Mutton Adulteration Identification Based On Hyperspectral Imaging Technology Research

In recent years,driven by economic interests in meat products adulteration accidents occur frequently,seriously threatening the economic interests of consumers and restricting the development of the meat products industry in China.Lamb meat is tender,unique flavor,and rich in protein,amino acids and minerals and other important nutrients,popular among consumers.With the increasing consumption of mutton,the per capita consumption of mutton has reached 3.97 kg/person until 2022,however,unscrupulous traders use inexpensive meat such as chicken,duck to adulterate mutton and even replace it for profiteering.Among them,minced meat adulteration eliminates the morphological characteristics of the muscle,making it difficult to visually identify the source of meat,such forms of adulteration have seriously endangered the meat market order.Traditional testing methods are tedious,time-consuming and difficult to achieve the public demand for meat product testing.Therefore,it is important to establish a rapid and non-destructive method to identify adulterated meat to protect the rights and interests of consumers and stabilize the market order.In this study,the adulteration of lamb with cheap chicken and duck meat as well as the mixture of chicken and duck meat was studied,and the adulteration in lamb was identified using HSI combined with data fusion strategy at the same time,and the main research results are as follows:(1)Study on the identification of lamb adulteration based on HSI.The lamb was adulterated with chicken and duck meat and mixed with chicken and duck meat with adulteration gradients of 0%,10%,20%,30%,40%,50%,60%,70%,80%,90%and 100%,respectively.The samples were collected using Vis-NIR and NIR HSI,respectively,and the raw spectral information was obtained by ENVI 4.8 software.Six pre-processing methods,OSC,Baseline,MSC,MA,S-G and De-trending,were selected to pre-process the data,and five feature wavelength selection algorithms,CARS,SPA,UVE,VCPA and iVISSA,were used to extract the feature wavelengths from the data,and the PLS-DA algorithm was used to establish.the lamb adulteration identification model.The results showed that the Baseline-UVE-PLS-DA model worked best after the Vis-NIR data were processed,and the accuracy of the model calibration set and prediction set were 95%and 93%(chicken adulteration),94%and 94%(duck adulteration)and 93%and 94%(mixed adulteration),respectively.After the NIR data were processed,the De-trendingiVISSA-PLS-DA model worked best,with accuracy of 93%and 93%(chicken adulteration),94%and 92%(duck adulteration),and 96%and 90%(mixed adulteration)for the model calibration and prediction sets,respectively.The Vis-NIR-HSI system was found to have a better identification performance after comparison.(2)Study on the identification of different adulteration types of lamb by fusing spectral data from different spectral ranges.Data fusion techniques stitch together data matrices from different analytical techniques with the aim of obtaining more comprehensive information from different analytical techniques to produce more accurate predictive performance.In this chapter,data-level and feature-level data fusion methods are used to fuse the spectral data of Vis-NIR and NIR and to build a PLS-DA discrimination model.The discrimination models built by data-level fusion yielded 89%and 87%accuracy for the correction set and prediction set(chicken adulteration),86%and 84%(duck adulteration),and 86%and 86%(mixed adulteration),respectively.The identification models built by feature-level fusion had 96%and 95%(chicken adulteration),95%and 95%(duck adulteration),and 96%and 94%(mixed adulteration)accuracy for the correction and prediction sets,respectively.The results indicate that the fused data based on the data level contains too much redundant information,resulting in poorer modeling results than those of single spectra(Vis-NIR and NIR regions).However,the fused data models obtained by feature extraction methods to filter feature wavelengths for feature-level fusion yielded more accurate discrimination than the single spectral data models.(3)Study of lamb adulteration identification by map deep feature fusion.The deep features of Vis-NIR spectra and images were extracted by using SAE network,and PLS-DA model was established based on the fusion of spectral deep features,image deep features and map deep features,respectively.The deep features of the spectral data were extracted by SAE to reduce the 125-dimensional spectral variables to 23 dimensions,and the accuracy of the correction set and prediction set of the discriminant model established based on it were 96%and 94%(chicken adulteration),95%and 95%(duck adulteration)and 95%and 94%(mixed adulteration),respectively.The image data were reduced from 1296 dimensions to 50 dimensions by SAE extraction of deep features,and the accuracy of the correction and prediction sets of the discriminant model based on it were 70%and 68%(chicken adulteration),69%and 68%(duck adulteration),and 72%and 70%(mixed adulteration),respectively.The deep features of the map fusion data were extracted by SAE,and the 1421-dimensional variables were reduced to 66 dimensions,and the accuracy of the correction set and validation level of the discriminant model built based on them were 98%and 96%(chicken adulteration),97%and 96%(duck adulteration),and 96%and 96%(mixed adulteration),respectively.The spectral data can better reflect the information related to lamb adulteration than the image data,and the spectral fusion data combined with the deep learning feature extraction algorithm can further improve the accuracy of the discriminative model.