Study On Feature Extraction And Classification Of Terahertz Spectrum

Physical and chemical information is embodied plentifully within the terahertz spectrum. It has been provided with some outstanding features about high sensitivity direct at compound crystal, single photon with low energy and high signal-noise ratio. The spectral classification task is extremely difficult because of physical limitations of terahertz radiation detectors, inspector’s prior knowledge, shielding by intervening cargo materials, accuracy of classification algorithm and the presence of background noise. As a consequence, accuracy and efficiency of spectral recognition is still at low level. In this paper, feature extraction and classification of spectra samples is studied to improve detectability of terahertz spectrum. Through the experiment and analysis, the following results are obtained.First, the method of Convex combination Kernel Principal Component analysis(C-KPCA) is studied to extract terahertz spectrum feature. In the terahertz spectrum data set, such method to have a better performance than Kernel Principal Component analysis(KPCA) in the aspect of selection of kernel function. The Convex combination kernel function is core for the method and we can get the parameters through the kernel function evaluation. Kernel function evaluation is a solving process of high-dimensional nonlinear programming equation. The experimental results show that features of spectrum was extracted triumphantly by such method. The method is proved to be better than KPCA by analyzing of the relationship between clusters.Secondly, the method of Logic Discriminant Analysis(LgDA) is studied to extract terahertz spectrum feature,. Regularization multivariate logistic regression of terahertz spectrum is core and it was utilized as a posteriori probability for optimal nonlinear discriminant analysis(ONDA). The experimental results show that features of spectrum was extracted triumphantly by such method. The method is proved to be better than LDA by analyzing of the relationship between clusters.Thirdly, Support Vector Machine (SVM), which is based on convex combination kernel function, is utilized for classification of THz pulse transmission spectra. To give difference to C-KPCA and LgDA, peaks and valleys of terahertz spectrum and Term Frequency Inverse Document Frequency(TF-IDF) are regarded as features. TF-IDF can conclude weight of each sampling point from the information theory. The weight represents the possibility that sampling point becomes feature. Kernel evaluation is utilized as an evaluation method for obtaining the parameters of optimal convex combination to achieve a better accuracy. When the optimal parameter of kernel function is determined, the model to process of classification and prediction is composed. Compared with the single kernel function, the method can combine with spectroscopic features with classification model iteratively. We carried out experiments using different samples. The results demonstrate that the new approach is on par or superior in terms of accuracy and much better in feature fusion to SVM with single kernel function.