Research On Detection Method Of Maize Seed Mildew Based On Hyperspectral Imaging Technology

China is a big country in corn production and sale. At the same time, China is also the country with the largest population, so it is very important to improve the yield of maize. But the disease will affect the yield of maize, and it was especially meaningful to test the corn seed disease. In recent years, the frequent occurrence of a large area of crop production due to crop mildew cases, pests and diseases related to the development of the whole seed of agriculture. Therefore, the research on the detection technology of maize seed mildew has a very important role in improving the quality of maize seed, promoting the increase of maize yield and improving the competitiveness of seed enterprises in china. The object of this experiment was to study the healthy and moldy maize seeds.The method of spectral data preprocessing was researched. Hyperspectral image of maize seed was collected by hyperspectral imaging system. Then, three kinds of pretreatment methods was used to pretreat spectral data, respectively, smoothing, multivariate scattering correction, standard normal variable. By comparing the three pretreatment methods, it is concluded that the standard normalization is the best method to preprocess the spectral data.The feature extraction of hyperspectral image of maize seed was researched. Spectral feature extraction and texture feature extraction was compared. Three methods was used to extract spectral feature, which were the principal component analysis, sccessive projections algorithm, the manifold distance method. The principal component analysis method was used to extract the characteristic wavelength of 608nm. The manifold distance method was used to extract the characteristic wavelength were 548nm, 768nm.The sccessive projections algorithm was used to extract the characteristic wavelength were 490nm, 572nm, 735nm, 845nm. The comparison of the 3 feature extraction methods shows that the sccessive projections algorithm is the best method for spectral feature extraction. The main features of texture are energy,entropy, moment of inertia and correlation. It wss found that spectral features combined with texture features were the best feature extraction methods for hyperspectral images.Modeling and identification of maize seed mildew was researched. Four modeling methods were compared, which were linear discriminant regression, independent soft mode,support vector machine, BP neural network. Based on the analysis of the effect of different recognition models on the identification of maize seed mildew, it was concluded that the BP neural network with the combination of texture features and spectral features was the best model. The correct recognition rate of positive samples and healthy samples were 93.33% and 100%, respectively.To sum up, BP neural network was the best modeling method, combining standard normalize variate, sccessive projections algorithm, spectral feature and texture feature. This paper can be used as a basis for hyperspectral detection of corn seed mildew, and will play a certain role in the detection of corn seed disease by hyperspectral imaging technology in the future.