Application Technology Of Near-Infrared Hyperspectral Imaging In Flue-Cured Tobacco Classification

Flue-cured tobacco is one of the important cash crops,and the difference in quality brought by different quality flue-cured tobacco to tobacco products is difficult to make up for in any subsequent processing,and the origin and part of flue-cured tobacco is one of the reasons affecting the quality of flue-cured tobacco.According to the national standard GB2635-1992 standard,the current grading work of flue-cured tobacco still mainly relies on manual identification,which not only has the problem of subjective factors affecting the large classification error,but also has the situation of complex process and high cost.Because hyperspectral technology has the characteristics of integrated maps,fast detection speed,no invasion of samples,no pollution,etc.,hyperspectral technology is more and more favored by scholars from all over the world,and is committed to using hyperspectral technology to replace traditional detection technology.In view of the classification of flue-cured tobacco production areas and parts,this paper selects 200 flue-cured tobacco samples from Yunnan,Jilin and Hunan and 200 samples from the upper,middle and lower parts of Jiyan No.9,a total of 1200 flue-cured tobacco samples as research objects,and models the classification of flue-cured tobacco production areas and parts.The main research contents and results are as follows:(1)Firstly,the visible light-near-infrared hyperspectral imaging system is used to collect the imaging data of each flue-cured tobacco sample,and the flue-cured tobacco sample database is established according to different origins and different parts,and the original hyperspectral data information of the flue-cured tobacco sample is preprocessed,which can better represent the actual characteristics of the flue-cured tobacco spectral reflectance;(2)In order to solve the SVM overfitting phenomenon,the five-fold cross-validation method is used,the radial basis function is used as the kernel function,and the classification model based on FFCV-SVM flue-cured tobacco origin and parts is established,and the classification results are visualized.Evaluation indexes of flue-cured tobacco origin:Accuracy=89.58%,Precision=90.52%,Recall=89.58%,F1=90.05%,Evaluation index of flue-cured tobacco part: Accuracy=90.57%,Precision=90.61%,Recall=90.63%,F1=90.62%,compared with the SVM flue-cured tobacco origin and part classification model,the evaluation index has improved;(3)In order to expand the dataset on the basis of small samples,this paper proposes to use the triplet network combined with CNN neural network to establish a classification model based on Trip CNN flue-cured tobacco origin and part,and visualize the classification results.Evaluation index of flue-cured tobacco origin: Accuracy=97.92%,Precision=98.04%,Recall=97.92%,F1=97.98%,Accuracy=96.23%,Precision=96.39%,Recall=96.19%,F1= 96.29%,compared with the FFCV-SVM flue-cured tobacco origin and part classification model and CNN-based flue-cured tobacco origin and part classification model,the classification results are further improved.In summary,based on the FFCV-SVM flue-cured tobacco origin and part classification model,the CNN-based flue-cured tobacco origin and part classification model and the Trip CNN-based flue-cured tobacco origin and part classification model can achieve a good effect on the classification of flue-cured tobacco origin and part,but Trip CNN can achieve better results in the case of small sample data,and is more universal and generalized.