Detecting Vitreous Wheat Kernels Using Reflectance and Transmittance Image Analysis

Cereal Chem. 81(5):594–597 The proportion of vitreous durum kernels in a sample is an important grading attribute in assessing the quality of durum wheat. The current standard method of determining wheat vitreousness is performed by visual inspection, which can be tedious and subjective. The objective of this study was to evaluate an automated machine-vision inspection system to detect wheat vitreousness using reflectance and transmittance images. Two subclasses of durum wheat were investigated in this study: hard and vitreous of amber color (HVAC) and not hard and vitreous of amber color (NHVAC). A total of 4,907 kernels in the calibration set and 4,407 kernels in the validation set were imaged using a Cervitec 1625 grain inspection system. Classification models were developed with stepwise discriminant analysis and an artificial neural network (ANN). A discriminant model correctly classified 94.9% of the HVAC and 91.0% of the NHVAC in the calibration set, and 92.4% of the HVAC and 92.7% of the NHVAC in the validation set. The classification results using the ANN were not as good as with the discriminant methods, but the ANN only used features from reflectance images. Among all the kernels, mottled kernels were the most difficult to classify. Both reflectance and transmittance images were helpful in classification. In conclusion, the Cervitec 1625 automated visionbased wheat quality inspection system may provide the grain industry with a rapid, objective, and accurate method to determine the vitreousness of durum wheat. Vitreousness is an important international grading attribute in assessing the quality of durum wheat. Higher vitreousness indicates higher protein content, a harder kernel, coarser granulation, higher yield of semolina, superior pasta color, improved cooking quality, and opportunity for premium sales pricing (Dowell 2000). Vitreous kernels are glassy and translucent, while nonvitreous kernels are chalky and opaque. Some minor defects such as bleached, cracked, or checked hard vitreous kernels are considered vitreous. In contrast, sprouted kernels, foreign materials, scabby kernels, etc., are considered nonvitreous. The current standard method of evaluating vitreousness of durum wheat in the United States is by manually inspecting a 15-g sample that is free of shrunken and broken kernels (USDA 1997). Because it is a subjective method, inspectors may disagree on classification. Therefore, an objective, automated, reproducible, and rapid method for determining durum wheat vitreousness is needed. Such a grading method should greatly reduce grain inspectors’ subjectivity and labor, and benefit wheat producers, processors, and handlers. Researchers have investigated various methods for evaluating durum wheat vitreousness. A perfect match with inspector classifications of obviously vitreous or nonvitreous durum wheat using near-infrared spectroscopy (NIRS) was reported by Dowell (2000). Classification rates of 91.1–97.1% were reported when studying dark, hard, vitreous, and nonvitreous hard red spring wheat using NIRS (Wang et al 2002). A Perten 4100 single-kernel characterization system was applied to detect wheat vitreousness (Sissons et al 2000; Nielsen et al 2003). This method was faster than the NIRS method, but classification capability was limited. Sissons et al (2000) reported ≈25–35% error of prediction with SKCS. Machine vision techniques have been used to determine grain quality and classify grain cultivars based on color and geometry features (Zayas et al 1996; Ruan et al 1998; Luo et al 1999; Majumdar and Jayas 2002a,b). Image analysis has also been applied to study wheat vitreousness (Symons et al 2003; Wang et al 2003). Wang et al (2003) developed an artificial neural network (ANN) model using reflectance images captured by a real-time image-based grain-inspection machine (Foss GrainCheck-310). Classification rates of vitreous and nonvitreous subclasses were 85–90% (Wang et al 2003). Symons et al (2003) found significant agreement between inspector-determined and machine-determined percentages of hard vitreous kernels in samples. Transmittance images of individual kernels were imaged with a monochrome camera. It was hypothesized that using both reflectance and transmittance images would improve classification rate of vitreousness. The objective of this study was to evaluate an automated machine vision inspection system for detecting wheat vitreousness using both reflectance and transmittance images captured by a new grain inspection system. MATERIALS AND METHODS Sample Preparation The Grain Inspection, Packers, and Stockyards Administration (GIPSA) of the United States Department of Agriculture collected samples for this study. Durum wheat samples were classified into two subclasses: hard and vitreous of amber color (HVAC) and not hard and vitreous of amber color (NHVAC). For this study, the HVAC subclass was further separated into three categories and NHVAC into six categories (Table I). Samples were visually reinspected by the Board of Appeals and Review (BAR) to check correct segregation. Samples in each category were divided into 1 Biological and Agricultural Engineering Department, Kansas State University, 147 Seaton Hall, Manhattan, KS 66506. 2 Corresponding author. E-mail: [email protected] 3 USDA-ARS, Grain Marketing and Production Research Center, 1515 College Avenue, Manhattan, KS 66502. *The e-Xtra logo stands for “electronic extra” and indicates that the online version contains a color version of Fig. 1 not included in the print edition. Publication no. C-2004-0728-01R. This article is in the public domain and not copyrightable. It may be freely reprinted with customary crediting of the source. American Association of Cereal Chemists, Inc., 2004. TABLE I Summary of Sample Categories and Sizes Used in This Study Subclass and Category Calibration Validation Reproducibility