Research And Application Of Intelligent Tobacco Classification Based On Deep Learning

At present,in the commercial acquisition and threshing and redrying stages,the grading of tobacco leaves mainly uses on-site technicians based on tobacco grading standard samples and tobacco grading technical specifications,and relying on personal vision,hand feeling and work experience to perform tobacco grading and quality judgment.This method of grading tobacco leaves is susceptible to subjective factors that cause fluctuations in the qualification rate of tobacco grading,and consumes manpower,material resources and financial resources.Therefore,research on efficient and accurate tobacco grading methods is a prerequisite to ensure product quality and has important application value.Aiming at the drawbacks existing in the current tobacco grading process,this article mainly conducts research from the aspects of near-infrared spectroscopy,image recognition and fusion.The main research work is as follows:（1）Aiming at the problems of high dimensionality,high noise,redundancy and nonlinearity in the near-infrared spectrum,a tobacco leaf classification model based on Joint Matrix Partial Preserving Projection（JMLPP）is proposed.The cluster analysis method is used to effectively select the characteristics of the near-infrared spectrum information of the tobacco leaves,and N feature matrices M₁,M₂,…M_N,selected according to N different clustering methods are obtained,and the joint matrix M is obtained by the intersection operation.The local preserving projection algorithm is improved from two aspects:the introduction of geodesic distance instead of Euclidean distance to construct the neighborhood distance matrix;in order to avoid the loss of effective information,the edge weight matrix is improved.The JMLPP algorithm is used to reduce the dimensionality of the joint matrix,and the support vector machine is used as a classifier to construct an intelligent grading model of near-infrared spectroscopy tobacco.（2）Aiming at the problems of poor real-time performance,large minimum input size,and low accuracy in the application of convolutional neural network to tobacco leaf grading process,an improved H-Mobile Net-v2 network model is proposed.H-Mobile Net-v2 has better recognition accuracy than Mobile Net-v2.The network model improves high-precision calculations by improving the inverted residual structure,which can retain more information and improve the accuracy of the network.An inverted residual structure h-bottleneck that can effectively improve the accuracy of the network is proposed.In the selection of classifiers,for tobacco leaf classification such as image information with higher similarity and insignificant feature gaps,AM-Softmax is used as the classifier,which is more suitable for the classification of tobacco leaf grades with a smaller gap than the Softmax classifier.Experiments show that the H-Mobile Net-v2network model has better recognition performance in tobacco leaf classification.（3）After constructing tobacco leaf classification models based on near-infrared spectroscopy and image recognition,the idea of model fusion was adopted to construct a tobacco leaf intelligent grading model based on the fusion of near-infrared spectroscopy and image recognition,which effectively reduced the generalization error of the model.Three different models with better performance,Inception Net-v3,JMLPP-SVM,and H-Mobile Net-v2,were selected as individual learners,and two fusion models were constructed using the integrated strategy of weighted voting.Experimental results show that the classification performance of the two fusion models is better than that of a single model.The best tobacco classification model is the fusion model based on JMLPP-SVM,H-Mobile Net-v2 and Inception Net-v3.