基于图像和光谱融合的脐橙货架期高光谱成像无损检测研究

水果货架期是影响水果品质的重要因素之一，快速无损检测货架期是消费者、食品加工企业日益关心的问题，为了探讨水果不同货架期的预测判别方法的可行性，以不同货架期脐橙为实验样品，运用高光谱成像技术并结合化学计量学方法对不同货架期脐橙进行了预测判别。分别采集脐橙货架期第0天、第7天、14天后的脐橙样本高光谱图像，并进行高光谱图像校正。从光谱角度，提取脐橙样本的平均光谱，每条光谱有176个波长点；从图像角度，先提取脐橙样本的RGB和HSI颜色空间中R，G，B，H，S和I特征值，得到6个分量的均值，然后提取灰度共生矩阵的能量、熵、对比度、逆差矩、相关性的5个图像纹理信息，一共11个图像特征值，并将图像特征进行归一化处理；结合光谱和图像信息，即176个原始光谱和11个图像信息一共187个特征值。利用光谱信息、图像信息、光谱和图像融合信息进行建模，分别建立偏最小二乘支持向量机（LS-SVM）和偏最小二乘判别（PLS-DA）模型。当原始176个光谱变量作为输入变量，核函数为LIN-Kernel时，LS-SVM模型预测效果最佳，预测集误判率为5.33%。当11个图像特征变量作为输入变量，核函数为LIN-Kernel时，LS-SVM模型预测效果最佳，预测集误判率较高为20%。当原始176个光谱变量和11个图像特征变量的融合特征作为输入变量，核函数为LIN-Kernel时，LS-SVM模型预测效果最佳，预测集误判率为1.33%。实验结果表明，以光谱和图像融合信息建立LS-SVM模型效果最优，提高了对不同货架期脐橙识别的正确率，可实现对不同货架期的脐橙准确有效分类识别，误判率为1.33%。利用高光谱成像技术对不同货架期脐橙进行快速判别，对消费者购买新鲜水果和水果深加工企业具有一定程度的理论指导，也为后期相关仪器研发奠定了基础。

Research on Non-Destructive Testing of Navel Orange Shelf Life Imaging Based on Hyperspectral Image and Spectrum Fusion

Fruit shelf life is one of the important factors affecting fruit quality. Rapid non-destructive testing of fruit shelf life is an increasingly concerned issue for consumers and food processing enterprises. In order to explore the feasibility of prediction and discrimination methods for different shelf life of fruits, navel oranges with different shelf life were used as experimental samples, and hyperspectral imaging technology combined with chemometric methods were used to predict and discriminate navel oranges with different shelf life. The hyperspectral images of navel orange samples on day 0, day 7 and day 14 of the shelf life of navel orange were collected and corrected. From the spectral point of view, the average spectrum of navel orange samples was extracted, each spectrum had 176 wavelength points ; from the perspective of image, the R, G, B, H, S and I eigenvalues of navel orange samples in RGB and HSI color space were extracted, and the mean values of six components were obtained. Then, five image texture information of energy, entropy, contrast, inverse moment and correlation of gray level co-occurrence matrix were extracted, and a total of 11 image eigenvalues were extracted, and the image features were normalized. Combining spectral and image information, namely 176 original spectral and 11 image information, a total of 187 eigenvalues. Partial least squares support vector machine ( LS-SVM ) and partial least squares discriminant analysis ( PLS-DA ) models were established by using spectral information, image information, spectrum and image fusion information. When the original 176 spectral variables are used as input variables and the kernel function is LIN-Kernel, the LS-SVM model has the best prediction effect, and the misjudgment rate of prediction set is 5.33%. When 11 image feature variables are used as input variables and the kernel function is LIN-Kernel, the LS-SVM model has the best prediction effect, and the misjudgment rate of prediction set is 20%. When the fusion features of the original 176 spectral variables and 11 image feature variables are used as input variables and the kernel function is LIN-Kernel, the LS-SVM model has the best prediction effect, and the misjudgment rate of the prediction set is 1.33%. The experimental results show that the LS-SVM model based on spectral and image fusion information has the best effect, which improves the accuracy of navel orange recognition in different shelf life, and can realize accurate and effective classification and recognition of navel oranges in different shelf life. The misjudgment rate is 1.33%. The rapid identification of navel oranges in different shelf life by hyperspectral imaging technology has a certain degree of theoretical guidance for consumers to purchase fresh fruit and fruit deep processing enterprises, and lays a foundation for the development of related instruments in the future.