Author's personal copy

Computers and Chemical Engineering 45 (2012) 2737

Contents lists available at SciVerse ScienceDirect

Computers and Chemical Engineering

journal homepage: www.elsevier.com/locate/compchemeng

Automated drop detection using image analysis for online particle size
monitoring in multiphase systems
Sebastian Maa a, , Jrgen Rojahn a,b , Ronny Hnsch b , Matthias Kraume a
a
b

Technische Universitt Berlin, Strae des 17. Juni 135, Sekr. MA 5-7, Chair of Chemical and Process Engineering, 10623 Berlin, Germany
Technische Universitt Berlin, Franklinstrae 28/29, Department of Computer Vision and Remote Sensing, 10587 Berlin, Germany

a r t i c l e

i n f o

Article history:
Received 2 November 2011
Received in revised form 24 March 2012
Accepted 24 May 2012
Available online xxx
Keywords:
Particle size distribution
Image analysis
Online monitoring
Sauter mean diameter
Dispersion
Automatic particle recognition

a b s t r a c t
Image analysis has become a powerful tool for the work with particulate systems, occurring in chemical
engineering. A major challenge is still the excessive manual work load which comes with such applications. Additionally manual quantication also generates bias by different observers, as shown in this
study. Therefore a full automation of those systems is desirable. A MATLAB based image recognition
algorithm has been implemented to automatically count and measure particles in multiphase systems.
A given image series is pre-ltered to minimize misleading information. The subsequent particle recognition consists of three steps: pattern recognition by correlating the pre-ltered images with search
patterns, pre-selection of plausible drops and the classication of these plausible drops by examining
corresponding edges individually. The software employs a normalized cross correlation procedure algorithm. The program has reached hit rates of 95% with an error quotient under 1% and a detection rate of
250 particles per minute depending on the system.
2012 Elsevier Ltd. All rights reserved.

1. Introduction
The competitive pressure in the chemical industry makes it necessary to take measures that enable processes to be drastically
improved in order to remain competitive also in the future (Ruscitti
et al., 2008). Product quality control is more complex in particulate than in conventional chemical processes. The key properties of
the product are often related to the particle size distribution (PSD)
which is inuenced by the operating conditions and the history
of the process (Zeaiter, Romagnoli, & Gomes, 2006). Disturbances
in operating conditions may irreversibly change the quality of the
product. Quantitative real-time measuring is needed to enable
feedback control. Monitoring and control of such processes have
evoked interest in the use of image-based approaches to estimate product quality in real time and in situ (Zhou, Srinivasan, &
Lakshminarayanan, 2009).
During the last decades extensive research has been performed
to establish and improve technologies which measure particle
properties, such as size, shape, composition, and velocity. Concerning the interpretation of particle size distributions using different

Corresponding author. Tel.: +49 30 314 78609.

E-mail addresses: sebastian.maass@tu-berlin.de, sebastian.maass@sopatec.com
(S. Maa).
0098-1354/$ see front matter 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.compchemeng.2012.05.014

physical principles there is still a considerable lack of understanding (Leschonski, 1986).

Various authors found unsatisfying results, analyzing spherical
drops in different liquid/liquid systems, using laser optical measurement techniques based on back scattering (Boxall, Koh, Sloan,
Sum, & Wu, 2010; Greaves et al., 2008; Honkanen, Eloranta, &
Saarenrinne, 2010; Maa, Wollny, Voigt, & Kraume, 2011). These
authors are questioning the reliability of these online probes in general and reafrming the need to use image analysis (IA) instead as
the particle surface unpredictably inuences the signals.
A further limitation, according to different authors (MartnezBazn, Montans, & Lasheras, 1999; Niknafs, Spyropoulos, &
Norton, 2011; Pacek, Moore, Nienow, & Calabrese, 1994), is the use
of external physical sampling. This never can guarantee that the
particle size does not change during measuring. Even for sampling
times less than a second, signicant measurement errors can occur.
In order to get reliable drop size distribution (DSD) measurements
the technique needs to be chosen carefully. This work is focused
on a MATLAB based image recognition algorithm, which is able
to automatically measure particles robust in different multiphase
systems.
The paper is structured as follows. The use of image analysis for
sizing uid particles is shortly reviewed in Section 2 followed by an
introduction of the used experimental set-ups in Section 3. Image
pre-processing and image analysis size measurements are given in
Sections 4 and 5. The results achieved by that method are compared
with manual results in Section 6.

Author's personal copy

28

S. Maa et al. / Computers and Chemical Engineering 45 (2012) 2737

Nomenclature
Symbols
Bi
B i
bj
d32
D
d
H
K
n
n
nB
q0
Q
Q0
Rj
r
S
S
s
T
T
T u,v
V
X

individual original picture

processed image
average gray value
Sauter mean diameter (m)
stirrer diameter (m)
Euclidean distance (pixel)
liquid level height (m)
convolution core
number of particles (#)
refraction index ()
number of images in one batch ()
number density distribution (1/m)
self quotient image
cumulative number distribution ()
absolute radius of an individual particle (m)
control variable
cumulative average image from a whole sequence
average pixel value of S
stirrer distance (m)
pattern
vessel diameter (m)
average gray value from the according pattern
result for NCC
convoluted image
angle ( )
dispersed phase fraction ()

Abbreviations
CP
correlation procedure
NCC
normalized cross correlation

2. Image analysis in multiphase systems

Although process characterization based on image analysis (IA)
can be intensely time consuming, it needs to be applied to almost
every dry and wet particulate system. A short summary of published applications in liquid systems is given by Guevara-Lpez
et al. (2008) and for dry systems by Andrs, Rginault, Rochat,
Chaillot, and Pourcelot (1996). Emerging applications show that
utilization of image analysis can facilitate the creation of new
and sophisticated models for the control of particulate systems
(Williams & Jia, 2003). In this paper we only focus on the characterization of the size of particles based on IA.
The most extensive available review on this specic eld is
given by Junker (2006). A detailed description of the technical
and historical developments can be found there. This review also
shows that the photo-optical in situ measurement of particle sizes
in multiphase systems is already well established. Many applications have been reported in literature with different set-ups (Aakre,
Solbakken, & Schller, 2005; Alban, Sajjadi, & Yianneskis, 2004;
Fantini, Tognotti, & Tonazzini, 1990; Galindo et al., 2005; Hossain,
Yang, Borgna, & Lau, 2011; Junker, Maciejak, Darnell, Lester, &
Pollack, 2007; Kamel, Akashah, Leeri, & Fahim, 1987; Khalil et al.,
2010; Mickler, Didas, Jaradat, Attarakih, & Bart, 2011; ORourke
& MacLoughlin, 2005; Roitberg, Shemer, & Barnea, 2006; Torabi,
Sayad, & Balke, 2005), all of which worked well for the applications
investigated. They are based on digital, high-speed, high resolution
modular camera systems and the images are analyzed with commercial or self developed image analysis software and standard
statistical methods.

Junker (2006) gives an organized overview of the applied

photographic techniques used in literature. They stated that CCDcameras are the optimum for the effort/cost ratio. Fig. 1 shows an
example image gallery from such a standard camera. Pictures are
taken at a maximum frequency of 50 frames per second with this
technique which is equal to a recording time, also called data acquisition time (DAT) of 1/50 s per image. Already Leschonski (1986)
and also Junker (2006) emphasize the necessity of short DAT and
additionally a short measurement acquisition time (MAT). The MAT
is the DAT plus the time needed to extract the necessary information from a sample image and translate this information into the
particle size distribution. Ideally the MAT equals the DAT (Crawley
& Malcolmson, 2004).
To get statistically reliable data sets 100s or 1000s of particles
have to be measured. Particle systems are very different and so are
the number of particles on one image (see Fig. 1) and therewith
the number of frames required for a single measurement. A minimum of 250 drops need to be captured within some seconds to
achieve real time DAT. Junker (2006) reports in her studies a variation between 2 and 400 frames. The required number of objects per
measurement becomes important for storage reasons, if all images
need to be reviewed and therefore saved. This problem should
become less and less signicant with the ongoing developments
of computer hardware.
Usually for drop sizing applications the MAT (560 min) is much
greater than the DAT (Junker, 2006). These time ranges are unsuitable for process monitoring and control. The manual evaluation of
such images is highly time-consuming and therefore automation
of image analysis should be employed to speed up the MAT at least
one to two magnitudes. Another disadvantage of manual evaluation
was rstly reported by Gwyn, Crosby, and Marshall (1965). Due to
the subjective nature of manual particle counting the measured distributions are human biased. The signicant statistical variations
mostly occur especially at the tails of the distributions. Boxall
et al. (2010) also showed the inuence of human bias with an average difference of 5.1% in the average drop size for two different
analyzers. Automated quantication would avoid bias by different
observers.
Simplied image analysis for the discussed multiphase systems
can be achieved by using only low dispersed phase fractions in
which no overlapping occurs. Several examples for this approach
are reported in literature (Khalil et al., 2010; Mickler et al., 2011;
Scherze, Knofel, & Muschiolik, 2005). These successful implementations of image analysis algorithms all fail for highly concentrated
(phase fraction larger than 1015%) dispersions. Additionally no
commercial software for the analysis of such systems, which would
be needed for industrial relevant applications is currently available
(Brs, Gomes, Ribeiro, & Guimares, 2009). More promising are the
works of Alban et al. (2004) and Brs et al. (2009).
The image analysis technique employed by Alban et al. (2004)
includes several steps of arc and circle center detection and pattern
matching. The number of particles detected in the image depends
on the image quality. Detection levels vary from 10 to 90%. Large
particles are accurately detected even when they are obstructed
by smaller particles. Errors in detection occur especially for very
small particle sizes when thick particle peripheries are present.
This does 
not signicantly
affect the used Sauter mean diame 2
di3 /
di ), due to emphasis of large particles in this
ter (d32 =
calculation. Brs et al. (2009) are using a two step approach that
automatically identies the contour of existing drops and classies
them according to their diameter. In the rst step, they detect the
edges of the drops in the original image by monitoring the values
of the image gray values as well as the descending thickness and
by creating an output image with those contours. In the second
phase, they detect the drops in this contour image, using the Hough
transformation (Cohen & Toussaint, 1977) to evaluate the circles.

Author's personal copy

S. Maa et al. / Computers and Chemical Engineering 45 (2012) 2737

29

Fig. 1. Example image gallery: representing samples taken with the endoscope technique using different lenses according to the expected particle size for different systems.

Although both works use automated drop size analysis, both are
not capable for process monitoring or control as both have too large
MAT. One goal of this work was to decrease the MAT signicantly
to achieve online monitoring of the particle size distribution in a
multiphase system.

have one advantage. The drops are so dense packed, that all visible particles are at the vicinity of the focal plane of camera lens.
Therefore no correction for possible out of focus depths (Kashdan,
Shrimpton, & Whybrew, 2003) are necessary.
3.2. Measurement procedure

3. Materials and methods

The objective of this investigation was to develop an image
analysis algorithm, which allows precise automatic particle size
measurement in real time. The accuracy and the online capability of the particle size monitoring were tested in various stirred
applications.
3.1. Experimental set-up
Fig. 2 shows the main features of the used experimental set-up.
The images are taken intrusively from inside the vessel by placing
an endoscope in front of a CCD camera as a microscope lens (see
Fig. 2). A strobe ash is guided by a ber optic cable surrounding the
endoscope to ensure sharp pictures even in vicinity of the stirrer
(Maa, Wollny, et al., 2011, see also www.sopatec.com).
A broad evaluation is necessary to ensure the reliability of the
image algorithm software. Often the software is developed and
optimized only according to the application in its founding working group. The major challenge is to develop a software program
which can be adapted to many differing multiphase systems. Therefore, four different dispersed phases were used for the validation
of the developed software. Deionised water always served as the
continuous phase. An overview of the work program is given in
Table 1. The dispersed phase fraction was varied from 2 to 45%
to analyze the inuence of the concentration, which inuences the
optical properties dramatically. Dilute systems like the one shown
in Fig. 1(a) have almost no particle overlapping and are therefore
easier to analyze. However, the more concentrated systems ( > 5%)

Every image batch according to the individual system was analyzed twice; the rst counting was always carried out manually
to build a base for evaluation of the second counting, which was
carried out automatically with image analysis software. The time
needed for one data point is around 30 min. The inuence of human
bias on drop size measurements was tested rst. The deviations
between four different observers with each other and also with
themselves through a repeated second counting of the Sauter mean
diameter d32 are shown in Fig. 3. The results of the d32 are presented on the left ordinate and the number of counted particles n
on the right ordinate. All participants counted the same frame set
twice. An example image from the chosen analyzed frame set is
given in Fig. 1(a). The image quality is very high due to the high
refraction index of the used system (see Table 1) and the low dispersed phase fraction ( = 2%). That is why almost no overlapping
effects occur. Even with those simple images the four different
observers have a deviation of 5% for the Sauter mean diameter
and of 15% for the number of particles found in the whole set
(presented as the numbers in the graph). Surprisingly these deviations are increased by repeating the same counting for both the d32
(10%) and the number of identied particles (30%). The standard
deviation of one observer quantifying manually the same frame set
is around 5%. These deviations are called human bias in this work
as they show a distortion which is not systematic but inuenced
by the observing human and his experience and mental state. The
newly developed image analysis algorithm will overcome these
weaknesses of human bias. All distortion will be systematic and
therewith interpretable or even avoidable.

Author's personal copy

30

S. Maa et al. / Computers and Chemical Engineering 45 (2012) 2737

Fig. 2. Experimental set-up and dimensions of the used stirred tank.

3.3. Computational specications

Sauter mean diameter d32 [m]

500

The calculations were carried out on an Intel Quad-Core i7 with

2.6GHz and 12GB RAM. The algorithms were implemented in Matlab 7.9 (R2009b). On this PC running Windows Vista, the developed
software was able to reach a processing speed of 190 ms/frame
for simple images. 50150 images were taken for every investigated point in time and analyzed as complete batches. Following
the guidelines given by Yan, Sayad, and Balke (2009a), Section
4 describes the image operations carried out for noise removal,
sharpening and contrast enhancement.

475

450

n = 205
n = 177

n = 247

n = 223

425

n = 267
n = 498

n = 364
400

n = 433
375

4. Image processing

first counting
second counting

350
one

two

three

four

manual counting results by individual observer

Fig. 3. Analysis of the variance of the Sauter mean diameter and the number of
observed drops by human bias in a single image batch at constant process parameters.

The image processing and detection steps are described in Sections 4 and 5. For all following comparisons between the automated
and the manual counted drop size distribution results, the observer
(four) with the high reproducibility is used.

The quality and variability in quality of the images obtained by

the camera computer system greatly inuences the success of the
automated interpretation (Yan, Sayad, & Balke, 2009b). The rst
step of the used method essentially consists of enhancing image
information to facilitate its analysis which is the second step and is
described in Section 5.
Images were at rst processed by modifying their intensity distribution to enhance certain features of the image. The processing
technique involves several steps of contrast enhancement. In order
to ensure best possible drop detection, the given image series are
pre-ltered to remove irrelevant and misleading image information. The redundant information, like lighting patterns and dirt or

Table 1
Overview of the working program, testing the image algorithm software.
Dispersed phase

Refraction index n ()

()

T (mm)

H/T

D/T

Origin of image data

Toluene
Petroleum
n-Butyl chloride
Glass beds

1.496
1.428
1.402
1.485

0.02, 0.10
0.10
0.45
0.10

150
150
155
200

1.0
1.0
2.3, 5.0
1.0

0.33
0.33
0.6
0.31

This study and Maa, Wollny, et al. (2011)

Maa and Kraume (2012)
This study and Maa et al. (2010)
Angst and Kraume (2006)

Author's personal copy

S. Maa et al. / Computers and Chemical Engineering 45 (2012) 2737

particles on the lens only given by the optic system, can be removed
by subtracting it from every image. Therefore the cumulative pixel
by pixel average image is formed out of the whole batch of nB
images Bi (see Eq. (1), with nB = 50 for this example):
1 
S=
Bi
nB

31

simplication treated with polar coordinates (r,):

bj (r) =

1 
Bj (r, ),

2r

1
= [0 , 360 [ ; r = 1, 2, . . . , 1 Rj
2

(5)

nB

(1)

i=1

The corrected image Bi is therewith the subtraction of the original image Bi (see Fig. 4(a)) with the arithmetic mean image S of the
whole sequence (see Fig. 4(c)), which is an approximation of the
redundant information.
Bi = Bi S + S,

Sthe
average intensity of S

(2)

The result of Eq. (2) is presented in Fig. 4(d).

Every pixel at position (x,y) in the pixel matrix of the ltered
image Bi is multiplied with the quotient of 255 and the maximum
gray value of this matrix to use the whole spectrum of a gray value
image. This standard amplication is shown in Eq. (3) and Fig. 5(b).
B i (x, y) =

 255

max(Bi )

Bi (x, y)

B i (x, y)
X(x, y)

with X = B K

Fig. 7 displays the results of bj for all example drops from Fig. 6.
For comparison the distribution of bj (rj Rj ) over the radius for different drop sizes, r is always normed to the absolute radius Rj of the
corresponding toluene drops. Therefore the drop border is always
at r/Rj = 1. The distribution are qualitatively similar in the region of
r/Rj = 1 but different for the rest of r. The absolute values for bj at a
constant r for different drop sizes also vary with no certain rule.
The qualitative similarity at the boundaries of the different
drops makes it possible to compute a pattern with an ideal distribution of bj over r. Such a distribution is shown in Fig. 7 for the
example drop pattern T(r,) with a radius R = 102 pixel. Such patterns are then generated automatically for every possible drop size
in a user dened diameter range and are used for correlation with
the gray value distributions of all processed images.

(3)
5.2. Pattern matching

The robustness of automatic detection of structures in noisy pictures can be increased by the self quotient image method (SQI)
(Gopalan & Jacobs, 2010; Shashua & Riklin-Raviv, 2001). These
operation norms the intensity of every local pixel based on the
local environment. This is carried out by a division of the processed
image B i pixel values by those of a smoothed version X of itself (see
Fig. 5(c)):
Q (x, y) =

(4)

where K is a 2-dimensional Gaussian kernel and * the convolution

operator.
The quantication of the inuence of every neighbor particle at
every coordinate is related to the convolution kernel K. The image Q
(Fig. 5(d)) is created by emphasizing the changes of the intensities.
This image is now not dependent on any illumination variation (see
Fig. 5) and processed to a stage where it can be analyzed.

Although the description of the signature curves bj in polar coordinates is more intuitive, the computation is carried out in Cartesian
coordinates as images are organized in discrete elements following
a Cartesian order. Therefore the following procedures are explained
as they are implemented.
There are different image algorithms for structure detection
available. In this project, two different principles are used for the
analysis: a voting and a correlation procedure (CP). A frequently
used CP is the normalized cross correlation (NCC). It is dened by
Lewis (1995) and shown in Eq. (6). The pattern T(x,y) denes a
sought-after example structure in the image B(x,y), in our case a
gray value matrix which includes a particle boundary with a certain radius r. The distribution of bj for a pattern T(x,y) of the radius
r = 102 pixel is given in Fig. 7. It is compared with the distribution
of the example drops and also given in direct comparison with the
correlating image segment B(u,v) for a radius r = 102 pixel. Those
are correlated with the NCC:

5. Image analysis
V (u, v) =
The particle recognition, presented in this section, essentially
consists of three steps: The pattern recognition by correlation of
pre-ltered images with search patterns, the pre-selection of plausible circle coordinates and the classication of each of these by an
exact edge examination. These individual steps include the compilation of the search patterns for correlation.
5.1. Pattern compilation
A necessary condition for coherent results of the algorithm is a
precise knowledge of the expected gray values to design meaningful search patterns. From example drops, which the user gives to the
program, these drop patterns are automatically computed by the
software for every representative optical system. It is necessary to
provide example drops over the whole spectrum of expected sizes,
since the drop appearance cannot be assumed to be equal or linear
scalable over the radius r. Such example drops are given in Fig. 6.
After the denition of example particles, the intensity signature curves of these examples are computed. The signature curve
is the average value bj of the gray values in the image Bj of a chosen object j with the radius Rj . It is computed over a circumference
within the image coordinates B(x,y) at every radius r which are for


x,y

x,y

[B(x, y) Bu,v ][T (x u, y v) T u,v ]

[B(x, y) Bu,v ]


x,y

[T (x u, y v) T u,v ]

(6)
2

Here Bu,v is the average of the gray values below the sample pattern
T and T u,v the average value from the sample pattern. The result
V(u,v) is a similarity measure for the wanted sample pattern T of
each local picture segment Bu,v , which is centered around (u,v). The
solution range is limited to 1 to 1. Maxima suggest an agreement
with the sample pattern, negative values an opposed match.
The search for circular patterns for a range of possible radii is
the most time consuming and critical process of the detection. The
results of V(u,v) for sample pattern T with the radius 20 pixel is
given in Fig. 8(a). The intensities behave proportional to the center
probabilities, white being the lowest and the strongest tone (black)
being the highest probability.
The best possible center points (these are local probability maxima) and radii are then checked in detail to nally be classied. This
is done by decomposing the possible drop circle into segments and
checking these for local validity of being part of a real drop. Note
that thanks to the contrast enhancement, borders can be evaluated with image illumination invariance. The percentage of features
needed to pass as circles are very much dependent on the optical
characteristics of the system. The results for the detection after the

Author's personal copy

32

S. Maa et al. / Computers and Chemical Engineering 45 (2012) 2737

Fig. 4. Image processing steps to remove redundant information during pre-ltering: (a) and (b) example images from a whole set, (c) integrated average image from this
set, (d) subtracted image (a) minus (c). Used system is n-butyl chloride/water; dispersed phase fraction = 45%, d32
= 80 m.

Fig. 5. Image process steps under SQI: (a) one example original image, (b) image after image processing steps illustrated in Fig. 4 plus a contrast amplication, (c) convolution
of image (b) with a 2 dimensional Gaussian kernel, (d) result of the self quotient (b)/(c).

Author's personal copy

S. Maa et al. / Computers and Chemical Engineering 45 (2012) 2737

33

Fig. 6. Example of different toluene drops over the whole size spectrum.

classication are given in Fig. 8(b) and compared with the processed
original image Fig. 8(c).

to facilitate online monitoring. The overall time for acquisition,

processing and analysis of the whole batch of images is below 180 s.

6. Results and discussion

6.2. Analysis of liquidliquid systems

6.1. Analysis of a solidliquid system

6.2.1. Dilute liquidliquid system

The already introduced simple system with a low dispersed
phase fraction was used (toluene/water; = 2%; see Fig. 1(a) as an
example image).
After optimizing the parameter set of the algorithm, quite
promising results could be achieved. The cumulative number distributions Q0 (dP ) shown in Fig. 10 are almost identical for the manual
(man.) and the automatic (autom.) detection. They are presented by
the dotted curves over the particle diameter for the right ordinate.
For a better differentiation the number density distribution
q0 (dP ) is plotted on the left ordinate, presented as straight lines. The
broadness of the distribution is slightly different and this becomes

Solid spherical glass particles were analyzed by the presented

technique and the spheres were detected by the image processing and analysis techniques. The number of particles detected in
one image was roughly 10 (see Fig. 1(b) as an example picture).
Fig. 9 shows the comparative results between estimated distributions based on manual and automated detection and the number
of detected particles n in the whole frame set of 82 images. Both
show almost the same results. The only differences are observed
for the larger spheres which lead to a deviation of the Sauter mean
diameters of around 20%. The processing speed was maximized

Fig. 7. Signature curves bj of all example drops from Fig. 6. Additionally the pattern T(r,) for R = 102 pixel based on an interpolation of several drop examples is given in
comparison with the example drop image B(r,).

Author's personal copy

34

S. Maa et al. / Computers and Chemical Engineering 45 (2012) 2737

Fig. 8. Original image (a) and the self quotient image for comparison (b); the results for the normalized cross correlation with a 20 pixel radius pattern (c) and results of the
drop detection for all radii (d), the resulting from T(r = 20) are highlighted by a dotted line.

obvious by comparing the number density distribution of the automatic and the manual detection. The automatic counting shows a
slightly more narrow distribution. That concurs with a more than
doubled number of detected particles for the automatic detection
(n = 553) in comparison to the average value of the manual counted
(n = 301). Those deviations lead to a difference of 2.7% in the Sauter
mean diameter.
6.2.2. High concentrated liquidliquid systems
The size of drops in agitated vessels is the result of two opposing
phenomena: drop breakage and coalescence. When agitation starts,
the bulk of the dispersed phase is pulled down into the continuous
phase and is broken into small drops. The rate of drop break-up
in this stage is much larger than the rate of drop coalescence. As

a result the size of drops shows an exponential decrease during

the transient stage (see the example for petroleum/water or nbutyl chloride/water in Fig. 11). The rate of drop coalescence rises
with the increasing number of drops with time during the transient stage. Eventually a steady state is reached where the rates of
drop break-up and coalescence are balanced. To reect this transient behavior, 200 pictures were taken at several points in time
over 60 or 80 min of stirring. The algorithm is able to count more
than 250 drops on the digital images within 2 min at each time of
measurement. For every different optical system, parameters for
the pattern have to be adapted. To analyze the images effectively a
single parameter optimization varying from 10 min (petroleum) to
4 h (n-butyl chloride) was carried out. The algorithm also delivers
the opportunity of suggesting those parameters based on 1050

2.0

80

672

1.4

0.90

1.2

0.75

1.0

0.60

0.8

0.45

0.6

0.30

0.4

0.15

0.2

0.00
100

200

300
particle diameter d

400
p

0.0
500

[m]

Fig. 9. Comparison of manual and automatic detected particle size distributions for
bimodal glass beds from a frame set with 82 images.

number density distribution q0 [1/mm]

1.05

640

cumulative number distribution Q0 [-]

number density distribution q0 [1/mm]

1.6

Q0 [-]
n [#]

Q0 [-]

1.8

q0 [1/mm]

1.20

1.0

man. autom.

man. autom.
1.35

60

q0 [1/mm]
d32 [m]

425

436

n [#]

301

553

0.8

40

0.6

20

0.4

0.2

10

100

cumulative number distribution Q0 [-]

1.50

0.0
1000

particle diameter dP [m]

Fig. 10. Comparison of manual and automatic detected drop size distributions of a
dilute toluene/water system with a dispersed phase fraction = 2%.

Author's personal copy

S. Maa et al. / Computers and Chemical Engineering 45 (2012) 2737

35

Fig. 11. Transient behavior of petroleum and n-butyl chloride drops after 1 and 60 min of stirring.

manually counted example particles. This automatic parameter

generation was used for the toluene/water system.
In Fig. 12 the results for transient Sauter mean diameter are
shown, for both the manual and the automatic counting of drops,
for all three systems. It shows clearly that, despite numerous overlapping of drops in the images obviously occurring in such highly
concentrated dispersion, the automatic algorithm works very well.
It obtains almost the same values for the transient Sauter mean
diameter that the manual observation for the three different systems attained.
The results for the toluene/water system are outstanding, taken
into account that only once a little number of example drops had
to be marked and all further data points were computed fully
automated. The measured values of the algorithm correspond to
the expected development of the Sauter mean diameter whereas

700

the manual counting shows unexpected uctuations. This can be

explained by the fact that the absolute number of particles was
higher with the automated analysis and therewith the condence
level of the data set.
Example pictures of the petroleum/water system (see Fig. 11)
show, how difcult it is even for humans to identify the drops.
The drops both reect strongly and are low transparent at the
same time. Thus chaotic and ambiguous shade effects occur. In
each picture there are more than 50 drops (see Fig. 11(b)), but only
approximately 5% of them can be clearly analyzed by humans. However, the results of the manual and automatic counting are in good
agreements. Notice that the automatic counting only needed 15%
of the manual counting time. The best results were achieved for
the n-butyl chloride/water system. The results show lowest deviation between the two methods (man. and autom.). The steady
state values for n-butyl chloride/water after 60 min of mixing have
a deviation of 4% (d32;man. = 175.8 m; d32;autom. = 182.9 m) for
petroleum/water 6.8% after 80 min stirring and for toluene/water
4.6%, also after 80 min of stirring.

Sauter mean diameter d32 [m]

600

6.3. Drop size monitoring

500
400
300
200

[-] man. autom.

n-butyl chloride/water 0.45

100

petroleum/water 0.10
toluene/water 0.10

0
0

10

20

30

40

50

60

70

80

90

time t [min]
Fig. 12. Comparison of manual and automatic counted Sauter mean diameter for
three different chemical systems.

Finally a drop size monitoring over 120 min with the n-butyl
chloride/water system (d = 45%) was carried out. The results of this
mixing experiment are shown in Fig. 13. Two endoscope probes
were used in parallel. The data from endoscope A are analyzed
manually while the data from endoscope B are determined with
the presented image algorithms. No additional parameter adaptation was carried for the algorithms as the values from the former
n-butyl chloride experiment were used. Two stirrer speeds have
been investigated. The rst period of mixing over 60 min was carried out under a stirrer speed of 350 rpm while the second 60 min
were agitated with a lower stirrer speed of 250 rpm.
With the use of this set-up, the spatial distribution of the drop
sizes, the inuence of the stirrer speed and the quality of the automated image algorithm for drop size monitoring was evaluated.

Author's personal copy

S. Maa et al. / Computers and Chemical Engineering 45 (2012) 2737
200

pH 7, = 45%
blade impeller

Sauter mean diameter d32 [m]

180
160

H/T = 5.0
D/T = 0.6

n-butyl chloride/water
A

140
120
100
80

sampling
points

60

40

350 rpm

250 rpm

20

endoscope A manual
endoscope B autom.

0
0

20

40

60

80

100

120

140

time t [min]
Fig. 13. Comparison of manual counting and automatic monitoring of the transient
Sauter mean diameter for n-butyl chloride/water system in a slim reactor with an
aspect ratio of 5.0. The automated monitoring was located at the bottom of the vessel
while the data for the manual counting were sampled at the top.

The discussion of the results will be subdivided as the two stirrer

speed regimes. The mixing during the rst 60 min showed clearly
two opposed dispersions. The expected oil in water (o/w) emulsion was observed at the bottom of the reactor but a water in oil
(w/o) emulsion occurred at the top. Due to the density differences of
the two used organics, the oil phase became the continuous phase
at the top, while the water became the continuous phase at the
bottom. However, the stirring over time was able to distribute the
dispersed phase homogeneously throughout the reactor. That led
to a change in the emulsion characteristics which were monitored
at sampling point A. The w/o dispersion turned into a water in oil in
water (w/o/w) double emulsion. That is reected by an increase in
the Sauter mean diameter, starting at minute 5 for the next 20 min
in Fig. 13 for the upper sampling point. Only the outer drop sizes
of the o/w/o drops were counted for the determination of the d32 .
With increasing stirring time, the double emulsion broke up and
became the expected o/w dispersion.
As the oil phase was not homogenously distributed in the beginning of the process, low oil concentrations at the bottom can be
assumed. Therefore, the drop sizes are smaller than expected in
the beginning of the process at sampling point B. After a homogeneous distribution of the oil phase is reached (around the point in
time of 40 min), the drop sizes at the bottom of the reactor increase.
This seemingly surprising result can easily be explained with the
local increase in dispersed phase fraction. This reason for a drop size
increase even for coalescence hindered systems is in good agreement with other studies in our research group (Kraume, Gbler, &
Schulze, 2004; Maa, Paul, & Kraume, 2012) and also with studies
from other scientists (El-Hamouz, 2009; Khakpay & Abolghasemi,
2010; Razzaghi & Shahraki, 2010).
The results for the drop sizes after the stirrer speed change are
not controversy with other results. The decrease in the energy dissipation rates leads to a decrease of the breakage rate in the system
under a critical value. Therefore, no further drop size decrease was
observed. As the system was stabilized using 1 mg polyvinyl alcohol per 1 g n-butyl chloride, also no coalescence occurred. This is
in good agreements with recent studies (Maa, Rehm, & Kraume,
2011). The manual as well as the automated analysis reects this
behavior. The size deviation between both counting methods is
lower than 10%. The preciseness of the algorithm results is in the
bandwidth of the expectable measurement error range described
in literature (Maa, Wollny, et al., 2011; Ritter & Kraume, 2000).
Although the automated detection introduces a small deviation to

automatically determined Sauter mean diameter d32 [m]

36

800

R2=0.97
+ 10%

600

- 10%
400

200

200

400

600

800

manually determined Sauter mean diameter d32 [m]

Fig. 14. Parity comparison of all results obtained for different liquidliquid systems.

the manual determination, it allows a continuous drop size monitoring. With the help of the automated analysis the computer
generated drop size results for every minute throughout the process. 100s of drops have been measured for each data point. That
would not be possible with manual work.
All results obtained for the liquidliquid systems are summarized in a parity plot in Fig. 14. Almost all deviations between
manual and automated detection are lower then 10%.

7. Summary and outlook

Image analysis has become a powerful tool for the work with
particulate systems, occurring in chemical engineering. A major
challenge is still the manual work load which comes with such
applications. Additionally manual quantication also generates
variance by different observers, which is shown in this study. Therefore the full automation of those systems is desirable. In this work
we have introduced an image recognition tool to automatically
count and measure particles in multiphase systems. Up to now this
software is limited to spherical particles but it will be extended to
irregular shapes as well.
In each of the processes, parameter optimization is necessary to
save computing time and improve error quotient and hit rate. This is
achieved by marking a few drops manually and letting the software
extract their optical characteristics. These include the steepness
and regularity of the particle border as general optical behavior.
These values vary strongly from system to system but in one particular system solely depend on the drop radius.
The automatic drop recognition is most effective with high contrast backlight images with a high drop density. The high contrast
ensures a good hit rate and low error quotient while the density
ensures that the computation renders more drops recognized per
minute. Nevertheless the proposed software was proven to be very
robust and able to compete with the human eye in complex images.
The computation time is nearly proportional to the pixel-quantity
being processed and is depending on the images twice to fty times
faster than manual counting. The user is able to dene thresholds
for the error quotient and the hit rate in order to guide the program
into further parameter optimization.
The program has reached hit rates of 95% with an error quotient under 1% and a detection of 250 drops per minute for simple
images and hit rates of 40% with an error quotient of under 5% and
a detection of 10 drops per minute for some difcult images. A parallelization on different cores in a multi-core PC will decrease those

Author's personal copy

S. Maa et al. / Computers and Chemical Engineering 45 (2012) 2737

calculation times dramatically. Therefore online process observation and control is possible (visit also www.sopatec.com).
Acknowledgement
The support by Johannes Herden with the experimental work is
gratefully acknowledged.
References
Aakre, H., Solbakken, T., & Schller, R. B. (2005). An in-line NIR/endoscope technique
for observations in real hydrocarbon multiphase systems. Flow Measurement and
Instrumentation, 16(5), 289293.
Alban, F. B., Sajjadi, S., & Yianneskis, M. (2004). Dynamic tracking of fast liquidliquid
dispersion processes with a real-time in situ optical technique. Chemical Engineering Research and Design, 82(A8), 10541060.
Andrs, C., Rginault, P., Rochat, M. H., Chaillot, B., & Pourcelot, Y. (1996). Particle-size
distribution of a powder: Comparison of three analytical techniques. International Journal of Pharmaceutics, 144(2), 141146.
Angst, R., & Kraume, M. (2006). Experimental investigations of stirred solid/liquid
systems in three different scales: Particle distribution and power consumption.
Chemical Engineering Science, 61(9), 28642870.
Boxall, J. A., Koh, C. A., Sloan, E. D., Sum, A. K., & Wu, D. T. (2010). Measurement
and calibration of droplet size distributions in water-in-oil emulsions by particle video microscope and a focused beam reectance method. Industrial and
Engineering Chemistry Research, 49(3), 14121418.
Brs, L. M. R., Gomes, E. F., Ribeiro, M. M. M., & Guimares, M. M. L. (2009). Drop
distribution determination in a liquidliquid dispersion by image processing.
International Journal of Chemical Engineering, 6 (Article ID 746439)
Cohen, M., & Toussaint, G. T. (1977). On the detection of structures in noisy pictures.
Pattern Recognition, 9(2), 9598.
Crawley, G., & Malcolmson, A. (2004). Online particle sizing as a route to process
optimization. Chemical Engineering (New York), 111(9), 3741.
El-Hamouz, A. (2009). Drop size distribution in a standard twin-impeller batch mixer
at high dispersed-phase volume fraction. Chemical Engineering and Technology,
32(8), 12031210.
Fantini, E., Tognotti, L., & Tonazzini, A. (1990). Drop size distribution in sprays by
image-processing. Computers and Chemical Engineering, 14(11), 12011211.
Galindo, E., Larralde-Corona, C. P., Brito, T., Cordova-Aguilar, M. S., Taboada, B., VegaAlvarado, L., et al. (2005). Development of advanced image analysis techniques
for the in situ characterization of multiphase dispersions occurring in bioreactors. Journal of Biotechnology, 116(3), 261270.
Gopalan, R., & Jacobs, D. (2010). Comparing and combining lighting insensitive
approaches for face recognition. Computer Vision and Image Understanding,
114(1), 135145.
Greaves, D., Boxall, J., Mulligan, J., Montesi, A., Creek, J., Sloan, E. D., et al.
(2008). Measuring the particle size of a known distribution using the focused
beam reectance measurement technique. Chemical Engineering Science, 63(22),
54105419.
Guevara-Lpez, E., Sanjuan-Galindo, R., Crdova-Aguilar, M. S., Corkidi, G., Ascanio,
G., & Galindo, E. (2008). High-speed visualization of multiphase dispersions in a mixing tank. Chemical Engineering Research and Design, 86(12A),
13821387.
Gwyn, J. E., Crosby, E. J., & Marshall, W. R. (1965). Bias in particle-size analyses by count method. Industrial and Engineering Chemistry Fundamentals, 4(2),
204208.
Honkanen, M., Eloranta, H., & Saarenrinne, P. (2010). Digital imaging measurement
of dense multiphase ows in industrial processes. Flow Measurement and Instrumentation, 21(1), 2532.
Hossain, M. I., Yang, Y. H., Borgna, A., & Lau, R. (2011). Depth-of-eld model for size
measurement using a borescope imaging technique under high object concentrations. Industrial and Engineering Chemistry Research, 50(9), 58245830.
Junker, B. (2006). Measurement of bubble and pellet size distributions: Past and
current image analysis technology. Bioprocess and Biosystems Engineering, 29(3),
185206.
Junker, B., Maciejak, W., Darnell, B., Lester, M., & Pollack, M. (2007). Feasibility of
an in situ measurement device for bubble size and distribution. Bioprocess and
Biosystems Engineering, 30(5), 313326.
Kamel, A. H., Akashah, S. A., Leeri, F. A., & Fahim, M. A. (1987). Particle-size distribution in oilwater dispersions using image-processing. Computers and Chemical
Engineering, 11(4), 435439.
Kashdan, J. T., Shrimpton, J. S., & Whybrew, A. (2003). Two-phase ow characterization by automated digital image analysis. Part 1: Fundamental principles and
calibration of the technique. Particle and Particle Systems Characterization, 20(6),
387397.

37

Khakpay, A., & Abolghasemi, H. (2010). The effects of impeller speed and holdup on
mean drop size in a mixer settler with spiral-type impeller. Canadian Journal of
Chemical Engineering, 88(3), 329334.
Khalil, A., Puel, F., Chevalier, Y., Galvan, J. M., Rivoire, A., & Klein, J. P. (2010). Study
of droplet size distribution during an emulsication process using in situ video
probe coupled with an automatic image analysis. Chemical Engineering Journal,
165(3), 946957.
Kraume, M., Gbler, A., & Schulze, K. (2004). Inuence of physical properties on drop
size distributions of stirred liquidliquid dispersions. Chemical Engineering and
Technology, 27(3), 330334.
Leschonski, K. (1986). Particle characterization, present state and possible future
trends. Particle & Particle Systems Characterization, 3, 99103.
Lewis, J. P. (1995). Fast template matching. In Vision Interface 95 Quebec City, Canada,
(pp. 120123).
Maa, S., Metz, F., Rehm, T., & Kraume, M. (2010). Prediction of drop sizes for
liquidliquid systems in stirred slim reactors Part I: Single stage impellers.
Chemical Engineering Journal, 162(2), 792801.
Maa, S., Wollny, S., Voigt, A., & Kraume, M. (2011). Experimental comparison of
measurement techniques for drop size distributions in liquid/liquid dispersions.
Experiments in Fluids, 50(2), 259269.
Maa, S., Rehm, T., & Kraume, M. (2011). Prediction of drop sizes for liquidliquid systems in stirred slim reactors Part II: Multi stage impellers. Chemical Engineering
Journal, 168(2), 827838.
Maa, S., & Kraume, M. (2012). Determination of breakage rates using single drop
experiments. Chemical Engineering Science, 70, 146164.
Maa, S., Paul, N., & Kraume, M. (2012). Inuence of the dispersed phase
fraction on experimental and predicted drop size distributions in breakage dominated stirred liquid-liquid systems. Chemical Engineering Science, 76,
140153.
Martnez-Bazn, C., Montans, J. L., & Lasheras, J. C. (1999). On the breakup of an air
bubble injected into a fully developed turbulent ow. Part 1. Breakup frequency.
Journal of Fluid Mechanics, 401, 157182.
Mickler, M., Didas, S., Jaradat, M., Attarakih, M., & Bart, H.-J. (2011). Droplet swarm
analysis by image processing for simulation of extraction columns with population balances. Chemie Ingenieur Technik, 83(3), 227236 (in German).
Niknafs, N., Spyropoulos, F., & Norton, I. T. (2011). Development of a new reectance
technique to investigate the mechanism of emulsication. Journal of Food Engineering, 104(4), 603611.
ORourke, A. M., & MacLoughlin, P. F. (2005). A comparison of measurement techniques used in the analysis of evolving liquidliquid dispersions. Chemical
Engineering and Processing, 44(8), 885894.
Pacek, A. W., Moore, I. P. T., Nienow, A. W., & Calabrese, R. V. (1994). Video technique
for measuring dynamics of liquidliquid dispersion during phase inversion.
AIChE Journal, 40(12), 19401949.
Razzaghi, K., & Shahraki, F. (2010). On the effect of phase fraction on drop size distribution of liquidliquid dispersions in agitated vessels. Chemical Engineering
Research and Design, 88(7A), 803808.
Ritter, J., & Kraume, M. (2000). On-line measurement technique for drop size distributions in liquid/liquid systems at high dispersed phase fractions. Chemical
Engineering and Technology, 23(7), 579582.
Roitberg, E., Shemer, L., & Barnea, D. (2006). Application of a borescope to studies of
gasliquid ow in downward inclined pipes. International Journal of Multiphase
Flow, 32(4), 499516.
Ruscitti, O., Franke, R., Hahn, H., Babick, F., Richter, T., & Stintz, M. (2008). Application of particle measurement technology in process intensication. Chemical
Engineering and Technology, 31(2), 270277.
Scherze, I., Knofel, R., & Muschiolik, G. (2005). Automated image analysis as a control
tool for multiple emulsions. Food Hydrocolloids, 19(3), 617624.
Shashua, A., & Riklin-Raviv, T. (2001). The quotient image: Class-based re-rendering
and recognition with varying illuminations. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 23(2), 129139.
Torabi, K., Sayad, S., & Balke, S. T. (2005). On-line adaptive Bayesian classication
for in-line particle image monitoring in polymer lm manufacturing. Computers
and Chemical Engineering, 30(1), 1827.
Williams, R. A., & Jia, X. (2003). Tomographic imaging of particulate systems.
Advanced Powder Technology, 14(1), 116.
Yan, S., Sayad, S., & Balke, S. T. (2009a). Image quality in image classication: Design
and construction of an image quality database. Computers and Chemical Engineering, 33(2), 421428.
Yan, S., Sayad, S., & Balke, S. T. (2009b). Image quality in image classication:
Adaptive image quality modication with adaptive classication. Computers and
Chemical Engineering, 33(2), 429435.
Zeaiter, J., Romagnoli, J. A., & Gomes, V. G. (2006). Online control of molar mass
and particle-size distributions in emulsion polymerization. AIChE Journal, 52(5),
17701779.
Zhou, Y., Srinivasan, R., & Lakshminarayanan, S. (2009). Critical evaluation of image
processing approaches for real-time crystal size measurements. Computers and
Chemical Engineering, 33(5), 10221035.