Classification and Identification of Tobacco Varieties after Curing Based on Hyperspectral Imaging and Deep Learning

In order to realize rapid identification of different varieties of cured tobacco leaves, hyperspectral images of Honghua Dajinyuan, K326 and Yunyan 105 tobacco leaves were collected, and the region of interest was selected by threshold segmentation and edge extraction algorithm to obtain spectral data of tobacco leaves. The raw spectral data were preprocessed by four common methods: convolution smoothing combined with first-order derivative, standard normal variable transformation, multivariate scattering correction and centralization. On this basis, Multilayer Perceptron (MLP) network was established. MLP, Convolutional Neural Network (CNN) and Support Vector Machine (SVM) models were used to compare the classification of varieties using cured tobacco leaves. The results showed that all the established models could effectively distinguish the varieties of tobacco leaves, among which the first-order derivative pretreatment performed better than the other three pretreatment methods in several models, and the classification model based on CNN performed better than the other models. The combined performance of the two models was the best, and the correctness and accuracy of the test set were 97.9% and 98%, respectively. Hyperspectrum combined with CNN could realize the qualitative identification of tobacco varieties. The results of this study provide theoretical basis and technical support for the subsequent development of rapid detection system for cured tobacco based on hyperspectral technology.