Study On Data Processing Methods Of Hyperspectral Detection In Meat Adulteration

Hyperspectral imaging technology is gradually emerging in the field of food safety detection due to it combines the spectroscopy and imaging technology so that can realize rapid and non-destructive detection.To achieve accurate detection of meat quality,the data analysis meathods have been utilized to establish mapping models between hyperspectral data and meat quality.However,the increased spectral bands has been the main chanllenge for hyperspectral data analysis,which could reduce the detection accuracy and increase the calculation.In order to ensure the detection efficiency without increasing the equipment cost,the algorithms of data anlaysis has been the kernel of HSI research in order to realize accurate and fast detection.Thus,the paper studied three kinds of methods in the field of hyperspectral data analysis,and their results and innovations are as follows:(1)In order to improve the computation speed of neural network in the filed of meat adulteration detection using hyperspectral imaging,our study introduced a fast learning algorithm,named extreme learning machine(ELM).The second-order differential method was proved as the best spectral preprocessing technique.Then,the results of hypotheses testing verified that ELM had higher accuracy and stability in identifying meat adulteration.The accuracy and specificity of beef adulteration identification were 96.28 % and 99.60 %,and their coefficients of variance were 0.06 and 0.01,respectively.The accuracy and specificity of pork adulteration identification were 98.56 % and 99.48 %,and their coefficients of variance were 0.03 and 0.01,respectively.(2)Aiming at solving the problem of multicollinearity effect which casuing by redundant bands,our study proposed a new type of quantatitive method.The method used ELM to reduce feature dimension and implementend the regression analysis based on PLS.In prediciting the chicken content of adulterated beef samples and the fat-to-thin ratio of pork sampl,PLS models were estabilsied using the fullspectrum with 950 bands,the root mean square error of prediction(RMSEP)were 1.40 % and 2.37 %,respectively.Then,the feature dimension were decreased to 40 and 10 by ELM,the RMSEP obtained were 1.56 % and 2.57 %,respectively.(3)After the above studies,to sovle the problem of the visualization of adulterants' distribution in adulterated beef samples,our study proposed a new visualization algorithm,and realized the visualization of distribution and the quantitation of gradients for the adulterants.By establishing the linear regression equation between the hyperspectral images of different types of pure samples,the identification of single pixel was identified according to the probability distribution of the regression coefficients.The proposed model realized the distribution of adulterants for beef samples,whose adulteration gradients were severally 0 %,10 %,20 %,30 % and 40 %.In addition,the best prediction results were 4 %,11 %,25 %,27 % and 40 %,and the average of absoluted error was 2.8 %.The data analysis methods could effectively improve the detection accuracy of meat based on the spectral features and image features of hyperspectral data.It has been proved that the data analysis methods can improve the detection accuracy of hyperspectral imaging technology effectively,and it is of great significance to promote the development of hyperspectral detection technology in the field of food safety detection.