Research On Detection Of Surimi Adulteration Based On Hyperspectral Reconstruction

Surimi products are cheap and simple to make,and occupy a relatively high proportion in the consumption of fish products in my country.The proportion of starch in surimi mixture affects the taste and quality grading of surimi.The existing surimi adulteration detection methods are cumbersome,and difficult to achieve rapid and timely detection.In response to this phenomenon,this paper explores the feasibility of using hyperspectral technology to detect surimi adulteration.In addition,because the acquisition of hyperspectral data is complex and expensive.In this paper,different ways of spectral reconstruction through color image were explored.Further more,the reconstructed spectral were used for surimi adulteration detection research.The samples of surimi tested in this paper were different ratio adulterations of surimi and starch.The main research conclusions were as follows:(1)Determine the optimal quantitative detection model of surimi adulterations based on hyperspectral technology.By comparing and analyzing the full-wavelength and simplified models of surimi adulteration established by different pretreatment methods and wavelength selection methods,it showed that models established could all achieve great results of surimi adulteration.The results showed that the prediction coefficient of determination R~2 in each model was better than 0.90,and the model with the best detection effect was the ELMR full-wavelength model based on SG preprocessing,with R_C~2=0.9938,R_P~2=0.9977,RMSE_C=1.9989%,RMSE_P=1.7983%.(2)The feasibility of spectral reconstruction of color images based on different spectral reconstruction methods was verified.The pseudo-color images synthesized by the spectral response tristimulus values were used as the model training set,and the pseudo-color images and color images were used as the prediction sets,respectively.When applying the traditional spectral reconstruction algorithm,the reconstruction effects of partial least squares regression and multivariate polynomial regression were compared,and the multivariate polynomial regression was finally determined as the spectral reconstruction algorithm.In order to determine the optimal polynomial,the reconstruction effects under different degrees and terms were analyzed and compared,and it was determined that the optimal basic matrix degree of multivariate polynomial regression was 4,and the optimal total number of terms was 22.When applying the deep learning method HRNet network to realize spectral reconstruction,the performance was deeply mined,the reconstructed spectrum of the pseudo-color image and the color image were compared.(3)The feasibility of detecting surimi adulteration with color image and pseudo-color image reconstruction spectrum was verified.The effects of the full-wavelength model and simplified model for surimi adulteration detection established by different image reconstruction spectra were analyzed and compared.In the reconstructed spectrum of pseudo-color image,the BP full-wavelength model without preprocessing established based on deep learning reconstructed spectral had the best effect,with R_C~2=0.9939,R_P~2=0.9981,RMSE_C=1.9085%,RMSE_P=1.6483%.In the reconstructed spectrum of color images,the ELMR simplified model under the SPA wavelength selection method has the best detection effect,with R_C~2=0.9976,R_P~2=0.9867,RMSE_C=1.9048%,RMSE_P=2.6093%.The results showed that the surimi adulteration detection model based on hyperspectral technology had excellent performance.At the same time,the use of pseudo-color images and color images can achieve both great spectral reconstruction results under the multivariate polynomial regression algorithm and deep learning method.The surimi adulteration detection model established based on the reconstructed spectrum could achieve good performance.This result will bring great convenience to the popularization and application of hyperspectral technology.