基于高光谱与太赫兹多维信息融合的立体成分缺陷识别

In order to realize stereo composition defect recognition in the intelligent manufacturing process of foods, pharmaceuticals and materials, a new detecting method based on the cooperation of near-infrared hyperspectral and terahertz imaging is put forward, and the approach of multidimensional information fusion and defect feature extraction is conducted by deep learning methods. The effective wavelengths with minimized collinearity and maximized defect expression ability are selected by improving the successive projection algorithm based on pairwise constraints. Several kinds of three-dimension convolution neutral network and recurrent-convolution network are proposed to realize multidimensional information fusion, including spectral information fusion of near-infrared and terahertz, spatial information fusion of surface and interior, and spectral-spatial information fusion. The stereo composition defect recognition is achieved by the multidimension information fusion. Finally, a semi-supervised transfer learning method is proposed to solve the overfitting and unbalance problem in small sample size modeling. In the proposed transfer learning method, the pretraining model not only provides the trained parameters, but also helps to pre-decide the label of unlabeled samples in small data set, which are used to fine-tune the parameters later. This project has important academic values and practical significances on nondestructive defect detecting technology.

针对食品、药品、材料等产业智能制造中无损快速立体成分缺陷检测的需求，项目提出基于反射式近红外高光谱与太赫兹实时成像协同的新检测方法，原创性探索基于深度学习的多维信息融合与立体成分缺陷特征提取的关键科学问题。提出基于成对约束评分的改进连续投影算法，进行高光谱特征选择，使被选特征共线性最小化的同时具有最强的缺陷表达能力；采用多种结构的三维卷积神经网络与递归-卷积神经网络进行包括高光谱与太赫兹的“谱信息融合”、表面与内部的“空间信息融合”及“空谱信息融合”的多维信息融合，实现基于深度信息融合的立体成分缺陷识别。最后，提出基于半监督迁移学习的小样本模型训练方法，以解决其过拟合与分类精度不均问题。在该迁移方法中，预训练模型不仅用于输出模型参数，还将用于预分类参与训练新模型的无标签样本。项目研究对立体成分缺陷识别技术应用及发展具有重要学术价值和实际意义。

针对食品、药品的质量无损检测与质量缺陷识别问题，本项目探索基于高光谱与太赫兹传感的深度学习建模方法，重点开展了以下四个方面的研究：（1）高光谱波段选择方法（2）高光谱“图谱信息融合”建模方法（3）高光谱与太赫兹联合建模方法（4）半监督学习算法。.提出了基于权值注意力机制的高光谱波段选择方法，采用神经网络权值评价近红外波段重要性，进而实现轻量化网络，降低高光谱数据的冗余，提升算法精度与效率；提出了基于二分支卷积网络的高光谱数据分类方法，以一维卷积网络和二维卷积网络分别提取高光谱的光谱信息和图像信息，并基于深度神经网络实现“图谱信息融合”，在草莓瘀伤识别、中药品种识别、咖啡豆品种鉴定等多个数据集上获得比传统算法更好的分类精度与稳定性；研究基于多维卷积网络的高光谱与太赫兹数据融合方法，实现胶囊等被测品的内部成分分析；提出基于生成对抗网络的半监督高光谱建模方法，将无标签和生成的虚拟样本加入训练，实现小样本建模，缓解过深度学习中的拟合问题，在食品、药品成分定量分析中取得良好效果。