-
成像光谱仪中分光器件的特性是影响其光谱检测能力的关键因素。作者首先利用LCTF器件的光谱透过率特性建立LCTF成像光谱仪光谱检测的数学模型。LCTF的分光原理是基于向列液晶材料的双折射现象,以电压控制改变其利奥波片组件的双折射特性,实现其光谱透过率可调谐[4-5]。如果仅以调谐波长所对应的输出光谱分布为采样值,拟合LCTF成像光谱仪的光谱输出分布,由于LCTF光谱透过率函数对入射光谱的卷积作用,成像光谱仪的输出光谱与入射光谱相比,会降低对光谱细节的分辨。因此,利用成像光谱仪获得的采样输出光谱分布反求入射光光谱是改善其光谱检测能力的主要方法之一。
-
就成像光谱仪的分光工作原理而言,通过已知测量光谱和仪器函数反求入射光输入光谱是一个抽样恢复问题,是典型的数学物理反问题。在解决该类问题时,需要先建立入射光输入光谱与输出光谱间关系的数学模型。理想情况下,成像光谱仪测量波长为λ处的光谱强度Ei(λ)是入射光输入光谱在波长λ处的取值,考虑到由分光器件特性决定的仪器函数,由成像光谱仪得到的光谱强度Eo(α)是一个与中心波长α有关的积分值,即:
$ {E_{\rm{o}}}\left( \alpha \right) = \int_{{\lambda _2}}^{{\lambda _1}} {{E_{\rm{i}}}\left( \lambda \right)T\left( {\lambda, \alpha } \right){\rm{d}}\lambda } $
(1) 式中, 积分区间[λ1,λ2]表示的是LCTF的透过率窗口,T(λ, α)为透过率函数,其形式取决于光谱仪的分光器件。对应LCTF成像光谱仪而言,透过率函数取决于LCTF的光谱透过率,然而生产商通常并未提供器件的光谱透过率函数。因此,本文中采用了AvaSpec公司的商用光纤光谱仪对其光谱透过率进行测量,拟合得到了每一中心波长下透过率函数的解析形式为:
$ T\left( {\lambda, \alpha } \right) = A{\rm{exp}}\left[{\frac{{-{{\left( {\lambda-\mu } \right)}^2}}}{{2{\sigma ^2}}}} \right] $
(2) 式中, A, μ, σ为参量, 在每一中心波长α下均有不同的取值,在720nm~980nm中心波长的透过率测量范围内,A, μ, σ与中心波长α的关系可以用一次或二次多项式拟合。利用AvaSpec光纤光谱仪测得的LCTF在720nm~980nm调谐波长下1nm的波长间隔的光谱透过率曲线,均表现为高斯函数曲线形式,如图 1所示。
在LCTF的光谱透过率函数T(λ, α)确知的情况下,LCTF成像光谱仪的测量过程即是对(1)式的一个求解过程,显然该求解过程是一个反卷积过程。(1)式属于第1类Fredholm积分方程[14],此类方程通常是不适定的。对于LCTF成像光谱仪而言,(1)式中T(λ, α)经实测形式为高斯函数,易证(1)式的解,即所求入射光光谱分布,具有存在性,但定解问题需要引入额外约束条件[14]。考虑到LCTF成像光谱仪对调谐波长下光谱分布测量值反映了入射光场的光谱信息,因而本文中以其作为约束条件,采用正则化方法将(1)式的求解问题转换为可稳定求解的近似问题,在偏差允许的范围内可用近似问题的解代替原问题的解,也即求解出入射光场的光谱分布。相比于通过LCTF成像光谱仪测得的调谐波长下的光谱数据Eo(α)而言,该方法所求得的入射光谱近似解与实际入射光光谱之间的偏差较小。近似光谱与入射光光谱的接近程度衡量了该方法对实际入射光光谱的分布Ei(λ)的逼近程度,反映了该方法对LCTF成像光谱仪光谱检测性能的改善。
-
截断奇异值分解(truncated sigular value decomposition, TSVD)正则化方法是解决第1类Fredholm积分方程求解的常用方法[15],它把积分方程求解问题转换为最优化求解问题[16],通过调整系数矩阵完成正则化近似求解,需将(1)式中的光谱测量积分方程离散化为矩阵形式:
$ {\mathit{\boldsymbol{E}}_{\rm{o}}} = \mathit{\boldsymbol{T}}{\mathit{\boldsymbol{E}}_{\rm{i}}} $
(3) 式中, 矩阵核函数T为LCTF的光谱透过率函数离散化得到的。经验证直接利用TSVD方法所得近似解与入射光光谱的偏差较大,且解的稳定性不佳。作者参考TSVD正则化方法的最优化求解过程,考虑入射输入光谱与该方法近似光谱解的一致性作为约束条件,提出了一种采用了先验约束的正则化求解方法。新方法设输出光谱Eo插值为Eo′,使Eo′向量长度与所求输入光谱Ei向量长度保持一致,以β为正则化参量定义正则化项β‖Ei-Eo′‖,则由(1)式转换而来的最优化问题可以表述为$\mathop {{\rm{min}}}\limits_{{\mathit{\boldsymbol{E}}_{\rm{i}}}^\prime } \left\{ {{{\left\| {{\mathit{\boldsymbol{E}}_{\rm{i}}}^\prime } \right\|}_2}} \right\}$满足Γ的最小值:
$ \begin{array}{l} \mathit{\Gamma = }\sum\limits_{\alpha = {\lambda _1}, \mathit{t }= 10}^{{\lambda _2}} {\left[{\left\| {\sum\limits_{\lambda = {\lambda _1}, t = 1}^{{\lambda _2}} {{E_{\rm{i}}}^\prime \left( \lambda \right)T\left( {\lambda, \alpha } \right)t-} } \right.} \right.} \\ \;\;\;\;\left. {{E_{\rm{o}}}\left( \alpha \right)\left\| {^2} \right. + \beta {{\left\| {{E_{\rm{i}}}^\prime \left( \lambda \right)-{E_{\rm{o}}}^\prime \left( \alpha \right)} \right\|}^2}} \right] \end{array} $
(4) 式中, Ei′(λ)表示入射光输入光谱的近似解,t代表的是近似光谱解的离散化步长,通过调整离散化步长增加Ei′(λ)的向量长度,使Ei′(λ)相较Eo(α)更为细化,使得求解LCTF光谱测量过程的积分方程问题转化为了一个可以稳定求解的最优化求解问题。求解过程如下, 令,Ei′=[λ1 … λn]T,Eo=[α1 … αm]T, (4)式转换为:
$ \mathit{\Gamma } = {\left\| {\mathit{\boldsymbol{T}}{\mathit{\boldsymbol{E}}_{\rm{i}}}^\prime-{\mathit{\boldsymbol{E}}_{\rm{o}}}} \right\|^2} + \beta {\left\| {{\mathit{\boldsymbol{E}}_{\rm{o}}}^\prime-{\mathit{\boldsymbol{E}}_{\rm{i}}}^\prime } \right\|^2} $
(5) 令:
$ \frac{{\partial \mathit{\Gamma }}}{{\partial {\mathit{\boldsymbol{E}}_{\rm{i}}}^\prime }} = {\mathit{\boldsymbol{T}}^{\rm{T}}}\left( {\mathit{\boldsymbol{T}}{\mathit{\boldsymbol{E}}_{\rm{i}}}^\prime-{\mathit{\boldsymbol{E}}_{\rm{o}}}} \right) + \beta \left( {{\mathit{\boldsymbol{E}}_{\rm{i}}}^\prime-{\mathit{\boldsymbol{E}}_{\rm{o}}}^\prime } \right) = 0 $
则输入光谱最优化近似解为:
$ {\mathit{\boldsymbol{E}}_{\rm{i}}}^\prime = {\left( {{\mathit{\boldsymbol{T}}^{\rm{T}}}\mathit{\boldsymbol{T}}{\rm{ + }}\beta \mathit{\boldsymbol{I}}} \right)^{-1}}\left( {{\mathit{\boldsymbol{T}}^{\rm{T}}}{\mathit{\boldsymbol{E}}_{\rm{o}}} + \beta {\mathit{\boldsymbol{E}}_{\rm{o}}}^\prime } \right) $
(6) 式中,I为单位矩阵。由于正则化项的引入,(6)式中的(TTT+βI)矩阵可逆,则该式能够稳定求解。所得近似解Ei′与输入光谱Ei的近似程度由正则化参量β决定;(6)式中的T是透过率函数矩阵与参量t的乘积,t决定了近似光谱Ei′的向量长度,反映了Ei′的细化程度,β和t可由数值仿真中最优化近似解与输入光谱的逼近程度决定。
-
为了验证上述LCTF成像光谱仪光谱检测模型和采用先验约束条件的正则化求解方法,在实测LCTF光谱透过率曲线基础上,以模拟光谱输入的数值仿真验证正则化求解方法的可行性,确定正则化参量β和离散化步长t,并采用氘卤连续谱光源和汞氩光源进行实验验证。
-
选用CRi公司的VariSpec近红外型LCTF作为成像光谱仪分光器件进行数值仿真和实验验证,LCTF的波长调谐分辨率为10nm,其实测光谱透过率函数T(λ, α)如图 1所示。以MATLAB仿真生成入射光光谱数据模拟光谱检测模型输入,以(1)式生成仿真光谱测量值,后利用T(λ, α)求解入射光光谱近似解,并与仿真入射光光谱对比以确定采用该LCTF时的正则化参量β和离散化步长t。数值仿真的输入光谱数据采用分离双峰、重叠双峰、重叠三峰3种形式,由于一般光谱曲线符合高斯函数形式,因此,仿真输入光谱曲线应为高斯峰f(λ)、基线漂移g(λ)、随机白噪声k(λ)3种成分合成的结果[17],即:
$ {E_{\rm{i}}}\left( \lambda \right) = f\left( \lambda \right) + g\left( \lambda \right) + k\left( \lambda \right) $
(7) $ \left\{ \begin{array}{l} f\left( \lambda \right) = {a_0}{\rm{exp}}\left[{-\frac{{{{\left( {\lambda-{b_0}} \right)}^2}}}{{2{c_0}^2}}} \right] + {a_1} \times \\ \;\;\;\;\;\;\;\;\;\;\;\;\;{\rm{exp}}\left[{-\frac{{{{\left( {\lambda-{b_1}} \right)}^2}}}{{2{c_1}^2}}} \right] + {a_2}{\rm{exp}}\left[{-\frac{{{{\left( {\lambda-{b_2}} \right)}^2}}}{{2{c_2}^2}}} \right]\\ g\left( \lambda \right) = {g_0} + {g_1}\lambda + {\left( {{g_2}\lambda } \right)^2}\\ k\left( \lambda \right) = {k_0}{N_{\rm{w}}}\left( \lambda \right) \end{array} \right. $
(8) 式中,不同下标的a, b, c, g, k均为生成光谱的参量常数; Nw(λ)为白噪声函数。仿真输入光谱与仿真测量光谱分别为图 2a~图 2c中的实线与虚线。根据LCTF成像光谱仪光谱检测模型和采用先验约束条件的正则化求解方法过程,考虑到入射光输入光谱与正则化方法的近似解光谱的一致性这一先验约束条件,取t=5,β=0.001,其它取值验证过程冗长,这里不再赘述。对于3种形式输入光谱,以(6)式求得的光谱近似解如图 2d~图 2f所示,图中实线为入射光光谱,虚线为本文中方法所得光谱近似解。图 2a和图 2d中双峰间隔为100nm,模拟大间隔分离双峰输入;图 2b、图 2e与图 2c、图 2f为峰间隔10nm的重叠双峰和重叠三峰输入,模拟具有一定精细结构的待测输入光谱。对比图 2中3种形式输入下的输入光谱近似解曲线可见,先验约束条件正则化方法求解所得光谱与输入光谱的接近程度得到了显著改善。光谱边缘处的震荡现象并非输入光谱中噪声导致,是由于有限数据数值迭代求解的普遍现象。作者以光谱强度差的标准差衡量仿真结果中光谱近似解与模拟输入光谱的吻合程度,表 1中给出了光谱强度差的标准差对比,显见采用先验约束条件正则化方法的近似解更为接近真实光谱。
Table 1. Comparison of standard deviation of spectral intensity difference after numerical simulation
standard deviation of difference between Ei(λ) and Eo(α) standard deviation of difference between Ei(λ) and Ei′(λ) bimodal (non-overlapping) spectrum 6.28×103 1.30×103 bimodal (overlapping) spectrum 5.02×103 1.59×103 tri-peak spectrum 5.20×103 2.14×103 -
在数值仿真基础上,以气卤混合(deuterium-halo-gen combined, DHc)光源作为连续光谱输入,以DHc光源与汞氩校准用(calibration, CAL)光源耦合后的混合光源作为分离光谱输入,实验验证了LCTF成像光谱仪光谱测量模型与先验约束条件正则化方法的有效性, 所得实验结果分别如图 3和图 4所示,两次实验光谱强度差的标准差对比见表 2。采用AvaSpec光纤光谱仪测得的输入光谱曲线如图 3、图 4中实线所示。经LCTF成像光谱仪后的光谱测量值如图 3a、图 4a中虚线所示。利用光谱测量值和对应调谐波长的T(λ, α),以先验约束条件正则化方法进行求解,光谱近似解曲线如图 3b、图 4b中虚线所示。
Table 2. Comparison of standard deviation of spectral intensity difference after experiment
standard deviation of difference between Ei(λ) and Eo(α) standard deviation of difference between Ei(λ) and Ei′(λ) the continuous spectrum 6.68×103 1.04×103 the coupled spectrum 5.91×103 2.32×103 实验结果分析可见,通过先验约束条件正则化方法解得的Ei′(λ)与直接光谱测量值Eo(α)相比更吻合输入光谱Ei(λ)。特别是采用DHc光源作为输入时,对于利用LCTF成像光谱仪直接进行光谱检测中存在的被误识别的光谱峰,本文中方法的光谱近似解结果与原始输入光谱更为吻合;对于DHc光源与CAL光源耦合后的光源光谱,多个因相隔较近而未测量出的光谱峰在经算法改进后变为可分离状态,说明该算法对于改善LCTF成像光谱仪光谱检测能力有较好的效果。
一种改善成像光谱仪光谱检测能力的新方法
A novel method to improve spectral detection capability of imaging spectrometers
-
摘要: 为了改善成像光谱仪的检测能力,在不改变硬件结构的情况下,采用光谱细化最优化的新方法,利用液晶可调谐滤光片式成像光谱仪,取得了入射光近似光谱数据,进行了理论分析和实验验证。结果表明,在3组数值仿真数据中,相较于成像光谱仪测量光谱,该方法得到的近似光谱与入射光真实光谱的光谱强度差的标准差分别减小了79.3%,68.3%和58.8%;在两组实验数据中,标准差分别减小了84.4%和60.7%;求解得到的近似光谱与入射光真实光谱的近似程度得到了显著提高,较好地分离了相隔较近的光谱峰。这一研究改善了成像光谱仪的光谱检测能力。Abstract: In order to improve spectral resolution of an imaging spectrometer without changing its hardware structure, a novel method of spectral refinement was adopted. An imaging spectrometer with liquid crystal tunable filter was used to obtain the approximate spectral data of the incident light for theoretical analysis and experimental verification. In three sets of numerical simulation data, the standard deviations of the spectral intensity difference between the approximate spectra and the true spectral were reduced by 79.3%, 68.3% and 58.8%, compared with the spectra measured with an imaging spectrometer. In two sets of experiment data, the standard deviations were decreased by 84.4% and 60.7%. The results show that the approximation degree between the approximate spectrum and the real spectrum of the incident light is improved and the spectral peaks are separated very well. It is helpful to improve the spectral detection capability of imaging spectrometers.
-
Key words:
- spectroscopy /
- spectral resolution /
- optimization method /
- imaging spectrometer
-
Table 1. Comparison of standard deviation of spectral intensity difference after numerical simulation
standard deviation of difference between Ei(λ) and Eo(α) standard deviation of difference between Ei(λ) and Ei′(λ) bimodal (non-overlapping) spectrum 6.28×103 1.30×103 bimodal (overlapping) spectrum 5.02×103 1.59×103 tri-peak spectrum 5.20×103 2.14×103 Table 2. Comparison of standard deviation of spectral intensity difference after experiment
standard deviation of difference between Ei(λ) and Eo(α) standard deviation of difference between Ei(λ) and Ei′(λ) the continuous spectrum 6.68×103 1.04×103 the coupled spectrum 5.91×103 2.32×103 -
[1] ZHENG C Y, GUO Zh H, JIN L. Measurement of total viable count on chilled mutton surface based on hyper spectral imaging technique[J]. Laser Technology, 2015, 39(2):284-288(in Chinese). [2] DU P Sh, GUO J, DONG Q M. Application of liquid crystal tunable filter in multispectral imager[J]. Infrared, 2007, 28(11):4-8(in Chinese). [3] SU L J. Study on the beam splitting technology of the imaging spectroscopy[D]. Xi'an: Xi'an Institute of Optics and Precision Mechanics of the Chinese Academy of Sciences, 2006: 37-56(in Chinese). [4] KOPP G, DERKS M, GRAHAM A. Liquid crystal tunable birefringent filters[J]. Proceedings of the SPIE, 1996, 2830:345-350. doi: 10.1117/12.256127 [5] MILLER A P J, HOYT C C. Multispectral imaging with a liquid crystal tunable filter[J]. Proceedings of the SPIE, 1995, 2345:354-365. doi: 10.1117/12.198889 [6] DONG Y. Study on resolution enhancement of Fourier transform spectroscopy[D]. Xi'an: Xi'an Institute of Optics and Precision Mechanics of the Chinese Academy of Sciences, 2001: 17-38(in Chinese). [7] DONG Y, XIANG L B, ZHAO B Ch. Theoretical study on Fourier self-deconvolution (FSD) of fourier transform spectrum[J]. Acta Photonica Sinica, 2002, 31(7):841-846(in Chinese). [8] WEI H L, WU Ch J, MA Zh J, et al. A new method for improving the measurement spectral resolution of atmospheric absorption spectra[J]. Acta Optica Sinica, 2002, 22(2):165-169(in Chinese). [9] KAWATA S, MINAMI K, MINAMI S. Superresolution of Fourier transform spectroscopy data by the maximum entropy method.[J]. Applied Optics, 1983, 22(22):3593-3601. doi: 10.1364/AO.22.003593 [10] XIANG L B, ZHAO B Ch. A new method for supperresolution of Fourier transform spectra[J]. Acta Optica Sinica, 1995, 15(11):1529-1533(in Chinese). [11] KAUPPINEN J K, MOFFATT D J, HOLLBERG M R, et al. A new line-narrowing procedure based on Fourier self-deconvolution, maximum entropy, and linear prediction[J]. Applied Spectroscopy, 1991, 45(3):411-416. doi: 10.1366/0003702914337155 [12] SU G, XU R, WANG J Y. Simulation of resolution enhancing algorithms for spectra measured by AOTF[J]. Infrared, 2016, 37(10):23-29(in Chinese). [13] ZOU M Y. Deconvolution and signal recovery[M]. Beijing:National Defence Industry Press, 2001:32-42(in Chinese) [14] WU H X. The study of numerical algorithms for Fredholm integral equations of the first kind[D]. Xi'an: Xi'an University of Technology, 2008: 21-32(in Chinese) [15] LIANG F. Research on numerical solution methods for some inverse problems[D]. Shantou: Shantou University, 2010: 56-63(in Chinese) [16] NOCEDAL J, WRIGHT S J. Numerical optimization[M]. Berlin, Germany:Springer, 2006:8-13. [17] FENG F, WANG F B, XIE F, et al. Implementation of spectral overlap resolution based on wavelet transforms and Gaussian fitting[J]. Acta Photonica Sinica, 2015, 44(6):630001(in Chinese). doi: 10.3788/gzxb