Research On Feature Extraction Methods Of Hyperspectral Scattering Image And Its Application In Nondestructive Detection Of Wheat Flour Quality

Hyperspectral scattering technology exhibits excellent performance in rapid non-destructive detection of agricultural and sideline products and food quality because it contains both spectral and spatial band information,and image information.The absorption and scattering information carried by the spectral curve can be used to reflect the strength of the internal absorption and scattering of the material,where the absorption information is mainly related to the chemical composition within the tissue and the scattering information is affected by the physical properties of the tissue,such as density,structure and so on.It is possible to provide nondestructive detection of internal quality attributes of agricultural products or foods based on the acquired information that carries and reflects the spectral information within the tissue.Simultaneously,particle size as one of the wheat flour quality indicators has important guiding significance for flour quality grading.While,flour has a certain flow characteristics,flour density as a very important physical parameter has a huge impact on the fluidity of flour,so the study of flour bulk density and particle size,are of great significance for flour processing,packaging,transportation,storage and so on.In the past,the detection of the density and particle size is time-consuming,laborious,and not comprehensive.Therefore,this paper combined with the advantages of hyperspectral scattering technology,focusing on the feature extraction methods of hyperspectral scattering image,and predicting the bulk density and classifying the particle size of wheat flour.The main contents of this article as follows:1.The mean spectrum feature extraction method of hyperspectral scattering imaging was used for predicting the bulk density and classifying the particle size of wheat flour.The scattering profiles of hyperspectral scattering images between 500 and 1000 nm were acquired and the mean spectral characteristics were extracted for each band.The features combined with partial least squares(PLS)algorithm were predicting the bulk density and combined with partial least squares discriminant analysis(PLSDA)algorithm were classifying the particle size of wheat flour.In addition,the experiments were compared with the visible and short-wave near-infrared spectroscopy(Vis/SWNIR),the results showed that the mean spectral method was excellent in quality inspection of wheat flour,and hyperspectral scattering technology could simultaneously detect the bulk density and particle size of flour,compared with Vis/SWNIR.2.In order to further improve the accuracy of the model and to use the shape characteristics of the scattering curve,a feature extraction algorithm based on wavelet transform-Volterra coefficients was studied.At the same time,features combined with different modeling algorithms had carried on the experimental study for bulk density and particle size of wheat flour.In addition,in order to verify the superiority of the algorithm,the experiments were compared with the single wavelet transform,the Volterra coefficient method and the mean spectrum method.The research results showed that in the prediction model of bulk density,compared with the mean spectral method,wavelet transform method and Volterra coefficient method,the prediction accuracy of wavelet transform-Volterra feature extraction algorithm had been greatly improved,and the model more robust.However,compared with the particle size classification model,the feature extraction method based on wavelet transform-Volterra coefficients was more prominent and more discriminative in the prediction model.3.In order to use the image information of the hyperspectral scattering image itself,the feature extraction method based on NMF+wavelet transform-Volterra by incorporating the image information into the spectral information was studied and the features were used for the prediction of bulk density and the classification of particle size of wheat flour.The research results showed that in the prediction model of bulk density,the NMF+wavelet transform-Volterra algorithm was more superior to the wavelet transform-Volterra algorithm,the generalized Gaussian distribution method and the four-parameter Lorentz method,the prediction or classification accuracy had been improved,in addition,the generalization ability of the model was greatly improved.But compared with the particle size classification model,the NMF+wavelet transform-Volterra algorithm performed better in the prediction model.In this paper,the research results not only enrich the feature extraction algorithm of hyperspectral scattering image,but also provide a new idea for the physical properties of powder materials.