| The blending ratio of cut stem,expanded tobacco silk,Tobacco silk and reconstituted is an important factor affecting the taste and physical indicators of tobacco.The type identification of the four types of tobacco is an important preliminary work for the determination of tobacco components.Therefore,it is important to achieve high accuracy and efficiency in identifying the types of tobacco yarn to investigate the design of formulations and to check the quality of tobacco products.In the production process of tobacco products,the current use of Near Infrared Spectroscopy(NIR)and physicochemical methods to achieve the differentiation of different types of tobacco yarn has problems such as high equipment costs and poor effectiveness.With the rapid improvement of industrial camera resolution and frame rate,as well as the development and widespread use of big data and artificial intelligence,the use of machine vision combined with deep learning has important research significance and application value to achieve rapid and accurate classification and identification of tobacco filaments.In this paper,we proposed a classification and recognition method for tobacco based on machine vision and improved convolutional neural network model.The main research elements are:(1)Construction of a tobacco image dataset.According to the characteristics of tobacco samples,the hardware module of the vision system is selected according to the classification and recognition needs and application scenarios,and the tobacco image acquisition platform is built to collect tobacco images.The HSV colour space-based tobacco image segmentation method is proposed,and a tobacco data set of 4118 samples in 4 categories is built to provide a data base for the study of tobacco classification and recognition models and algorithm optimisation.(2)An improved VGG16-based tobacco type recognition method is proposed.By reducing the number of convolutional kernels in VGG16,increasing the residual module and using global pooling instead of fully connected layers,the number of parameters and training time of the model were significantly reduced and the classification accuracy was significantly improved,with an accuracy of 95.51% in the self-built tobacco dataset.Accuracy,recall,and F1 score are used as model evaluation metrics to compare and analyze the classification performance with the benchmark convolutional neural network models Alex Net,VGG13,and Goog Le Net on specific categories of tobacco.(3)A tobacco type recognition method based on improved Mobile NetV2 and knowledge distillation is proposed.The I-Mobile NetV2-kd model is proposed to address the problem of complex and inaccurate traditional classification models.By adding a multi-scale feature fusion module based on improved depth-separable convolution to Mobile NetV2,the backbone network is adapted to obtain I-Mobile NetV2 and used as the student network,while the knowledge distillation method is used to use Res Net50-TL after migration learning as the teacher network.the I-Mobile NetV2-kd model’s classification accuracy was effectively improved,with an accuracy of 95.86% in the self-built tobacco dataset.And the prediction time for a single image was only 62 ms.(4)The design and implementation of the tobacco classification and recognition system was completed.The system was developed using the Pycharm IDE and Py Qt5 software tools,including image acquisition,image pre-processing,model training and tobacco recognition,etc.The human-computer interaction interface and background logic code design were completed,and the effectiveness and practicality of the system were verified through experiments. |
