Modeling And Application Of Terahertz Spectrum Based On Pattern Recognition Technology

Terahertz(THz)radiation occupies the electromagnetic spectrum between 0.1 and 10 THz.As a new front in-cross science,it bridging the gap between electromagnetic wave theory and micro quantum theory.As a beneficial supplement for infrared spectroscopy,terahertz wave shows many excellent characteristics,such as spectroscopic fingerprinting ability,perceptivity for the nonpolar and nonmetallic materials and strong safety,etc.In view of this,as the new technology of spectrum and imaging it has attracted the attention of many scholars at home and abroad and become a hot researching field of the non-destruction detectioninformation security.Over the last twenty years,research in terahertz time domain spectroscopy has progressed to such a great extent that terahertz is finding potential use in nondestructive detection of products/food,pesticide residues and identification of transgenic speciesand so on.While,owing to the infancy of the technology,much work has to be carried out to improve its utility and reliability,especially for the detection of complex agricultural products and food.In order to make a useful exploration of the application of THz spectroscopy in the field of agricultural product testing,this Thesis brings together the fields of THz spectroscopic analysis,pattern recognition,and digital image processing to advance the state of the art in quantitative and qualitative detection of materials based on the THz spectroscopy and THz pulsed imaging.The major result and contributions are list as follows:1.The sources of error existing in a terahertz time-domain spectrometer and throughout the parameter estimation process are systematically analyzed.Base on that an improved deconvolution algorithm for eliminating F-P effect in liquid transmission detection is adopted for assessing the uncertainty in THz-TDS measurements.The systematic errors and random errors introduced during the THz signal generation,transmission,detection and throughout the parameter extraction process are systematically analyzed from the aspects of system composition,sample measurement and error propagation.On this foundation,in order to solve spurious oscillations caused by the Fabry-Perot effect in the detection of liquid,an innovative liquid optical parameter extraction algorithm for THz time-domain spectroscopy is proposed.Considering the nonlinear absorption of THz wave by samples,containers,emitters or detectors,a THz time-domain trace containing echo signals can be represented as a convolution of the primary peak,some delta functions and nonlinear transfer functions.By analyzing equations,spurious oscillations in the THz spectra result from echo signals can be removed or reduced effectively.Experiment results show that this method can effectively improve the measurement accuracy of liquid optical parameters in THz band,and make a base for the following pattern recognition.2.Under the framework of Clifford algebra a THz signals analysis method is developed.Based on this,this department provide a concise mathematical means for attacking qualitative identification problem of THz spectra.From the point of view of multidimensional signal analysis,a new mathematical tool,Clifford algebra,is introduced in the processing of THz spectral signal.Under the framework of Clifford algebra,the THz spectral signal in frequency domain is expressed as a multidimensional real vector in Clifford vector space,and then,the effective information contained in THz spectrum is sufficiently utilized.Based on the theory of Clifford algebra,the geometric distribution and algebraic relation of THz signal vector are studied,and the relationship between the optical parameters of the sample and the THz signal vector is revealed.It was proved that,in ideal conditions,the transfer vectors obtained from the transfer function of one samples of same substance but with different thickness are co-planar and belong to a unique subspace corresponding to the complex refractive index of matter,the elative deviation is zero.Given that the signal is affected by inevitable noise and other factors in a terahertz time-domain spectroscopy system,so the vectors will deviate from their original orientation and move away from the subspace.Therefore,with the projection and rejection concept of the Clifford algebra,the relative deviation parameter are defined as the index of identification of THz spectra.Then a substance identification method based on the minimum value of elative deviation is demonstrated.Next terahertz spectroscopy measurements on four substances,melamine,tartaric acid,lactose,and glucose with six different thickness are performed.This data is used to validate the approach by comparing this classifier performance against a conventional support vector machine(SVM)classifier.3.The regularized extreme learning machine algorithm(RELM)is introduced to the quantitative and qualitative analysis of THz spectroscopy.Additionally,a fast LOO-RELM algorithm is adopted for the qualitative identification of genetically modified soybean oil.For the first time,RELM is introduced into the quantitative and qualitative detection of materials based on the THz spectroscopy.A novel algorithm based on the efficient RELM with LOO cross validation approach is proposed,which can improve the performance as well as reduce the complexity of the quantitative and qualitative models.In this algorithm,as to reduce the computational complexity of LOO error with every regularization parameter and automatically select the optimal model with limited user intervention,the singular value decomposition is employed to decompose the hidden layer output matrix H,and the corresponding pseudocode for the core algorithm is presented as well.The validity of the algorithm is demonstrated by two aspects: theoretical analysis,quantitative and qualitative experiments based on THz spectroscopy.Then this new approach is applied to identification of transgenic soybean oil based on THz spectroscopy,which exhibited the similar absorption spectra with its parent,and the identification results were compared with the well-known SVM algorithm in terms of generalization performance,recognition accuracy,parameter sensitivity and training time,etc.It was shown that both SVM algorithm and RELM algorithm can achieve high recognition accuracy and good generalization ability.However,it is also deserving to point out that the generalization performance of RELM is less sensitive to the training parameters.Thus,compared to SVM,RELM models are more efficient and more effective by avoiding tedious and time-consuming parameter tuning.4.A new cluster validity evaluation index VSO(-)is established.In this part of the thesis,the various techniques are briefly reviewed prior to a thorough discussion of the traditional cluster validity indexes.Then,a novel cluster validity index was presented that determines the optimal partition and optimal number of clusters for the THz pulsed imaging.Based on the measurement of the intra-cluster variation,the inter-cluster separation and the overlap between clusters,the proposed validity index considers not only the variation of the pixel within a cluster and the isolation of clusters from one another but also the pixels in THz pulsed imaging that have similar membership degrees to more than two clusters.The proposed VSO index is then tested and validated using several THz imagines cluster segmentation.The results of the comparison with other popular validity indexes show the superior effectiveness and reliability of the proposed index,especially applicable for the clustering of THz pulse images with large different in densities and sizes.5.Considering that the higher dimension of THz images can be a hurdle to their display,their analysis and their interpretation,in this department a THz pulse image clustering analysis algorithm based on PCA-FCM is proposed.As a new imaging modality,THz pulsed imaging has been used with great potential in many applications.Owing to its specific properties,THz imaging can supply a large amount of spectra features for every pixel and present a unique fingerprint of the material,but the higher dimension of THz images also bring challenges and vulnerabilities to the segmentation.In order to eliminate the interference caused by the noise and remove the redundant information in THz spectra signals and achieve the best trade off between the clustering accuracy and the low computational cost,we proposed a novel clustering approach,PCA-FCM,to realize the segmentation of THz pulsed images.This clustering approach is tested on two synthetic THz images and one real THz pulsed image.By the using of principal component analysis(PCA),the features in THz image that have less contribution to the variances are considered to be less descriptive and therefore removed.Then,according to the maximum membership the reconstructed image data set are segmented into different clusters based on the Fuzzy C means cluster algorithm.Experimental results show that,with the optimal number of clusters determined by the VSO(-)validity index,this approach is more efficient,and not only improves the convergence speed,but also makes the edge of image segmentation more clear than FCM without PCA and classical K-mean algorithm with or without PCA.