Research On Intelligent Tobacco Grading Algorithm And System Design And Implementation

Tobacco industry plays an important role in China’s economy,and tobacco grading is an important key link in the process of transforming tobacco agriculture into industry.So how to ensure the quality of tobacco grading is very important.The existing tobacco classification methods mainly rely on manual experience,which has strong subjectivity.With the advent of the intelligent era,more and more artificial intelligence technologies are used to achieve tobacco grading.However,most of the researches are focused on the basic theory,which is far from practical application.From the point of practical grading system,the algorithms are studied and a preliminary intelligent tobacco grading system is designed and implemented in this dissertation,which made tobacco grading step forward from algorithm to practical application.The main work is as follows:1.The tobacco leaves were graded by feature screening and XGBoost method.Firstly,104 tobacco color features,shape features and texture features were extracted and normalized by max-min method.Secondly,in order to improve the accuracy and speed of tobacco leaf classification,a cascade method of discrete difference value and adaptive genetic algorithm was proposed to screen features.The number of optimal feature combinations was reduced to 35,and the accuracy of XGBoost tobacco leaf classification algorithm is increased to 90.01%from 83.63%.The tobacco grading speed gets to 9,090 pieces/s from 2,381 pieces/s.Finally,data enhancement is employed to further improve the generalization ability.The accuracy is improved to 90.27%.The experimental results show that the accuracy and grading speed of tobacco leaf classification were improved effectively by cascade screening of tobacco leaf characteristics and data enhancement of tobacco leaf image.2.An improved ResNet network is proposed to grade tobacco leaves.Firstly,the traditional ResNet network model is constructed to grade the tobacco leaf images.Secondly,atrous convolution is employed to replace the original traditional convolution and pooling modules.Therefore,the loss of feature information caused by pooling operation is avoided and the image receptive field is increased.Thirdly,gated fusion is used to change the connection mode of residual learning unit and realize the complementary advantages between different inputs in the modified convolution module.The experimental results showed that the accuracy rate is increased to 90.5%from 88.02%,and the tobacco leaf classification speed becomes 926 pieces/s from 280 pieces/s.Finally,the accuracy rate of tobacco grading of ResNet network can reach 93.46%by data enhancement.3.An intelligent tobacco leaf grading system is designed and implemented.Under the Windows operating system,a preliminary intelligent tobacco grading system is implemented by taking python and front-end language as the core development language,using Pycharm integrated development environment,Django development framework and MTV design mode.The system includes image acquisition,data storage,intelligent classification,classification results and feature information display modules.The experimental results showed that the method of feature screening and XGBoost tobacco leaf classification was faster in speed,the hardware implementation cost was lower,but the classification accuracy was slightly lower.The improved ResNet network tobacco leaf classification model has better classification accuracy but slower classification speed and higher hardware implementation cost.The intelligent grading system designed and implemented in this dissertation includes these two modules at the same time.Users can choose corresponding grading module according to the actual demand.The system has better extension and the necessary modules are easily to be added into the system.