基于可见/近红外光谱和机器视觉信息融合的小麦呕吐毒素污染在线检测机理与方法研究

The deoxynivalenol (DON) contamination caused by Fusarium head blight (FHB) of wheat has become a major threat to the safety of wheat product. Visible/near infrared spectroscopy and machine vision techniques are regarded as important tools for quick sorting and continuous monitoring of DON contamination in wheat. However, due to the unclear mechanism for DON detection based on spectral and image analysis, their application are restricted by theoretical evidence. Concerning this issue, this project tends to study the mechanism and methods for on-line detection of DON contamination by data fusion based on spectral and image analysis. Firstly, association analysis between DON contamination, specific changes of internal and external quality caused by DON contamination and response spectral/image information will be taken to identify sensitive spectral/image features raised by DON contamination. Secondly, correlation analysis between spectral and image features and comparison of different feature combinations will be studied. The optimal coupling mode between the features can be confirmed. Finally, throughout the analysis of influence factors towards model performance and different modeling strategies, theoretical basis for enhancing the prediction of DON contamination and robustness of models should be discussed. The project aims to clarify the mechanism for DON contamination detection based on spectral and image analysis, and to reveal coupling mechanism between spectral and image features. The completion of this project will provide basic theory and technical support to realize quality classification of wheat and ensure the safety for storage and utilization.

小麦赤霉病产生的呕吐毒素DON污染已成为小麦质量安全的主要威胁，可见/近红外光谱和机器视觉技术被认为是小麦DON污染快速分级和连续监测的重要手段。然而目前，DON光谱和图像无损检测机理尚不明晰，应用缺乏理论依据。针对这一问题，本项目拟在青年基金研究基础上，进一步开展基于光谱和图像信息融合的小麦DON污染在线检测机理与方法研究。首先，通过DON污染水平、DON污染引起内外部品质特定变化与光谱/图像信息的关联分析，提取DON灵敏光谱和图像特征；其次，通过DON光谱/图像特征相关性分析与不同组合方式融合效果对比，研究光谱和图像特征最佳耦合方式；最后，通过模型性能影响因素分析与建模策略对比，探讨在线检测模型DON检测水平和稳健性提升的理论依据。项目旨在阐明小麦DON污染光谱和图像无损检测机理，揭示信息融合过程光谱和图像特征的耦合机制，为实现小麦品质分级、储运加工安全提供基础理论和技术支撑。

由小麦赤霉病产生的呕吐毒素DON污染已成为小麦增产和质量安全的主要威胁。可见/近红外光谱和机器视觉技术由于其检测速度快、测量便捷的优势被认为是小麦DON污染快速分级和连续监测的重要手段。然而目前，小麦DON污染光谱和图像无损检测机理尚不明晰，应用缺乏理论依据。为此，本项目开展了基于光谱和图像信息的小麦、玉米等粮食有害霉菌侵染及DON污染检测机理与方法研究。结果表明：（1）侵染不同有害霉菌小麦在储藏72 h 后在线可见/近红外光谱总体识别率达91.7%，不同霉变程度识别率达90.0%。小麦中菌落总数的PLSR预测模型Rp2为0.890，RMSEP为0.369 log CFU/g，RPD达3.03。此外，利用可见/近红外光谱和机器视觉融合信息对侵染不同霉菌及不同霉变程度玉米的判别正确率达到100%和92.2%，比单独应用光谱和图像至少提高5.2%。基于光谱和图像信息融合的玉米菌落总数PLSR模型的预测误差比应用单一光谱和图像特征分别减少25.0%和17.4%。（2）建立了基于光谱和图像信息的小麦籽粒和小麦粉DON污染的在线定性定量分析方法。对于小麦籽粒而言，光谱和纹理特征融合效果最佳，对DON污染程度的识别率达到92.2%，比单独应用光谱或图像提升5%以上。对于小麦粉而言，光谱和颜色特征融合取得最佳效果，对DON污染程度识别率达到93.55%，降低了单独应用光谱时的不确定度。DON定量分析结果不够理想。（3）运用高光谱图像技术验收了光谱和图像信息融合对于DON检测的有效性。对于小麦籽粒而言，光谱和纹理特征同样取得最佳效果，对DON污染程度的识别率达到96.52%。对于小麦粉而言，光谱和颜色特征融合取得最佳效果，对DON污染程度的判别率达到91%。（4）通过特征分析和扫描电镜结果，光谱和图像信息融合能够更加准确描述由于DON污染所引起的小麦内外部性质的综合改变，显著提高对小麦DON污染的快速判别能力，具有实际应用潜力。