构建了高光谱反射、 透射和交互作用成像系统对蓝莓的硬度和弹性模量进行无损检测, 并对比不同成像模式的预测准确率。 反射高光谱图像采用以大津法为核心的算法进行分割, 而透射和交互作用高光谱图像利用基于区域增长的算法进行分割。 对提取平均光谱分别进行标准正态变量变换(SNV)和一阶SG卷积平滑(Der), 并构建相应光谱的最小二乘支持向量机模型。 在全波段模型中, 基于SNV预处理反射光(Reflectance-SNV)模型对蓝莓硬度的预测相关系数(Rp)=0.80, 相对预测误差(RPD)=1.76; 基于SNV预处理透射光(Transmittance-SNV)模型对蓝莓弹性模量的Rp(RPD)=0.78 (1.74)。 随机蛙跳算法(Random Frog)可以有效地减少了建模所需的波段数, 同时还提高了大部分模型的预测准确率。 基于随机蛙跳选择的Der交互作用光(Random Frog-interactance-Der)模型对蓝莓硬度的Rp(RPD)=0.80(1.83), 但该模型建模所需的波段数为140; 基于随机蛙跳算法的SNV透射光(Random Frog-Transmittance-SNV)模型对蓝莓弹性模量的Rp(RPD)=0.82(1.83), 同时该模型建模所需的波段数为20。

In this study, a imaging system with hyperspectral reflectance, transmittance and interactance was constructed for estimate the firmness and elastic modulus of blueberry. The comparisons of these three imaging modes were carried out. This hyperspectral system could also be applied for scattering modewhile this mode was not suitable for small fruit such as blueberry. The reflectance hypercubes were segmented with the algorithm based on the Otsu method, and the transmittance and interactance hypercubes were processed with the algorithms based on region growing approach. Subsequently, the extracted spectra were pretreated with the Standard Normal Variate (SNV) and Savitzky-Golay of the first derivative (Der), and least squares-support vector machine was applied for the establishment of the corresponding prediction models. The obtained results demonstrated that -reflectance-SNV model could predict blueberry firmness with correlation coefficient of prediction sample set (Rp) of 0.80 and the ratio of percent deviation (RPD) of 1.76 among the models using full spectra. The elastic modulus of blueberry was better estimated by the full transmittance spectra subjected to SNV pretreatment with Rp (RPD) of 0.78 (1.74) than the other models. Furthermore, Random Frog selection approach could to some extent reduce the uninformative wavelengths while increasing the prediction accuracy of the established models. Random Frog-Interactance-Der model achieved Rp (RPD) of 0.80 (1.83) for blueberry firmness, but the number of wavelength was 140. In the case of blueberry elastic modulus, random frog-transmittance-SNV showed the relatively superior performance compared to the other models, with Rp (RPD) of 0.82 (1.83) and fewer wavelength number of 20.