Sådhanå (2019) 44:114 Ó Indian Academy of Sciences

https://doi.org/10.1007/s12046-019-1100-6 Sadhana(0123456789().,-volV)FT3
](0123456789().,-volV)

Application of image enhancement and mixture of Gaussian approach

in combustion research
LITU ROUT1, RAJESH SADANANDAN2 and DEEPAK MISHRA1,*
1
Department of Avionics, Indian Institute of Space Science and Technology, Thiruvananthapuram,
Kerala 695 547, India
2
Department of Aerospace, Indian Institute of Space Science and Technology, Thiruvananthapuram,
Kerala 695 547, India
e-mail: [email protected]; [email protected]; [email protected]

MS received 10 September 2018; revised 19 February 2019; accepted 21 February 2019; published online 13 April 2019

Abstract. Chemiluminescence is one of the most commonly used optical diagnostic techniques in combustion
research where a line-of-sight projected information is generated from spatial fields. The exactness and
uniqueness of reconstruction along with ease of implementation gives Abel inversion an edge over the other
existing single-view reconstruction techniques for efficient estimation of spatial field from line-of-sight pro-
jections. Though there exist many such algorithms, the primary focus of these has been to ensure tractable in-
version through a systematic regularization by imposing a smoothness constraint on discrete data points. But
these techniques do not have the provision to process the input image prior to deconvolution in order to prevent
accumulation of noise infiltrated during data acquisition. Another major limitation of these algorithms is to adopt
the changes in characteristics of the input data points while maintaining optimal storage and time complexity. To
address these issues, we have proposed a new image processing technique using standard Abel inversion for the
application in combustion research. It provides a suitable model to ensure regularized inversion by imposing a
smoothness constraint on acquired raw data. The new algorithm has been implemented to yield the physically
significant chemiluminescence emission from hydroxyl radicals in flames from line-of-sight integrated images.
The effectiveness of this algorithm is highlighted using exemplary OH chemiluminescence images captured
from a standard swirl stabilized research burner.

Keywords. Abel deconvolution; image enhancement; mixture of Gaussian; swirl combustion; optical
diagnostics; chemiluminescence.

1. Introduction (hydroxyl radical) in an electronically excited state (termed

OH*). In combustion research OH* is, therefore, one of the
In this study, we develop an image processing algorithm most commonly used radicals to identify the spatial
using Abel inversion for application in combustion dimensions of the flame and the heat release behavior. OH*
research, especially in 2D analysis of chemiluminescence has a very short lifetime i.e. on the order of a few
images from flames. The chemical reactions in flames nanoseconds. From investigations in flames with gaseous
generate energy and the electronic excitation of molecules fuels, there are indications that the intensity of the OH*
takes place. The excited molecules undergo deactivation to chemiluminescence is related to the reaction rate or heat
lower energy levels by the emission of photon, and the release rate. So by capturing the OH* chemiluminescence
process is termed as chemiluminescence. The primary emissions from the flame (using an intensified CCD cam-
source of chemiluminescence in hydrocarbon flames con- era), information can be obtained about the instantaneous
stitute the radiative emissions from electronically excited size and shape of heat release zone. Since the measurement
species that are formed chemically, such as CH*, OH*, C2 * technique is line of sight, with the OH* emissions collected
and CO2 * [1]. Under fuel lean conditions chemilumines- along the line of sight, the spatial resolution is compro-
cence can, therefore, be taken as an indicator of the reacting mised. However, in case of flames from axisymmetric
conditions in the flame since these species are mainly burners, Abel deconvolution can be used to reconstruct the
produced in the reaction zone [2] [3]. Of these, the emis- underlying 2D scalar fields or radial distributions. The 2D
sions in the UV range stem predominantly from OH information so obtained are also vital for the modeling,
validation and optimization of numerical simulation codes.
*For correspondence

1
114 Page 2 of 12 Sådhanå (2019) 44:114

Various Abel deconvolution algorithms have been

developed over the years with diverse applications such as
heat conduction, infrared tomography and many more
[4–9]. But most of these algorithms focus primarily on
developing a strategy to enhance degree of smoothness with
minimal consideration given to pre-processing the acquired
raw results. In this paper, we propose to implement Abel
deconvolution under noisy environment by taking utmost
care while acquiring the results (raw images), modeling the
noise and implementing a Mixture Of Gaussian (MOG)
model to ensure higher degree of smoothness. An extensive
investigation has been conducted in deciding the suit-
Figure 1. Projection of a transparent, emitting scalar field with
able curve that is preferred over the others while fitting the circular symmetry as per the field of view of an observer along the
discrete data points. Based on this, a new way of imple- blue arrows.
menting Abel deconvolution, which is robust to noisy and
discrete experimental data, has been proposed. The fol- Z 1
lowing sections give a brief description about the mathe- 1 dF dy
f ðrÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð2Þ
matical aspects and geometrical interpretation (section 2), p r dy y2  r 2
related works (section 3), followed by the implementation
of the algorithm (section 4) in a 2D-OH* chemilumines- For the ease of understanding, we have demonstrated the
cence image acquired from a standard swirl stabilized physical interpretations of Abel transformation with math-
burner. ematical equations in light of circular symmetry. However,
the general equations (1) and (2) of Abel transformation
with certain modifications can be extended to other
2. A brief review of Abel transform instances, such as spherical and cylindrical symmetry,
without the loss of generality.
In this section 2, a very brief discussion about the forward Of particular interest, the projection of a cylindrical
(2.1) Abel Transform and Abel Deconvolution (2.2) is symmetric function f ðq; zÞ onto a plane parallel to Z axis
given. More details on Abel Transform can be found at can be obtained by taking Abel transform of f ðq; zÞ as
[10]. shown in the following equation.
Z 1 Z 1
f ðq; zÞqdq
Fðy; zÞ ¼ f ðq; zÞdx ¼ 2 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð3Þ
2.1 Forward Abel transform 1 y q 2  y2

The forward Abel transform of a function f(r) is defined by where q2 ¼ x2 þ y2 is the cylindrical radius and F(y,z) is
the following equation as the projection on YZ plane. The inverse Abel transform of
Z 1 such projection can be obtained by the following equation.
f ðrÞrdðrÞ Z
FðyÞ ¼ 2 pffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; ð1Þ 1 1 dF dy
y r 2  y2 f ðq; zÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð4Þ
p q dy y2  q2
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
where r ¼ ðx2 þ y2 Þ represents the radial distance of the
points on the line of sight of integration and y represents the Since our application provides the projected image F(y, z),
distance of the observer from X axis as shown in figure 1 we aim to employ inverse Abel transform, as given in
[10]. Eq. (4), to obtain the cylindrical symmetric flame f ðq; zÞ.

2.2 Abel deconvolution 3. Related works

The Abel deconvolution is usually used to obtain the Abel transform and its inverse integral equation finds
symmetric function from it’s projection i.e. a scan or a numerous applications in various branches of science
photograph [10]. Assuming that f(r), f 0 ðrÞ (‘0’ represents especially in plasma diagnostics and X-ray radiography,
first derivative) drops to zeros more quickly than 1r , the where it is used to yield information about the density or
inverse Abel transform is given by Eq. (2). The emission temperature distribution [4]. In combustion studies, the
function f(r) can be obtained from its projection F(y) with deconvolution algorithms are used for extracting the
the help of this inverse Abel transform. tomography of spectroscopic emissions from axisymmetric
objects, for example flames in axisymmetric burners. One
Sådhanå (2019) 44:114 Page 3 of 12 114

of the most commonly used algorithms in this case is the technique [13] that uses less number of transcendental
onion peeling method [5]. The ill posed sets of equations operations while using the spline based smoothing proce-
generated by the onion-peeling scheme are transformed into dure. The FAI algorithm computes coefficients which are
well posed sets with the help of Tikhonov regularization independent of the data points and hence, these can be
scheme [6]. A comparative study of Abel three-point computed once, and used for many dataset [13]. Though
deconvolution, onion-peeling and Tikhonov regularization this independent approach is considered to be significantly
is given [8] for a known spatial field from a line-of-sight impressive, there is a need to change the coefficients
integrated projection data. slightly in order to adapt to the changes in data points.
Other approaches to solving this singular integral equa- Therefore, a smoothing scheme which can adapt to the
tion obtained from axisymmetric flame and parallel scan- changes with less number of arithmetic operations, espe-
ning measurements are generalized quadrature method, cially transcendental operations, and a high degree of
spline smoothing and Tikhonov regularization along with smoothness is desirable.
infrared tomography [9]. Though the smoothing procedure eliminates the abrupt
For an axisymmetric object the measured intensity, in transition in intensities, we have observed that noise due to
gray level units, is related to the object’s radial optical various sources still persists in the captured images.
density profile through Abel’s integral equation. So, in Therefore, in this paper, we have experimented various
order to extract the physically significant radial density standard image enhancement schemes to further alleviate
profile the measured raw data (or image) needs to be Abel the adverse effects of image acquisition in real world
inverted, which can then be used for non-destructive eval- applications.
uation of the object. In other words, the forward Abel
transform is used for the projection of an axially symmetric
function onto a plane, whereas inverse Abel transform is 4. Proposed methodology
used for extracting the axially symmetric function from its
projection i.e. a scan or a photograph. Computational issues In the proposed algorithm we assume that the line-of-sight
arise when the raw data contains noise because the efficacy projections of the spatial field are equally spaced over the
of inversion algorithm greatly depends on the suppression whole field and each of these projections are obtained with
of noise. Since the expression used for inversion is an the help of a narrow beam. It has been proved that sampling
integral equation, it is certainly affected by the accumula- has a great influence on the quantity of noise which gets
tion of small drifts due to noise. Also, the derivative used in added to the spatial field during the process of deconvolu-
Abel integral equation tends to amplify the errors. Over the tion [5].
years, many Abel inversion techniques have been proposed Therefore one of the most important steps in deconvo-
[9, 11–13]. One of the successful algorithms was proposed lution is to eliminate oversampling and smooth the pro-
by Deutsch and Beniaminy [11], that uses derivative free jected discrete data prior to deconvolution. While the
implementation of Abel inversion. One of the key features oversampling issue is resolved by taking utmost care during
incorporated in this derivative free inversion is based on the image acquisition, the Mixture of Gaussian (MOG) model,
least square smoothing of discrete experimental data. as proposed in this paper, has been used to approximate the
Though the algorithm produces high quality images after discrete noisy experimental data.
inversion, its time complexity is Oðn1:5 Þ and also it requires In the following sections, we elaborate on the proposed
high storage space. In spline based representation of the methodology with detailed experimentation and analysis.
input data, the implementation of spline fit avoids ampli- The image acquiring process and image enhancement
fication of noisy data. But the quality of this fitting proce- (section 4.1) techniques have been discussed prior to esti-
dure deteriorates at the end points of the sampling interval. mation of smoothing function (section 4.2) and Abel
Therefore, a plausible solution to address this issue would inversion (section 4.3). Thereafter, the post-processing task
be to reflect part of the data relative to y-axis before fitting such as spatial distribution of deconvoluted data points has
the data using a spline fit [13]. been discussed (section 4.4), followed by the essential
The analytic spline Abel inversion (ASAI) is another algorithms (section 4.5) namely coefficient extraction and
successful implementation of Abel deconvolution that has Abel deconvolution.
been proved to be extremely effective under noisy and
noise free environment. Despite the effectiveness of ASAI
algorithm, it requires Oðn2 Þ number of arithmetic opera- 4.1 Image acquisition and enhancement
tions that includes four transcendental operations per each
Space and time complexities are two of the important
set of data and also, its coefficients are uniquely associated
factors in most image processing techniques. Storage and
with the data points. Hence these coefficients can’t be
retrieval without loss of valuable information is nonetheless
stored in advance. In order to address this issue Gueron and
challenging task in various scenarios. In order to deal with
Deutsch have proposed a fast Abel inversion (FAI)
such problems, we have taken utmost care during data
114 Page 4 of 12 Sådhanå (2019) 44:114

acquisition. The OH* emission images from a standard

swirl burner, as used in this work, is captured using an
intensified CCD camera. The raw images are recorded in
.im7 format (type: graphic) which is a proprietary file for-
mat from LaVision GmbH. For the post processing of the
images the .im7 images are exported into .jpeg format.
After storing the data, the next vital step is to remove noise
using image enhancement techniques as a regularization
strategy to make Abel inversion more tractable.
In real world applications, it is almost intractable to
acquire images without noise. In image processing per-
spective, it can be justified by central limit theorem that
these unwanted signals, often referred to as noise, tend to
follow Gaussian distribution when several random numbers
are added [14]. Thus, the unavoidable noise, since it affects
all the pixels of the captured image, is reasonable to be
modelled as Additive White Gaussian Noise (AWGN).
Hence, an edge preserved smoothing mechanism will be a
great option to eliminate the AWGN without corrupting the
data. However, it is extremely difficult to compute the
latent variables of such distribution and eliminate the noise
completely. Therefore, one of the plausible solutions to
tackle such an issue is to employ various image enhance-
ment techniques on the raw pixels and analyze the impact
of such methods qualitatively. Since neither the noise-free
image nor the actual noise distribution is at hand, we have
preferred qualitative analysis over quantitative measures
such as SNR. In order to validate our assumption of
AWGN, we have experimented with variance and loga-
rithmic transform based enhancement techniques. After
exploring these methods, we have observed that edge
Figure 2. Qualitative analysis of various image enhancement
adaptive smoothing using Gaussian kernels turns out to be
techniques. (a) Noisy discrete experimental data. (b) Edge pre-
more effective when compared with other image enhance-
served Gaussian smoothing. (c) Unsharp masking using adaptive
ment algorithms such as unsharp masking and retinex. Gaussian window. (d) Retinex algorithm with 10 iteration.
Therefore, this comparison validates the additive property (e) Unsharp masking using global window. (f) Image sharpening.
of the noise in our raw data up to some extent. One can find
a detailed description with notable comparisons among
these image enhancement techniques in [14]. The qualita- In local unsharp masking a local window whereas in
tive analysis of the aforementioned techniques is shown in global unsharp masking a global window is considered
figure 2. As figure 2(b) depicts, the edge preserved while computing a low pass filtered image. The low pass
smoothing using Gaussian kernel outperforms the rest filtered version of the image is then subtracted from itself
mechanisms, and hence, it has opted for further processing. and multiplied by an amplification factor. The difference is
In the following few paragraphs, we briefly discuss some amplified only when it exceeds a certain threshold which is
key enhancement techniques that we have experimented in used to suppress small high frequency fluctuations due to
this study. noise. Thus, the amplification factor is chosen according to
In edge preserved Gaussian smoothing, a small window the local variance. This transformation can be achieved
Waround a pixel is taken. The variance of all the possible using the following equation [14].
windows w of the same dimension in the local neighbour-
Fðy; zÞ ¼ A½Fðy; zÞ  mðy; zÞ þ mðy; zÞ ð5Þ
hood that consists of at least one pixel from the window
W is computed. Then the central pixel of W is replaced with kM
where A ¼ rðy;zÞ , k is scalar, M is the average grey value of
a smoothed value of the window w that has minimum
the whole image, rðy; zÞ variance of the local window,
variance. For a 5x5 window W, there are 25 possible local
m(y, z) is the low pass version of F(y, z) and F(y, z) is the
windows out of which 1 window w has a minimum variance
line-of-sight integrated image with y and z representing
which is used for computing the value of the center pixel
columns and rows respectively.
[14].
Sådhanå (2019) 44:114 Page 5 of 12 114

The results are compared with single scale retinex or Fourier series approximation is given by the following
logarithmic transform which is usually used to isolate equation.
unwanted signals due to variable illumination that may
manifest during image acquisition. It has two basic X
8
Fðy; zÞ ¼ a0 þ ðan cosðnxÞ þ bn sinðnxÞÞ; 8z ð8Þ
ingredients:
n¼1
1. grey value normalization by dividing local mean value;
where x is the fundamental frequency, an and bn are the
2. conversion into a logarithmic scale which will enhance
coefficients of cosines and sines of nth harmonic respec-
dark pixels more and bright pixels less.
tively, a0 is the offset value. The polynomial model is given
This transformation can be achieved by using the following by the following equation.
equation.
X
8

Fðy; zÞ ¼ lnðFðy; zÞ þ 1Þ  lnðFðy; zÞÞ ð6Þ Fðy; zÞ ¼ ðan yn Þ; 8z ð9Þ

n¼1

where Fðy; zÞ is the local mean value of F(y, z). The mean where an represents the coefficients of the polynomial
value is obtained using a large Gaussian filter. A deep interpolation. Here, the coefficients used in these models
insight into all these image enhancement techniques can be are within 95% confidence bounds [15–17]. A line-of-sight
found in [14]. Here, we essentially employ AWGN filter to integrated image is shown in figure 3. Polynomial, Gaus-
enhance the required details of the captured image due to its sian, and Fourier approximation for a fixed row(here, 368
close approximation to eliminate noise to some degree. from top) of the line-of-sight integrated image is illustrated
in figure 4. As described above, a proper goodness of fit
decides the accuracy of the spatial field. Therefore, the
4.2 Function estimation Gaussian mixture model has been selected for further pro-
After removing the unwanted signals from the acquired cessing, as it is found to outperform the rest with minimal
image, the discrete data points are approximated by a error, high goodness of fit and 95% confidence bound in the
continuous function. Here, we discuss various approxi- region of interest i.e. from column 100 to 200. For better
mating functions and their corresponding goodness of fit understanding, we have visualized the individual Gaussian
(GOF). In order to select the right fit for the discrete distributions along with MOG model in figure 5.
experimental data, we have experimented various functions A comparative study of GOF of top performing function
including polynomial, Gaussian, Fourier, smoothing spline, approximators has been detailed in Table 1. We have used
the sum of sine, exponential, power, rational, and Weibull. Regression Squared(R-squared), Adjusted R-squar-
Based on their performance and required criteria, the fea- ed(AdjR-squared) and Root Mean Squared Error(RMSE) as
sible functions have been narrowed down to the top three the essential statistical measures for assessing the quality of
approximators which are mixture of Gaussian (MOG), fit. The details of these metric can found at [18].
Fourier series expansion, and higher order polynomial. The graphical interpretation (figure 4), as well as quan-
Also, we have observed that there is an insignificant change titative measure (Table 1) clearly indicates that the mixture
in the accuracy of approximation as the order becomes of Gaussian model has exceedingly high goodness of fit in
higher and higher. However, the computational cost grows the region of interest. Outside this region there are dark
significantly as the order increases. So, the order of the
approximating function (here, 8) has been decided such that
it provides an optimal residual sum of errors with minimal
computational complexity.
In our notation, F(y, z) represents the intensity at (y, z)
location of input image in spatial domain. We aim to
approximate each row (z) of F(y, z) with appropriate con-
tinuous functions and perform Abel inversion to obtain ele-
ments of f ðq; zÞ in that particular row. The proposed mixture
of Gaussian distributions, which is used to approximate
discrete F(y, z), is given by the following equation.
8 
X 
ybn 2
Fðy; zÞ ¼ an expð cn Þ ; 8z ð7Þ
n¼1

where an is the maximum amplitude, bn is the mean posi-

pffiffiffi Figure 3. An example of a line-of-sight integrated image.
tion and cn is 2 times standard deviation(r). Similarly, the
114 Page 6 of 12 Sådhanå (2019) 44:114

sufficiently large number of Gaussian basis functions are

considered. This hypothesis is supported by the central
limit theorem and is well investigated in many signal pro-
cessing applications including image enhancement and
interpolation [14]. On the other hand, the polynomial and
Fourier series approximation does not fit the data as effi-
ciently as MOG with limited number of free parameters.
Therefore, we have opted the MOG model for estimating
the continuous approximation of discrete data. The
parameters of MOG model i.e. an ; bn and cn are computed
by optimizing the squared sum of errors. In this process, the
parameters are initialized with random values and then
successively iterated until convergence within a prescribed
limit. The partial derivatives used to compute the incre-
ments are given by
oF yb1 2
¼ eð c1 Þ ;
oa1
oF ð y  b1 Þ ðyb 2
c1 Þ and
1
¼ 2a1 e
Figure 4. Comparison among Polynomial, Gaussian and Fourier ob1 c21
approximation of discrete data points. It is observed that the oF ð y  b1 Þ2 ðyb 2
e c1 Þ :
1
Gaussian mixture model approximates the data significantly better ¼ 2a1 3
than Polynomial and Fourier approximation. oc1 c1

This technique could obviously be generalized to multiple

Gaussian basis functions, although the convergence prop-
erty gets worse as the number of parameters increases.
However, it is necessary, in some scenarios, to increase the
number of Gaussian basis functions for efficient approxi-
mation of noisy experimental data. Though the order 8 of
MOG works reasonably well in case of our data, it is still a
hyper parameter that needs to be tuned for approximating
diverse data sets. For first convergence, we have used
optimized existing algorithms [19] to compute the param-
eters of MOG during function estimation.

4.3 Abel deconvolution integral

The coefficients of Gaussian Model, obtained in the func-
Figure 5. Mixture of eight individual Gaussian distributions. tion estimator stage are used to construct a continuous
function F(y, z) of the specific order for the variation of
intensity in each row (z). These coefficients of MOG are
computed by non-linear least square fitting method. In
Table 1. Statistical inference of Goodness Of Fit (GOF).
order to obtain the emission function f ðq; zÞ, the integral
Statistical inference Polynomial Fourier MOG form of inverse Abel transform [10], as given by equation
(4), has been used. Thus, the estimated continuous function
Regression squared (R-squared) 0.8544 0.9750 0.9994 F(y, z) is Abel deconvoluted, and the data corresponding to
Adjusted R-squared (AdjR- 0.8506 0.9736 0.9993 emission function in each row are acquired [12, 13]. For
squared)
deconvolution to be meaningful cylindrical symmetry is
Root Mean Square Error (RMSE) 722.7 303.6 47.9
mandatory. Though the experimental results are not 100%
symmetric we actually ignore the asymmetry as the varia-
tions are minimal. This is also the reason why we do only
patches of background information which is not of much deconvolution on one side. Assuming cylindrical symmetry
importance in such applications. The underlying hypothesis the deconvolution process, irrespective of whether it is
of function approximation is that an unknown distribution done on left or the right half, will give the same result.
can be well approximated by mixture of Gaussian when Therefore, in order to address this issue, only one side of
Sådhanå (2019) 44:114 Page 7 of 12 114

the projected image F(y, z) is deconvoluted to obtain one Therefore, f ðq; zÞ can be reduced further to
side of f ðq; zÞ and then it is reflected along the vertical axis
2 Z 1 2
(Z) for complete reconstruction of f ðq; zÞ. In practice, the eq 2eu
f ðq; zÞ ¼ pffiffiffi pffiffiffi du
deviation from symmetry is minimal when compared with p 0 p
the average counts, and hence, it is reasonable to tackle the q2
e
asymmetry issue in such manner. This procedure is fol- ¼ pffiffiffi erf ð1Þ
lowed for all the rows of the enhanced image to obtain p
2
spatial field information of the entire image. eq
¼ pffiffiffi ; *erf ð1Þ ¼ 1
The analytical solution of Abel inversion can be derived p
as following. For reducing the complexity, we have shown
yb 2 This approach can be extended to higher order Gaussian
the derivation for one Gaussian basis i.e.Fðy; zÞ ¼ aeð c Þ basis functions in the similar manner.
with a ¼ 1; b ¼ 0; c ¼ 1. Now, f ðq; zÞ can be obtained from
F(y, z) by
Z 4.4 Spatial distribution of data
1 1 dF dy
f ðq; zÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; 8z
p y¼q dy y2  q2 One of the important aspects of this processing is to arrange
Z    the data in spatial coordinate frame. After performing Abel
1 1 ðyb Þ
2 yb 1 dy
¼ ae c ð2Þ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi deconvolution, the deconvoluted image is obtained in pixel
p y¼q c c y  q2
2
Z 1 coordinate frame which needs to be converted into spatial
2 y 2
domain in order to analyse better in real world units such as
¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ey dy;
p y¼q y  q
2 2
cm, mm, etc. In order to achieve this, we have used the
substituting a ¼ 1; b ¼ 0; c ¼ 1 stored relevant scaling information to convert from pixel
2 Z 1 coordinate frame to spatial coordinate frame. The scaling
2eq 2
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
¼ eu du; substituting y2  q2 ¼ u information is obtained by estimating the pixel resolution
p u¼0 from images of standard calibration targets. By proper
utilization of this information, Abel inverted image is rep-
Gaussian error function is a special integral given by the resented in spatial coordinate frame. The pseudo codes for
following equation. coefficient extraction and Abel deconvolution are given in
Z 2 Algorithm 1 and Algorithm 2 respectively.
x
2et
erf ðxÞ ¼ pffiffiffi dt ð10Þ
t¼0 p
114 Page 8 of 12 Sådhanå (2019) 44:114

4.5 Proposed algorithm

Sådhanå (2019) 44:114 Page 9 of 12 114

Figure 6. (a) Photograph of flame from a standard swirl burner. (b) Time averaged 2D-velocity distribution along the axial plane.
(c) Time averaged OH chemiluminescence image. (d) Abel deconvoluted image. (e) Multi-layered velocity field overlapped with Abel
deconvoluted OH* image as contour plot.

4.6 Computational complexity adapt this change as the coefficients involved in our model
depend on input data. Also, it performs less but not the least
One major advantage of using mixture of Gaussian is the number of transcendental operations. It has storage space
number of transcendental operations involved in computing
complexity of Oðn2 Þ in order to store the resulting image.
the coefficients of fitted curve. This model requires three
The coefficients of the model can be pre-computed and
transcendental operations for computing an ; bn ; cn (Eq. (7)),
stored prior to the execution of the Abel integration which
in contrast with ASAI [13] which requires four transcen-
helps to boost the speed. The Abel integration algorithm is
dental operations between two discrete data points. The
also of O(n) time complexity which can be computed in the
worst case time complexity of the proposed algorithm (4.5)
similar manner as in coefficient generation. Thus, the
is O(n) because it takes O(1) time in computing the coef-
overall time complexity of the algorithm is O(n) by the
ficients of each n row of an nm image where as the time
addition rule of measuring worst case time complexity [20].
complexity of ASAI is Oðn2 Þ [13]. Though FAI algorithm Above all, the proposed algorithm meets the requirement of
[13] uses less number of transcendental operations than the using minimal transcendental operations with the ability to
proposed algorithm, it doesn’t adapt to the changes in capture necessary changes in the input data.
characteristics of data points. The proposed algorithm can
114 Page 10 of 12 Sådhanå (2019) 44:114

(a) (b) (c)

Figure 7. Abel deconvolution without and with image enhancement (IE) are shown in (a) and (b) respectively. It is evident that the
edge adaptive smoothing of input image alleviates the adverse effect of unwanted noise signals in the deconvoluted image. (c) Intensity
profile along a fixed row (here, 313) of deconvoluted images with and without IE.

5. Experimental details and analysis

The intended application of the newly developed algo-

rithm is explained in figure 6. The exemplary images shown
correspond to optical and laser diagnostic investigations
conducted in a swirl stabilized, standard burner with
methane as fuel [21]. Figure 6(a) shows the photograph of
the flame, and the corresponding time averaged OH* image
captured using an ICCD camera (La Vision, Model:
Nanostar, 12 bit, 1280 x 1024 pixels) is shown in fig-
ure 6(c). The burner is operated with an air and fuel flow Figure 8. Qualitative comparison of the proposed method
(b) with one of the state-of-the-art methods (a) from literature.
rates of 200 lpm and 18.9 lpm respectively, resulting in a
The exemplary image shows the deconvoluted chemiluminescence
thermal power of 11.1 kW. The burner operated in non- image from a swirl stabilized burner. Since any one side is
premixed mode at atmospheric conditions produced a tur- reflected to reconstruct the complete image, one can compare
bulent flame at the investigated conditions. The OH* ima- either side of both outcomes to draw inferences regarding the
ges are captured with an exposure time of 100 efficacy of the method.
microseconds, and 300 such instantaneous images are used
to generate the time averaged image. In general the domi-
nant noise source in scientific grade CCD and ICCD cam- same camera and intensifier settings with the flame extin-
eras are shot noise (the variations in the photoelectrons guished. By uniform illumination in the experimental setup
generated at each pixel) and the dark current read noise. we try to minimize noise due to variable illumination,
The dark current read noise is negligible in our case owing though not completely, and by edge adaptive smoothing we
to the short exposure and readout times. Post processing of try to remove the AWGN noise up to some extent. It is
the raw images was done to correct the camera chip sen- evident from the flame photograph that the flame is lifted
sitivity or shot noise and background noise (signals coming from the burner exit. One of the goals is to understand the
from unwanted background light). Shot noise usually fol- influence of underlying local turbulent velocity field on the
low Poisson distribution. For large numbers, the Poisson flame stabilization and heat release o r in other words, the
distribution tend to follow normal distribution, and the underlying physical mechanism that enables the flame to
elementary events such as photons, electrons etc. are no stabilize against such high velocity flows. One way is by
longer observed individually. Typically this makes shot pixel-to-pixel comparison of the flame data with the cor-
noise in actual observations indistinguishable from true responding velocity information. This implies that spatially
Gaussian noise, and hence, the necessity of adaptive resolved information about the velocity field along with the
smoothing in image enhancement process. Variations in the information about the flame/reactant species is
camera chip sensitivity is normalized using an ensemble- indispensable.
average image of a uniformly illuminated screen. The The 2D vector field is obtained using a laser based
ensemble averaged background image is obtained using the diagnostic technique called Particle Image Velocimetry
Sådhanå (2019) 44:114 Page 11 of 12 114

(PIV) [22], where the displacement vector is calculated towards the axis of symmetry is expected due to the
using auto-correlation or cross correlation techniques. influence of heat release from the burner exit. The state-
Using the time between the laser shots and the physical of-the-art method, as shown in figure 8(a), does not have
pixel resolution of the camera the velocity is then calcu- the provision to regularize inversion by imposing
lated from the displacement vector. As the PIV data is smoothness constraint on acquired raw data. This results
spatially resolved in 2D, the missing element in this anal- in abrupt fluctuations in the reconstructed scalar field.
ysis is the spatially resolved information about the reactive However, the proposed reconstruction, as shown in fig-
species. So, the Abel deconvolution algorithm is employed ure 8(b), alleviates this effect through constrained regu-
on the time averaged OH* images to obtain the 2D spatially larization. Due to unavailability of continuous
resolved information as shown in figure 6(d). Due to the reconstructed profile in the state-of-the-art method, we
axial symmetry only one half of the image is shown. The have quantized our continuous reconstructed profile and
time averaged and planar vector field corresponding to the compared with the available quantized version of the
flow conditions is shown in figure 6(b). Here the red and state-of-the-art method. The efficacy of inversion can be
black colour vectors indicate the maximum and minimum suitably inferred from this comparison, as shown in fig-
velocities respectively. As OH* is a measure of the heat ure 8, by focusing on each quantized region of both the
release, by overlapping the Abel deconvoluted OH* image images. The proposed method has fewer fluctuations in
on the 2D-PIV image a deeper insight into the turbulence- each quantized region of the state-of-the-art method that
chemistry (heat release) interactions can be obtained. Such leads to marginally superior inversion. However, as the
a multi-layered image is shown in figure 6(e) where the final reconstruction is greatly influenced by the quality of
planar velocity field is overlapped with Abel deconvoluted the raw data, which again depends on the experimental
OH* images in contour plot. The image gives a good settings, it is not possible to have a concrete conclusion on
impression of the flow field conditions in which the flame is the superiority of the new method from the images here.
stabilizing. Just ahead of the flame front there is an apparent What is clear from figure 9 is that the proposed MOG
deceleration of the flow field and also the OH* contours of based Abel inversion through constrained regularization is
maximum heat release are located in regions of relatively proved to be equally good from the application of point of
lower local velocities. Here the reader is reminded that 2D view and hence, can be suitably used as an alternative to
images do not give the complete picture as the turbulence - the conventional methods.
flame interactions in such flows are highly three dimen- The obtained results asserts that the proposed technique
sional in nature. However the example clearly highlights using standard Abel inversion for the application in com-
the importance and the benefits of the Abel deconvolution bustion research is useful where limited infrastructure is
algorithm in analysis of such combustion related available for acquiring the expensive laser diagnostics
phenomena. methods like planar laser induced fluorescence of CH rad-
In order to validate the relevance of image enhance- ical (CH-PLIF) image. Our findings indicates, the applica-
ment, we have compared the Abel deconvolution process tion of MOG prior to deconvolution approximates the
with and without the proposed pre-processing i.e. edge discrete data points relatively better than the other com-
adaptive Gaussian smoothing in an exemplary deconvo- pared methods. Based on the intensity profile and qualita-
luted flame image as shown in figure 7. The image tive comparison among various standard methods, we
enhancement process imposes additional degree of believe that the proposed framework is an alternative way
smoothness on the input image which is reflected in the to achieve the desired results and it works satisfactorily
deconvoluted images. From this comparison, we report well compared to other standard techniques in the adopted
that the edge adaptive smoothing lessens the effect of experimental setting.
unwanted signals in such applications. From figure 7(c), it
is apparent that the MOG model still lacks smooth tran-
sition in intensity, mainly due to high fluctuation in raw 6. Conclusion
data as a result of unconstrained noisy environment.
However, enhancing the degree of smoothness prior to Abel transformation finds its applications in many research
fitting MOG model resolves this issue up to some extent, fields including plasma diagnostics, X-ray radiography etc.
as can be inferred from figure 7(b). Further, we have In this paper, we critically analyze the inherent issues of
compared our approach with one of the state-of-the-art Abel deconvolution and propose a generic framework to
methods [23] dealing with similar issues as ours. As fig- tackle these issues up to some extent. The proposed algo-
ure 8 indicates, the state-of-the-art method suffers from rithm gave maximum efficacy with the adopted image
adverse effects of noisy environment, however, the pro- acquisition settings. Moreover, the proposed framework
posed algorithm could suppress these unwanted signals offers a provision to eliminate a certain amount of
and hence, it could impose higher degree of smoothness unwanted signals which is essential to be removed in many
by simultaneously leveraging the image enhancement and scenarios. Of particular interest is the extraction of 2D and
Gaussian mixture model. The variation of intensity
114 Page 12 of 12 Sådhanå (2019) 44:114

spatially resolved electronically excited emission from the [9] Sizikov V S, Evseev V, Fateev A and Clausen S 2016 Direct
hydroxyl radical (OH* chemiluminescence) in axisym- and inverse problems of infrared tomography. Appl. Opt.
metric flames. The deconvoluted OH* image overlapped 55(1): 208–220
with the time-averaged 2D velocity plot from the same [10] Abel N H 1827 Recherches sur les fonctions elliptiques. J.
burner highlights the importance and effectiveness of the fur die reine und angewandte Math. 2: 101–181
[11] Deutsch M and Beniaminy I 1982 Derivative-free inversion
algorithm in combustion research. Above all, the algorithm
of abel’s integral equation. Appl. Phys. Lett. 41(1): 27–28
meets the requirement of performing Abel inversion in O(n) [12] Deutsch M and Beniaminy I 1983 Inversion of abel’s integral
time and Oðn2 Þ space complexity through regularization by equation for experimental data. J. Appl. Phys. 54(1):
imposing smoothness constraint over acquired raw images. 137–143
[13] Gueron S and Deutsch S 1994 A fast abel inversion algo-
rithm. J. Appl. Phys. 75(9): 4313–4318
[14] Petrou M and Petrou C 2010 Image processing: the funda-
References mentals. Hoboken: Wiley
[15] De Moor B 1994 Total least squares for affinely structured
[1] Trupti K 2011 Reaction kinetics modeling of OH*, CH*, and matrices and the noisy realization problem. IEEE Trans.
C2* chemiluminescence. Ph.D. thesis Signal Process., 42(11): 3104–3113
[2] Najm H N, Paul P H, Mueller C J and Wyckoff P S 1998 On [16] Press William H, Flannery Brian P, Teukolsky Saul A and
the adequacy of certain experimental observables as mea- Vetterling William T 1989 Numerical recipes, volume 3.
surements of flame burning rate. Combust. Flame 113(3): Cambridge University Press, Cambridge
312–332 [17] Dyn N 1989 Interpolation and approximation by radial and
[3] Ikeda Y, Kojima J and Nakajima T 2005 Basic aspects of related functions. Approximation Theory VI, 1(s 1): 211–234
oh(a), ch(a) and c2(d) chemiluminescence in the reaction [18] Ross S M 2014 Introduction to probability and statistics for
zone of laminar methane-air premixed flames. Combust. engineers and scientist. India: Academic Press, New Delhi, 5
Flame 140(1): 34–45 edition
[4] Boehly T R 1996 Abel inversion of emission and back- [19] MATLAB Optimization Toolbox 2014b The mathworks inc.
lighting images. LLE Rev. 66: 66–72 Natick, MA
[5] Dasch C J 1992 One-dimensional tomography: a comparison [20] Allen W M 1996 Algorithms, data structures, and problem
of abel, onion-peeling, and filtered backprojection methods. solving with C??. Boston: Addison Wesley Longman
Appl. Opt. 31(8): 1146–1152 Publishing Co., Inc
[6] Tikhonov A N 1975 Inverse problems in heat conduction. J. [21] Jarpala R, Burle N V S A, Voleti M and Sadanandan R 2017
Eng. Phys. Thermophys. 29(1): 816–820 Effect of swirl on the flame dynamics and pollutant emis-
[7] Simons S N and Yuan Z-G 2003 The filtered abel transform sions in an ultra-lean non-premixed model gt burner. Com-
and its application in combustion diagnostics. 1995 Fall bust. Sci. Technol. 189(10): 1832–1848
Meeting, Western States Section, sponsored by The Com- [22] Willert C E and Gharib M 1991 Digital particle image
bustion Institute, Stanford, California, October 30–31, 1995 velocimetry. Exp. Fluids 10(4): 181–193
[8] Daun Kyle J, Thomson Kevin A, Liu Fengshan and Small- [23] Sadanandan R, Kutne P, Steinberg A and Meier W 2012
wood Greg J 2006 Deconvolution of axisymmetric flame Investigation of the syngas flame characteristics at elevated
properties using tikhonov regularization. Appl. Opt. 45(19): pressures using optical and laser diagnostic methods. Flow
4638–4646 Turbul. Combust. 89(2): 275–294