Damage of plant is an important issue in agriculture. There are lots of factors involving weather, fungi, artificial drying, and mechanical damage during harvest and storage which can cause damage. NIR spectroscopy for classifying sound and damaged soybean seeds is very useful. NIR spectrometer is used to collect the spectra of single seed then PLS and neural network are used for classification of sound and damaged seeds. Near infrared spectroscopy is used because machine vision cannot provide information related to chemical composition because it is only useful for visible regions. NIR spectroscopy is useful for both physical and chemical properties. Seeds of six categories are used which are sound, weather damaged, frost-damaged, sprout-damaged, heat damaged and mold damaged. Grams/32 software is used for changing reflectance of spectrum in color space L, a , b. L ranges from 0(black) to 100(white), a ranges from -100(green) to 100(red)and b ranges from -100(blue)to 100(yellow).
NIR spectrometer is used to collect spectra at a rate of 30/s. Spectrum of 700 sound seeds and 900 seeds damaged by other factors were measured. Two class and six class models are used for classification of sound and damaged soybean seeds with the help of Partial Least Square (PLS) software. Two class model is used for classifying sound and damaged soybean seeds whereas six class model is used for classifying sound seeds and weather, frost, sprout, heat and mold damaged seeds.
In order to develop a neural network model for classification of sound and damage soybean seed the Neural Works Professional II/Plus software package is used. The neural network model package is based on back propagation networks.
In back propagation networks increment or decrement in weights is needed because of the errors. At first weights are randomly allocated but after every trial weights are adjusted until the errors are reduced to acceptable values. Physical and chemical properties of sound and damaged soybean seeds are totally different so by using only visible wavelength region results in poor classification. By using near infrared region important information can be obtained. Highest classification accuracies can be obtained by using full wavelength region (490-1690nm). By using visible and NIR wavelength region alone results in lower classification accuracy.
The best classification can be obtained by using neural network without hidden layer. PLS gives higher classification accuracy if two class classifications is used but in case of classification of six categories classification neural network (NN) gives higher accuracy results.
Computer Vision Based Weed Identification Under Field Conditions Using Controlled Lighting
Identification of weeds in crops is done by using methods of digital image analysis. Different kind of weeds often grow up with crop and its difficult to differentiate crop and weed so digital image analysis are used which are useful for differentiating both.
Images were captured through MatroxMeter(RGB) ,this device provides controlled lighting.
Two crops cabbage and carrot are used in greenhouse and open field experiments . For greenhouse experiments weed was added but for open field experiments natural weed population
In digital images, it is difficult to differentiate between crop plants and weeds especially when they reached on advanced growth stage. Segmentation algorithm is used for differentiating crops and weeds and soil.
This algorithm is based on union of two sets of each image which are S(soil) and V(vegetation). V has two components C (crop) and W (weed). The crop image data used in this research was the image of cauliflower at four different growth stages which are grouped. Experiments were performed on the 12 images. Colour is an important distinguishing feature and used as one component of the selection algorithm.
Noise can occur in images through which misclassification occur and can deal by using square morphological closing filter. in a large bright region this filter can reduce the noise by removing small dark holes. Erosion is used for suppressing small bright region and removed pixels from the outer edge of large bright regions. The central position of each plant is located by processing Pv with a large erosion filter. The output of erosion filter is bright central position of crop plant. At growing stage 4 this approach is modified, centroid of soil region has to be found instead of centroid of crop plant. weeds are the main source of bias in the location of grid point.
Segmentation algorithm is used to identify crop plant pixels but there was a higher probability of weed pixels being classified as crop plant pixels because they are very close to crop plants and grew in rows. So the difference between the size and texture of crop plant and weed is used in order to find the location of crop plant boundary. Morphological opening filter separates crop from adjoining areas of weed. The last stage involved thresholding the output of opening filter.
Improving Plant Discrimination In Image Processing By Use Of Different Colour Space Transformation
Image processing is becoming popular in different agricultural applications. color images taken by a digital camera stored in RGB colour space. Colour cameras can deal with large variety of situation for differentiating single object from an image. Thresholding is applied on each colour channel. Separation of object can be improved by transforming RGB by weighting each channel in different way in order to emphasize specific features. Different colour transformations were performed and then compared them. 40 images of RGB colour spaces are used discriminant analysis, canonical transformation, i1i2i3, HSV, HSI and Lab colour spaces were used for transformation. Thresholding is performed on transformed image to convert it into binary images in order to differentiate plant and soil. Manual and automatic thresholding was for i1i2i3,thresholding according to hemming was used for HSV,HSI and Lab colour spaces. Discriminant analysis consists of colour transformation and binarisation. Thresholding was not needed in discriminant analysis.
Linear and logarithmic discriminant functions were used. Logarithmic discriminant analysis is the most effective in discriminating but it takes much time for processing of one image.
HSV,HSI am Lab colour spaces gave better results but not in open field.i1i2i3 were recommended for plant detection . This transformation is more useful if the reflection occur due to high solar radiation or some water on leaf.
Image pattern classification for the identification of disease causing agents in plants
For the identification of plant diseases machine vision system is used. Different images of cotton crops which shows diseased region were used, enhanced, segmented and the feature extraction is performed. The extracted features were then used as inputs to SVM classifier and then testing will be performed to choose the best classification model.
Different features such as shape, texture, greylevel, connectivity etc were extracted from segmented region. Co-occurrence matrix was used in order to calculate the image texture. This method is used to measure occurrence of greylevels between a specific position in image and neighboring pixels according to distance and direction.
Fractal dimension is a feature which measured dimension of object and box counting algorithm is used to estimate this measurement. Lacunarity was a multiscaled method which measures texture associated with spatial dispersion and gliding box algorithm was used to calculate lacunarity.
Different features extracted from 117 images of cotton crops were labeled according to disease they belonged. SVM used Radial Basis Function kernel. There are different problems in classification if it involves more than two classes are used then multiple classes classification was used which uses one-against-one method.
Different approaches were used to identify best classification model. Each feature is used as a single input to classifier .the groups of feature were used as inputs to classifier and then all the features except one is used as input. All features were not give the same amount of information so 7 fold cross validation is used.
Fall Armyworm Damaged Maize Plant Identification Using Digital Images
An algorithm is developed to identify damaged maize plant by the fall armyworm at simplified lighting conditions using digital color images. Eight different stages of diseased and non-diseased maize plant were taken in three different light intensities. This algorithm involves processing and image analysis. First, the binary images were created by segmentation and then the images were divided into blocks and classified as diseased or non-diseased.
The algorithm starts by converting original RGB image into greylevel image then by iterative threshold method it is converted into binary image then by applying 383 median filter its is converted into binary filteres image. These steps are part of first stage which is image processing.
For next stage image analysis binary image is subdivided into 12 blocks. Blocks were selected from the subdivided image and then by object identification and counting damaged and non-damaged blocks were classified
A review of advanced techniques for detecting plant diseases
Diseases in plants are major issue in field of agriculture as they result in major production and economic losses. There is a mechanism called scouting is used for this purpose but this is not only expensive but also time consuming so there is need for a mechanism which is rapid, cost-effective so there are different technologies spectroscopic and imaging based and volatile profiling based plant disease detection methods.
In the spectroscopic and imaging techniques, fluorescence spectroscopy, visible IR spectroscopy, fluorescence imaging and hyperspectral imaging involved.
In VOC profile-based metabolite analysis released by healthy and diseased plants as a tool for identifying diseases. These methods can accurately detect plant diseases.
Automatic Identification of Weed Seeds
Image processing techniques were used to obtain seed size, shape, color and texture characteristics. Large database of images were used. Naïve bayes classifier was used for evaluation. It gives excellent results. Not only the color images were used but also the black and white images of weed seeds were used.
By using morphological and textural characteristics as classification feature, it would reduce the complexity and cost. Naïve bayes classifier and Artificial Neural Network (ANN) were used for weed seed identification but naïve bayes has an excellent performance as compared to ANN.
Identification of citrus disease using colour texture features and discriminant analysis
Machine vision and AI techniques are used to achieve intelligent farming including early detection of diseases. Colour co-occurrence method is used to determine whether HIS color features in conjunction with statistical classification algorithms would be used to identify diseased and normal citrus leaves under laboratory conditions.
Greasy spot, melanose, normal and scab are four different classes of citrus leaves used. By using image processing techniques, algorithms were designed for feature extraction and classification. Colour cooccurence methodology is used for feature extraction. It used colour and texture to get unique features. SAS discriminant analysis is used to evaluate the potential classification accuracies and this can be achieved by a traditional statistical classifier.
Image texture feature dataset appeared as the best data model for citrus leaf classification, it uses reduced hue and saturation feature set. It gets high classification accuracy, less computation time and the elimination of intensity features which is beneficial in highly variable outdoor lighting conditions.
Fast and accurate detection and classification of plant diseases
First the images were acquired using a digital camera then the image processing techniques were applied to extract features which are useful. Then the classification is performed.
The algorithm starts by acquiring RGB images. In the next step colour transformation is applied on RGB images. Images were then segmented using K-means clustering techniques. Green pixels are masked by using Ostu's method. Pixels with zeros red, green, blue values and boundary pixels of infected objects were removed. The infected cluster is then converted into HIS from RGB.
In the next step SGDM matrix were generated for H and S. For calculation of features GLCM function is used. Neural Network is used as a classification tool.
Statistical and neural network classifier for citrus disease detection using machine vision
Image data sets of common disease of citrus were collected and then CCM is used for detection of diseases. Different strategies and algorithms were developed for classifications which were based on feature obtained from CCM and then compared the classification algorithm in order to check accuracies.
After acquiring images image processing algorithms for feature extraction and classification were developed. Feature extraction used CCM methodology. SAS discriminant analysis was used to evaluate the classification accuracies. Classification tests were applied on different classification algorithms. Statistical classifier using Mahalanobis minimum distance method achieved 98% classification accuracy. Neural network classifier using back propagation algorithm and neural network classifier using Radial Basis Function achieved 95% accuracy rate so the Mahalanobis minimum distance method is the best for classification.
Rice disease identification using pattern recognition techniques
For the identification of rice disease , software prototype system is described. Image segmentation techniques used to detect infected parts of the plants. These infected parts were further used for classification using Neural Network. For feature extraction first the segmentation is performed and for this entropy based bi-level thresholding method is used.
After segmentation boundary detection algorithms were applied this uses 8- connectivity method. In the next step spot detection is applied for the normalization of spot size and interpolation method is used for fractional zooming. After this when all the uniform size spots were obtained, unsupervised learning technique Self Organinzing Map is used.
Classification of grapefruit peel diseases using colour texture feature analysis
Colour texture feature were used for detection of citrus peel disease.images of normal and five common peel diseases which are canker, copper burn, greasy spot, melanose and wind scan were used. Using colour cooccurence method, 39 image texture features were determined. Before applying CCM, RGB is transformed into HSI. SGDM( Spatial Gray level Dependence Matrix)was used to develop color cooccurence texture analysis. Texture feature were then calculated by SGDM. SAS procedure STEPDISC can find variables which are important for discriminating samples and it will use for texture feature selection. SAS procedure DISCRIM creates a discriminant function which was used to develop classification model. This is also used to test the accuracies of classification models.
Plant leaves classification based on morphological features and a fuzzy surface selection technique
Artificial vision system is designed to extract special features from plant leaves. Feature selection approach is used to identify significant image features and for the classification test Neural Network is used.
In morphological feature extraction, morphological and geometrical features were extracted from plant leaves. These features provide critical information. Feature selection is very important task which is needed to determine the most relevant features for pattern recognition. Neural Network take features as inputs and perform classification.
Weed seeds identification by machine vision
There is a need of fast and reliable method for the identification and classification of seeds. Seeds of 57 weed species were used. Different features extractes were used as classification parameter . 12 classification parameters were used in which 6 morphological, 4 colour and 2 textural were involved. With the help of these parameters naïve bayes and Artificial Neural Network were compared for the identification of seed species. ANN performed better than naïve bayes.