高光谱成像的三七粉质量等级无损鉴别

三七粉是三七的主要消费和商品形式，市场上存在以次充好、甚至是掺假的现象，由于是粉状物料，难以用肉眼判别，为了实现对不同质量等级的三七粉进行无损鉴别。将30头、40头、60头和80头的三七主根研磨成粉，制备样本。采用可见近红外高光谱成像系统（400.68～1 001.61 nm）采集4种不同头数三七粉，共计384个样品的高光谱图像，提取高光谱图像感兴趣区域(ROI)的平均光谱值作为样本原始光谱。将384个三七粉样本按2∶1的比例划分训练集和测试集。采用卷积平滑（SG）、多元散射校正（MSC）和标准正态变量变换（SNV）3种预处理方法对三七粉样本光谱信息进行预处理并建立支持向量机(SVM)分类模型，通过比较基于3种预处理方法的SVM模型测试集分类准确率，确定SNV为最优预处理方法。采用迭代保留信息变量（IRIV）、变量组合集群分析（VCPA）和变量组合集群分析混合迭代保留信息变量（VCPA-IRIV）3种特征选择方法提取SNV预处理后光谱的特征波长并建立基于特征光谱和原始光谱的SVM分类模型，通过比较基于3种特征选择方法得到的特征波长建立的SVM模型测试集分类准确率，发现将VCPA与IRIV相结合的VCPA-IRIV为最优特征选择方法。VCPA-IRIV提取了18个特征波长代替全光谱数据参与建模，该算法在降低模型复杂度的同时保持了模型的分类精度。为了提高模型的分类精度，采用引力搜索算法（GSA）对SVM模型中惩罚因子c和核参数g进行寻优，并与网格搜索（GS）的结果进行比较，结果表明，VCPA-IRIV-GSA-SVM模型分类效果最好，测试集分类准确率达到100%。可见，利用可见近红外高光谱成像对三七粉进行质量等级无损鉴别是可行的，为市场上三七粉的质量等级鉴别提供了参考。

Study on Nondestructive Identification of Panax Notoginseng Powder Quality Grade Based on Hyperspectral Imaging Technology

Panax notoginseng powder is the main consumption and commodity form of panax notoginseng. There are shoddy or even adulterated phenomena in the market. As panax notoginseng powder is a powdery material, it is not easy to distinguish with the naked eye. In order to identify the quality grade of panax notoginseng powder, visible near-infrared hyperspectral imaging technology was used to identify the panax notoginseng powder with different quality grades. The taproots of panax notoginseng of 30 heads, 40 heads, 60 heads and 80 heads were ground into powder to prepare samples. The hyperspectral image of 384 samples of four quality grades was acquired by using a visible near-infrared hyperspectral imaging system(400.68～1 001.612 nm). Region of interest (ROI) was extracted from the hyperspectral image, and the average spectral value of samples was calculated. 384 samples of panax notoginseng powder were divided into training sets and test sets in a ratio of 2∶1. The original spectra of panax notoginseng powder were preprocessed using multiplication scatter correction (MSC), Savitzky-Golay (SG) and standard normal variable (SNV), and the support vector machine (SVM) was employed to form the classification models based on MSC, SG and SNV. By comparing the classification accuracy of SVM models based on MSC, SG and SNV, it was found that SNV had the best effect on preprocessing. Iterative reserved information variable (IRIV), variable combined cluster analysis (VCPA) and variable combined cluster analysis and iterative reserved information variable (VCPA-IRIV) were adopted to extract feature wavelengths from the spectra after SNV pretreatment, and the SVM was employed to form the classification models based on feature spectra and original spectra. By comparing the range of feature wavelengths and the classification accuracy of SVM models based on IRIV, VCPA and VCPA-IRIV, it was found that VCPA-IRIV, which combines VCPA and IRIV, had the best effect on feature selection. VCPA-IRIV extracted 18 feature wavelengths to participate in the modeling instead of the full spectra, and the algorithm can reduce the complexity of the model while maintaining the model’s classification accuracy. In order to improve the classification accuracy of the model, the gravitational search algorithm (GSA) was introduced to search the optimal parameters(c,g) in the SVM model and compared with Grid Search (GS). The results indicated that the VCPA-IRIV-GSA-SVM model has the best classification effect, and the classification accuracy of the test set reached 100%. Thus, it is feasible to use visible near-infrared hyperspectral imaging technology to identify the quality grade of panax notoginseng powder. This method references the quality grade identification of panax notoginseng powder in the market.