Application Of Intelligent Classification Method For Flue-Cured Tobacco Leaves Based On Convolutional Neural Network(CNN)

China is the world’s largest producer and consumer of cigarettes.Cigarette is not only an important trade product but also an important consumer product of our country.The number of cigarettes produced in China is one third of that in the world,while China needs to consume more than 15 million cigarettes every year.Tobacco is growing in many parts of our country because of the large demand for cigarettes.However,compared with other crops,tobacco is not only unable to use large-scale mechanized cultivation,and then roasting,grading,acquisition,hand-over and other links need to be manually involved.In the production practice of flue-cured tobacco leaf grading,the national tobacco leaf grading standard(GB2635-1992)was introduced as a reference for tobacco leaf grading.The standard evaluated the grade of tobacco leaf through the position,color and other factors,and finally could be divided into 42 types of tobacco leaf grades.However,manual grading relies too much on the experience of grading personnel and is too subjective,which often leads to inaccurate or even unqualified grading results,which not only affects the work efficiency,but also easily leads to the loss of interests of tobacco companies or farmers in the process of purchasing tobacco leaves due to too high or too low grading.In recent years,the intelligent auxiliary grading technology of flue-cured tobacco leaves based on Computer Vision(CV)and Artificial Intelligence(AI)technology has been gradually applied in the tobacco industry,which has achieved the effect of improving the accuracy of tobacco leaves grading and ensuring the grading stability.There are many methods for intelligent grading of tobacco leaves.This paper is mainly based on Convolutional Neural Networks(CNN)to classify flue-cured tobacco leaves so as to realize intelligent auxiliary grading of tobacco leaves.The specific work arrangement is as follows:1.Pretreatment of flue-cured tobacco leaf image data.OpenCV was used to extract color features from flue-cured tobacco images to reduce the influence of background on image classification training.RGB and HSV color space conversion,texture feature extraction,feature fusion and other methods were used to extract image features.The median filter is used to enhance the image.Finally,the processed data set is divided.In order to increase the diversity of training data and improve the generalization ability of training model,the data of training set is expanded.2.Select or design different convolutional neural network models for training.This paper uses two different CNN models for training,one is the classic VGG16,the other is the model proposed in this paper TbcNet.The processed RGB image,HSV image,texture feature image and feature fusion image were used to train the classification model respectively,and the performance of the classification model and the influence of different features on the training results were evaluated.The comparative analysis shows that using RGB image as input image and TbcNet model for training results in the best classification model,the accuracy of verification set can reach 81%,and the accuracy of generalization ability can still reach 80%when tested with test set.3.Landing application.Flue-cured tobacco leaf images were collected at the tobacco leaf purchase site for 23 days.The collected flue-cured tobacco leaf images included upper and middle tobacco images.Images of different types of flue-cured tobacco leaves were screened to ensure that the final number of images of different types was equal.Then,the tobacco images of flue-cured tobacco were processed,the data set was made,the data of the training set was expanded,and the data was sent into the TbcNet model designed in this paper for training.The classification accuracy of the middle tobacco was 85%,and the classification accuracy of the upper tobacco was 87%.Field staff saved the collected flue-cured tobacco images to the server through the collaborative working mode of end-cloud,and used the trained TbcNet model to classify flue-cured tobacco images so as to achieve the effect of auxiliary grading,which was recognized by professional tobacco makers.