Study On Nondestructive Detection Of Freshness Of Mutton Based On Machine Vision

Mutton is tender,delicious and nutritious,which has also been one of the main meats consumed by the residents of the western in Xinjiang and Ningxia in daily life.In recent years,it has been favorited by consumers.With the improvement of living standards,people's demand for the freshness of mutton is also increasing.At present,traditional detection methods for the freshness of meat always require complex sample preparation,which is very time consuming and laborious,and it is difficult to satisfy the need for rapid and non-destructive testing of meat freshness..Therefore,it is important to explore a rapid and accurate method for the detection of mutton freshness for mutton quality testing and grading.Machine vision technology is fast,nondestructive,low-cost,safe and reliable,and is gradually replacing traditional defect detection methods.In recent years,with the progress of science and technology,it has been gradually improved and promoted,including color images and hyperspectral images,etc.This article takes Xinjiang mutton with distinctive local characteristics as the research object,comprehensive use of hyperspectral imaging technology and color image technology combined with image processing technology and pattern recognition methods,to carry out rapid nondestructive testing of mutton.The main research contents and results are as follows:(1)The freshness detection of vacuum-packed cooled mutton using the color image.For lamb color images,the R,G,and B images were processed using median filtering and homomorphic filtering to extract the gray level co-occurrence matrix(GLCM)and Gabor texture features,and a PLSDA model was established to compare the effects.It was found that the use of homomorphic filtering to extract the combined texture feature modeling effect is better,the accuracy of the model calibration set,cross validation set,and test set was 88.19%,83.46%,and 85.71%,respectively.On this basis,GA,SPA,and CARS were used to screen feature variables,and SVM,ELM,BP,and adaboost-BP models based on the feature variables were established.Among them,the three-layer BP neural network model established by GA is the optimal mutton freshness detection model.The accuracy of model calibration set,cross validation set and test set was 100%,92.94% and 95.24%,respectively.(2)The freshness detection of vacuum-packaged cooled mutton using hyperspectral RGB images and feature band images.Hyperspectral RGB band images and feature band images were used to detect the freshness of vacuum packaged lamb.Among them,the RGB band images and the feature band images was averaged from five band images around 640.1 nm,550.4 nm,459.9 nm and 476.9 nm,562.7 nm,625.0 nm,650.2 nm,816.7nm,respectively.For RGB band images and feature band images,the PLSDA mutton freshness detection effect of different image filtering methods combined with different texture feature extraction methods was discussed.On the basis of optimal full-variable PLSDA modeling,GA,SPA,and CARS were used to select feature variables,and SVM,ELM,BP,and adaboost-BP comparative analysis models based on feature variables were established.For hyperspectral RGB band images and feature band images,both the results of CARS-ELM modeling were optimal.The accuracy of the model calibration set,cross verification set,and test set were 98.43%,91.36%,90.48% and 93.70%,91.36%,95.24%,respectively.(3)The freshness detection of non-vacuum packaged mutton using the hyperspectral image information.First of all,the hyperspectral image of the sample was subjected to band operation,binarization,and masking to remove image background,fat,etc,and the pure muscle area of the sample was used as a region of interest.Subsequently,one and multiple spectra(2,4,8,10)were extracted from the pure muscle area of each mutton sample,and then the comparative study of partial least squares discriminant analysis(PLSDA)models was carried out.It was found that the optimized 8 spectral modeling effects was the best.The accuracy of the calibration set,cross validation set,test set and verification set was 94.34%,91.51%,95.31% and 95.65%,respectively.Second,under the optimal modeling spectrum,different variables screening methods and modeling methods were used to identify the freshness of mutton.Among them,the CARS-ELM model had the best performance.The accuracy of the calibration set,cross validation set,test set,and verification set was 97.64%,97.24%,100% and 100%,respectively.It shows that the CARS-ELM model can detect the freshness of mutton better.