Research On Wheat Seed Classification Based On Hyperspectral Imaging

Wheat is the main food crop in China.Its high-quality production is not only the basic guarantee for people to pursue a better life.And it is of great significance to the steady development of the national economy.High-quality wheat seeds are the key to wheat production,so it is very important to explore a practical method of wheat seed classification.For ten kinds of wheat:Dwarf Kang 58,Huai Mai 30,Kai Mai 20,Zhong You 9507,Zhou Mai 28,Zhou Mai 22,Luo Mai 24,Xi Nong 509,Zheng Mai 366,Yu Mai 49,we get neat and stacked hyperspectral image data in the visible-near-infrared(VNIR)and near-infrared(NIR)spectra.Combining the characteristics of spectral information and image information,this paper uses cross-validation(CV)-support vector machine(SVM),Elastic Net and genetic algorithm(GA)-BP neural network classification model to study wheat seed classification.The main contents of this paper are as follows:(l)This article uses the spectral information in the acquired data for classification.Comparing the characteristics of VNIR and NIR spectral information of wheat seeds,we find that the average line of VNIR is relatively smooth and lacks characteristic peaks,while the NIR average line characteristic peak is obvious.After comparison with classification experiments,we find that the classification of spectral information under NIR is better.The method of principal component analysis(PCA)for spectral information extraction is used to extract the principal components as features,which is more conducive to classification.(2)The feasibility of stacked sampling data for classification is studied.The stacked sampling data is used for classification in SVM.The average accuracy of three classifications is more than 95%,the classification of four is about 87.5%,and the accuracy of six classifications is 75%.The results show that when using spectral information for classification,stacked sampling data can be used for classification.(3)This paper focuses on the classification model.We use the spectral information in the NIR band to classify wheat seeds.It is found that using PCA as the main component extracted from spectral data as input,it is more accurate to establish SVM classification model than three,four,six and ten categories using Elastic Net method.In addition,under the same conditions,the accuracy of GA-BP neural network model is lower than that of SVM model.(4)This paper focuses on the optimal classification model SVM.In terms of model parameter determination,it is found that the K-CV and particle swarm optimization algorithms are used for parameter optimization,and the classification results are basically the same.When the spectral information feature is used as the model input,the normalization effect is not good.In terms of fixed model migration parameters,we propose a method to determine the optimal parameters c and g between groups,so that the model has convenient and efficient migration ability.(5)This paper combines image features for ten classification and parental similar interspecies classification.Based on the spectral feature information,we use Halcon software to extract the four features of the seed image area,perimeter,long axis and short axis under the feature band as the input feature SVM model.In the end,the recognition rate of the ten categories of seeds is 90.47%,and the close recognition rates of the parents is 92%.