Nondestructive Detection Study Of The Hyperspectral Image Of Wheat Grain Hardness

As a staple food of Chinese,wheat's quality directly affects the safety of our daily diet,and its grain hardness is an important parameter to evaluate the quality of grain.Therefore,how to detect the hardness of wheat grain quickly?effectively and accurately is of great significance.In this paper,three different hardness varieties of Wenmai 6,Zhongmai 895 and Xi Nong 979 were used as samples,and the theoretical mechanism of dividing of the different wheat grains hardness based on near-infrared hyperspectral image was studied.Which can characterize the effective area of wheat grain hardness were extracted,and the near-infrared hyperspectral image data of 22 different varieties of wheat grains were pretreated.The intelligent measurement model based on the regression analysis technique of radial basis function extreme learning machine was established to realize the automatic nondestructive detection of wheat grain hardness.(1)Study on the nondestructive detection mechanism of hyperspectral image of wheat grain hardness based on principal component analysisA total of 1540 wheat grains of 22 different varieties were collected by hyperspectral image acquisition system.The actual hardness values of different varieties wheat grains were determined according to the national standard method.The region of interest which can effectively characterize the grain hardness was extracted by image preprocessing,and the average spectral curve in ROI of wheat grain was used as the near-infrared characteristic spectrum.After the pretreatment analysis,the original spectral characteristic wavelength can be reduced from 256 in the range of 871.6-1766.3nm to 232 active wavelengths in 902.1-1699.6nm.The hyperspectral image were analyzed by PCA method.The contribution rate of PC1,PC2 and PC3 was more than 99.15%.Therefore,the first three principal components can be used to characterize the original image information.According to the principle of density degree distribution of pixel distribution area based on different hardness of wheat grains,explained the chemical differences between the wheat grain samples,and determined to use the score image of combination of PC2 and PC3 and study the feasibility of wheat grain hardness classification.By establishing a classification verification model based on partial least squares discriminant analysis,and the correct recognition rate was 100%,which proved the feasibility of nondestructive detection mechanism of hyperspectral image of wheat grain hardness based on principal component analysis.(2)Wavelength selection of wheat grain hardness based on artificial population optimization algorithmAiming at the characteristics of the large amount,the large redundancy of information,and the large mixing degree of three-dimensional data of the near-infrared hyperspectral of wheat grain,the characteristic wavelength was optimized by using the artificial bee colony optimization algorithm.For the shortcomings of early convergence of ABC algorithm,easy to fall into the local optimal value and slow search speed in the close to global optimal solution,a hybrid algorithm based on chaotic artificial bee was proposed.The results show that the wavelengths of 105 in the range of 902.1-935.9nm,968.7-992.6nm,1042.7-1072.4nm ranges from the wavelength of 232 in the range of 902.1-1699.6nm,and the number of wavelengths was reduced by 54.7%.Compared with ABC algorithm,the optimization model of the improved CABC algorithm reduced the running time by 45.3%,reduced the MSE by 0.042%,and increased the SCC by 0.55%.(3)The intelligent prediction model of wheat grain hardness based on RBF-ELM algorithmIn order to solve the problem of system instability caused by manual setting parameters in ELM algorithm,using GSM to determine the input of the parameters in the RBF-ELM algorithm automatically,and using the regression analysis technique based on RBF-ELM algorithm to establish the intelligent detection model of wheat grain hardness.The results showed that compared with the ELM model,the RBF-ELM prediction model has little difference in the running speed of the model,but the accuracy and the correlation coefficient was respectively improved by 3.31% and 7.36%.Compared with the SVR prediction model,the training and prediction time were shortened by about three orders of magnitude,the prediction accuracy was reduced by 0.53% and the correlation coefficient was increased by 0.96%.Therefore,using the regression technique based on the RBF-ELM intelligent measurement model effectively implemented the automatic non-destructive detection of wheat grain hardness.