基于扩散映射的太赫兹光谱识别

发布时间：2018-01-01 23:16

本文关键词：基于扩散映射的太赫兹光谱识别　出处：《光谱学与光谱分析》2017年08期 　论文类型：期刊论文

【摘要】：特征提取对于太赫兹光谱识别来说至关重要。传统方法是通过人工选取太赫兹光谱中差异性较大的吸收峰作为特征进行光谱识别,但当部分物质在太赫兹波段没有明显波峰、波谷等光谱图形特征时,这种方式便不再适用。为此,研究人员利用统计学习与机器学习方法对高维太赫兹光谱数据进行降维和特征提取。由于物质的太赫兹光谱数据各维度呈现非线性,尤其是当不同物质的太赫兹光谱曲线整体非常相似时,线性处理方法易产生较大误差。针对这一问题,提出了一种基于扩散映射(DM)的太赫兹光谱识别方法。扩散映射能在保持数据内在几何结构的同时对其进行非线性降维,提取的流形特征区分度较高,对数据还有聚类效果。首先用S-G滤波器对Alloxazine等10种物质的太赫兹光谱样本进行滤波,并用三次样条插值法对截取相同频段后的光谱样本进行统一分辨率处理;然后利用DM将高维太赫兹光谱数据映射到低维特征空间并提取太赫兹光谱的流形特征;最后用多分类支持向量机(M-SVM)对十种物质的太赫兹透射光谱进行分类。实验结果表明,相比于主成分分析(PCA)和等距映射(ISOMAP),使用DM提取的太赫兹光谱流形特征具有更高的区分度,而且DM可以直接得到太赫兹光谱数据本征维数的估计值,这为相似太赫兹光谱的快速精准识别提供了一条新的途径。
[Abstract]:Feature extraction is crucial for the recognition of terahertz spectrum. The traditional method is through artificial selection of difference absorption peaks of terahertz spectrum larger as the characteristic spectrum recognition, but as part of the material with no obvious peaks in THz band, spectral trough graphic features, this method is no longer applicable. Therefore, researchers used statistical learning and machine learning methods of feature extraction and dimension reduction of high dimensional data. Because of the dimensions of terahertz spectroscopy THz spectroscopy data of material non-linear, especially when different material spectral curves of terahertz overall very similar, linear processing method is easy to produce large errors. To solve this problem, propose a method based on diffusion maps of terahertz (DM) the spectral identification method. Diffusion maps can keep the data inherent geometric structure and the nonlinear dimension reduction, extraction Manifold feature discrimination is high, the data and the clustering results. Firstly using S-G filter THz spectroscopy samples of Alloxazine and other 10 kinds of substances are filtered using three cubic spline interpolation method to intercept the same spectral band of uniform sample resolution; and use DM to map high dimensional terahertz spectral data to the manifold features low the feature space and extract the terahertz spectrum; finally, the multi class support vector machine (M-SVM) classification of THz transmission spectra of the ten compounds. The experimental results show that compared with the principal component analysis (PCA) and distance mapping (ISOMAP), terahertz spectroscopy manifold features extracted using DM has higher differentiation moreover, DM can obtain the THz spectral data to estimate the intrinsic dimension of value, which provides a new way for the rapid and accurate recognition of terahertz spectrum is similar.

【作者单位】：昆明理工大学信息工程与自动化学院;昆明理工大学材料科学与工程学院;
【基金】：国家自然科学基金项目(61671225,61302042)资助
【分类号】：O433.5
【正文快照】： 引言有机化合物及生物大分子在太赫兹波段具有指纹特性,而且太赫兹波的光子能量很低,用于物质检测时不会发生有害的光致电离[1]。随着太赫兹辐射光源及探测器等硬件方面的迅速发展,太赫兹在物质识别和无损检测等领域的应用也逐渐增多。国内外许多实验表明,很多化学或生物物质

【相似文献】