Rapid Detection Method Of Cooked Beef Freshness Using Hyperspectral Imaging Technigue

Meat products,as an integral part of the human diet,are favored by most consumers.However,the freshness of meat products and the degree of spoilage in storage directly affect consumers' physical and mental health and desire to buy,which requires higher requirements for the detection and evaluation methods of meat industry.At present,the freshness detection of meat products is realized mainly through some conventional methods,such as sensory evaluation,physicochemical indexes and microbiological experiments.However,there are many problems such as strong subjectivity,complex sample processing,large consumption of reagent and long detection cycle,which obviously cannot satisfy the rapid development demand of modern meat industry.Therefore,it is urgent to develop a fast,accurate and efficient method for the detection of fresh meat products.Currently,hyperspectral imaging technology is now a scientific and effective tool for meat quality and safety detection,taking into account both the internal and external attributes of samples.In this paper,the cooked beef samples with different refrigerated time were taken as the research object,and the hyperspectral imaging technology was used to detect the freshness of beef quickly and effectively.A series of quantitative analysis models of volatile base nitrogen(TVB-N)content,total colony count(TVC),biogenic amine content(TBA),water content and refrigeration time in cooked beef was established.At the same time,a qualitative analysis method for judging the grade of freshness was established.The safety and freshness of cooked beef can be foreseen in advance.It has important research and practical significance to realize the scientific guidance of meat processing and production and to meet the consumer's demand for meat quality.The main reserch contents and results are as follows:Modeling and analysis based on full wavelength spectrum for the freshness of cooked beef.The measured data of TVB-N content,TVC value,total amount of TBA and water content in cooked beef were statistically analyzed.The results met the requirement of modeling.Further,the full band spectral characteristics of the cooked beef samples were analyzed.The combination of four quantitative analysis methods(the partial least squares regression(PLSR),BP neural network(BP-ANN),least squares support vector machine(LS-SVM)and extreme learning machine(ELM))and different pretreatment methods(multiplicative scatter correction(MSC),wavelet transform(WT),standard normal variate(SNV),two derivative(2ND))were used to a series of prediction models of freshness indexes based on full band spectrum,which performances were compared and analyzed.The results showed that the MSC-ELM model was the most suitable predictors of TVB-N content,and the WT-PLSR model was the most suitable predictors of TVC value.RAW-ELM model was the most suitable predictors of total TBA,while SNV-BP-ANN and SNV-PLSR models were the most suitable predictors for water content and refrigerated time.Modeling and analysis based on feature wavelength spectrum for the freshness of cooked beef.On the basis of spectral pretreatment,further using four kinds of characteristic wavelength selection algorithm(variable combination population analysis(VCPA),random frog(RF),clonal selection algorithm(CSA),sparse representation(SR))to select out the feature wavelengths that were most relevant to freshness index.And a series of simplified models were constructed,the feasibility of improving the performance of the models were discussed.The analysis showed that the simplified MSC-VCPA-ELM,WT-RF-PLSR,RAW-SR-ELM and SNV-CSA-BP-ANN models were the best prediction models for TVB-N content,TVC value,total TBA,water content and refrigeration time respectively.Compared with the full band modeling,the determination coefficient()of the simplified models in test set was increased by 1.9%,1.9%,3.7%,0.5%,4.2%,respectively.The root mean square error(RMSE)decreased by 9.4%,36.6%,17.3%,5.9% and 17.3% respectively.And the amount of calculation is greatly reduced.It can be seen that the selected feature bands can replace the full spectral variables to characterize the freshness of cooked beef.Modeling and analysis based on image feature information for the freshness of cooked beef.Starting from the image domain,according to the feature wavelengths with different fresh indexes to select out the corresponding feature images.Tamura algorithm,discrete wavelet transform(DWT)singular value decomposition algorithm and discrete cosine transform(DCT)coefficient decomposition algorithm are used to extract texture feature variables of each sample respectively.And using PLSR,BP-ANN,LS-SVM and ELM algorithms to establish a series of prediction models of freshness indexes based on texture feature information.Analysis showed that the optimal models of TVB-N content,TVC value,total TBA,water content and refrigeration time in cooked beef were DCT-ELM(=0.742;RMSEP=3.592)? Tamura-PLSR(=0.802;RMSEP=3.264)? DWT-ELM(=0.805;RMSEP=3.128)? Tamura-BP-ANN(=0.805;RMSEP=2.988)?DWT-BP-ANN(=0.883;RMSEP=2.168),respectively.But the accuracy and stability of the models are both poor,and the prediction ability is less than the simplified model based on the spectral feature wavelength of the corresponding freshness indexes.As a result,the prediction model based on the single image texture feature information is not suitable for the detection method of freshness index in cooked beef.Based on the two color spaces of RGB and HSV,the color feature variables are extracted from the sample images,and the prediction models of freshness index of cooked beef are established based on color feature information.The analysis shown that,in addition to the water content,the performance of other models is better than that of the corresponding indexes based on the texture feature information models.But at the same time,it is still less than the simplified model established based on spectral characteristic wavelengths.In addition,the prediction model for storage time of cooked beef based on texture or color features is basically equivalent to the prediction ability based on spectral feature modeling.Modeling and analysis based on feature information fusion for the freshness of cooked beef.The feature spectral information,image texture and color feature information were fused by standardized manner.And using PLSR,ELM,BP-ANN,LS-SVM respectively to establish a series of prediction models of freshness indexes based on different information fusion in cooked beef.The analysis showed that the established ELM model by the fusion of spectral and color features is the best model for predicting the content of TVB-N(=0.958,RMSEP=1.027);PLSR model established by fusion spectral and image texture features as the best model for predicting TVC value(=0.955,RMSEP=1.104);After the fusion of spectral and color feature information is most related to the total amount of TBA,the performance of the established ELM model is optimal(=0.961,RMSEP=1.011);In addition,the established BP-ANN model based on spectral and texture information fusion is the optimal prediction model for moisture content,theis 0.911 and RMSEP is 2.079 of test set;The BP-ANN model based on the fusion of three features is the best model for predicting the cold storage time of cooked beef(the is 0.942 and RMSEP is 0.927 in test set).The performance of all the model after the fusion of feature information is superior to the model based on single spectral,image texture,or color feature information.Thus it may be known,the fusion of feature information can effectively improve the prediction accuracy and stability of the model.The predicted values were further converted to the image of cooked beef samples,and the visualized distribution maps of TVB-N content,TVC value,total TAB,water content and refrigerated time were generated.The dynamic changes of the internal quality parameters of cooked beef during the cold storage and the authenticity of the shelf life weree understood in a more intuitive and clear way.Discriminant analysis for the freshness grade of cooked beef.The freshness grade of cooked beef samples was divided into three categories by using the measured values of the total colony counts.Using VCPA,RF,CSA and SR four kinds of characteristic wavelength selection algorithms to select out feature wavelength variables that can represent the essential information of cooked beef,respecttively.The new classification model of cooked beef freshness was established by using two methods of soft independent modeling of class analogy(SIMCA)and fuzzy neural network(FNN).By comparison,we can see that the established RF-FNN model based on 6 characteristic wavelengths(645nm,746 nm,812nm,851 nm,926nm and 965nm)selected by RF algorithm combined with the FNN method showed strong stability and classification ability.The classification accuracy(CCR)of the correction set and the predictive set is 95.71% and 97.14%,respectively,and the Sensitivity and Specificity results in the range of 0.90 to 1.Furthermore,the feature images of the spoilage beef samples were extracted,and the feature images were fused by principal component analysis(PCA).Further,the threshold segmentation method was used to identify the surface and the scattered point pollution area.The results showed that the RF-FNN model could effectively discriminate the freshness grade of cooked beef.PCA combined with threshold segmentation technology can accurately identify the contaminated area of spoilage meat,and the classification and recognition achieved satisfactory results.