neural network multi-criteria optimization image reconstruction
TRANSCRIPT
-
7/25/2019 Neural Network Multi-criteria Optimization Image Reconstruction
1/12
Neural network multi-criteria optimization image reconstructiontechnique (NN-MOIRT) for linear and non-linear process
tomography
W. Warsito, L.-S. Fan *
Department of Chemical Engineering, The Ohio State University, 140 West 19th Avenue, 121 Koffolt Laboratories, Columbus, OH 43210-1180, USA
Received 20 May 2002; received in revised form 25 October 2002; accepted 25 October 2002
Abstract
In this work, an analog neural network is utilized to de velop a new image reconstruction technique for the linear as well as the
non-linear process tomography. The ultrasonic computed tomography (CT) and the electrical capacitance tomography (ECT) are
chosen to represent the linear and the non-linear tomography. The image reconstruction technique is based on a multi-criteria
optimization, namely neural network multi-criteria optimization image reconstruction technique (NN-MOIRT). The optimization
technique utilizes multi-objective functions: (a) the negative entropy function, (b) the function of the least weighted square error of
projection (integral) values between the measured data and the estimated projection data from the reconstructed image, and (c) a
smoothness function that gives a relatively small peakedness in the reconstructed image. The optimization image reconstruction
problem is then solved using the Hopfield model with dynamic neural-network computing. The technique has been tested using
simulated and measured data; this technique has shown significant improvement in accuracy and consistency compared with other
available techniques for both linear and non-linear tomography.
# 2003 Elsevier Science B.V. All rights reserved.
Keywords: Hopfield neural network; Image reconstruction; Linear tomography; Electrical capacitance tomography; NN-MOIRT
1. Introduction
In the past several years, process tomography has
emerged as a powerful technique in the study of a wide
range of multiphase flow systems. The technique
provides both local and global information on the
dynamic behavior of the flow system, which is of
considerable importance to the system design, control
and monitoring.The process tomography involves the task of recon-
structing integral (projection) measurement data from
remote sensors mounted on the periphery of the conduit
of the flow system to generate a cross-sectional image of
the vessel. The process tomography system basically
consists of three parts: (1) a sensoring system to acquire
the measurement data, (2) an electronic system for data
acquisition, and (3) a computer system for measurement
control, image reconstruction and displaying the result
[1,2]. Fig. 1 shows a schematic representation of an
electrical capacitance tomography (ECT) system.
To obtain reliable images of the flow system studied, a
successful implementation of process tomography lies in
the selection of the sensor system deployed and the
image reconstruction algorithm. Based on the sensor
physics, process tomography may be differentiated into
two types: linear tomography, also referred as to hard
field tomography, and non-linear tomography, also
referred as to soft field tomography. The first type
includes tomography techniques based on electromag-
netic radiations, such as X-ray computed tomography
(CT), gamma-ray tomography, and acoustical tomogra-
phy. This type represents the early stage of process
tomography development with a wide range of applica-
tions from medical systems to chemical processes.
Nuclear radiation based tomography techniques provide
very high spatial resolution up to 1% of column
* Corresponding author. Tel.: /1-614-292-7907; fax: /1-614-292-
1929.
E-mail address: [email protected](L.-S. Fan).
Chemical Engineering and Processing 42 (2003) 663/674
www.elsevier.com/locate/cep
0255-2701/03/$ - see front matter# 2003 Elsevier Science B.V. All rights reserved.
doi:10.1016/S0255-2701(02)00204-0
mailto:[email protected]:[email protected] -
7/25/2019 Neural Network Multi-criteria Optimization Image Reconstruction
2/12
diameter. However, the scanning speed is too slow (few
seconds per frame) to perform real time measurement
[3]. The non-linear tomography is represented by the
more recently developed electrical-based tomography,
such as ECT that is used to image the dielectric
properties; and electrical resistance tomography (ERT)
that is employed to image the conductivity of a multi-
phase media. The ECT and ERT have been gaining
Fig. 1. Schematic diagram of ECT system.
Fig. 2. Comparisons of reconstructed results for ultrasonic CT.
W. Warsito, L.-S. Fan / Chemical Engineering and Processing 42 (2003) 663 /674664
-
7/25/2019 Neural Network Multi-criteria Optimization Image Reconstruction
3/12
acceptance as laboratory and industrial tools for the
purpose of multiphase flow imaging because of their
high-speed capability, low construction cost, high safety
and suitability for small or largevessels. The high-speed
capability is one of the most relevant factors in multi-
phase flow imaging for a real time monitoring. ECT is
particularly attractive to industrial systems as manychemical processes use non-conductive organic liquids
as the liquid phase[4]. Currently, commercially available
ECT systems can acquire a speed up to 100 image
frames per s.
One of the most important issues for tomography
applications in chemical processes is spatial resolution.
The process tomography systems as compared with
medical tomography systems is of relatively lower
spatial resolution due to the limited integral measure-
ment data obtained, which consequently influences the
accuracy of the reconstructed image. The relatively
smaller number of the integral measurements is commonas the result of limitations in the number of the sensors
that can be employed. The need for high-speed imaging
also prompts the number of the sensors to be limited as
few as possible. Consequently, an increase in the number
of the integral measurements consequently would in-
crease the data acquisition time, and thus would
decrease the real time capability. The issue raised is
how to acquire reconstruction results with sufficient
accuracy for process quantification from the limited
experimental integral data without sacrificing the time
resolution for reliable real time measurements.
The choice of the image reconstruction technique is
crucial as it determines the quality of the imageproduced. For linear tomography, the commonly used
filtered back projection technique (FBPT) requires a
large number of integral measurement data to obtain a
reliable accuracy[5]. In medical imaging, the number of
projection data is up to 256/256 (number of projection
directions/paths) compared with only 10/20% of that
number in chemical process applications. For multi-
phase flow imaging with limited projection data iterative
techniques based on iterative version of FBPT [5,6],
entropy maximization [7] or algebraic reconstruction
technique (ART) [8] are usually employed to obtain
more accurate reconstruction result. The iterative tech-niques implement a single criterion, i.e. the least square
error between the measured and the calculated projec-
tion data to be minimized or the entropy function to be
maximized. The iterative techniques described above are
a type of optimization technique aimed at finding a least
squared error between the measured and the estimated
projection data (capacitance data in the ECT). However,
the image reconstruction is an ill-posed problem. In flow
imaging especially, there are much fewer independent
measurements (projection data) than unknown pixel
values (field distribution). Thus, such a criterion does
not necessarily determine the image vector, and there
may be more than one possible alternative image.
Another reason why such a solution is not necessarily
good is that the least squared criterion does not
contain any information regarding the nature of a
desirable solution [9]. Therefore, more than one
objective function is required to be minimized simulta-
neously to choose the best compromise solution among
possible alternatives.
In the case of the electrical-based tomography, in
addition to the issue noted above, a non-linearity
problem due to the soft field effect of the electrical
field distribution exists for which no analytical solution
so far is available for the inverse problem. In the
electrical-based tomography, linearization using the so
called sensitivity model is usually applied before the
same algorithm as in the linear tomography is utilized
[7,8]. The commonly used image reconstruction techni-
que for the electrical tomography is based on the linear
back projection (LBP), which is essentially the same as
the back projection technique used for the linear
tomography. The technique provides only a rough
estimation of the image reconstructed. A great demand
for a high quality image generated by electrical tomo-
graphy has led many researchers to develop reconstruc-
tion techniques based on iterative algorithms. More
comprehensive reviews on the reconstruction technique
for ECT are presented by Isaksen [10]and Warsito and
Fan [11]. Most iterative techniques for electrical tomo-
graphy are of the ART type as applied for linear
tomography, in which the least square error between
the measured and the calculated integral data (current in
ERT or capacitance in ECT) is minimized. The main
problem of this criterion is the sensitivity to noise, in
which convergence is difficult to reach when the
measurement data contain noises. This is the drawback
for the imaging of highly fluctuated multiphase flow
systems.
Thus, there is a need for development of a robust
reconstruction technique for linear and non-linear
process tomography, which is capable of yielding a
reconstructed image with reliable accuracy and meeting
the required chemical process applications. In this work,
a new image reconstruction technique for linear and
non-linear process tomography is developed based on
the Hopfield analog neural network multi-criteria opti-
mization, namely the neural network multi-criteria
optimization image reconstruction technique (NN-
MOIRT). The analog neural network technique has a
major advantage over the feed forward neural network
algorithm because prior knowledge of the image pattern
to be reconstructed for network training is not required.
The performance of the image reconstruction technique
of this study is compared with other commonly used
techniques.
W. Warsito, L.-S. Fan / Chemical Engineering and Processing 42 (2003) 663 /674 665
-
7/25/2019 Neural Network Multi-criteria Optimization Image Reconstruction
4/12
2. Theory
2.1. Forward problem in integral measurement
The tomography technique can be divided into two
steps: the integral measurement and the image recon-struction. The integral measurement is a forward
problem which transforms the field property X(x , y )
into the measured integral data, y , which is easily
measurable, e.g. time-of-flight of an ultrasonic wave or
intensity of an electrical signal. In the linear tomography
the relationship between the field property and the
measured integral data becomes linear and can be
expressed as:
Y(s; u) gL(s; u)
X(x; y)dl (1)
where L (s , u ) is the projection line as a function of thedistance from the origin, s and the angle, u . In the case
of the ECT, the field property distribution corresponds
to the permittivity, o(x , y), while the measured para-
meter is the capacitance data. The forward problem is
based on the following Poissons equation[12,13]:
9o(x; y)9f(x; y)0 (2)
where o(x , y ) is the permittivity distribution, f (x , y ) is
the electrical field distribution. This is second-order,
elliptic-type partial differential equations with the typi-
cal boundary conditions of Dirichlet or Neumann type.
The measured capacitance C(s , u) is obtained by
integration ofEq. (2):
C(s; u) 1
DV(s; u) GG(s; u)
o(x; y)9f(x; y)dl (3)
where DV(s , u ) is the voltage difference between the
source and the detector electrodes. G (s , u) is the curve
enclosing the detector electrode. Eq. (3) relates the
dielectric constant (permittivity) distribution, o(x , y),
to the measured capacitances C(s , u). That is, for a
given medium distribution o(x , y ) and the boundary
conditions, the capacitances can be calculated using, for
example, the finite element method (FEM).
2.2. Inverse problem in image reconstruction
The image reconstruction process is an inverse
problem involving the estimation of the field property
distribution X(x , y ) from the measured parameter Y(s ,
u). For linear tomography, the commonly used image
reconstruction technique is the FBPT and ART. In the
FBPT, the field property distribution X(x , y) is
obtained analytically by integrating the convolution of
the measured integral data, Y(s , u), and a filter
function, g(s ):
X(x; y)gp
0
[Y(s; u)g(s)]sx cos uy sin udu (4)
The major problem of the technique is that it requires a
large number of projection data, which is not suitable
for process applications, because in most cases thenumber of collectable projection data is limited. An
iterative technique is usually applied for this case. Here,
the reconstructed image (field distribution) is used to
predict the projection data, which is then compared with
the measured projection data to correct the recon-
structed image iteratively.
The ART type is based on the series expansion ofEq.
(1).In matrix formulation it can be written as:
YAX (5)
where Y is the measurement vector, X is the image
vector (field distribution), and A is the projection
matrix. The image reconstruction using the ART is theproblem of estimating image vector X such that the
estimated integral projection data Y*(X)5/Y, given a
measurement vector Y. The ART can be expressed by
the following form
X(k1)
X(k)a(k)AT(YY(X(k))) (6)
Here, AT is the transpose matrix of A, X(k) is the
estimated image vector in the k-th iteration, a is a
relaxation factor, also called as gain factor or weighting
factor. The ART requires less number of projection data
than IFBPT; it is, however, generally sensitive to noise.
For the ECT, as seen from Eq. (3), it is difficult to
obtain an explicit expression that relates the measured
capacitances to the fraction and position of the dielectric
components inside the measured domain. Since there is
no general method of solution even for the forward
problem, in which the equation becomes linear, approx-
imation methods are usually used. The most common
method is the use of the so-called sensitivity model
[12,14]. Based on this model, Eq. (3) can be written in
matrix expression as:
CSG (7)
where C is the measured capacitance matrix, G is the
image vector (permittivity distribution) and S is the so-called sensitivity map. The image vector G is obtained
from the following matrix multiplication:
GSTC (8)
where ST is the matrix transpose ofS. The reconstruc-
tion technique is referred to as the LBP. The recon-
structed image using the LBP is blurred, showing a
smoothing effect on the sharp transitions between the
different dielectric constants. To obtain more accurate
reconstruction results, an iterative procedure as in the
linear tomography is usually applied. Eq. (7) corre-
sponds toEq. (5)for the linear tomography. Therefore,
W. Warsito, L.-S. Fan / Chemical Engineering and Processing 42 (2003) 663 /674666
-
7/25/2019 Neural Network Multi-criteria Optimization Image Reconstruction
5/12
an iterative technique used for the linear tomography is
also applicable to this problem.
2.3. Multi-criterion optimization in image reconstruction
problem
Considering the error due to the measurement in-
accuracy and the truncation in the series expansion and
linearization, bothEqs. (5) and (7) can be written as
YAXe (9)
where e is an M-dimensional error vector. A corre-
sponds to the projection matrix for the linear tomogra-
phy, and the sensitivity matrix for the ECT. Then, the
reconstruction problem is to find methods for estimating
the imagevector (field property distribution) X from the
measurement vector Y, and minimize the error e. Since
we obviously do not know e , the problem is to find acompromise solution of the system under certain
conditions (criteria), such that
AX5Y (10)
The criteria that have been used for the image recon-
struction problem are usually of the form: choose as the
solution of Eq. (10) an image vector X for which the
value of some function fi(X) is minimal, and if there is
more than one X which minimizes fi(X) choose among
these one for which the value of some other function
fj(X) is minimal [9]. The multi-criteria optimization
based solution method is usually employed to find an
image which (a) has the largest entropy; (b) has the least
weighted square error between the measured data set
and the estimated value calculated from the recon-
structed image; (c) is smooth and has a relatively small
peakedness. The corresponding objective functions are,
respectively:
f1(X)g1
XNj1
Xj lnXj: (11)
f2(X)1
2g2AXY
2
1
2g2
XMi1
XNj1
AijXjYi
2(12)
f3(X)1
2g3(X
TNXX
TX) (13)
where g1, g2 and g3 are normalized constants between 0
and 1. N is N/N non-uniformity matrix. Again, the
multi-criteria optimization for the reconstruction pro-
blem is to choose an imagevector for which thevalue of
the multi-objective function F(X)/[f1(X), f2(X), f3(X)]T
is minimized simultaneously. This can be realized by the
weighted sum technique stated mathematically as:
minimizeXP
F(X)X
i
wifi(X); (i1; 2; 3)
so that AXY50:
( (14)
where wiis a weight operating on the objective function
fi(X) and can be interpreted as the relative weight or
worth of the objective compared with the other
objectives, i.e.:
X3i1
wi1 : (15)
P is the feasible set of constrained decisions defined by
linear constraints:
PfX RNAX5Y; X]0; Y RMg (16)
2.4. Solution using Hopfield dynamic neural networks
Here, we use an analog neural network based on the
Hopfield model to solve the optimization problem (Eq.
(14)). More details of the reconstruction technique are
described elsewhere [11]. To solve the optimization
problem using the neural network technique, the gray
level (attenuation coefficient distribution for X-ray
tomography, sound velocity distribution for acoustical
tomography, or permittivity map for ECT) Xj in the
image vector is mapped into the output variable vj for
neuron j bounded by 0 and 1. The output variable is a
continuous and monotonic increasing function of the
internal state of the neuron:
XjvjfS(uj); (17)
where uj is the internal state variable of the neuron j.
Function fS is called activation function with typical
choice of the form as:
fS(uj) 1
1 exp(buj); (18)
where b is a steepness gain factor that determines the
vertical slope and the horizontal spread of the sigmoid-
shape function. The non-linear sigmoid function forces
the neuron output to converge to binary output 1 or 0.
However, when an output value between 0 and 1 isequally expected as the extreme values, as in the
application of tomography for concentration measure-
ment, we use the following linear activation function:
fS(uj)
0 if uj5j
b
bujj if j
bBujB1
j
b
1 if 1j
b5uj
8>>>>>>>>>>>>>>>:
(19)
whereb andj become the slope and the intercept of the
W. Warsito, L.-S. Fan / Chemical Engineering and Processing 42 (2003) 663 /674 667
-
7/25/2019 Neural Network Multi-criteria Optimization Image Reconstruction
6/12
line, respectively. In the Hopfield model dynamics
neural network the neural, activities of the output vjas well as the input ujdepend on time t . The evolution
of the network is determined by the N-coupled ordinary
differential equations[15,16]:
C0j dujdt@E(
v)
@vj(20)
whereC0jis an associated capacitance in thejth neuron
in the Hopfield nets with time constant t (t/R0jC0j,
R0j is the associated resistance), and E is the system
energy of the Hopfield networks. The corresponding
neural networks energy function to the optimization
problem inEq. (14) can be written as:
E(X)v1g1
XNj1
Xj lnXj1
2v2g2AXY
2
12v3g3(X
TNXX
TX)
XM
i1
C(zi)X
N
j1
1
Rj
gXj
0
f1S (X)dX (21)
where Nis the number of neurons in the Hopfield nets,
which is equal to the number of pixels in the digitized
image. Mis the number of integral data measurements.
The first three terms in Eq. (21) are the interactive
energy among neurons based on the objective functions.
The forth term is related to the constraint violation,
which must be minimized. The fifth term encourages thenetwork to operate in the interior of the N-dimensional
unit cube {05/Xj5/1} that forms the state space of the
system. The constraint function C(zi) is defined as:
dC
dzid(zi)
0 if zi50
azi if zi0
(22)
where,
ziXNj1
AijXjYi (23)
Herea is a large penalty parameter. The function C(zi)provides for the output of the neuron j to be a large
positive value when the corresponding constraint equa-
tion is being violated. Since the Hopfield network is a
monotonic descent technique as described above, it will
converge to the first local minimum it encounters. By
introducing the penalty function, temporary increases in
the descending evolution of the Hopfield network
energy is permitted in order to allow escape from local
minima. The form of penalty parameter a is chosen as:
a(t)a0zexp(ht) (24)
Here a0, z and h are positive constants. This modified
Hopfield network provides a mechanism for escaping
local minima by varying the direction of motion of the
neurons. In the initial steps, the ascent step is taken
largely by the penalty function, but less likely as the
algorithm proceeds.
For simplicity, choosing R0j/R0 and C0j/C0, and
redefining R0C0, g1v1/C0, g2v2/C0and g3v3/C0as, t , g1,g2and g3, respectively, the time evolution of the internal
state variable of neurons in the networks becomes:
u?(t)u(t)
t
[g1(1ln X(t))g2ATz(t)g3(NX(t)X(t))
ATd(z(t))] (25)
where,
u(t) [u1(t); u2(t); . . . ; uN(t)]T;
X(t) [X1(t); X2(t); . . . ; XN(t)]T
;z(t) [z1(t); z2(t); . . . ; zN(t)]
T:
The image vector Xj becomes the output of jth
neuron, which is calculated from the sigmoid function
(Eq. (18)orEq. (19)):
Xj(t)fS(uj(t)); j1; 2; . . . ; N (26)
Unlike in the multi-layer feed forward (MLF) neural
network, the parameter uj is calculated here by solving
Eq. (25) using, for example, Eulers method. Here, we
set C0/1.0, until the system converges to a stationary
state. Take t/1.0 so that time is measured in units t ,
and time step length Dt/0.01. The iteration process ofthe internal state variable of jth neuron and the
corresponding pixel value of the image are, respectively,
given by:
uj(tDt)uj(t)u?j(t)Dt (27)
Xj(tDt)fS(uj(tDt))Xj(t)f?S(u)u?j(t)Dt (28)
where fS? (u )/dfS(uj)/duj and uj?(t )/duj(t )/dt . The
stopping rule is used when the changes in the firing
rates become insignificant, i.e. for all pixelsjGj(t/Dt)/Gj(t)j/1.
3. Simulations and experiments
The image reconstruction algorithm proposed in this
work is tested by reconstructing simulated and actual
projection measurement data based on ultrasonic CT
and ECT. The simulated projection data for ultrasonic
CT is calculated based on Eq. (1) with a known field
property distribution. The projections are conducted
from 18 directions and eight paths with the same spacing
for each direction. The simulated capacitance data for
ECT is calculated based on Eq. (3) with a given
W. Warsito, L.-S. Fan / Chemical Engineering and Processing 42 (2003) 663 /674668
-
7/25/2019 Neural Network Multi-criteria Optimization Image Reconstruction
7/12
permittivity distribution using the FEM. A 12-electrode
capacitance sensor model is used for the simulation. The
images are reconstructed on 32/32 digitized image
pixels from 66 measured capacitance data.
An experiment to test the reconstruction technique for
real capacitance data is carried out in 5 cm I.D. column
using a 12-electrode sensor with a dielectric material inthe column. The schematic diagram of the ECT is given
in Fig. 1. The data acquisition system used is from
Process Tomography Limited (UK) and is capable of
capturing image data up to 100 frames per s. Sixty-six
measured capacitance values from the 12 electrodes for
one frame are used to reconstruct the image on 32/32
digitized image pixels using the algorithm described
above, i.e. the resolution (pixel size) is about 3% of the
column diameter. The reconstruction results using the
new algorithm are also compared with those using other
techniques based on the same experimental data. No
experiment is performed to test the reconstructiontechnique for the linear tomography.
An experiment to test the reconstruction technique for
flow imaging of gas/liquid/solid flows in bubble
column is also performed. The experiment is carried
out in a 10 cm diameter column using the same ECT
system as described above. The gas distributor is a single
nozzle with a diameter of 0.5 cm. Air (dielectric
constant/1), Norpar 15 (paraffin, density/773 kg/
m3, viscosity/0.253 mPa/s, dielectric constant/2.2)
and glass beads (density/2470 kg/m3, average dia-
meter/200 mm, dielectric constant/3.8) are used as
the gas, liquid, and solid phases, respectively. The
experiments are conducted in a semi-batch mode with-out inlet liquid flow with the liquid levels maintained at
60/80 cm during fluidization and the total solids
loading of 40% by volume. A twin-plane capacitance
sensor with time resolution of 50 frames per s per plane
is used to perform a real time imaging of the three-phase
flow system. The NN-MOIRT reconstructs the permit-
tivity map from the collected measurement capacitance
data off-line. A three-phase capacitance model com-
bined with a two-region model solids concentration in
the emulsion phase of the three-phase system is used to
attain the gas and the solids concentration distributions.
The two-region model assumes a uniform solids con-centration distribution in the solid/liquid emulsion
phase. More details on the models are described else-
where[17].
For linear tomography, the data obtained by Warsito
et al. [6] for air/water/glass beads system based on
ultrasonic tomography is used. The column used is 14
cm I.D. and 140 cm height. The gas distributor is six
nozzles located at the center of the column. The average
diameter of the glass beads is 260 mm with density of
2470 kg/m3 with the solids loading less than 2% and the
gas superficial velocity less than 2 cm/s. The ultrasonic
tomography technique could simultaneously differenti-
ate the gas and the solids concentrations in the three-
phase system based on measurements of the time-of-
flight and the attenuation of the ultrasonic wave
transmitted through the three-phase media. It is, how-
ever, only the time average measurement data available,
as the technique is too slow to perform a real time
imaging. The NN-MOIRT reconstructs the time-of-flight and the attenuation data to obtain the time
average cross-sectional distributions of the gas and the
solids concentration distributions.
4. Results and discussion
Fig. 2shows the reconstruction results from simulated
projection data based on the ultrasonic tomography
using the NN-MOIRT as well as commonly used
reconstruction techniques for linear tomography: ART
and iterative filtered back-projection technique(IFBPT). No noise is added to the simulated projection
data. Clearly, the reconstruction results from the limited
projection data by the IFBPT are severely distorted. The
ART gives better results than IFBPT; it is, however,
difficult to reach convergence especially when noise
exists (data not shown). Over 200 iterations are needed
for the ART for the noise-free data as compared with
less than ten iterations for the IFBPT. More iteration is
needed for the ART depending on the magnitude of the
noise added. The NN-MOIRT gives the best results
among the three reconstruction techniques and con-
verges in less than half of the iterations required for the
ART.Fig. 3 shows the reconstruction results for the ECT
using the NN-MOIRT in comparison with commonly
used techniques, including the LBP and the iterative
version of LBP (ILBP). No noise is added in the
simulated capacitance data for these results. From Fig.
3, it is seen that LBP only gives rough estimations of the
original images. Clearly, the ILBP and the NN-MOIRT
give much better results than the LBP, and the NN-
MOIRT gives the best accuracy and consistency. When
a Gaussian noise (up to 20 dB) is added, the ILBP no
longer reconstructs the original images accurately (Fig.
4). Again, the LBP only reconstructs the original imagesvery roughly. The reconstructed results by the ILBP are
severely distorted from the original images. Meanwhile,
the NN-MOIRT shows its robustness to the noise and
still gives a good accuracy and consistency despite of the
existence of the noise. The robustness of a reconstruc-
tion technique with respect to noise is crucial as
measurement data always contains noise varying in
magnitude.
Reconstruction results from actual measurement data
using the ECT are shown in Fig. 5(a and b). Fig. 5(a)
shows the reconstruction results of a static tube placed
in a column and measured by the capacitance sensor.
W. Warsito, L.-S. Fan / Chemical Engineering and Processing 42 (2003) 663 /674 669
http://-/?-http://-/?-http://-/?-http://-/?-http://-/?-http://-/?- -
7/25/2019 Neural Network Multi-criteria Optimization Image Reconstruction
8/12
The 3-dimensional image is obtained by stacking 50
image frames taken during 1 sec in the same position.
Fig. 5(b) shows the reconstruction results of a tube
circling with the period of one circle per s in the column
measured by the capacitance sensor. The 3-dimensional
image is obtained by stacking 125 image frames taken
during 2.5 s in the same position. Clearly, both results
reveal the accuracy of the NN-MOIRT in reconstructing
the actual experimental data.
Fig. 6 shows time average distributions of gas and
solids concentrations in a dilute gas/liquid/solid flow
system at a different level of the columns based on the
Fig. 3. Comparisons of reconstructed results for ECT from noise free simulated capacitance data.
Fig. 4. Comparisons of reconstructed results for ECT from noise contented simulated capacitance data.
W. Warsito, L.-S. Fan / Chemical Engineering and Processing 42 (2003) 663 /674670
http://-/?-http://-/?- -
7/25/2019 Neural Network Multi-criteria Optimization Image Reconstruction
9/12
ultrasonic tomography measurement. In the gas/liquid
system, there is no remarkable change in the gas holdup
profile with increasing the column level. In the gas/
liquid/solid system, the time averaged gas holdup
profile changes from a doughnut-shape distribution in
the lower region to a center-peak distribution in the
upper region. Meanwhile, the solids concentration
profile turns a doughnut-shape distribution with a
peak in the center region to a well-defined wall-peak
distribution in the upper region. Warsito et al. [6]
described that the doughnut-shape distribution in the
lower region for the gas holdup is due to the spiral
motion of the gas bubbles, and the solids concentration
profile is a result of bubble/particle interaction and
interaction between large scale liquid vortices and
particles.
Fig. 7 shows the quasi-3D real time distributions of
the gas bubbles and the solids concentrations in the two-
and the three-phase systems at gas velocity of 5 cm/s and
column levels of 20 and 55 cm above the gas distributor
Fig. 5. Reconstruction results of dielectric material using ECT.
Fig. 6. Time averaged cross-sectional distributions of gas and solids holdup in G/L and G/L/S systems based on ultrasonic tomography (UG/2
cm/s).
W. Warsito, L.-S. Fan / Chemical Engineering and Processing 42 (2003) 663 /674 671
-
7/25/2019 Neural Network Multi-criteria Optimization Image Reconstruction
10/12
based on ECT. The quasi-3D image is constructed by
stacking 200 tomography images captured in 4 s for the
same plane but different in time. The vertical coordinate
is taken as time by choosing the length scale for the time
coordinate (Z-coordinate) as the bubble rise velocity/
time. The quasi-3D images show the flow structures of
the gas/liquid flow along with the effect of the gas
velocity at levels 20 and 55 cm above the distributor. A
cutoff boundary for the gas holdup of 0.2 is chosen from
the images to determine the bubble surface and yields
reasonable bubble images, i.e. only sufficiently large
bubbles are imaged.
From the figure, for the gas/liquid system, at a lower
level (z/20 cm) the bubble swarms rise in spiral motion
while rocking back and forth from the sides of the wall
with period of 2/3 s. At a higher level (z/55 cm), the
spiral motion becomes insignificant, and the bubbles are
more concentrated in the center region of the column.
Comparing the images at the lower and the higher levels
shows that there are more relatively large bubbles (up to
Fig. 7. Quasi-3D gas and solids holdup distributions in G/L and G/L/S systems based on ECT (UG/5 cm/s).
W. Warsito, L.-S. Fan / Chemical Engineering and Processing 42 (2003) 663 /674672
-
7/25/2019 Neural Network Multi-criteria Optimization Image Reconstruction
11/12
5 cm diameter, indicated by dark red color) mixed with
small bubbles (indicated with light blue clouds) at the
upper level than that at the lower level. Meanwhile, for
the gas/liquid/solid system, a significant difference in
the gas bubble flow structures in the upper region as
compared with those in the lower region is clearly
observed. In the upper region, the bubble size is moreuniform, the bubbles are more concentrated in the
center region, and there is no relatively large bubbles
observed. This indicates that the interaction between the
bubbles and the solid particles causes significant bubble
break-ups, resulting in smaller size and more uniform
bubbles in the upper region. For the solids concentra-
tion distribution, the solids concentration in the region
where the gas holdup is nearly zero (almost no gas
bubble, indicated by bright red color) is almost uniform
throughout the column. This supports thevalidity of the
assumption made in the two-region model employed in
this work. The solids concentration profile is in opposi-tion to the gas holdup profile.
Fig. 8shows the cross-sectional profiles of the gas and
the solid holdups in the gas/liquid and the gas/liquid/
solid systems at a different level of the columns and
various gasvelocities averaged over a period of 4 s. The
time averaged gas holdup profiles in both the gas/liquid
and the gas/liquid/solid systems have a similar trend,
i.e. a doughnut-shape distribution in the lower region
and a center-peak distribution in the upper region. This
result indicates that the spiral motion becomes less
significant as the level of the column increases.
This phenomenon is the same as seen in Fig. 6 based
on ultrasonic tomography measurement regardless of
the types of liquid used, the gas distributor, the solids
loading and the column size. By increasing the column
level, the solids concentration profile turns from a
double-ring distribution or a single ring
(wall-peak distribution) with some amount of solid
particles in the center in the lower region to a well-
defined wall-peak distribution in the upper region. This
is also similar to those observed by the ultrasonictomography.
5. Concluding remarks
A new image reconstruction technique based on
analog neural network optimization, namely NN-
MOIRT is developed in this work for linear as well as
non-linear tomography. The reconstruction technique
implements multi-criteria optimization solved by a
modified Hopfield neural network. Comparisons withthe other commonly used reconstruction techniques for
both linear and non-linear tomography revealed much
improvement of the technique in the accuracy, the
consistency and the robustness to noise. The technique
is also applicable for any linear or non-linear tomogra-
phy using the appropriate projection (or sensitivity)
matrix. The technique also has the advantage of much
reduced computation time due to the possibility for
hardware implementation and inherent parallelism in
the computation. Some examples of the use of the
reconstruction technique to image multiphase flows in
three-phase bubble column based on time average modeultrasonic tomography and real time ECT have also
been presented.
Fig. 8. Time averaged cross-sectional distributions of gas and solids holdup in G/L and G/L/S systems based on ECT atUG/5 cm/s (left) and 15
cm/s (right) (color map is the same as in Fig. 1).
W. Warsito, L.-S. Fan / Chemical Engineering and Processing 42 (2003) 663 /674 673
http://-/?-http://-/?- -
7/25/2019 Neural Network Multi-criteria Optimization Image Reconstruction
12/12
Appendix A: Notations
A projection matrix
C capacitance
C0 associated coefficient in Hopfield neuron
E network energy
e error matrixg(s ) filter function
k number of iteration
L , l projection line
M number of integral measurement
N smoothness matrix
N number of image pixel
R0 associated resistance in Hopfield neuron
S sensitivity matrix
t time
Dt time step length
s projection distance from origin
u neuron internal statevariableuG gas superficial velocity
DV voltage difference
v, v neuron output vector, neuron output variable
w1,2,3 weight coefficient
X, X image matrix, gray level
x variable
y variable
Y, Y projection matrix, projectionvalue
z parameter defined inEq. (23)
Symbols
a penalty factor
a0 initial penalty factorb steepness gain factor
o permittivity
f electrical field distribution
G curve enclosing the detector electrode
g1, g2, g3 normalized coefficients
h coefficient inEq. (24)
P set of constraint decision defined inEq. (16)
u projection angle
t unit time
j coefficient inEq. (19)
z coefficient inEq. (24)
References
[1] M.S. Beck, Selection of sensing techniques, in: Process
Tomography*/Principles, Techniques and Applications, Butter-
worth-Heinemann, Oxford, UK, 1995, pp. 41/48.
[2] M.S. Beck, R.A. Williams, Process tomography: a European
innovation and its applications, Meas. Sci. Technol. 7 (1996) 215/
224.[3] S.B. Kumar, D. Moslemian, M.P. Dudukovic, Gas-holdup
measurements in bubble columns using computed tomography,
AIChE J. 43 (1997) 1414/1425.
[4] L.-S. Fan, G.Q. Yang, D.J. Lee, K. Tsuchiya, X. Luo, Some
aspects of high-pressure phenomena of bubbles in liquids and
liquid/solid suspensions, Chem. Eng. Sci. 54 (1999) 4681/4709.
[5] N. Reinecke, G. Petitsch, D. Schmitz, D. Mewes, Tomographic
measurement techniques*/visualization of multiphase flows,
Chem. Eng. Technol. 21 (1998) 7/18.
[6] W. Warsito, M. Ohkawa, N. Kawata, S. Uchida, Cross-sectional
distributions of gas and solid holdups in slurry bubble column
investigated by ultrasonic computed tomography, Chem. Eng.
Sci. 54 (1999) 4711/4728.
[7] L.A. Shepp, R.F. Logan, The Fourier reconstruction of a head
section, IEEE Trans. Nucl. Sci. NS-21 (1974) 21/43.
[8] R. Gordon, R. Bender, G.T. Herman, Algebraic reconstruction
techniques (ART) for three-dimensional electron microscopy and
X-ray photography, J. Theor. Biol. 29 (1970) 471/481.
[9] G.T. Herman, Image Reconstruction from Projections, Academic
Press, New York, USA, 1980.
[10] . Isaksen, A review of reconstruction techniques for capacitance
tomography, Meas. Sci. Technol. 7 (1996) 325/337.
[11] W. Warsito, L.S. Fan, Neural network based multi-criterion
optimization image reconstruction technique for imaging two-
and three-phase flow systems using electrical capacitance tomo-
graphy, Meas. Sci. Technol. 12 (2001) 2198/2210.
[12] C.G. Xie, N. Reinecke, M.S. Beck, D. Mewes, R.A. Williams,
Electrical tomography techniques for process engineering applica-
tions, Chem. Eng. J. 56 (1995) 127/133.
[13] C.G. Xie, S.M. Huang, B.S. Hoyle, R. Thorn, C. Lean, D.
Snowden, M.S. Beck, Electrical capacitance tomography for floe
imaging system model for development of image reconstruction
algorithms and design of primary sensor, IEE Proc. G 139 (1992)
89/98.
[14] S.M. Huang, A. Plaskowski, C.G. Xie, M.S. Beck, Tomographic
imaging of two-component flow using capacitance sensors, J.
Phys. E: Sci. Instrum. 22 (1989) 173/177.
[15] J. Hopfield, Neurons with graded response have collective
computation properties like those of two-state neurons, Proc.
Natl. Acad. Sci. 81 (1984) 3088/3092.
[16] J. Hopfield, D. Tank, Neural computation of decisions in
optimization problems, Biol. Cybern. 52 (1985) 141/152.
[17] W. Warsito, L.S. Fan, ECT imaging of three-phase fluidized bed
based on three-phase capacitance model, Chem. Eng. Sci. (2003),
in press.
W. Warsito, L.-S. Fan / Chemical Engineering and Processing 42 (2003) 663 /674674