Study On The Detection Method Of Meat Adulteration Using Spectral Technology

Raman spectroscopy and VNIR hyperspectroscopy,as new techniques for the characterization and identification of food components,have the advantages of fast,nondestructive and simple pretreatment,and have a wide application prospect in the online detection of meat adulteration.In view of the disadvantage that the traditional spectral classification model is not suitable for the detection of small samples and cross-field without labels,a method was developed for the identification of animal-derived adulteration of meat based on spectral technology.The main research contents of this paper are as follows:(1)The method was proposed for detection of pork adulteration in beef ground based on Raman spectroscopy.Based on baseline correction and principal component analysis,five Raman characteristic shifts were selected for the establishment of adulteration detection model:605 cm-1,1646 cm-1,1416 cm-1,1708 cm-1 and 2952 cm-1.The Raman spectral classification model of meat based on PCA-SVM method was established.The detection limit of PCA-SVM model is 10%adulteration mass fraction,and the accuracy of PCA-SVM model in test set is 98.9%.(2)A Meat Spectral classification model MSNN(Meat Spectral Neural Network)based on Neural Network was proposed.Aiming at the problems of small sample detection difficulty and slow model speed,On the basis of MSNN,an Improved Meat Spectral Neural Network(IMSNN)was built with three methods,namely,Spectral eigenvector pyramid,segmentation maximum pooling and deep separable convolution.In the two scenarios with small samples simulated in this paper,the classification accuracy of the improved IMSNN model is improved by 33.8%and 12.5%compared with the PCA-SVM model.The inference speed of IMSNN model is 122.6%higher than that of MSNN model.(3)Improved Meat Spectral Neural Network Based on Domain Adaptation(IMSNN-DA)was developed for unlabeled samples with a large number of newly adulterated methods in the data set.The accuracy of the IMSNN-DA model trained on the target domain T(x,y)reached 92.4%,which was able to better deal with the interference caused by domain differences and accurately identify adulterated meat.