The Color Detection Of Cold Fresh Tan Mutton Based On Deep Learning And Hyperspectral Technology

Yanchi Tan sheep is an advantageous animal breed in Ningxia agriculture.It has unique qualities such as tender meat,no smell of mutton,and low fat content,and is deeply loved by consumers.As the most intuitive visual indicator for evaluating the freshness of mutton,meat color determines the commodity value of meat.Traditional meat color detection methods have problems such as rough accuracy,damage to samples,time-consuming and labor-intensive,high cost,and difficulty in realizing batch detection,which cannot meet the rapid development needs of the meat industry.Therefore,an emerging green non-destructive testing method that is fast,economical,efficient,and all-round measurement is urgently needed.This paper takes the nine-month-old Yanchi Tan mutton as the research object,and uses the Visible Near Infrared Reflection(Vis/NIR)hyperspectral imaging technology and deep learning model to analyze the color index parameters(L*,a*,b*)of the cold fresh Tan mutton.Quantitative research was conducted to provide a scientific and effective theoretical reference for quickly identifying meat color freshness and meat color changes.The main findings are as follows:(1)Research on color and luster detection of whole-band hyperspectral cold fresh Tan mutton.Use Vis/NIR hyperspectral experimental equipment to collect Tan mutton hyperspectral images,use CM2300d colorimeter to measure Tan mutton color index(L*,a*,b*)reference values,and correlate the spectral information of the sample with the instrument measurement indicators,using the Monte Carlo method to detect the chemical outliers of the color index of the sample;using the Spectral Physical and Chemical Value Symbiosis Distance(Sample set Partitioning based on joint X-Y distance,SPXY)algorithm to divide the calibration set and test set samples according to the ratio of 3:1;using different The preprocessing method is used to process the original full-band spectral data and analyze the linear partial least squares(Partial Least Squares Regression,PLSR)modeling.The results show that for Tan mutton color index L*,the best preprocessing method is the baseline calibration(Baseline)algorithm,the model calibration set Rc=0.944,the prediction set Rp=0.896;for Tan mutton color index a*,the original The model built by the spectrum has the best effect,its Rc=0.892,Rp=0.853;for the Tan mutton color index b*,the best preprocessing method is the de-trending algorithm,the model calibration set Rc=0.902,the prediction Set Rp=0.819.(2)Research on the color detection of mutton in the characteristic band hyperspectral cold fresh tan.Five different methods and quadratic combination algorithm were used to extract the characteristic wavelengths of the preprocessed spectra,and the quantitative analysis models of linear PLSR and nonlinear Least Squares-Support Vector Machine(LSSVM)were established.The results show that the best model of L*index characteristic band is the LSSVM model established by 12 characteristic wavelengths extracted by the continuous projection method(Successive Projections Algorithm,SPA),and its Rc=0.953,Rp=0.915;the a*index model has the best performance.is an LSSVM model established by 32 characteristic wavelengths obtained by the Competitive Adaptive Reweighted Sampling(CARS)algorithm,with Rc=0.914 and Rp=0.856;the best prediction model for b*is the 12 extracted by the SPA algorithm.The LSSVM model established by the characteristic bands,and Rc=0.887,Rp=0.834.(3)Deep learning research on image recognition of cold and fresh mutton.The RGB images of Tan mutton were collected by smart phones and fill light circles,and the artificially collected image data set was pre-processed.Model training and visual representation of images from different convolutional layers.The results show that the self-built MLP-CNN network model has better predictive analysis performance than the improved AlexNet network model.Among them,the prediction result of the improved AlexNet network model for the L*indicator RMSEP=1.701,the prediction result of the a*indicator RMSEP=0.851,and the prediction result of the b*indicator RMSEP=0.814;The effect RMSEP=1.563,the predicted effect of a*has RMSEP=0.657,and the predicted result of b*has RMSEP=0.689.