Research On Maize Moldy Recognition Method Based On Hyperspectral Technology

Our country is a big country in maize production and sales.Maize is an important food crop in my country.It is particularly important to ensure the quality of maize.The safety of maize is related to food safety issues and the development of many industries.Maize is rich in dietary fiber,multiple vitamins and trace elements.It is not only an important commercial food,but also the main raw material in livestock and poultry feed formulations.As an important food material and chemical raw material,maize is very prone to mildew during storage and transportation.After maize is mildew,a large amount of mycotoxins will be produced,and the surrounding healthy maize will also be mildew.Livestock and poultry will eat this for a long time.Moldy maize will affect the growth of livestock and poultry,and mycotoxins will accumulate in the organs and muscle tissues of livestock and poultry.The consumption of such livestock and poultry by humans will endanger human health.Traditional methods and chemical detection methods are time-consuming and laborious and will destroy maize samples.Hyperspectral imaging technology,as a new,rapid and non-destructive detection technology,can be applied to maize mildew detection,which can well identifies early mildew Maize.Research on the preprocessing method of hyperspectral image data.First,programmable constant temperature and humidity test chamber was used to conduct moldy experiments on maize,and then hyperspectral imaging technology was used to collect hyperspectral images of moldy maize.The hyperspectral image data were preprocessed by three preprocessing methods:Savitzky-golay algorithm,multiplicative scatter correction algorithm and standard normal variate algorithm,and the classification model was established by using the preprocessing results.The classification model accuracy of the three preprocessing methods is analyzed and compared,and it is concluded that the optimal data preprocessing method of the early hyperspectral image recognition and detection model of maize mildew is the standard normal variate algorithm.Research on Feature Selection Method of Hyperspectral Image Data.First,the original hyperspectral image data is preprocessed by the standard normal variate algorithm,and then the hyperspectral image data is characterized by three characteristic variable selection methods:competitive adaptive reweighted algorithm,successive projections algorithm,and uninformative variable elimination algorithm.The selection of variables,and finally the establishment of a BP neural network classification model.Competitive adaptive reweighted algorithm selects 21 characteristic wavelength variables;successive projections algorithm selects 7 characteristic wavelength variables;uninformative variable elimination algorithm selects 218 characteristic wavelength variables,which are respectively distributed among 400 ? 423.7 nm,536.3 ? 548.1 nm,554.8 ? 568.1 nm,679.3 ? 762.2 nm,825.9 ? 932.6 nm and 943.7 ? 981.5 nm.By comparing the three characteristic wavelength variable selection methods,it is concluded that the optimal characteristic variable selection method of the early hyperspectral image of maize mildew is the uninformative variable elimination algorithm.Research on the classification model method of hyperspectral image data.First,the hyperspectral image data is preprocessed by the standard normal variate algorithm,and the feature variable selection is performed by the uninformative variable elimination algorithm.Then the ACO-BP neural network modelled is constructed using the spectral reflectance spectrum of the hyperspectral image data,and compared with the four classification models of partial least squares regression algorithm,K nearest neighbor algorithm,support vector machine algorithm and BP neural network algorithm.By comparing the overall classification accuracy and average classification accuracy of the four classification models,it is concluded that the ACO-BP neural network modelled is the optimal classification model for hyperspectral image data.The overall classification accuracy of the verification set reached 94.4%,the average classification accuracy of all samples was not less than 92%,and the classification accuracy of mildly mildew maize was up to 96%.In summary,this paper concludes that the optimal early detection and recognition modelled after moldy maize hyperspectral images is standard normal variate algorithm + uninformative variable elimination algorithm + ant colony optimization algorithm + BP neural network.This article can provide theoretical guidance of the discovery and timely cleaning of mold-producing maize in the process of storing maize,so as to prevent a large number of molds from being stored in maize.