Study On The Extraction And Grouping Method Of Digital Elements Of Tobacco Leaf Color

The tobacco grading indexes in the national standards are all qualitative descriptions,which are difficult to measure quantitatively.At present,the grouping of tobacco leaf color in China mainly uses human senses to make subjective assessment of tobacco leaves.Although the existing there are problems in tobacco color classification such as no suitable color space for tobacco leaves,no consideration of dividing tobacco leaves into positive and sub-groups,and lack of interpretation of grouping rules.In response to the above problems in the tobacco grouping algorithm,this thesis has conducted in-depth research and exploration in two main aspects,namely,tobacco positive and sub-group classification model and leaf color grouping model,and the main work includes the following aspects.(1)To address the problem of difficulty in recognizing and characterizing the color of tobacco leaves in their natural state,the sensory features of tobacco leaves are represented using digital features to achieve the quantitative characterization of the color.By collecting tobacco leaf images using a dedicated device,a sample set of tobacco leaves was created and a dedicated color space for tobacco leaves was constructed.Different color spaces were used to establish a multidimensional feature digital basic sample,which prepared for the subsequent correlation analysis between the features and the grouping of tobacco leaf colors.(2)For shallow,regional and revealed feature extraction characteristics,a network for classification of natural state tobacco positive and sub-groups is designed,based on Shuffle Net V2 network.Firstly,the training speed of the network model is accelerated by reducing the convolution depth of the network and evolving the activation function.Secondly,the channel attention SE module is introduced to enhance the feature differences between channels,improve the characterization ability of the network model,and avoid the group misclassification caused by the regionalization of the leaf parts of positive and sub-groups.Finally,the aggregation of the contextual information of positive and negative groups of tobacco is enhanced by embedding the pyramid pooling module PPM to fully integrate the revealed features of tobacco with global information.(3)To address the problem of redundant elements interfering with tobacco color classification and the difficulty of regularizing national standards.The accuracy of color classification of each element of tobacco leaf using BP neural network classification algorithm measured the correlation of each element.The CART algorithm was applied to segment the color of tobacco leaves under different groups according to the element correlation segmentation rules and compared with deep learning and machine learning algorithms.The experimental results showed that the accuracy of CART＿Rule algorithm model reached 92.10% and achieved high classification results.Compared with RF,LR,KNN,and NBC,the classification accuracy is improved by 10.28%,20.58%,13.87% and 24.59%,respectively.The accuracy is significantly improved compared to Squeeze Net and Mobile Net V2 models,and the accuracy of CART＿Rule model is improved by 9.16%,9.15% and12.06% compared to three networks,Resnet34,Dense Net and Shuffle Net V2,respectively.