许多不同的稻谷品种看起来很相似, 但它们的化学成分和最终产品质量却有很大差别, 每年因品种混淆而造成巨大的经济损失, 对稻谷品种的鉴别是发展优质粮食工程的现实需要, 为此提出了一种采用高光谱成像技术实现稻谷品种无损快速鉴别的方法。 主要研究内容和结果如下: (1)在全波段388～1 000 nm范围内采集5个品种共150粒的稻谷高光谱反射率数据, 筛选出差异明显的波段(600～800 nm), 将此波段内每个品种的反射率进行Stacked计算和curve-smoothing平滑处理以增加其区分度。 (2)对5种稻谷经平滑处理后的反射率数据做主成分分析, 找到权值系数最大的波长位于680 nm, 将其作为特征波长。 加载特征波长下的纹理图像, 计算每粒稻谷样品的纹理特征参数: 均值(Mean)、 方差(Variance)、 信息熵(Entropy)和偏差(Skewness)。 利用阈值分割的方法将目标与背景区分开, 计算每粒稻谷形态特征参数: 面积像素数/pixels2、 边界的周长/pixels、 长轴长度/pixels、 短轴长度/pixels。 结合稻谷的纹理特征参数和形态特征参数, 比较Fisher判别分析模型、 偏最小二乘回归模型(PLSR)和人工神经网络模型(ANN)对稻谷品种鉴别的效果。 (3)结果显示, Fisher判别分析中函数1和函数2的累计方差贡献率达到93%, 能够较好地解释稻谷的品种信息。 将样本的函数值与组质心的平方马氏距离(Mahalanobis)做比较, 值相近的作为同一分组类别, 对稻谷品种的整体识别正确率能达到95.3%; 偏最小二成回归模型: Y品种=0.03X均值-0.36X方差-0.24X信息熵+0.37X偏差+0.31X面积-0.32X周长-0.39X长轴长度+0.45X短轴长度, 该回归模型相关系数r=0.98, 校正均方根RMESS=0.29, 交叉验证均方根PMESSCV=0.32, 对稻谷的品种鉴别正确率能达到95%; 构建的ANN模型为具有sigmoid隐含和softmax输出神经元的双层前馈网络, 对150个样品按70%∶15%∶15%的比例随机划分训练集、 测试集、 验证集, 选择共轭梯度法(scaled conjugate gradient)作为训练算法, 以交叉熵(cross-entropy)作为模型的评价指标, 对稻谷品种鉴别的正确率可达到98%。 稻谷品种鉴别的ANN模型在分类精度上优于Fisher判别和PLSR, 选择特征波长下的图像信息建立稻谷品种识别的ANN模型, 对稻谷品种的无损快速鉴别具有重要指导意义。

Many different varieties of rice look very similar, but their chemical composition and final product quality vary greatly, which causes huge economic losses each year as a result of variety confusion. Identification of rice varieties is the practical requirement for developing high quality grain engineering. In this paper, a fast and non-destructive method for rice variety identification using hyperspectral imaging technology was proposed. The main research contents and results were as follows: (1) Average spectrawere extracted from the region of total 150 samples with wavelength from 388～1 000 nm. In the full band, the reflectance was most obvious at 600～800 nm, which was calculated by Stacked stacking and curve-smoothing for increasing its differences. (2) Principal component analysis (PCA) was used to analyze the reflectance data smoothed. It was found that the wavelength with the largest weight coefficient was located at 680 nm and used as the characteristic wavelength. Loading the texture image of the characteristic wavelengths, the texture characteristic parameters of each rice sample were calculated as follows: Mean, Variance, Entropy and Skewness. Meanwhile, the thresholding method was used to separate the target from the background, and the morphological parameters of each grain werecalculated as follows: areas/pixels2, perimeter/pixels, length of long axis/pixels, length of short axis/pixels. Based on the texture characteristics and morphological characteristics, the Fisher discriminant analysis model, partial least squares regression (PLSR) mode and Artificial neural network model (ANN) were established respectively for rice variety identification. (3) The results showed that the cumulative variance contribution rate of function 1 and function 2 established by Fisher discriminant analysis reached 93%, which could better explain the rice variety information. Comparing the function value of the sample with the square Mahalanobis distance of the group centroid, the individuals with similar values were taken as the same category. The overall recognition accuracy of the five rice varieties could reach 95.3%. The PLSR model: Yvarieties=0.03Xmeans-0.36Xvarious-0.24Xentropy+0.37Xskewness+0.31Xarea-0.32Xperimeter-0.39Xlength of long axis+0.45Xlength of short axis, with correlation coefficient (r)=0.98, corrected root mean square (RMESS)=0.29, cross validation root mean square (RMESSCV)=0.32, the accuracy of rice varieties identification could reach 95%. The neural network model is a two-layer feedforward network with sigmoid hidden and soft max output neurons, which randomly divides 150 samples into training samples, validation sets and test sets according to the ratio of 70%∶15%∶15%. With training algorithm of conjugate gradient method and evaluation index of Cross-Entropy method, the accuracy of rice variety identification can reach 98%. The overall results show that the neural network model of rice variety identification is superior to Fisher discriminant and PLSR in classification accuracy, which has an important guiding significance for rapid and non-destructive identification of rice varieties.