Study On Pork Freshness Detection By Near Infrared Spectroscopy Based On Deep Learning

Pork,rich in nutrients,is the main source of meat and animal protein intake for Chinese people.However,it is easy to breed bacteria and deteriorate in processing,transportation and sales.In order to protect the interests of consumers and prevent food safety problems,it is necessary to strictly supervise the pork freshness.Traditional methods for pork freshness detection mainly include sensory evaluation and physicochemical analysis.The sensory evaluation has strong subjectivity cumber and the physicochemical analysis is time-consuming,which are incompatible with the requirements of real-time detection.Near infrared spectroscopy(NIRS)technique is a rapid and nondestructive analysis method,which has been successfully used in many fields such as food industry and agriculture.Considering that deep learning has the advantage of extracting internal features from data automatically,this research combines NIRS and deep learning to study the detection of pork freshness.The main contents of the research are as follows:Given the outliers and noise interference in the raw data,outlier elimination and spectral information recovery are studied to improve the quality and reliability of data set.Firstly,the outlier elimination methods based on Euclidean distance and Mahalanobis distance are employed respectively.Afterwards,the classification accuracies of one-dimensional residual network(1D-Res Net)on the test set before and after outlier elimination are compared to judge their effectiveness.Then,the effects of standardization,smoothing,multivariate scattering correction and derivative on spectral information recovery are studied.The experimental results show that standardization and smoothing can improve the quality of data set and raise the accuracy of 1D-Res Net on the test set by 1.03% and 0.32% respectively.Considering the qualitative analysis of pork freshness,frequently-used classification models such as support vector machine(SVM),random forest(RF)and partial least-squares discrimination analysis(PLS-DA)are established,in which the accuracy of SVM is the highest(90.82%).Then the Squeeze-and-Exception(SE)module is introduced into the 1D-Res Net to improve the network performance from the aspect of channel.The classification accuracy of the improved model is 95.53%.The experiment results show that SE module can effectively improve the generalization of model through adjusting the weights of channels and indicate that deep learning can realize the accurate correlation between pork spectrum and freshness.Finally,the effects of activation functions and pooling methods in SE module on network performance are studied through ablation experiments.Considering that there are few systems for pork freshness detection based on NIRS,a specialized system is developed from the aspects of software and hardware.The design of the software platform is based on C# and aims to realize easy interaction.The software platform covers the functions of user management,spectrometer focusing,spectral data acquisition,sample analysis and data query.The hardware module is designed to focus the spectrometer,mainly including controller,digital driver and stepping motor.The interface of the system is clear,intuitive and easy to operate,which can promote the practical application of the NIRS in the qualitative analysis of pork freshness.This thesis has 38 figures,9 tables and 87 references.