为无损检测桃的内部品质, 提出了结合高光谱成像技术和人工神经网络无损检测桃硬度的方法。采集了摘后贮藏了12 d的140个桃在900~1 700 nm的高光谱图像, 以每个桃高光谱图像中40 pixel×40 pixel的感兴趣区域的平均光谱作为桃的原始反射光谱; 利用Savitzky-Golay平滑和标准正态变量变换对光谱进行预处理; 基于x-y共生距离算法划分样本, 得到校正集样本105个和预测集样本35个。利用连续投影算法、无信息变量消除法和正自适应加权算法从全光谱的216个波长中分别提取了12个、103个和22个特征波长; 分别建立了基于全光谱和提取的特征波长预测桃硬度的支持向量机模型和BP网络模型。结果表明, 基于全光谱建立的BP网络模型具有最好的预测性能, 其预测相关系数为0.856, 预测均方根误差为0.931。本研究为基于桃内部品质的工业化分级提供了基础。

To explore a nondestructive method to measure peach internal quality, a hyperspectral imaging technology combined with Artificial Neural Networks (ANN) was applied to evaluate the firmness of intact peaches. The hyperspectral images of 140 peaches during 12 day storage were acquired from 900 nm to 1 700 nm, and the average reflective spectrum of interest region of 40 pixel×40 pixel in each image was calculated and was used as the original spectra. The spectra were preprocessed by Savitzky-Golay smoothing and the standard normal variate. The sample set was partitioned based on joint x-y into calibration sets (105) and prediction sets (35). Then the successive projection algorithm, uninformative variable elimination method and competitive adaptive reweighted sampling method were used to select characteristic wavelengths by 12, 103 and 22 from 216 wavelengths, respectively. A support vector machine and an error back propagation (BP) network model were established based on full spectra and selected characteristic wavelengths for predicting the firmness of intact peaches. The result shows that BP model based on full spectra has the best prediction performance with a correlation coefficient and a root-mean-square error of 0.856 and 0.931, respectively. This study offers the base for identifying internal qualities of peaches in industry.