Research On THz Spectrum Sparse Representation And Feature Extraction Method

Dimension reduction,feature extraction and prediction model building of Terahertz(THz)spectral data are the focus of much attention.Effective spectral dimension reduction and feature extraction can improve the effectiveness and accuracy of the model.Sparse representation can reduce the dimension of spectral data while retaining effective spectral information.Multi-feature extraction can represent THz spectral information of samples as much as possible.In this paper,a THz spectral data processing model based on sparse representation and multi-feature extraction was proposed to deal with problems such as high dimension and large amount of computation of THz spectral data.Qualitative and quantitative analysis was carried out on sugar as the research object,and the validity of the model was verified.This study improves the quality of terahertz spectral data and model detection accuracy,and enriches and develops the theory and method of terahertz detection.Specific research contents are as follows:The terahertz spectral data of sugar samples were obtained and analyzed.Savitzky-Golay smoothing,adjacent-averaging,FFT filtering,percentile filtering,first derivative,first derivative + SG,second derivative,second derivative + SG and wavelet denoising were used to preprocess terahertz spectral data respectively.The principal component regression model was constructed,and the experimental results show that the percentile filter is the optimal pretreatment method,which can effectively solve the error caused by the terahertz spectrum measurement process and system uncertainties.To solve the problems of large amount of THz spectral data,high dimension,complex structure and long time of model calculation,sparse representation was used to reduce the dimension of pre-processed THz spectral data.OMP was used to calculate the sparsity coefficient,and K-SVD was used to update the dictionary to achieve the sparse representation of terahertz spectral data.Compared with the four traditional dimensionality reduction methods,such as principal component analysis,linear discriminant analysis,local linear embedding and isometric mapping,a Support Vector Machine(SVM)classification model was established.Compared with other dimensional reduction methods,the time of the sparse terahertz representation model proposed in this paper is reduced to 0.132 s,and the experimental results verify the effectiveness of the sparse representation nonlinear dimensional reduction method to improve the model efficiency.In order to further improve the prediction accuracy and performance of the model,a method combining sparse representation and multi-feature extraction is proposed.SVM classification model for sugar,PLS quantitative model for sucrose-polyethylene mixture and PLS quantitative model for glucose-wheat flour mixture were established by selecting single and multiple features as the input of the model.The experimental results show that the performance of the four-feature model is significantly improved compared with that of the single-feature model.The accuracy of the four-feature SVM classification model for sugar was0.98.The determination coefficients of PLS quantitative model of sucrose-polyethylene mixture and glucose-wheat flour mixture were 0.938 and 0.867,and the root mean square errors were 0.067 and 0.051,respectively.In this paper,terahertz spectral data preprocessing,sparse representation and multifeature extraction modeling were studied,and a terahertz spectral sparse representation and multi-feature extraction model was built.The model was verified and evaluated,and the quality of terahertz spectral data,classification accuracy and performance of the model were improved.It provides a theoretical basis for the application of terahertz spectroscopy in the field of agricultural products detection and has important practical guiding significance.