| 段鹍,陈心秀,杨琦,丁美宙,陈洋,王海滨,陈顺辉,郭炳胜,潘磊庆.基于短波红外高光谱成像的再造烟叶成分指标无损检测与可视化分析[J].中国造纸,2026,45(9):113-122 |
二维码（扫一下试试看！） |
| 基于短波红外高光谱成像的再造烟叶成分指标无损检测与可视化分析 |
| Nondestructive Detection and Visualization Analysis of Component Indicators in Reconstituted Tobacco Based on Short-wave Infrared Hyperspectral Imaging |
| 收稿日期：2026-03-16 修订日期：2026-04-17 |
| DOI：10.11980/j.issn.0254-508X.2026.09.013 |
| 关键词: 再造烟叶 高光谱成像 卷积神经网络 无损检测 可视化 |
| Key Words:reconstituted tobacco hyperspectral imaging convolutional neural network non-destructive detection visualization |
| 基金项目:国家重点研发计划2022YFF1100801河南中烟工业有限责任公司科技项目AW2024047国家重点研发计划（2022YFF1100801）；河南中烟工业有限责任公司科技项目（AW2024047）。 |
| 作者 | 单位 | 邮编 | | 段鹍* | 1河南中烟工业有限责任公司,河南郑州,450000 | 450000 | | 陈心秀 | 2南京农业大学食品科技学院, 江苏南京,210000 | 210000 | | 杨琦 | 1河南中烟工业有限责任公司,河南郑州,450000 | 450000 | | 丁美宙 | 1河南中烟工业有限责任公司,河南郑州,450000 | 450000 | | 陈洋 | 1河南中烟工业有限责任公司,河南郑州,450000 | 450000 | | 王海滨 | 1河南中烟工业有限责任公司,河南郑州,450000 | 450000 | | 陈顺辉 | 1河南中烟工业有限责任公司,河南郑州,450000 | 450000 | | 郭炳胜 | 1河南中烟工业有限责任公司,河南郑州,450000 | 450000 | | 潘磊庆* | 2南京农业大学食品科技学院, 江苏南京,210000 | 210000 |
|
| 摘要点击次数: 529 |
| 全文下载次数: 77 |
| 摘要:本研究利用短波红外（SWIR）高光谱成像系统采集再造烟叶样品在波长1 000~2 000 nm范围的光谱数据，系统比较自动标准化（Autoscale）、一阶导数（1st D）、多元散射校正（MSC）的预处理方法，以及竞争性自适应重加权采样法（CARS）和无信息变量消除法（UVE）的特征波长筛选效果，再分别运用随机森林（RF）、偏最小二乘回归（PLS）、多元线性回归（MLR）和卷积神经网络（CNN）算法，建立了再造烟叶含水率、总氮含量和水溶性总糖含量的最优预测模型。结果表明，特征波长筛选可显著提升模型预测性能，UVE-1st D-CNN模型在含水率预测中表现最佳，性能偏差比（RPD）=5.574 1，UVE-MSC-CNN模型在总氮含量预测中取得最优效果（RPD=4.499 2），CARS-1st D-CNN模型在水溶性总糖含量预测中达到最高精度（RPD=6.070 8）。经最佳CNN模型反演后，各指标可视化分布图像的色彩变化与实测值具有良好对应关系，可通过颜色差异区分不同指标含量区间的再造烟叶样品，实现了指标含量的无损检测和空间分布的可视化成像。 |
| Abstract:This study utilized a short-wave infrared (SWIR) hyperspectral imaging system to collect spectral data of reconstituted tobacco samples within the wavelength range of 1 000 to 2 000 nm. The study systematically compared the pre-processing methods of automatic standardization (Autoscale), first derivative (1st D), and multiplicative scatter correction (MSC), as well as the feature wavelength selection effects of competitive adaptive reweighted sampling (CARS) and variable elimination without information (UVE). Then, the random forest (RF), partial least squares regression (PLS), multiple linear regression (MLR), and convolutional neural network (CNN) algorithms were applied to establish the optimal prediction models for the moisture content, total nitrogen content, and water-soluble total sugar content of reconstituted tobacco. The results showed that feature wavelength selection significantly improved the model prediction performance. The UVE-1st D-CNN model performed best in moisture content prediction, with a performance deviation ratio (RPD) of 5.574 1. The UVE-MSC-CNN model achieved the best effect in total nitrogen content prediction (RPD=4.499 2), and the CARS-1st D-CNN model reached the highest accuracy in water-soluble total sugar content prediction (RPD=6.070 8). After inversion by the optimal CNN model, the color changes of the visualized distribution images of each indicator were well correlated with the measured values, and different content intervals of reconstituted tobacco leaves samples could be distinguished by color differences, achieving non-destructive detection of indicator contents and visualization imaging of spatial distribution. |
| 查看全文
HTML
查看/发表评论 下载PDF阅读器 |
|
|
|