Research Of Online Grading For Flesh Jujube Based On Hyperspectral Imaging Abstract

Jujube is favored by people for its high medicinal value and nutrition value, and now people’s requests for the appearance and flavor of jujube are going up. The paper researched the fast and nondestructive detecting method for fresh pear jujube based on hyperspectral imaging, which built the detecting model and provided theory evidence for the online classifying technology.Conclusions of the research are as follows:(1)Analyse and compare the effect of different proceeding methods to the spectrum models. Proceed the spectrum information extracted from the hyperspectral image by16methods and compare the performance parameters by building PLS models. It proved that decomposing by using the dbl wavelet basis function for9times and the baseline emendation were best for the qualitative analysis, while decomposing by using the dbl wavelet basis function for9times and wavelet denoise were best for the quantity analysis.(2)The analysis used many simplifying and optimizing methods for the extracting of feature information which includes feature wavelengths, principle components, the wavelet approximation coefficients and the characteristic coefficients. In the qualitative analysis,9feature wavelengths,10principle components,43wavelet approximation coefficients and12characteristic coefficients were extracted with the decreasing of98.09%、97.88%、90.89%and97.46%respectively for data size, and in the quantity analysis,21feature wavelengths,10principle components,43wavelet approximation coefficients and16characteristic coefficients were extracted with the decreasing of98.09%、97.88%、90.89%、and97.46%respectively for data size.(3)Analysing the performance parameters and prediction results of different models based on the full-wave band, feature wavelengths, approximation coefficients and characteristic coefficients and considering the numbers of variables for modeling, it drew the conclusion that the model based on the linear combination of6parameters which were chosed form12feature parameters by stepwise method was best.(4) Analysing the performance parameters and prediction results of different models based on the full-wave band, feature wavelengths, approximation coefficients and characteristic coefficients and considering the numbers of variables for modeling, it drew the conclusion that the partial least squares model based on the spectrum characteristic coefficients was best.(5)Detect and classify the3varieties of samples by the spectral angle mapping algorithm to the reference spectrum which was created by minimum noise fraction(MNF) and pixel purity index(PPI). Compare the results and it suggested that the result based on the pixel purity index was better than minimum noise fraction. Therefore, choose the reference spectrum based on pixel purity index for the online detection and the threshold value were0.08、0.08and0.085for decay samples, disease samples and normal samples respectively.