Research On The Recognition Method Of Terahertz Spectrum Based On Manifold Learning

Organic compounds and biological macromolecules have the "fingerprinting"feature in the terahertz band.When using terahertz wavethe to test material,the harmful photoionization will not occur because of the photon energy of terahertz wave is very low.In recent years,with the rapid development of terahertz radiation source and detector,the applications in material identification and nondestructive testing have increased gradually.Accordingly,the number of terahertz spectroscopy is rapidly increasing.How to quickly and efficiently identify the terahertz spectroscopy is the important issue that scientific research workers have to face at this stage.The absorption peaks in the terahertz band are regarded as feature to identify the terahertz spectroscopy in traditional way.However,for many materials,there are no apparent spectrual graphics features in the terahertz band,such as peaks,valleys and etc,or their spectrum curves are very similar,which make it difficult to extract their terahertz spectroscopy feature and identify by traditional way.Due to the dimension of terahertz spectrum data is very high,it will be extremely difficult to identify the terahertz spectroscopy without dimensionality reduction and feature extraction.Aiming at this problem,this paper proposes a terahertz spectroscopic recognition method based on Diffusion Maps.From a pure data perspective,using Diffusion Mapps,a manifold learning method,to reducce the dimension of terahertz spectrum data and extract its low dimensional manifold features in keeping the intrinsic geometric structure of terahertz spectrum data at the same time.Avoiding the the influence of spectrum itself graphics information for the terahertz spectroscopic identification.The method can also directly determine the target dimension of dimension reduction,and the manifold features have a high degree of differentiation and clustering effect.In practice,the terahertz spectroscopy are not measured in the same batch.The traditional manifold learning methods can only handle the sample data from a same batch.The additional samples can be only combined with the original sample set,only after that can we utilize traditional manifold learning methods to reduce the dimension of sample collection.No doubt this will reduce the efficiency of recognition of terahertz spectroscopy,and can not identify terahertz spectroscopy dynamically.Aiming at this problem,this paper proposes a terahertz spectroscopic recognition method based on incremental Orthogonal Neighborhood Preserving Embedding.This method can effectively use the dimension reduction result of original sample set.Firstly,central points of all kinds of samples are selected as overlapping points from the original sample set and added to the additional sample set.Then integrating the dimension reduction results of two sample sets well based on the principle of overlapping points in the two sample sets having same low dimensional embedding coordinates.That way,it can greatly reduce the computational complexity of dimension reduction of terahertz spectroscopy,cluster and identify the terahertz spectroscopy dynamically.