Rapid Detection Of Meat Quality Using Reflectance Spectroscopy And Hyperspectral Imaging With Multivariate Analysis Methods

China is one of the largest meat consumption country.Meat products can provide a variety of nutrients needed by human body,rich in extremely high nutritional value.The rapid development of meat industry hides many problems,such as meat adulteration and meat origin.The traditional meat quality detection methods are inefficient,complicated operation and low accuracy.Therefore,the research of fast and real-time detection technology has become an insistent need of meat industry.In this paper,reflectance spectroscopy combined with deep learning was used to quickly detect the adulteration of minced beef.Then hyperspectral imaging combined with effective variables of multiple features was used to authenticate the mutton geographical origin.The research contents were as follows:(1)The reflectance spectra combined with multivariate analysis methods were used to distinguish the pork meat,beef heart and beef tallow in beef,and the content of the pork meat and beef heart in beef adulterated to 0%-48%was detected quickly.A spectral acquisition platform was designed and built to obtain the reflectance spectroscopy of minced beef in the band of 350-2500 nm.Standard normal variate and Savitzky–Golay smoothing were applied to reduce spectral interference and noise.Then,SVM(support vector machine),RF(random forest),PLSR(partial least squares regression),and CNN(convolutional neural network)were adopted for adulteration type identification and prediction level.Moreover,the feature wavelengths were extracted by PCA(principal component analysis),locally linear embedding,subwindow permutation analysis,and CARS(competitive adaptive reweighted sampling).PCA-CNN identified adulteration type with the accuracy above 99%.In adulteration level prediction,CARS-RF model achieved the best performance,with R_p~2(determination coefficient of prediction)of 0.973 and RMSEP(root mean square error of prediction)of 2.145.In the detection of beef heart content in adulterated beef,the optimal model was CARS-PLSR,with R~2_pof 0.960 and RMSEP of 2.758.(2)Hyperspectral imaging with effective variables of multiple features was used to identify the mutton geographical origin.Firstly,rapid identification of mutton geographical origin were achieved based on spectral information of hyperspectral image.MSC(Multivariate scattering correction)was adopted to remove the scattering caused by non-uniform spectral surface.PCA was applied to select single band corresponding to the three principal components with the largest feature weight as the effective wavelengths,and RF,SVM and KNN(K-nearest neighbor)models were established.The results showed that MSC can improve the accuracy of models compared with the raw spectral modeling.Best performances were obtained using RF model as evidenced by ACC_C(accuracy of correction set)and ACCp(accuracy of prediction set)of 89.78%and 86.56%,respectively.The accuracy of the model based on features wavelength was slightly higher than that of the full wavelengths model,and the best performance was still the RF model(ACC_C=94.18%,ACC_P=90.55%).Then the hyperspectral image information was used to realize the fast identification of the mutton origin.Grayscale gradient co-occurrence matrix method was applied to obtain the texture characteristics of mutton,and image feature parameters were analyzed by correlation to extract effective modeling feature parameters.The model was constructed based on the image and the spectral information of mutton respectively.The accuracy of the model based on the texture feature of mutton was 88.54%and 67.50%,respectively.The experimental results showed that the image information had a certain effect on the identification of mutton geographical origin,but it was far lower than the classification accuracy of the model based on the spectral information.Finally,the spectral and image information of hyperspectral images were fused to realize the fast identification of mutton geographical origin.The four texture feature parameters selected by correlation analysis and the ten spectral feature wavelengths selected by PCA were converted to the same order of magnitude by the normalization algorithm.Finally,the fusion model of spectra and texture features was established by RF,SVM and KNN.The results indicated that the effect of the three models has been greatly improved,among which the best classification performance was the RF model,with the recognition accuracy of the prediction set reaching over 98%.Reflectance spectroscopy has been widely used in food,medical and agricultural industries,while relatively few studies on the identification of meat authenticity.Therefore,this paper mainly investigates the application of reflectance spectroscopy with the deep learning method in meat adulteration.Due to the spectroscopy technology contains only spectral information,but does not contain image information,which will limit the accuracy of detection.While hyperspectral imaging technology includes both spectral and image information of samples,which provides the possibility of using hyperspectral imaging technology.In conclusion,reflectance spectroscopy and hyperspectral imaging combined with multivariate analysis methods have the advantage of fast and accurate detection of meat quality,and the above analysis offers some theoretical references and practical exploration for the design of an online,economic and rapid meat quality detection system.