Based On Machine Vision And Deep Learning For The Research Of Tobacco Leaf Grading

With the development of machine vision technology,image acquisition and recognition has become a research hot spot.In the tobacco industry,tobacco leaf grading technology is undergoing a transition from low efficiency manual grading to computer automatic grading.This thesis proposes research on tobacco leaf grading based on machine vision and deep learning,which has certain theoretical significance and application value.Based on machine vision and deep learning,tobacco leaf grading research is essential for image feature extraction and segmentation.In this paper,the improved K-mean algorithm is combined with the RGB color space spherical coordinate mixed distance method to realize the complete segmentation of tobacco image;the deep learning is applied to the recognition of tobacco image,and the recognition accuracy and speed are improved by optimizing the deep learning algorithm.The main works and innovations are as follows:1.Setting up the tobacco leaf image acquisition system,collect the tobacco leaf image and preprocess the image;compare the advantages and disadvantages of different filtering methods,select the appropriate filtering method to remove the image noises;analyze the shortcomings of the existing tobacco leaf segmentation method,and propose the RGB color space spherical coordinate mixed distance method and the improved Kmeans algorithm are used to segment the tobacco image and obtain complete foreground information.2.Based on the preprocessing,the tobacco leaf characteristic parameters are extracted.Using the support vector machine(SVM)pattern recognition theory and method,the tobacco leaf grading model based on SVM network is built by experimenting with the parameter data and training the network.The graded recognition rates of tobacco leaf B2 F,C3F and X3 F grades were calculated,which were 81.67%,90.67% and 86.67%,respectively.3.By comparing the SVM model with other machines,the convolutional neural network is optimized according to the nature of tobacco database,and the method of optimizing tobacco database is proposed to solve the over-fitting problem.By using the convolutional neural network to train and test the tobacco image,the grades of tobacco leaf B2 F,C3F and X3 F are graded,and the recognition rates are increased by 11.66%,1.67% and10.00%,respectively.