Blind spectral deconvolution algorithm for Raman
spectrum with Poisson noise
Hai Liu, Zhaoli Zhang, Jianwen Sun, and Sanya Liu*
National Engineering Research Center for E-Learning, Central China Normal University, Wuhan 430079, China
*Corresponding author: lsy5918@gmail.com
Received August 8, 2014; revised September 11, 2014; accepted September 13, 2014;
posted September 22, 2014 (Doc. ID 220670); published October 17, 2014
A blind deconvolution algorithm with modified Tikhonov regularization is introduced. To improve the spectral
resolution, spectral structure information is incorporated into regularization by using the adaptive term to dis-
tinguish the spectral structure from other regions. The proposed algorithm can effectively suppress Poisson noise
as well as preserve the spectral structure and detailed information. Moreover, it becomes more robust with the
change of the regularization parameter. Comparative results on simulated and real degraded Raman spectra are
reported. The recovered Raman spectra can easily extract the spectral features and interpret the unknown chemi-
cal mixture. © 2014 Chinese Laser Press
OCIS codes: (300.6450) Spectroscopy, Raman; (100.3190) Inverse problems; (300.6320) Spectroscopy,
high-resolution.
http://dx.doi.org/10.1364/PRJ.2.000168
Raman spectra often suffer from common problems of bands
overlapping and Poisson noise [
1,2]. Raman spectra measured
by a spectrophotometer can be mathematically modeled as
gvPoissonsv ⊗ hv; (1)
where gv and sv are the measured Raman spectrum and
actual spectrum, and hv stands for the instrument function,
which mainly collects the instrumental broadening. Poisson·
denotes the Poisson noise. The goal of blind spectral decon-
volution is to seek the best estimates of sv and hv based on
measured spectrum gv and prior information about the ac-
tual spectrum scene. Over the years, many deconvolution
methods have been developed, such as high-order statistic [
3],
Fourier self-deconvolution (FSD) [
4], the Wiener filtering
method [
5], maximum entropy deconvolution [6], and the
semi-blind deconvolution method [
7,8]. In [9], Katrasnik et al.
extended the Richardson– Lucy (RL) method to the spectros-
copy field. Nevertheless, owing to the ill-posed nature of in-
verse problems, this algorithm often lacks stability and
uniqueness. The amount of noise will increase with the iter-
ation process. And the algorithm must be stopped before
the noise rises above a certain level. Tikhonov regularization
(TR) was initially introduced as a regularizer for spectral
processing in [
10,11]. Since then it has been used extensively
and with great success for inverse problems because of its
ability to suppress noise.
The maximum a posteriori (MAP) technique is a commonly
used approach to estimate the actual spectrum sv given the
measured spectrum gv. This technique maximizes the
conditional probability of an actual spectrum given a certain
measured spectrum. Based on Bayes’s rule, MAP can be con-
verted to one likelihood probability multiplied by two priori
probabilities. In this paper, the MAP framework is employed
to construct the spectral deconvolution model. Considering
the cases in which the data are contaminated by Poisson
noise, the intensity of each pixel gv in the measured
spectrum is a random variable that obeys an independent
Poisson distribution. Hence the likelihood probability can
be written as
pgjs; h
Y
L
v
sv ⊗ hv
gv
gv!
expf−sv ⊗ hvg; (2)
where L denotes the spectral length.
To estimate s and h, iterative deconvolution algorithms can
be used. One option is to use the RL method [
9], which min-
imizes the following functional to maximize Eq. (
2):
E
1
s
X
v
−gv logsv ⊗ hv sv ⊗ hv: (3)
The RL algorithm does not converge to the solution be-
cause the noise is amplified after a small number of iterations.
In order to get better convergence, the RL method with Tikho-
nov regularization [
11] (TR-RL) was proposed, and the cost
functional to be minimized is then
E
2
s
X
v
−gv logsv ⊗ hv sv ⊗ hv αj∇sj
2
;
(4)
where ∇s s
i
− s
i1
∕2. Symbol α is the regularization
parameter, which plays a very important role, controlling
the constraint strength. If α is too small, the noise will not
be well suppressed; conversely, if it is too large, the spectral
structure will be removed. This means that the regularization
parameter α needs to be adaptively adjusted according to
spectral structure information. For better structure informa-
tion preservation, in the TR term, we added a nonnegative
168 Photon. Res. / Vol. 2, No. 6 / December 2014 Liu et al.
2327-9125/14/060168-04 © 2014 Chinese Laser Press