Detection Method For Beef Adulteration Based On Data Fusion Between Near-Infrared Spectroscopy And Bio-speckle Imaging

As one of the main meat products,beef has high food value and commercial value.With the increase of beef deep processing,the phenomenon of beef adulteration is increasing,which seriously harms health and economic interests of consumers.In this paper,beef adulterated with different proportions of spoiled meat were selected as experimental objects,Near-infrared spectrum and bio-speckle images of the samples were collected,and the beef adulteration detection models based on near infrared spectroscopy,biological speckle imaging,fusion between near infrared spectroscopy and biological speckle image were established respectively.The best model for beef adulteration detection was determined through comparison analysis.The results of the study were as follows:（1）The detection method for beef adulteration based on near infrared spectroscopy were determined.To optimize the detection model of beef adulteration based on near-infrared spectroscopy,artificial fish-swarm algorithm（AFS）was proposed and used to select the near-infrared wavelengths and spectral preprocessing method synchronously,and Partial Least Squares Regression（PLSR）mode for beef adulteration detection based on genetic algorithm（GA）,AFS and full wavelength were established.In order to study the effect of the sequence between wavelength selection and preprocessing on the performance of the simplified model,two type of model,i.e,pretreatment before wavelength selection（PW）and wavelength selection before pretreatment（WP）were established.Results showed that the performance of the PW simplified model（PW_AFS）based on AFS is slightly better than that of WP model based on AFS（WP_AFS）,R_c²,R²_cv R_p² of WP_AFS were 0.98,0.90 and 0.88,respectively.RMSEC,RMSECV and RMSEP of WP_AFS were 0.04,0.08 and 0.09,respectively.On the contrary,WP simplified model based on GA（WP_GA）have better performance than PW simplified model based on GA（PW_GA）.R_c²,R²_cv R_p² of WP_AFS were 0.97,0.91 and 0.81,respectively.RMSEC,RMSECV and RMSEP of WP_AFS were 0.05,0.09 and 0.09,respectively.Comparing the optimal models（PW_AFS and WP_GA）of AFS and GA,it was found that the PW_AFS and WP_GA performed better than the full-wavelength model.In addition,compared with WP_GA,PW_AFS has larger RPD（2.95）and a smaller RMSE,the performance of PW_AFS is better than that of WP_GA.The results show that NIR can be used to detect the beef adulteration;AFS can improve the performance of model;PW_AFS.（2）The detection method for beef adulteration based on bio-speckle imaging was determined.In view of the poor stability of conventional bio-speckle image processing methods,this paper proposed a method of processing bio-speckle images based on the moment of inertia spectrum（IM）.The characteristics of even column IM spectrum,odd column IM spectrum and all column IM spectrum were analyzed,and they were applied in the regression analysis of beef adulteration,respectively.The performance of the support vector regression machine（SVR）model based on all columns IM spectrum yielded the optimal result,where the correction coefficient（R²）and the root mean square error（RMSE）were 0.86 and 0.12 as well as 0.82 and 0.11 for calibration and prediction,respectively.This indicate that bio-speckle technique imaging combined with IM spectrum analysis method was feasible to detect the adulteration content of beef.This study also studied the application of genetic algorithms and artificial fish swarm algorithm in the extraction of IM spectrum features.The results showed that the two methods failed to improve the performance of models and therefore further research is needed.（3）The detection method for beef adulteration based on fusion of near-infrared spectra and bio-speckle images was determined.The near-infrared and bio-speckles were merged based on the feature layer and the decision layer,respectively.For feature layer fusion,the features of near-infrared spectra and IM spectrum was extracted by AFS.Then,these features are normalized and combined into joint feature vectors using interval intersections as the input of SVR.The model was compared with the direct merging feature model,where R_c²,R_p²,RMSEC,RMSEP are 0.91,0.84,0.08and 0.10,respectively.The results showed that the model based on joint feature vector performed better,indicating that feature joint combination can be used to the detection of beef adulteration.In the other hand,the performance of feature layer fusion model was better than that of the bio-speckle imaging models when it was used alone.However the accuracy of the feature layer fusion model was inferior to that of the NIR when it was used alone.For decision layer fusion,the prediction results of the best model based on near-infrared and bio-speckle imaging were used as fusion targets for decision-level fusion.The model output R_c²,R_p²,RMSEC and RMSEP of 0.96,0.95 and 0.07,0.06,respectively.It showed that the fusion model based on the decision layer was accurate and stable.we compared the near-infrared spectral model,the bio-speckle model,and the fusion model of the feature layer and decision layer based on the two technologies,it is found that the fusion model of the decision layer fusion has high accuracy and good stability.Finally,the decision layer fusion model is determined as the best model for beef adulteration detection.