Jitendra Paliwal

Cereal grain and dockage identification using machine vision

Algorithms were written to extract a total of 230 features (51 morphological, 123 colour, and 56 textural) from the high-resolution images of kernels of five grain types [barley, Canada Western Amber Durum (CWAD) wheat, Canada Western Red Spring (CWRS) wheat, oats, and rye] and five broad categories of dockage constituents [broken wheat kernels, chaff, buckwheat, wheat spikelets (one to three wheat kernels inside husk), and canola (rapeseed with low erucic acid content in the oil and low glucosinolate content in the meal)]. Different feature models, viz. morphological, colour, texture, and a combination of the three, were tested for their classification performances using a neural network classifier. Kernels and dockage particles with well-defined characteristics (e.g. CWRS wheat, buckwheat, and canola) showed near-perfect classification whereas particles with irregular and undefined features (e.g. chaff and wheat spikelets) were classified with accuracies of around 90%. The similarities in shape and size of some of the particles of chaff and wheat spikelets with the kernels of barley and oats affected the classification accuracies of the latter, adversely.

Fungal Detection in Wheat Using Near-Infrared Hyperspectral Imaging

Different species of fungi infect grain in the field and storage facilities. Contamination by fungi in grain is detected and quantified by traditional methods, such as microbial incubation and microscopic detection, which are subjective, labor intensive, and time consuming. An accurate and timely detection technique for fungal growth in grain is needed to prevent grain from spoiling and to reduce quality loss. In this study, the potential of near-infrared hyperspectral imaging to detect fungal infection in wheat was investigated. Wheat kernels infected with storage fungi, namely Penicillium spp., Aspergillus glaucus, and Aspergillus niger, were scanned using a hyperspectral imaging system, and a total of 20 image slices at evenly spaced wavelengths between 1000 to 1600 nm were acquired to form a hypercube. A multivariate image analysis (MIA) technique based on principal component analysis (PCA) was used to reduce the dimensionality of the image hypercubes. Two-class and four-class classification models were developed by applying k-means clustering and discriminant (linear, quadratic, and Mahalanobis) analyses. Two-class discriminant classification models gave maximum classification accuracy of 100%, and on average 97.8% infected kernels were correctly classified by the linear discriminant classifier. The four-class linear discriminant classifier correctly classified more than 95% of the kernels infected with Penicillium and 91.7% healthy kernels. However, the discriminant classifiers misclassified the kernels infected with A. niger and A. glaucus.

GRAIN KERNEL IDENTIFICATION USING KERNEL SIGNATURE

A machine vision algorithm was developed to distinguish the kernels of Canada Western Red Spring (CWRS) wheat, Canada Western Amber Durum (CWAD) wheat, barley, oats, and rye. The algorithm is based on the assumption that the shape of any of these kernels can be described by a radial function using the properties of Fourier descriptors which are invariant to translation, rotation, and scale. To evaluate the discrimination capability of the algorithm, color images of 2000 kernels for each type of grain (200 each from ten growing regions across Canada) were taken. For each kernel, three attributes viz. length, shape function (Fourier descriptors in polar coordinates), and color were extracted which were collectively called the kernel signature. Each attribute was then averaged for each grain type to form a training set. To identify any unknown kernel, all the three attributes were calculated for it and were compared against the corresponding values for each grain type in the training set using three different distance functions one each for each attribute. Classification was done by assigning different weights to the attributes, shape being the most important and length being the least. Robustness of the algorithm was tested by taking the test kernels from growing regions alien to the training set. Classification accuracies of 100, 94, 93, 99, and 95% were obtained for CWRS wheat, CWAD wheat, barley, oats, and rye, respectively.

Comparison of a Neural Network and a Non-parametric Classifier for Grain Kernel Identification

Abstract The performances of a four-layer backpropagation neural network and a non-parametric statistical classifier were compared for classification of barley, Canada Western Amber Durum wheat, Canada Western Red Spring wheat, oats, and rye. A total of 230 features (51 morphological, 123 colour, and 56 textural) from the high-resolution images of kernels of the five grain types were extracted and used as input features for classification. Different feature models, viz . morphological, colour, texture, and a combination of the three, were tested for their ability to classify these cereal grains. To make the classification process fast, the number of input features were reduced to 60 and 30. A set of features consisting of an equal number of morphological, colour, and textural features gave the best classification accuracies. The neural network classifier outperformed the non-parametric classifier in almost all the instances of classification.

Image Analysis of Bulk Grain Samples Using Neural Networks

Algorithms were developed to acquire and process color images of bulk grain samples of five grain types, namely barley, oats, rye, wheat, and durum wheat. The images were acquired using a video camera and were digitized using a frame grabber board. The images were stored on a personal computer from where they were accessed by an image processing program which extracted over 150 color and textural features. A neural-network-based classifier was developed to identify the unknown grain types. The color and textural features were presented to a back propagation neural network for training purposes. The trained network was then used to identify the unknown grain types. Results showed a classification accuracy of over 90% for all grain types.

Classification of fungal infected wheat kernels using near-infrared reflectance hyperspectral imaging and support vector machine

A classification algorithm was developed to differentiate individual fungal infected (Aspergillus niger, Aspergillus glaucus, and Penicillium spp.) and healthy wheat kernels. A near-infrared reflectance hyperspectral imaging system captured hyperspectral images at 20 wavelengths spaced evenly between 1000 nm and 1600 nm. Four statistical features (mean, variance, skewness, and kurtosis) were extracted from the hyperspectral image data of single kernels at each wavelength. The statistical features at all wavelength levels composed the pattern vector of a single kernel. The dimensionality of pattern vectors was reduced by principal component analysis. A multi-class support vector machine with kernel of radial basis function was used for classification. Using the statistical features, the wheat kernels infected by Aspergillus niger, Aspergillus glaucus, and Penicillium spp. and healthy wheat kernels were classified with accuracies of 92.9%, 87.2%, 99.3%, and 100%, respectively. Almost perfect classification was obtained under the infected vs. healthy model. There was 10.0% misclassification between Aspergillus niger and Aspergillus glaucus infected wheat samples.

Detection of Sprouted and Midge-Damaged Wheat Kernels Using Near-Infrared Hyperspectral Imaging

ABSTRACT Sprout damage which results in poor breadmaking quality due to enzymatic activity of α-amylase is one of the important grading factors of wheat in Canada. Potential of near-infrared (NIR) hyperspectral imaging was investigated to detect sprouting of wheat kernels. Artificially sprouted, midge-damaged, and healthy wheat kernels were scanned using NIR hyperspectral imaging system in the range of 1000–1600 nm at 60 evenly distributed wavelengths. Multivariate image analysis (MVI) technique based on principal components analysis (PCA) was applied to reduce the dimensionality of the hyperspectral data. Three wavelengths 1101.7, 1132.2, and 1305.1 nm were identified as significant and used in analysis. Statistical discriminant classifiers (linear, quadratic, and Mahalanobis) were used to classify sprouted, midge-damaged, and healthy wheat kernels. The discriminant classifiers gave maximum accuracy of 98.3 and 100% for classifying healthy and damaged kernels, respectively.

Fungal Damage Detection in Wheat Using Short-Wave Near-Infrared Hyperspectral and Digital Colour Imaging

Healthy and fungal-damaged wheat kernels infected by the species of storage fungi, namely Penicillium spp., Aspergillus glaucus, and A. niger, were scanned using a short-wave near-infrared hyperspectral imaging system in the 700–1100 nm wavelength range and an area scan colour camera. A multivariate image analysis was used to reduce the dimensionality of the hyperspectral data and to select the significant wavelength using principal component analysis. Wavelength 870 nm, which corresponded to the highest factor loading of first principal component, was considered to be significant. Statistical and histogram features from the 870 nm wavelength image were selected and used as input to statistical discriminant classifiers (linear, quadratic, and Mahalanobis). From the colour images, a total of 179 features (123 colour and 56 textural) were extracted and the top features selected from these features were used as input to the statistical classifiers. The linear discriminant analysis classifier correctly classified 97.3–100.0% healthy and fungal-infected wheat kernels, using the combined hyperspectral image features and the top ten features selected from 179 colour and textural features of the colour images as input.

Classification of cereal grains using a flatbed scanner

In the quest for an inexpensive machine-vision system (MVS) to identify and classify cereal grains, a flatbed scanner was used and its performance was evaluated. Images of bulk samples and individual grain kernels of barley, Canada Western Amber Durum (CWAD) wheat, Canada Western Red Spring (CWRS) wheat, oats, and rye were acquired and classification was done using a four layer back-propagation neural network. Classification accuracies in excess of 99% were obtained using a set of 10 color and textural features for bulk samples. For single kernel images, a set of at least 30 features (morphological, color, and textural) was required to achieve similar classification accuracies. Classification accuracies for single kernel samples varied between 96 and 99%.

Spectral Data Compression and Analyses Techniques to Discriminate Wheat Classes

Near-infrared spectroscopy (NIRS) was evaluated to differentiate six different classes of wheat grown in western Canada. The original spectral data consisted of 2594 wavelength variables, and three data compression techniques, namely, principal component analysis (PCA), discrete Fourier transform (DFT), and discrete wavelet transform (DWT), were compared to reduce the dimensionality of the original dataset. Five different classifiers, namely, linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), k-nearest neighbor (k-NN) classifier, radial basis function probabilistic neural network (PNN), and least-squares support vector machines (LS-SVM), were compared based on correct classification rates. Classification was performed on principal component scores, Fourier coefficients, and wavelet coefficients of the original spectral data. For dimensionality reduction, PCA was most efficient technique. It was also corroborated that linear end point baseline correction was necessary to achieve efficient data compression using DFT. Classification accuracies achieved using LDA or QDA combined with PCA as a pre-processing method were consistently better than the other three classifiers. Classification results based on Fourier or wavelet coefficients were less favorable than those directly obtained from principal component scores of the original spectra. LS-SVM did not perform well on test samples. Fishers criterion (FC) was used to select eight wavelength features, and it was demonstrated that LDA, k-NN, and PNN classifiers could effectively discriminate wheat classes based on reflectance spectra.