Research On Detection Method Of Restructured Beef Based On Raman Scattering Imaging Technology

Restructured beef technology can effectively bond the ground beef produced in the process of beef carcass splitting into intact beef,so as to improve the economic value of ground beef.However,in recent years,in order to seek illegal interests,illegal traders have used the characteristics that it is not easy to distinguish between the recombinant beef and the raw intact beef to replace the restructured beef label with the raw intact beef label,which has seriously damaged the interests of consumers.Raman scattering image not only has the characteristics of high specificity and sensitivity of conventional Raman spectra,but also have the ability to obtain the tissue characteristics of restructured beef samples and the spectral signals of subsurface substances.However,there are still some challenging problems in the detection of restructured beef based on Raman scattering image data.Raman scattering images have serious data redundancy problems,feature extraction is difficult,and the problem of beef spectral signal detection with package interference needs to be solved.This paper focuses on the above problems and establishes a high-precision and robust detection method for restructured beef.The specific research contents are as follows:(1)Aiming at the serious data redundancy of the restructured beef Raman scattering image data,a classification method of the restructured beef Raman scattering image based on band dimensionality reduction and feature extraction was studied.Firstly,the optimal band is selected according to the Mean feature of the scattering image,and then the key features are extracted by the scattering image feature extraction method.Finally,the least-squares support vector machine classification model(LS-SVM)is established.The experimental results show that the successive projection algorithm(SPA)band dimensionality reduction method can effectively reduce band redundancy,and 22 optimal bands are selected;Compared with mean reflection method(Mean),generalized Gaussian distribution feature method(GGD)and GGD-mean method,the modified Lorentz(MLD)feature extraction method is the most effective.The optimal classification model is established,and 95.8% recall,95.4% accuracy and 94.1%overall accuracy are achieved.This provides an effective method for rapid nondestructive testing of restructured beef.(2)Aiming at the problems of traditional modeling methods that feature extraction methods depend on human experience and there is a risk of losing effective information,a Raman scattering image transformer(RSITr)model based on transformer neural network is studied.RSITr introduces self-attention mechanism,which enables the model to focus on important bands or offsets in Raman scattering image data,so as to further improve the performance of the model.The experimental results show that compared with the traditional LS-SVM and the classical LSTM network,the RSITr model with self-attention mechanism has achieved the best model performance,in which the recall rate is 98.3%,the accuracy rate is94.4%,and the overall accuracy is 97.6%,which provides an effective way to establish an endto-end depth learning detection method for the restructured beef Raman scattering image.(3)In view of the widespread problem that adding packaging bags will seriously interfere with beef spectral signals,and a spectral signal detection method of beef with packages based on Self-modeling mixture analysis(SMA)and spatially offset Raman spectroscopy(SORS)technology was studied.This method is inspired by the principle of Raman scattering image technology.The SMA algorithm is used to decompose the SORS data to obtain the pure component spectra of internal and external substances.Finally,the decomposition effect is verified according to the decomposition curve and spectral angle(SA)parameters.The experimental results show that compared with principal component analysis(PCA)and nonnegative matrix factorization(NMF)algorithm,SMA achieves the best decomposition result,and has better decomposition performance in beef samples with different packing thickness.Therefore,the proposed method can effectively remove the interference of packaging on the spectral signal of beef puree.