水果新鲜度是反映水果是否新鲜、 饱满的重要品质指标, 为了探讨水果不同货架期的预测和判别方法, 以酥梨为研究对象, 利用高光谱成像技术, 结合偏最小二乘判别法(PLS-DA)和偏最小二乘支持向量机(LS-SVM)算法对酥梨货架期进行判别。 由光源、 成像光谱仪、 电控位移平台和计算机等构成的高光谱成像装置采集样品光谱, 装置光源采用额定功率为200 W四个溴钨灯泡成梯形结构设计, 光谱范围为1 000~2 500 nm, 分别率为10 nm。 选取优质酥梨30个, 货架期设置为1, 5和10 d, 对30个样品完成3次光谱图像的采集, 并矫正原始图像。 实验结果表明: 基于图像的酥梨货架期定性分析时, 对不同货架期样品的原始图像进行PCA压缩, 得到三种不同货架期的权重系数数据, PC1图像提取特征波长点为1 280, 1 390, 1 800, 1 880和2 300 nm, 以特征图像的平均灰度值作为自变量且以货架期作为因变量建立定性判别模型, 建模集68个, 预测集22个。 最小二乘支持向量机以RBF为核函数时, 预测集中样品的误判个数为1, 误判率为4.5%。 而当采用lin核函数时, 样品的误判个数为0, 误判率为0。 PLS-DA定性分析时RMSEC为1.24, R_c为0.93。 RMSEP为1, R_p为0.96, 预测集误判率为0。 特征图像对酥梨货架期判别LS-SVM中的lin核函数所建立的模型结果较好, 优于RBF核函数的建模效果, 也优于PLS-DA判别模型。 ENVI软件提取实验样品光谱后建立LS-SVM和PLS-DA判别模型, LS-SVM利用RBF和lin核函数误判率分别为4.5%和0。 与RBF核函数相比, lin核函数所建立的模型预测酥梨货架期的效果更好。 PLS-DA方法主成分因子数为12, RMSEC和RMSEP分别为0.48和0.78, R_c和 R_p分别为0.99和0.97, 建模集与预测集的误判率均为零。 LS-SVM中的lin核函数所建立的模型结果较好, 依然优于PLS所建立的检测模型。 酥梨的光谱信息结合LS-SVM可以实现对酥梨货架期的检测和判别。 基于图像建立酥梨的货架期预测模型与光谱相比, 都实现了酥梨货架期的判别, 而特征图像法, 选择区域较少流失部分信息, 计算量小, 建模结果相对略差。 酥梨货架期的高光谱成像检测模型研究为消费者正确评价水果新鲜度提供了理论指导, 也为后期果水果货架期检测仪器的开发提供了技术支持。

Fruit freshness is an important quality index reflecting whether the fruit is fresh and full. In order to explore the prediction and discrimination methods of different shelf life of fruits, this paper takes the pear as the research object, and uses hyperspectral imaging technology combined with partial least squares discrimination (PLS). DA and partial least squares support vector machine (LS-SVM) algorithm to distinguish the shelf life of pears. The spectrum of the sample is collected by a high-spectrum imaging device consisting of a light source, an imaging spectrometer, an electronically controlled displacement platform, and a computer. The device light source is designed with a ladder power of 200 W four bromine tungsten bulbs, and the spectral range is 1 000~2 500 nm. 10 nm. The material was selected from 30 high-quality pears, and the shelf life was set to 1 day, 5 days and 10 days. Three spectral images were acquired for 30 samples and the original image was corrected. The experimental results show that the image-based analysis of the shelf life of the pears is carried out by PCA compression of the original images of different shelf life samples, and the weight coefficient data of three different shelf periods are obtained. The wavelength points of PC1 image extraction are 1 280, 1 390 and 1 800 nm. 1 880 and 2 300 nm, with the average gray value of the feature image as the independent variable and the shelf life as the dependent variable to establish a qualitative discriminant model, 68 modeling sets and 22 prediction sets. When the least squares support vector machine uses RBF as the kernel function, the number of misjudgments in the predicted concentrated samples is 1, and the false positive rate is 4.5%. When the lin kernel function is used, the number of misjudgments of the sample is 0, and the false positive rate is 0. The RMSEC for PLS-DA qualitative analysis was 1.24, which was 0.93. The RMSEP is 1, which is 0.96, and the prediction set false positive rate is zero. The characteristic image has better model for the lin kernel function in the LS-SVM of the shelf life of the pear, which is better than the modeling effect of the RBF kernel function and better than the PLS-DA discriminant model. The LS-SVM and PLS-DA discriminant models were established by ENVI software to extract the spectra of the experimental samples. The false positive rates of RB-SVM using RBF and lin kernel functions were 4.5% and 0, respectively. Compared with the RBF kernel function, the model established by the lin kernel function predicts the shelf life of the pears better. The PLS-DA method has a principal component factor of 12, RMSEC and RMSEP of 0.48 and 0.78, respectively, and 0.99 and 0.97, respectively. The false positive rate of the modeling set and the prediction set are both zero. The model established by the lin kernel function in LS-SVM is better than the detection model established by PLS. The spectral information of the pears combined with LS-SVM can realize the detection and discrimination of the shelf life of the pears. Compared with the spectrum, the shelf life prediction model based on the image was used to distinguish the shelf life of the pear, while the feature image method, the selected area was less lost part of the information, the calculation amount was small, and the modeling result was relatively poor. The research on the hyperspectral imaging detection model of the shelf life of the pear provides theoretical guidance for consumers to correctly evaluate the freshness of the fruit, and also provides technical support for the development of the fruit shelf detection instrument in the later stage.