Research And Implementation Of Parameter Extraction And Classification Recognition System Based On Tobacco Image

Tobacco industry has always played an important role in China’s economy.In the process of tobacco purchasing and processing,how to quickly and accurately grade tobacco is of great importance.At present,the grading of tobacco leaves in the tobacco industry is mainly carried out by manual grading,which is not only time-consuming and laborious,but also cannot be effectively guaranteed in accuracy.In current research of tobacco grading algorithms,the classification is mainly carried out by identifying the difference in tobacco leaves’chroma,but the result is not satisfying when grading the tobacco leaves with similar color.In this paper,a tobacco parameter extraction and grading system based on traditional image processing algorithm and convolutional neural network is designed to improve the accuracy of tobacco grading.The main contents of this paper are as follows:First,this thesis studied a method of Angle extraction between main and branch veins based on the algorithm of vein skeleton refinement and regional growth.Firstly,through the improvement of the fast parallel thinning algorithm,the shorter branch veins on the main veins and the branch vein skeleton on the branch veins can be removed,and the number of comer points in the subsequent comer point detection algorithm can be reduced,so that the improved thinning algorithm is more suitable for the extraction of leaf vein skeleton of tobacco leaves.Secondly,by improving the initial pixel selection and growth rule of the region growth algorithm,the included Angle between the main pulse and the branch pulse can be extracted,and the running speed of the algorithm is effectively improved.Finally,according to the characteristics of the extracted tobacco leaves such as included Angle and chromaticity,the initial clustering center was manually selected to optimize the clustering algorithm,and the chromaticity characteristics of tobacco leaves were combined with morphological characteristics to improve the accuracy of grading tobacco leaves with similar chroma.Second,this thesis studied a convolutional neural network for tobacco grading.Because the characteristics of different types of tobacco are very similar,a convolution module of multi-scale feature extraction is proposed to improve the network,so that the network can learn the local details and global characteristics of tobacco images at the same time,and improve the network’s classification accuracy.Then the convolution module of the multi-scale feature extraction was put into the network,and the network structure with the optimal performance was obtained by comparing a large number of performance measurement indexes in the experiment,so as to further improve the accuracy of the tobacco grading algorithm.Third,this thesis designed a tobacco image parameter extraction and grading system.Firstly,the traditional image processing algorithm was used to extract the three characteristics of tobacco image length,the included Angle between the main vein and the branch vein,and the chroma value.Then,according to the deep learning algorithm,various features in tobacco images are extracted.Finally,the network structure was adjusted and these two characteristics were combined to improve the network’s ability to extract high-dimensional features and local detail features in tobacco images,and the overall system’s classification accuracy reached 84.52%.Fourth,this thesis designed and implemented a software system for extracting and grading tobacco image parameters.Through the software system,users can realize the detection of tobacco image more easily and quickly.