基于深度学习的亚像素级黄曲霉毒素高光谱图像定量检测

Aflatoxin is a kind of highly toxic and carcinogenic substance, which widely found in peanuts, corn and their products. Based on its ultraviolet fluorescence characteristics and surface and superficial layer distribution characteristics, In previous studies, we extracted some key wavelengths by ultraviolet-induced hyper-spectral imaging and constructed a qualitative detection model. However, due to the limitations in imaging resolution and the complexity of the imaging mechanism, quantitative detection is especially difficult. We speculate that: the ubiquitous phenomenon of "mixed pixel" and "same substance different spectrum" is the key issue affecting quantitative detection. In this study, we intend to study the quantitative detection methods of aflatoxin using sub-pixel hyper-spectral image based on deep learning, which including: 1) Independent Component Analysis (IC) feature spectra of toxins based on independent component analysis (ICA) to reveal the mechanism of spectral variation; 2) Non-negative matrix factorization (NMF) is used to achieve mixed pixel detection of toxic kernels, and to ascertain the distribution pattern of toxin abundance; 3) to construct large data by reshape transform of hyperspectral image and realize in-situ labeled of toxic pixels. And then to establish a pixel-level toxin recognition model by convolution neural network (CNN) of deep learning algorithm, and to achieve quantitative inversion of toxic kernels by abundance weighting of each pixels. The project can provide a new idea for the development of intelligent sorting machines for agricultural products with quantitative detection of aflatoxin.

黄曲霉毒素是一种剧毒、强致癌物质，广泛存在于花生、玉米及其制品之中。基于其紫外荧光特性和表层、浅表层分布特征，我们前期通过紫外诱导高光谱成像提取了其关键波长，构建了定性检测模型；然而由于成像分辨率的限制和成像机理的复杂性，定量反演异常困难。我们推测“混合像元”和“同物异谱”现象的普遍存在，是影响定量检测的关键问题。本研究拟以深度学习为基础研究亚像素级黄曲霉毒素高光谱图像定量检测方法，包括：1）基于独立分量分析(ICA)重构毒素的独立分量（IC）特征光谱，揭示光谱变异的机理；2)通过非负矩阵分解（NMF）实现含毒籽粒的混合像元探测，探明毒素丰度分布规律；3) 构建高光谱图像Reshape变换大数据，实现含毒像素的原位标记，进而建立卷积神经网络（CNN）深度学习算法的像素级毒素识别模型,并通过丰度加权实现含毒籽粒的定量反演。项目可为研制具有黄曲霉毒素定量检测功能的农产品智能分选机提供新的思路。

黄曲霉毒素是一种剧毒、强致癌物质，广泛存在于花生、玉米及其制品之中。基于其紫外荧光特性和表层、浅表层分布特征，本研究构建了黄曲霉毒素光谱解析模型和亚像元分解方法，建立了毒素定量反演关键技术，进行了高光谱亚像元分解预测花生中的毒素、基于亚像素级深度学习的毒素定量检测和关键波长挖掘、基于一维改进时间卷积网络毒素像素级检测，亚像素CNN回归法定量检测毒素，基于像素级一维卷积神经网络的毒素检测，基于特征选择的玉米像素级黄曲霉毒素检测，利用波段指数检测毒素污染程度，基于深度学习像素级黄曲霉毒素检测，应用驱动的黄曲霉毒素检测关键波长挖掘方法等方面的研究。为进一步理论成果推上应用，本研究先后提出了基于深度学习的素智能检测方法和基于迁移学习的毒素智能检测方法，开发了基于深度学习和高光谱技术的谷物毒素检测系统软件，设计了一种使用多面转镜光谱检测花生毒素的光学检测装置，设计了一种具有素检测功能的农产品智能分选机。.发表项目密切相关SCI论文8篇，EI收录1篇，在相关领域发表标注SCI论文12篇，EI收录3篇，出版英文学术专著1部，申请中国发明专利3项（其中已授权2项），国际PCT申请1项。培养研究生2名（2020级朱宏飞、2018级髙霁月），其中朱宏飞同学先后3次获国家奖学金、硕士论文被推荐为高水平硕士论文，髙霁月同学获国家奖学金，已考取中国海洋大学博士。培养青年骨干3人，其中邓立苗老师获批山东省青创引育团队带头人，另外主持人韩仲志获批山东省泰山学者青年专家称号。主持人2019年在日本千叶大学食品工学科做访问学者，指导研究生学生髙霁月同学2019到澳大利亚默多克大学访学，疫情之前先后邀请澳大利亚格里菲斯大学高永晟教授，加拿大不列颠哥伦比亚大学王光玉教授、美国哥伦布州立大学范桂红副教授等来校交流并做学术报告。牵头举办中澳学术交流研讨会智慧农业中心论坛，疫情期间组织网络会议报告5次。