Detection Of Meat Freshness Based On Hyperspectral Imaging

Meat freshness is an important indicator of meat quality,closely relating to consumer safety.More importantly,as living standards improve,there is now a higher demand for food quality.So it is very interesting to study how to get fast,non-destructive and high precision methods to detect meat freshness.The subject makes full use of the theory of hyperspectral image,the theory of machine learning(minimum support vector machine,sparse autoencoder network,theory of evidence,etc.)and data fusion idea to establish a high accuracy,fast and nondestructive meat freshness detection method.The main work of the dissertation is as follows:1.Sparse autoencoder network(SAE)is used for multi-feature fusion,and then the model is established for detection of meat freshness.Hyperspectral imaging owns both image information and spectral information.First of all,the spectral features of hyperspectral,entropy characteristics of the hyperspectral image,and the average image characteristics of the four directions under the Gabor transform are obtained to form multiple features.Because Gabor function can obtain the related texture features at different scales and different directions in the frequency domain,Gabor transform is applied to hyperspectral image processing to obtain more image texture features.Due to the information redundancy problem caused by hyperspectral 94 bands and multiple features,the partial least squares method(PLS)is used to select the nine optimal wavelengths.Aiming at the multi-feature of the wavelength selection,a multi-feature fusion is carried out by using a sparse autoencoder network.Finally,a least squares support vector machine model is established to predict the meat freshness.The results of the prediction model are that correlation coefficient of prediction(R_P)is 0.948,root-mean-square errors of prediction(RMSEP)is 2.01mg/100g and residual prediction deviation(RPD)is 3.12.The results show that compared with the multi-features of the wavelength selection,the feature fusion method based on PLS wavelength selection combined with SAE can significantly improve the accuracy and stability of the prediction model.2.After using two-dimensional principal component analysis for multi-feature fusion,a prediction model for detection of meat freshness is established.Making full use of the feature information of the hyperspectral image in the optimal feature band is the key factor to improve the precision and practical application of the hyperspectral image detection method of pork freshness.There are too many parameters to be adjusted for SAE,which is used to fuse multi-feature.In this study,two-dimensional principal component analysis(2DPCA)is used to replace SAE for feature fusion.After selecting the 9 optimal bands,the average spectral characteristics,entropy characteristics,and average image characteristics of these bands are obtained.Then,2DPCA is taken advantage of multi-feature fusion to obtain the information between features.Finally,a least squares support vector machine model is established to predict the meat freshness.The results of the prediction model are that root-mean-square errors of prediction is(RMSEP)1.86 mg/100g,correlation coefficient of prediction(R_P)is 0.955 and residual prediction deviation is(RPD)3.35.It provides a feasible method for practical application of meat freshness detection technology using hyperspectral imaging.3.The research makes use of the theory of evidence to fuse the data at decision-making level.The most used modeling method is still a single model,but the single model has the problem of non-generalization and so on.Therefore,this paper proposes the use of D-S evidence theory for model fusion.First,two models are established:SAE-based least squares support vector machine prediction model and 2DPCA-based least squares support vector machine prediction model.Then D-S evidence theory is used to fuse the two models.The results of the prediction model are that correlation coefficient of prediction(R_P)is 0.961,root-mean-square errors of prediction(RMSEP)is 1.78 mg/100g and residual prediction deviation(RPD)is 3.60.The results show that compared with a single model,the proposed method can effectively improve the stability of the model and obtain high accuracy.