Study On Detection Of Chilled Mutton Freshness Based On Hyperspectral Imaging Technique

In the dietary structure,meat has long been an important role for our residents.Mutton is one of the most major meats in our country.Therefore,the quality and safety of mutton is a main event,which relates to the people’s livelihood,and a great practical significance for controlling mutton quality and protecting the safety of meat consumption.At present,most of the mutton quality detection depends on sensory evaluation and lab physical and chemical detection combined with microbiological indicators.But all of them are not conductive to the rapid detection of mutton circulation.In recent years,hyperspectral imaging technology has become a hotspot in the field of non-destructive detection for agricultural and livestock products because of its high resolution,easy operation and non-destructive characteristics.It has also made some achievements in the application of meat quality nondestructive detection.However,there is still some problems need further studied.Some factors impact the speed of evaluation model,such as cube data,large scales,high redundancy,and large amount of calculation during data processing and time-consuming.Data processing methods will affect the accuracy of the model.The sample varieties and other factors will affect the scope of application of the model.Therefore,the study on mutton quality non-destructive detection using hyperspectral imaging,can improve the detection progress and detection speed,enhance the adaptability of the detection model,and promote the further development.The improvement of mutton industry management ability has important scientific significance and broad application prospects.In this dissertation,chilled mutton acted as the detection objects.According to the national food safety detection standard,total volatile based on nitrogen(TVB-N),pH value,total bacterial count(TBC)and total coliform(MPN)were used as freshness indexes.The method of spectral data preprocessing,spectral feature extraction and chilled mutton freshness classification model for non-destructive detection of freshness of mutton with hyperspectral imaging were studied by stoichiometry,statistical analysis,machine learning and computer technology.The study contents and results are as follows:1)According to the national food safety detection standards,key physical and chemical and microbial indicators of the chilled mutton were determined to freshness evaluation.The freshness of chilled mutton was evaluated from two aspects of physicochemical and microbiological factors by TVB-N and pH as the indicators of physical and chemical detection,and the total number of colonies and coliform bacteria were used as indicators of microbial detection.Through laboratory detections and analysis results,it showed that the correlation between the four indexes was higher,and each index had significant difference with the three fresh grades of chilled mutton,which indicated the separability was better.2)A variety of data preprocessing methods,commonly used for hyperspectral imaging,were studied.The results showed that the two mixed methods of S-G convolution smoothing and multiple scattering corrections had good pretreatment effect on hyperspectral data of chilled mutton.3)In this dissertation,we propose a feature extraction method based on kernel sparse graph embedding canonical correlation analysis(KSGECCA).It is generally known that the hyperspectral image data has high dimensionality,non-linearity,redundancy and strong correlation between bands.Firstly,the non-linear mapping of Gaussian kernel is used to improve the nonlinear separability of hyperspectral data for chilled mutton.The sparse representation technique is used to express informations of the whole band in the feature space as a linear combination of the few sparse dictionaries and its coefficients.The data itself is excavated while reducing the redundancy between the bands,and the information of the dictionary is expressed.Through the nonlinear mapping and sparse representation for the original spectral data,the complexity of the calculation process is constantly improving.Thus,this dissertation uses the graphical embedding framework to expand the canonical correlation analysis.It can be a good mapping and sparse representation combined with a canonical correlation analysis method.At the same time,on the one hand the complexity of the entire calculation process can be reduced;on the other hand the hyperspectral data in the nonlinear manifold structure can be fully maintained.In this dissertation,the feature extraction method is analyzed by the eigenvalue curve,scatter plot and correlation analysis.While the method is contrasted PCA feature extraction method,the canonical correlation analysis method and its deformation,the results show that the proposed method is effective for the detection of chilled mutton.4)This dissertation proposes a random sampling spectrum clustering method based on sparse representation and a classification method based on adaptive BP neural network.The classification model,based on hyperspectral non-destructive detection,for the freshness of chilled mutton was discussed in this dissertation,from the perspective of unsupervised and supervised classification,respectively.Because of the high complexity of graph construction in spectral clustering,it is difficult to cluster analysis large data of spectral scale by the spectral regression method.At the same time,as the feature extraction results are associated with sparse representation of hyperspectral information,this dissertation proposes sparse representation evaluation model of freshness classification of spectral clustering in characteristic space.This dissertation not only studies the classification model of chilled mutton in clustering method,but also researches the application of BP neural network in the classification of chilled mutton freshness,which is supervised classification method.Due to the large size of the hyperspectral data,learning time in training processing by the traditional BP neural network algorithm is too long and easy to fall into the local minimum value,which will affect the learning efficiency and classification accuracy.Therefore,this dissertation proposes an adaptive BP neural network to analyze the freshness of chilled mutton.In order to test the validity of the two kinds of chilled mutton freshness classification evaluation model,firstly,this dissertation classifies the four standard values of the freshness of the conventional test through the two classification methods,respectively.The standard accuracy of the four indexes can reach 88% using stochastic spectral clustering method,and the overall accuracy is 100% by the adaptive BP neural network classification method.It is shown that the overall accuracy of the supervised classification method is higher than that of the unsupervised clustering method.During experiments,the parameters setting and optimization of the two methods are detailed tested simultaneously.Based on the experiment of four freshness indexes measured by the conventional measurement method,the hyperspectral characteristics of chilled mutton under different feature spaces were applied to the two classification models respectively,and the parameters setting and optimization were further adjusted and tested during the operation of the model.Through the experiments,compared with the standard results of the conventional measurements,it is concluded that the two methods,proposed in this dissertation,can be used to establish the freshness classification evaluation model,which is effective and feasible.Meanwhile the results show that the overall classification accuracy,obtained the supervised adaptive BP neural network classification method,is higher than that of the unsupervised clustering method.