (1), where N Green (G) and Blue (B).Mean and standard deviations are energy calculated by using the Histogram. Shape features are is mapped as the number of pixels having distinct color extracted using the DRLSE algorithm. The algorithm having values in the image, number of grey levels is equivalent to two step (1) edge detection (2) shape detection. length L and mass of the image (M) is calculated by Edge and Shape Detection aggregating the mass of each pixel (i, j) using Eq. (2). ij rgb m H r , g ,b pq The RGB images are given as an input for edge r g b (2) detection. Output of the process is edge of the infected region. Initially images are read from the specified folder Where Hr,g,b is the number of pixels having a and that image can be converted into grey scale image. The particular intensity with r, g and b grey level values computer generated curves that move within images to find corresponding to Red, Green and Blue color planes object boundary (Gradient Vector Flow).After the gradient respectively and p q is of the image computed, pixels with large gradient values the size of the image. Energy becomes possible edge pixel. In DRLSE Level Set method E(i, j) at (i, j)th pixel position is calculated using Eq. (3) and represents the outline of the irregular shape and also compared with the threshold value EF for segmenting the controls the changes in shape. Such as splitting and merging in a natural and efficient way. The level set method infected region of the image. represents a closed curve using an auxiliary function 2 Ei, j Er , g ,b If , 2 h r 2 g 2 b2 ij 2 2 mrgb L called the level set function. (3) level set of Ei, j EF then the pixel (i, j) is treated as part is represented as the zero by {( x , y ) | ( x , y ) 0} (4) of the infected region, otherwise in background region. To reduce computational complexity constants h , and L are is implicitly manipulated by level set method eliminated from Eqs. (1) and (2) as the values are compared. through the function C. Feature Extraction values inside the region .The function takes positive and negative values outside. x,y Change of color of the plant leaf due to infection, are the coordinates of the infected region. The advantage of shape of the spot (infected region) of the leaf is used as this method is,it can perform number of computations features to classify the diseases. There are several ways to involving curves and surfaces on a fixed Cartesian grid detect plant pathologies (diseases). Some diseases have without visible symptoms in leaf. Change of color of the infected formulation with a distance regularization term and an region is compared with the background is considered as external energy term derives the motion of the zero level one of the important features for disease identification. contour towards the desired location. The level set evolution Color features are determined by calculating mean and is a gradient flow that reduces this energy function. The standard deviation of the infected and background pixels as level set evolution, uniformity of the LSF is maintained by a well as change of color of the infected region in comparison forward and backward diffusion that can be derived from 293 parameterize these objects. The level set International Journal On Engineering Technology and Sciences – IJETS™ ISSN(P): 2349-3968, ISSN (O): 2349-3976 Volume I, Issue VII, November - 2014 the distance regularization term. The distance regularization infected by corresponding diseases. In this method to term with a potential function forces the gradient magnitude perform number of computations involving curves and of the level set function to one of its minimum point, so it surfaces on a fixed Cartesian grid without parameterize maintains the desired shape of the level set function. these objects. D. Rough Set Theory IV. PERFORMANCE COMPARISON AND RESULT ANALYSIS Rough set theory is a mathematical concept dealing with uncertainty in data. The rough set is defined by the A. Data Set The data set has been taken from tuple ( PX , PX ).It is determined by the upper and lower www.shutterstock.com which contains Blast, Sheath rot and Brown spot images. Six images are taken for each disease, boundary of a set. A discernibility matrix is constructed to three images taken for training dataset and three images represent the family of discernibility relations. Each cell in a taken for testing dataset in each disease. The intensity, color discernibility matrix (M) contains the features for which two and shape are the major attributes of the image objects are discernible. The element mij of M is defined by B. Result Analysis Eq. (5). mij {ac : a(xi ) a(xj ) (d D, d(xi ) (dxj ))}, i,j =1,2,3,……n (5) GT = ground truth (7) Fig. 4. Performance Evaluation of Segmentation Algorithm Where d(xi ) labels of objects xi and and xj d(x j ) represent the class respectively. An entry containing minimum number of features implies that the features are sufficient to distinguish associated diseases and considered as the most important The Figure 4 shows that Fermi energy method features or core. The core feature set, say CR is defined by contains less noise in segmentation when compared with the Eq. (6)and remaining are treated as noncore features, say Otsu and K-means. The gray scale image is input to the (NC). Otsu and K-Mean method but color component of the image CR {mij || mij || 1, i, j 1,2,3...., n} is input to the Fermi energy method. Because of the color (6) component is used as an input, infected images are correctly E. Feature Matching segmented. From the result analysis, the Fermi energy based segmentation performs better than traditional segmentation. Minimal feature subset or reduct is considered for Computational time is reduced due to core feature selection. Feature matching where feature match is represented as IF- If the disease prediction system uses Otsu and K-means THEN form consisting of antecedent and consequent. IF the Segmentation method, produces 50% and 45% accurate color and shape of trained dataset images is equal to the result respectively, but results 75% of accuracy when it uses color and shape of test dataset image then the plant leaf is Fermi Energy segmentation method. In traditional method 294 International Journal On Engineering Technology and Sciences – IJETS™ ISSN(P): 2349-3968, ISSN (O): 2349-3976 Volume I, Issue VII, November - 2014 greyscale images taken for segmentation. In uninfected Communication Program and her ideas and suggestions, region some of the pixel treated as infected and in infected which have been very helpful in the project. I am so deeply region some of the pixel treated as uninfected. So the grateful for her help professionalism and valuable guidance infected region not correctly segmented. throughout this project. 1) Fermi Energy Based Segmentation Fig.5. Segmentation Algorithm REFERENCES [1] Abdullah.SL.S, Hambalia.H.A, and Jamil.N (2012), ‘Segmentation of Color component of the images are input to the Fermi Natural Images Using an Improved Threshold Based technique,’ energy method. So the infected region correctly segmented. International Symposium of Robotics and Intelligent Sensors, Vol. 50,No. 3, pp. 938-944. CONCLUSION AND FUTURE WORK [2] Camargoa.A, and Smith.J.S (2009), ‘An Image Processing Based Algorithm to Automatically Identify Plant Disease Visual Symptoms,’ International Journal of Biosystem Engineering, Vol. 102, No. 1, pp. 9-21. Predictive system has been developed for the prediction of rice plant leaf diseases using the symptoms [3] Das.A.K, Phadikar.S, and Sil.J (2012), ‘Classification Rice Leaf Diseases Based on Morphological Changes,’ International Journal of Information created by the diseases. Fermi energy based region and Electronics Engineering, Vol. 20, No. 2, pp. 80-95. extraction method correctly segments the infected region. [4] Das.A.K, Phadikar.S, and Sil.J (2013), ‘Rice Diseases Classification using Color features are determined by calculating mean and Feature Selection and Rule Generation Techniques,’ International Journal of Computers and Electronics Engineering, Vol. 90, No. 3, pp. 76-85. standard deviation of the infected and background pixels as [5] Li.S.Z, Theo.E.K, and Xue.Z (2003), ‘Bayesian model for facial feature well as change of color of the infected region .The desired extraction and recognition,’ International Journal of Pattern recognition, shape of the infected region is identified by using the Vol. 36, No. 12, pp. 2819-2833. DRLSE algorithm. By using rough set theory core features [6] Lu.R, and Mizushima.A (2013), ‘An Image Segmentation Method for Apple Sorting and Grading Using Support Vector Machine and Otsu’s are selected which minimizes loss of information and Method,’ International Journal of. Computer and Electronics in reduces the computational time. Finally, the testing dataset image features are compared with training dataset image Agriculture, Vol. 94, No. 4, pp.29-37. [7] Salamo.M, and Sanchez.M.L (2011), ‘Rough set based approaches to feature selection for case based reasoning classifiers, International Journal features, by using highest match features, diseases are of Pattern recognition, Vol. 60, No. 4, pp. 280-292. classified. The result shows that it performs well on nonuniform illumination images and also reduces the noisy Author Biography: features. The predictive system 75% correctly predicts the Kumar Charlie Paul, Principal of A.S.L Pauls College of Engineering & Technology. Had did many National and International Conferences and published many papers in journals. He also guided many students for their Ph.D project works. Having more than 23 years of experience in teaching field. rice plant leaf diseases. In future, it can be extended using more dataset from various kind of plant leaf and also can find risk factor of diseases. ACKNOWLEDGMENT First of all I would like to extend my sincere gratitude to my supervisor Dr.C.Nalini for providing me the opportunities of taking the part in Master of Computer and 295