【摘要】 茶叶是大众青睐的健康饮品之一，但茶叶在机器采收和加工过程中，容易混入茶梗和昆虫异物，污染茶叶、影响其质量安全，是未来应防范和检测的重点。X射线成像技术，根据食品基质和异物的密度差实施检测，广泛适用于金属异物并延伸至高密度塑料，但对于茶梗、昆虫这类低密度有机异物尚不适用，所以迫切需要研发新型无损检测技术和方法。针对片状茶叶重叠、遮掩异物的问题，提出了电磁振动上料辅助近红外光谱和荧光图像的检测方案，进行绿茶中的内源性异物茶梗和外源性异物昆虫的在线检测研究。通过电磁振动上料辅助近红外光谱和荧光成像系统，采集了600～1 050 nm范围的近红外光谱600条和RGB-N四通道图像各65幅。采用451条光谱进行建模，其余149条光谱作为预测集，评估模型的性能，比较了去趋势(Detrending)、多元散射校正(MSC)、标准正态变换(SNV)、变权重正态变换(VSN)、迭代自适应加权惩罚最小二乘法(airPLS)、不对称最小二乘法(ALS)、光程估计与校正(OPLEC)等不同光谱预处理方法的处理效果，其中OPLEC能较好地消除散射效应，偏最小二乘判别分析(PLS-DA)模型的正确识别率由78%提高到85%。结果表明，近红外光谱结合OPLEC能够更加准确地检出茶叶中的茶梗和昆虫异物。与近红外光谱的精确点测量相比，图像提供了较大范围的检测手段。对65幅蓝色(B通道)图像进行分析，采用最大类间方差法进行阈值分割后，对图像进行取反运算、中值滤波、连通区域标记和特征提取，提取出长轴长、短轴长、长短轴比、离心率四个特征变量，采用267个感兴趣目标，建立了线性判别分析(LDA)模型，用未参与建模的88个感兴趣目标来评价模型的预测能力，正确识别率达到了64%。实验结果表明，电磁振动上料辅助近红外光谱和荧光图像在线检测茶梗和昆虫异物具备可行性，为食品中有机异物在线检测提供了低成本的解决方案。更多还原

【Abstract】 Tea is one of the health drinks favored by the consumer, but during the process of tea machine harvesting and processing, it is easy to be mixed with tea stalks and foreign insect bodies. It resulted in pollution and influenced the quality and safety of tea products. In the future, we should focus on preventing and detecting of foreign bodies. X-ray imaging technology, based on the density difference between food substrate and foreign bodies, is widely applied to detect metal foreign bodies and extended to high-density plastic. However, it is not suitable for low-density organic foreign bodies such as tea stem insects, so it is urgent to develop a new and non-destructive detection technology and method. In order to solve the problem of overlapping and covering foreign bodies in tea leaves, a scheme of electromagnetic vibration feeding assisted near-infrared spectroscopy(NIRS), and fluorescence image was proposed to online detect endogenous foreign bodies of tea stalks and insects.A total of 600 NIRSranging from 600 to 1 050 nm, and 65 channel images including R, G, B and N were collected by electromagnetic vibration-assisted NIRS and fluorescence imaging system. Among them, 451 spectra were used to develop the model, and the remaining 149 spectra were used to evaluate model performance. The effects of different correction methods such as detrending, multiplicative scatter correction(MSC), standard normal variate transformation(SNV), variable sorting for normalization(VSN), adaptive iteratively reweighted penalized least squares(airPLS), alternative least squares(ALS),optical path length estimation and correction(OPLEC) were compared. OPLEC could eliminate the scattering effect better, and the correct recognition rate of the partial least squares discriminant analysis(PLS-DA) model of NIRS increased from 78% to 85%. The results showed that the calibration method of OPLEC combined with the PLS-DA model could-detect foreign bodies in tea more accurately.Compared with the accurate measurement of NIRS, imaging technologyprovided a wider range of detection means. Sixty-five clear blue(B) channel images were analyzed. Using threshold segmentation by maximum interclass variance method, inversing operation, median filtering, connected component labeling and feature extraction, we extracted four feature variables of long axis length, short axis length, short axis ratio and eccentricity, a total of 355 objects of interest.The linear discriminant analysis(LDA) model was established with 267 interesting targets, and 88 interested targets not involved in modeling were used to evaluate the model’s prediction ability. The correct recognition rate reached 64%.The experimental results show that electromagnetic vibration feeding assisted NIRS and fluorescence image is feasible for online detection of tea stalk and foreign insect bodies, providing a low-cost solution for online detection of organic foreign bodies in food.更多还原