Research And System Implementation Of Intelligent Grading Methods For Single And Overlapping Tobacco Leaves Based On The Combination Of Traditional And Deep Features

The tobacco industry has contributed huge taxes to our country,and improving the purchasing quality of tobacco leaves has important economic value and brand influence on cigarette production.Tobacco grading is the first step in cigarette production and the last step in tobacco production by tobacco farmers.The grading quality of tobacco leaves directly affects the enthusiasm and economic benefits of tobacco farmers.At present,the classification of tobacco leaves mainly relies on the subjective evaluation of the grading workers.Tobacco leaves are classified through visual observation,hand touch,ear hearing,and their years of experience in tobacco leaves grading.The manual classification of tobacco leaves has strong subjectivity and limitations,and the classification efficiency is low.In recent years,computer-vision-based classification methods have been proposed to tackle the issue,but there still exist the issues of low accuracy and less consideration of classiying overlapping tobacco leaves.Therefore,this thesis combine traditional image features and deep neural network features to achieve an intelligent classification of both single and overlapping tobacco leaves.The software engineering method is used to design and realize the tobacco leaves automatic grading system,which greatly improves the efficiency of tobacco leaves grading.The main contributions of this thesis are as follows:(1)This thesis proposed a novel tobacco leaf grading model by using the ensemble of three support vector machines.First,we extract traditional features of tobacco leaves,including color,texture,and shape,and uniformly quantify the color features according to the color distribution of the flue-cured tobacco.Then we use VGG16 classification network and fine-grained convolutional neural network to extract high-dimensional semantic features of tobacco leaves.Then,we use the improved F-score feature selection method to perform feature screening on the three different types of features above,then use the support vector machine(SVM)to train three different grading models for the three types of features,and finally use the boosting idea to get the next result based on the three different grading results.A confusing suppression strategy is proposed to modify the boosting results.The final classification accuracy is 74.07%,and the grouping accuracy of parts is 94.02%.The accuracy of the classification model is 3.98%～23.93% higher than some traditional classification models,and 23.28% higher than Res Net152 that is a end-to-end convolutional neural network classification model.(2)This thesis studies how to divide and classify tobacco leaves when two tobacco leaves overlap each other.Based on the observation that the intersection of two tobacco leaves is concave,and the tip and the base of tobacco is convex,this thesis proposes an elliptical tobacco leaf fitting model based on concave and convex points.According to the concave and convex points,the contours of overlapping tobacco leaves are segmented,and then the contour segments are elliptical fitting,and the position relationship of the overlapping tobacco leaves is judged according to the main vein.The pixel accuracy of the segmentation is 0.9096 and the segmentation time is 0.5370 seconds,which are better than traditional segmentation methods.The divided tobacco leaves are classified individually using the propsed single-leave classification method.The classification accuracy on a dataset of two overlapping tobacco leaves with five grades is 45.21%.(3)We conduct the requirements analysis and system overview of the tobacco leaf grading software are carried out,as well as the interface design of the software is given in detail.Also,the execution efficiency of the grading algorithm is considered,and the experiment shows that when the scaling factor of tobacco leaf is 0.5,the grading speed increases obviously,achieveing a good balance between the efficiency and accuracy.Each module and function design of the software interface are introduced in detail.Finally,the methods are applied to the tobacco leaf intelligent grading equipment,and prelimilarily used in real tobacco leaf grading site.The use of this equipment can effectively improve the accuracy and efficiency of tobacco leaves grading and help tobacco stations improve its digitalization and intelligence.