Nondestructive Detection Of Freshness Of Chilled Mutton With Multi-indicator Based On Hyperspectral Imaging Technology

Meat is considered a high-protein,nutritious,and delicious food by people.As the living standards of residents continue to improve,the consumption of meat products has been increasing year by year.Mutton is an important part of meat consumption and occupies a significant proportion.Therefore,ensuring the safety and quality of mutton is particularly important for guaranteeing the safety of residents’ diet and consumption.Freshness is a comprehensive indicator used to evaluate the safety and quality of meat products.People usually judge the freshness of meat based on appearance,odor,texture,shelf life,and other aspects.Traditional detection methods are complex and inefficient,and cannot meet the daily needs of people for mutton.Hyperspectral imaging technology is a fast,non-destructive,and efficient detection technology that can effectively detect changes in surface texture,color,and internal components of chilled mutton during the process of spoilage.However,hyperspectral image bands are numerous and have high data redundancy,which makes feature selection and model establishment difficult.In order to use hyperspectral imaging technology to evaluate the freshness of chilled mutton,this thesis has researched multi-indicator nondestructive detection methods for freshness of chilled mutton based on artificial neural network(ANN)and deep forest(DF)with the help of many physical,chemical and microbiological indicators.The main research work of this thesis is as follows:(1)Based on the national standard for the freshness of chilled mutton,this thesis designed an experimental plan and collected hyperspectral image data of chilled mutton samples stored at 4℃ in a refrigerator for 14 days.At the same time,the total volatile basic nitrogen(TVB-N),pH,total aerobic count(TAC),and approximate number of coliforms(ANC)of the samples were measured using chemometrics.(2)Through multiple experiments,this thesis selected S-G smoothing filter and multiple scattering correction algorithms to preprocess the raw hyperspectral data of chilled mutton samples,eliminating noise,clutter,and scattering effects.(3)This thesis researched the method of selecting feature bands in hyperspectral data to reduce the redundancy and facilitate the construction of freshness evaluation models.By comparing various dimensionality reduction methods,the successive projections algorithm is used to select feature bands of sample data,and finally 18 bands were selected as chilled mutton feature bands.(4)In order to construct a freshness evaluation model using multiple indicators,this thesis first improved the traditional single-indicator neural network classification model to a multi-indicator classification model applied to freshness evaluation.Considering the correlation between different freshness indicators,this thesis proposed to improve the error function in the feedforward neural network to increase the discrimination ability of multiple indicators and construct a multi-indicator freshness classification evaluation model for chilled mutton.(5)In order to simplify the model parameter settings and improve the computational efficiency of the model,this thesis further proposed a deep forest algorithm based on random tree construction.The algorithm applies multi-label metrics to deep forests,mines the correlation between multiple indicators of chilled mutton and hyperspectral imaging data through feature filtering mechanism,and uses layer-by-layer growth control to explore potential manifold structures in spectral data to achieve adaptive classification of freshness levels.