Williams Paul J, Kucheryavskiy Sergey
Department of Food Science, Stellenbosch University, Private Bag X1, Matieland (Stellenbosch) 7602, South Africa.
Department of Chemistry and Bioscience, Aalborg University, Esbjerg, Denmark.
Food Chem. 2016 Oct 15;209:131-8. doi: 10.1016/j.foodchem.2016.04.044. Epub 2016 Apr 20.
NIR hyperspectral imaging was evaluated to classify maize kernels of three hardness categories: hard, medium and soft. Two approaches, pixel-wise and object-wise, were investigated to group kernels according to hardness. The pixel-wise classification assigned a class to every pixel from individual kernels and did not give acceptable results because of high misclassification. However by using a predefined threshold and classifying entire kernels based on the number of correctly predicted pixels, improved results were achieved (sensitivity and specificity of 0.75 and 0.97). Object-wise classification was performed using two methods for feature extraction - score histograms and mean spectra. The model based on score histograms performed better for hard kernel classification (sensitivity and specificity of 0.93 and 0.97), while that of mean spectra gave better results for medium kernels (sensitivity and specificity of 0.95 and 0.93). Both feature extraction methods can be recommended for classification of maize kernels on production scale.
硬、中、软。研究了两种方法，即逐像素法和逐对象法，以根据硬度对玉米粒进行分组。逐像素分类为单个玉米粒的每个像素分配一个类别，但由于误分类率高，结果不尽人意。然而，通过使用预定义阈值并基于正确预测像素的数量对整个玉米粒进行分类，取得了改进的结果（灵敏度和特异性分别为0.75和0.97）。逐对象分类使用两种特征提取方法进行——得分直方图和平均光谱。基于得分直方图的模型在硬玉米粒分类方面表现更好（灵敏度和特异性分别为0.93和0.97），而平均光谱模型在中等硬度玉米粒分类方面效果更佳（灵敏度和特异性分别为0.95和0.93）。这两种特征提取方法都可推荐用于生产规模的玉米粒分类。
