You are on page 1of 23

Atomization and Sprays, 25 (9): 753–775 (2015)

IMAGE PROCESSING TECHNIQUES FOR


VELOCITY, INTERFACE COMPLEXITY, AND
DROPLET PRODUCTION MEASUREMENT IN
THE NEAR-NOZZLE REGION OF A DIESEL
SPRAY

K. Lounnaci,1 S. Idlahcen,1 D. Sedarsky,2 C. Rozé,1,∗


J. B. Blaisot,1 & F. X. Demoulin1
1
UMR 6614-CORIA, Normandie Université, CNRS, Université et INSA
de Rouen, BP12, Avenue de l’Université 76801 Saint-Étienne du
Rouvray Cedex, France
2
Applied Mechanics, Chalmers University of Technology, S-412 96,
Gothenburg, Sweden

Address all correspondence to C. Rozé E-mail: claude.roze@coria.fr

Original Manuscript Submitted: 03/10/2014; Final Draft Received: 09/27/2014

An ultrafast shadow imaging arrangement using a double-pulsed femtosecond laser system in con-
cert with a frame transfer CCD was used to record time-correlated image-pairs in the near-nozzle
region of a dissymmetric diesel jet issuing from a single-hole injector. A region-matching procedure
was applied to produce velocity maps of the spray. The time-correlated image data were binarized and
segmented to produce comparative data sets which isolate the jet core and the surrounding droplet
cloud. The velocity mapping process was applied to these segmented data sets, a variable characteriz-
ing the degree of atomization of the jet was defined by means of statistical analysis, and the curvature
scale space of the jet core edge was used to extract a measure of its complexity.

KEY WORDS: atomization, imaging technique, single-hole injector, spray char-


acteristics, high-pressure injection

1. INTRODUCTION

The compression-ignition engine cycle is attractive as one of the most robust, efficient,
and cost-effective means of chemical to mechanical energy conversion. However, the
liquid-fueled internal combustion technology currently in widespread use must be opti-
mized in order to remain viable in the face of increasingly stringent pollutant emission
regulations, environmental concerns, and consumer demand for increased fuel economy.

1044–5110/15/$35.00 © 2015 by Begell House, Inc. 753


754 Lounnaci et al.

Given that the injection and mixing of the fuel are major controlling steps of the combus-
tion process, liquid jet breakup and spray penetration in the combustion chamber have
increasingly been subject to special attention (Kohse-Höinghaus and Jeffries, 2002). In
nearly all current applications, the liquid fuel is supplied by a high-pressure injection
giving rise to a spray which can be characterized by some drop-size distribution and
specific morphology which strongly influences the evaporation and distribution of the
fuel. Since the development of the injection spray ultimately determines mixture frac-
tion in the engine, the control and understanding of fuel-injection parameters is a crucial
link to effective optimization of the combustion process.
The forms and distribution of liquid structures in the spray are the result of many
complex interdependent processes, e.g., competing hydrodynamic instabilities, fluctuat-
ing pressure conditions, turbulence interactions, and cavitation inside the injector (Du-
mouchel, 2008). The variety and interdependence of these effects complicate analysis of
spray phenomena and frustrate efforts to directly control spray formation. Thus, there is
a strong demand for high-quality spray measurements to validate numerical models of
breakup and improve the fundamental understanding of these processes.
The aim of the present work is to examine high-quality spatially and temporally re-
solved imaging data in the context of liquid structure velocities, jet composition, and the
deformation of the edges of the spray to reveal statistical trends of the jet behavior. The
sprays used in this work were produced by a single-hole, plain orifice injector assembly
dispersing fuel oil into ambient atmospheric conditions. The velocity data and shadow
images of the diesel spray used in the analysis were obtained by ultrafast imaging, which
provides high-resolution visualization of the spray edges and droplets, resolved within a
depth-of-field limited by the collection optics.
A variable characterizing the deformation of the jet was defined by means of statis-
tical analysis and applied to a nozzle which produces a dissymmetric jet and exhibits a
variety of contrasting breakup modes. These variations in jet morphology were used to
test the statistical analysis model, highlighting the effectiveness of the characterization.
A scale space analysis was also carried out in order to quantify the deformation and the
fluctuation of the interface.

2. EXPERIMENTAL SETUP

In-depth visual analysis of atomization in the near-field region of practical fuel-injection


sprays requires the use of high-resolution imaging techniques. Particular care must be
taken for the study of modern diesel injection applications, which generally require high
injection pressures (>100 MPa) and small nozzle orifice diameters (∼100 µm). These
pressures and orifice dimensions give rise to small-scale structure and high liquid ve-
locities (until 600 m/s). An ultrafast imaging system (Fig. 1) was developed to extract
high-resolution visual information from limited areas of the spray in the near-nozzle
region (Sedarsky et al., 2012, 2013).

Atomization and Sprays


Image Processing Techniques for Velocity 755

FIG. 1: Experimental setup (high-resolution ultrafast shadow imaging).

The light source used to image the spray was a double-pulsed ultrafast laser sys-
tem consisting of a pair of matched regenerative Ti-sapphire amplifiers (Coherent Libra)
seeded by a common oscillator (Coherent Vitesse). Here, the ∼120 fs pulses generated
by the mode-locked laser oscillator are stretched and separated into two beams, which
are subsequently amplified, compressed, and recombined to form a single beam con-
taining a matched pair of pulses. The specific oscillator pulses selected for amplification
determine the delay between the amplified output pulses, allowing adjustment of the
inter-pulse timing from 12.5 ns to 500 µs.
The adjustment step size is 12.5 ns and corresponds to the period of the oscillator.
For this work, the short duration of each output pulse is more than sufficient to freeze the
spray motion during the measurement such that time-resolved images were acquired. A
nanosecond laser could be used (typically less 10 ns), but a femtosecond laser will allow
complementary investigations by ballistic imaging (Purwar et al., 2014).
The laser system was carefully aligned to obtain similar profiles in each pulse-pair,
in terms of duration, amplitude, optical alignment, polarization, and spatial profile. The
beam diameter was ∼10 mm, with a low divergence, a center wavelength of 800 nm, and
energy per pulse of the order of 3.7 mJ.
The two-pulse beam was directed across a chamber housing the injector to illumi-
nate the spray. The chamber was ventilated and maintained at ambient conditions, with
optical access provided by thin fused-silica windows to prevent contamination of the
optical components by fuel.
The light transmitted by the spray was collected to form images on the face of a
double-image CCD (PCO, 2048 × 2048 pixels), and the imaging system was equipped
with a long distance microscope allowing ∼7× magnification of the spray images. The

Volume 25, Number 9, 2015


756 Lounnaci et al.

scale of the resulting spatial intensity data was 900 pixels/mm, and the resolution of the
complete imaging system was on the order of 7 µm.
The fluid used in the measurements was a calibration oil (Shell NormaFluid, ISO
4113) with properties similar to diesel fuel and precisely controlled viscosity, density,
and surface tension specifications (Table 1).
The injector used in this work was equipped with a single-hole test nozzle with
length to diameter ratio l/d = 5.66, and discharge coefficient, Cd = 0.84, which pro-
duces a distinctive dissymmetric flow, probably due to cavitation inside the injector.
Although it is not representative of commercial injectors, the small difference between
the two sides of the jet for the same injection conditions is used here to test the image
processing tools. The entire injection duration is 1 ms. The images for these tests were
recorded approximately 500 µs after the start of the injection corresponding to the max-
imum needle lift. Indeed this time corresponds to the stationary state and the jet is stable
along the vision field of the camera. Figure 2 shows a diagram of this test nozzle, which
was covered in more detail by Sedarsky et al. (2013). All images used for the present
study were recorded in the symmetric plane of the injector, which maximizes the differ-
ence between opposite sides of the jet. The injector was installed in a three-axis trans-
lation stage and fuel was supplied to the assembly through a common-rail accumulator

TABLE 1: Thermophysical properties of ISO 4113 oil


Density Viscosity Surface tension
821 kg·m−3 0.0032 kg·m−1 ·s−1 0.02547 N·m−1

FIG. 2: Schematic view of the diesel test injector nozzle [from Sedarsky et al. (2013)].

Atomization and Sprays


Image Processing Techniques for Velocity 757

delivering an injection pressure of 60 MPa. For all images, the vertical axis is adjusted
to coincide with the injector axis.

3. VELOCITY FROM CORRELATED IMAGE-PAIRS

The velocity measurements of liquid structures in a diesel jet are very complicated to
perform. The small scales characterizing this kind of flow related to the extreme velocity
of the phenomenon explains the difficulties. In addition, in the near field of the injector,
the flow is easily perturbable, which excludes the use of intrusive methods of measure-
ment. Although a number of optical techniques are commonly used for the study of fluid
motion on small scales and at high velocity, the most are not suitable for the study of
diesel injection that produces a flow optically dense. Indeed, single-point techniques,
such as laser Doppler velocimetry and other related methods (Bachalo, 1994) are based
on properties of spherical droplets which make them not usable in the near field of the
injector. Laser correlation velocimetry (LCV) is a local velocity measuring technique
which has proved applicable to flows with high optical thickness (Chaves et al., 2004).
However, the interpretation of the LCV information requires consideration of various
parameters (distance, time of flight, time window size, etc.) making the measurement
complicated. This technique proved to be quite effective in the near field of the injector,
but under some conditions of injection (Leick et al., 2004). The technique is also suit-
able for assessing the performance of the atomization but still impractical for a complete
characterization of the spray (Hespel et al., 2012). Particle image velocimetry (PIV), the
most used imaging technique for velocity measurement, requires that the region of in-
terest is seeded with particles (Raffel et al., 2007). This technique has been successfully
used to measure the velocity of gas flow around the spray, but it is difficult to use this
technique on the jet itself. However, the correlation methods applied in the PIV may be
adapted to calculate the speed from images of unseeded flows and then to use the light
scattering by the elements constituting the liquid spray (Tokumaru and Dimotakis, 1995;
Masataka, 2012). This approach, generally called image correlation velocimetry (ICV)
uses matching algorithms for velocity calculation, either by flow labeling with features
easy to follow (Krüger and Grünefeld, 1999) or by matching the motion of structures
constituting the liquid jet (Sedarsky et al., 2006), or predicatively morphing and validat-
ing the scalar field (Marks et al., 2010). High-quality results can be obtained with ICV
methods. However, it should validate the results of velocity obtained by the correlation,
because matching errors are more frequent than on a seeded flow (Fielding et al., 2001).
In this work we used a similar technique adapting to the hard conditions of the near-field
regions of a high-pressure jet.
A primary aim of the work presented here is to generate informative observations
from velocities of liquid structures in fuel-injection events.
A flat-field correction (Seibert et al., 1998) was applied to each image to remove
inhomogeneities, background artifacts resulting from the optical components, and

Volume 25, Number 9, 2015


758 Lounnaci et al.

irregularities in the source light spatial profile. Figure 3 shows original, background,
and processed image results illustrating this operation.
The calculations and analysis used to generate velocity data from the measurements
were implemented in an in-house image analysis code based on the OpenCV image pro-
cessing routines (Bradski and Kaehler, 2008). The code calculates velocity information
from spatially resolved regions of spray images by identifying and correlating image
features in successive time-resolved image-pairs. The delay between the pulse-pairs in
the femtosecond pulse train was adjusted so that the displacement of liquid structures
was apparent in the resulting image-pairs. Note that larger delays increase the perceived
distortion of the fluid features, making accurate correlation more difficult. Very small
delays, on the other hand, yield accurate correlations but result in smaller displacements
with intrinsically higher relative error. Figure 4 shows an example of data obtained with
an inter-frame time of 284 ns. In this case, only a small part of each image in the image-
pair is shown in order to give a detailed view of the small displacement of a fluid feature.
The velocity results generated for this work were obtained by region-matching select
image areas using normalized cross-correlation:

FIG. 3: Illustration of flat-field correction procedure: (a) original image, (b) flat-field
background, and (c) processed image.

FIG. 4: Example of an image region pair. The time delay between (a) and (b) is 284 ns.

Atomization and Sprays


Image Processing Techniques for Velocity 759

£ ¤ £ ¤
1 X S (x0 , y 0 ) − S̄ . T (x0 , y 0 ) − T̄
(x, y) = (1)
n−1 0 0 σS σT
x ,y
0 0
where S(x0 , y 0 ) corresponds to the value of the pixel located at (x y ) in a search field
selected from the second image and T (x0 , y 0 ) corresponds to the value of the same pixel
in a template image selected from the first image of a time-correlated image-pair. (xy)
are the coordinates of the center of the template. S̄ and T̄ are, respectively, the average
level of the search field and of the template, whereas σS and σT are the standard devia-
tions. The normalized cross-correlation approaches the value +1 when the two functions
are identical, –1 when the functions are reversed (inverted gray scale), and tends to zero
when there is little or no correspondence in spatial intensity. The velocity algorithm
searches translations (dx, dy) of the search field which maximize R(xy). The values
(dx, dy) which produce the strongest correlation are chosen as the best estimates of the
image displacement, indicating the motion of the imaged fluid structure.
This procedure is selectively applied over the spatial extent of the image-pair, yield-
ing a map of displacement vectors describing fluid motion. To increase the accuracy of
the matching procedure and eliminate systematic errors of the correlation approach, val-
idation criteria are imposed on the correlation results. For example, match results with
correlation values less than an appropriate threshold are discarded. Figure 5 shows an
image of validated displacement vectors.
In light of the limitations of the spatial information made available by 2D images
of the complex three-dimensional spray region, it can be difficult to draw broad conclu-
sions about spray formation from individual images. However, it is possible to compile
informative statistical velocity profiles by combining results from many images, aver-
aging, and mapping the velocity magnitudes across the spatial extent covered by the
image data. For the current work, image sets containing 200 time-resolved image-pairs
were recorded and the velocity vectors corresponding to each pair were calculated and
compiled to derive representative velocity profiles of the spray. The 2048 × 2048 pixel
images were binned into square cells of 20 × 20 pixels. Within each cell, average vector
magnitudes were computed and individual bins containing fewer than 20 vectors were
ignored in the calculation to be sure that computed averages are statistically significant.
It allowed also to eliminate fake vectors appearing far from the jet due to optical pollu-
tion or droplets carried by recirculation flow. This threshold has been adjusted so that it
does not affect the velocity profiles. An example of such a reconstructed field is shown
in Fig. 6.

4. SEGREGATING DISPERSED AND CONTINUOUS REGIONS

It is reasonable to expect liquid breakup and the formation of droplets to be related to


the complexity of the jet interface (Shavit and Chigier, 1995). To investigate this as-
sumption, the individual images of the data set were segmented into two classes, one

Volume 25, Number 9, 2015


760 Lounnaci et al.

FIG. 5: Spray image shown with displacement vectors estimated from the time-
correlated image-pair.

FIG. 6: Map of velocity field in m·s−1 .

Atomization and Sprays


Image Processing Techniques for Velocity 761

containing only droplets and one representing the intact liquid core which is still under-
going breakup.
To select the pixels for each class, it is most convenient to deal with binary images.
However, conversion of the gray-scale input data to binary images results in an undesir-
able loss of information. The severity of this effect on the spatial content of the image
data is related to the binarization process and its suitability for the form, range, and scale
of the input data. A preliminary study comparing threshold and pixel sorting algorithms
was conducted to find an optimal binarization process for the diesel spray used in this
work. Results from the comparison of a number of different algorithms (Otsu, 1979; Ri-
dler and Calvard, 1978; Kapur et al., 1985; Cheng and Chen, 1999; Li and Lee, 1993)
are shown in Fig. 7.
None of the standard global thresholding methods examined in the survey yielded
satisfactory results for separating the jet (high-contrast object) and droplets (low-contrast
objects) of the diesel spray images. We note, however, that the method of Li and Lee
(1993) [Fig. 7(e)] appears to adequately separate the structure of the main jet which ex-
hibits strong contrast with the background. This method groups the gray-scale pixels into
two classes based on the minimal cross-entropy. However, for low-contrast image struc-
tures, such as the drops, this threshold is insufficient. Given that the image is separated
in terms of a single threshold value, these objects which lie below the threshold will not

FIG. 7: Examples of global binarization methods: (a) Otsu (1979), (b) IsoData, (c) Ka-
pur et al. (1985), (d) Cheng and Chen (1999), (e) Li and Lee (1993), and (f) MinError.

Volume 25, Number 9, 2015


762 Lounnaci et al.

be detected by a global thresholding process. Another segmentation approach, based on


the “wavelet transform” (Yon, 2003) can detect changes in the slope of the gray level of
the droplets, and thus allows simultaneous detection of low-contrast image features as
illustrated by Fig. 8.
Adding the image data obtained by the wavelet transform with those obtained by
global thresholding (Li and Lee, 1993) results in a binary representation which closely
follows the content of the original image. An example image produced by this process
is shown in Fig. 9.
The velocity estimation for the binary image data is expected to closely resemble
the results from evaluation of the gray-scale images. Figure 10 shows a map of fluid
structure speed obtained from binary images for comparison with Fig. 6.
The discrepancy between the two approaches can be viewed by examining the ve-
locity profile along the horizontal axis. Figure 11 shows this profile at different distances
from the injector orifice. It is apparent from the results shown in Fig. 11 that there are
some differences in the velocities evaluated using binary as opposed to gray-scale im-
age data. However, the horizontal profiles exhibit similar shapes and follow the same
downstream evolution. Appropriate binarization of images allows effective separation of
the dispersed elements from the continuous flow, as illustrated in Fig. 12. This enables
accurate computation of separate velocity fields for the dispersed droplets and the con-
tinuous interface, respectively. The resulting maps from separated spatial data are given
in Fig. 13, where the jet core interface speed is shown in Fig. 13(a) and speed of the
droplet cloud is given in Fig. 13(b).

FIG. 8: Binary image produced by wavelet transform thresholding.

Atomization and Sprays


Image Processing Techniques for Velocity 763

FIG. 9: Binary image resulting from the combination of wavelet transform thresholding
and Li and Lee (1993) method.

FIG. 10: Map of velocity magnitude field in m·s−1 obtained from binarized images.

The evolution of the horizontally averaged velocity versus distance from the nozzle
inlet is plotted in Fig. 14 for each side of the flow. Given that the horizontal variation in

Volume 25, Number 9, 2015


764 Lounnaci et al.

FIG. 11: Velocity profile at 400 µm to the injector (a), 800 µm (b), 1200 µm (c), and
1600 µm (d).

FIG. 12: (a) Image of the flow, (b) main jet, and (c) dispersed liquid elements.

the measured velocity is small, this velocity is assumed to be directly related to the jet
core velocity. Note that with a suitable scaling, the two curves overlap.

Atomization and Sprays


Image Processing Techniques for Velocity 765

FIG. 13: Map of velocity magnitude field in m·s−1 , (a) of the interface and (b) of the
dispersed elements.

FIG. 14: Evolution of the interface velocity versus the distance to the injector.

The velocity profile of the droplet cloud versus horizontal distance at 2000 µm from
the injector is plotted in Fig. 15. Here, the velocities of the dispersed elements decrease
as distance from the jet core increases. Close to the core interface, the velocity appears
to trend asymptotically toward the interface velocity plotted in Fig. 14.
The dissymmetry of the flow heavily influences the velocities of the droplets as well
as the interface of the jet core. The following section extends the analysis of separated
spray components to characterize the behavior of the dispersed and continuous portions
of the spray, the droplet density, and the core interface.

Volume 25, Number 9, 2015


766 Lounnaci et al.

FIG. 15: Velocity profile of the droplets cloud versus the horizontal distance at 2000 µm
to the injector.

5. QUANTIFICATION OF THE ATOMIZATION


Using sets of separated spatial data, as shown in Fig. 12, three spatial classes can be
defined: No Liquid (NL), Liquid Core (LC), and Dispersed Phase (DP). In this context,
one can apply statistical analysis to an appropriate group of images to calculate the
probability, PDP , for each measured region (pixel) to be in state DP. Such analysis leads
to the probability field presented in Fig. 16. This result was obtained with 200 images.
Figure 16 reveals significantly different spatial distributions in the dispersed phase
between the left and right sides of the jet. The following discussion will address each
side of the spray individually. Figure 17 shows the evolution of the occurrence of the
dispersed phase versus radial direction for both sides of the spray, where the probability
has been integrated in the axial direction to include the full field of view.
Examining the aggregate behavior across the spray, it is apparent that the atomization
process develops more quickly on the right side of the jet.

6. ANALYSIS OF THE CORE JET INTERFACE COMPLEXITY


In order to quantify the deformation and movement of the jet interface, the external
contours of the jet core interface were extracted and a shape complexity analysis was
performed. A wide range of methods has been proposed in the literature (Yang et al.,
2008) to extract shape features of curves, in relation with pattern recognition. Among
them, curvature is one of the preferred approaches for shape analysis, and can be of
great interest in the atomization domain since interface curvature is strongly related to
surface tension effects.

Atomization and Sprays


Image Processing Techniques for Velocity 767

FIG. 16: Evolution of the probability of atomization versus the distance to the injector.

FIG. 17: Probability PDP to have a dispersed phase.

The apparent curvature depends on the scale at which the imaged curve is observed.
Liquid surface ripples yield a spectrum of high curvatures whereas large deviations lead
to small curvature values. The local distribution of curvatures is therefore an indication
of the complexity of the interface. A typical main jet contour is shown in Fig. 18.

Volume 25, Number 9, 2015


768 Lounnaci et al.

FIG. 18: Example of contours extracted from a main jet image.

Here, small-scale curvature can be observed together with larger structures, indicat-
ing that a multiscale approach is necessary for examining the jet interface complexity.
Selecting a class of curvatures may be achieved by a low-pass filter. A Gaussian convo-
lution can be applied to fulfill the required properties for filtering. The applied Gaussian
is governed by a scale parameter, σ, where the resulting processed curve tends to the un-
filtered curve as σ tends to zero. Mokhtarian et al. (2005) proposed an implementation
of this approach by parametrizing the curve by its arc length s, so that the coordinates of
each of its points are
Γ0 = [x0 (s) , y0 (s)] , s = 0 . . . L (2)
where L is the total arc length of the curve. The curvature at any position s is then
computed by
0 00 0 00
x y − y x0
κ(s, 0) = ¡ 00 0 0 0¢1.5 (3)
x02 + y02

Atomization and Sprays


Image Processing Techniques for Velocity 769

This formula gives the curvature at the smallest scale and may be very noisy when ap-
plied to a digitized curve. A smoothed version of the curve is given by

Γσ = [xσ (s) , yσ (s)] , s = 0 . . . L (4)

where xσ (s) and yσ (s) are computed by convolution of original coordinates with a
Gaussian centered on x0 and y0 , respectively, of width σ:

xσ (s) = x0 (s) ⊗ gσ (x0 , s) and yσ (s) = y0 (s) ⊗ gσ (y0 , s) (5)

The curvature distribution along the curve is given by


0 00 0 00
xσ yσ − yσ xσ
κ (s, σ) = 1.5 (6)
(x0σ2 + yσ0 2 )
where (Mokhtarian et al., 2005)
0 0
xσ = x0 (s) ⊗ gσ (x0 , s)
0 0
yσ = y0 (s) ⊗ gσ (y0 , s)
00 00
(7)
xσ = x0 (s) ⊗ gσ (x0 , s)
00 00
yσ = y0 (s) ⊗ gσ (y0 , s)
0 00
gσ and gσ are, respectively, the first and second derivative of the Gaussian with respect
to s. Note that derivatives of x0 and y0 , which are a source of noise when applied to
an experimental curve, are not used directly. Figure 19 illustrates smoothed curves for
different scales σ. When σ increases, the curve tends to a straight line.
Finally, the curve complexity can be completely described by the spectrum of the
curvatures computed over different scales, which is called “curve scale space” (CSS) by
Mokhtarian et al. (2005):

|κ (s, σ)| , s = 0 . . . L and σ = 0 . . . σm (8)


where σm is the smallest scale at which the curve remains as a straight line. Figure 20
shows the computed CSS of the left-side contour of the jet shown in Fig. 18.
Curvatures are detected on this contour from scales (or smoothing parameter) of 1
µm (i.e., almost no smoothing) to ∼120 µm (i.e., smoothing producing almost a straight
line). The range of curvatures starts from |κ| = 0.3 µm−1 to , which corresponds to
curvature radii from 3 µm to ∞. Small curvature radii, even smaller than the image
resolution may arise at small scales due to the pixelization of the images and are highly
dependent on the threshold value used for the binarization of individual images. They
can also appear at larger scales, as a result of smoothing on a highly perturbed zone of the
curve (for example, at y ' 1400 µm on Fig. 19). Finally, a general trend of complexity
evolution along the distance y can only be extracted from a statistical analysis.

Volume 25, Number 9, 2015


770 Lounnaci et al.

FIG. 19: Smoothing of the left contour of jet shown in Fig. 16 for the following scales
from left to right: original contour, σ = 17 pixels, 33 pixels, 63 pixels, and 129 pixels.

FIG. 20: CSS of the left-side contour of the jet of Fig. 18.

Atomization and Sprays


Image Processing Techniques for Velocity 771

For each image, the CSS was built from the contour of the core jet. As the length of
the contour varies, scales were localized by the distance y from the nozzle outlet rather
than from arc length s to compute the averaged curvature diagram. Thus, each image
was divided into horizontal strips located at y and an arithmetic mean was performed
on each strip over all arc-length values belonging to this strip. A global average is also
performed over the whole set of images:
1 1 Xni X
κ̃ (y, σ) = × |κi (s, σ)| (9)
ni ny i=0 s/y(s)∈[y−h/2,y+h/2]

where ni is the number of available images, |κi (s, σ)| is the CSS of image i, and ny =
Card {y (s) ∈ [y − h/2, y + h/2]}, i.e., the number of curve points in the horizontal
image strip of height h located at y. κ̃ (y, σ) is plotted in Fig. 21 for the contour of the
left and right sides of the jet. Here one can observe that the occurrence of large liquid
structures tends to increase with the distance from the nozzle and that the growth is larger
on the right side than it is on the left.
In Fig. 21, the curves κ̃ (y, σ) = constant are nearly homothetic in σ. When the
points corresponding to κ̃ (y, σ) = 0.01,...,0.09 by a step of 0.005 (color levels on
Fig. 19) are plotted with a suitable scaling as in Fig. 22, they gather in a single curve,
which follows a power function. This implies that the average curvature κ̃ depends on
the scale σ and distance y to the nozzle through the variable σ∗ = σy −α , with α = 0.38
on the left side and α = 0.62 on the right. This allows the maps of Fig. 21 to be grouped
into single curves κ̃ (σy −α ) as shown in Fig. 23(a). By the same argument, these two
resulting curves can be joined in a single plot by an appropriate scaling [Fig. 23(b)].
Finally, the two sides of the spray exhibit an evolution which may be characterized by
two values: (i) exponent α [Fig. 23(a)], which may be interpreted as a complexity growth

FIG. 21: Averaged left- and right-side curvature scale distributions for the jet contour.

Volume 25, Number 9, 2015


772 Lounnaci et al.

FIG. 22: Averaged left- and right-side curvature scale distributions for the jet contour.

FIG. 23: Averaged left- and right-side curvature scale distributions for the jet contour.

factor, larger at right than at left, (ii) the scale range of κ̃ (σ∗ ), shown in Fig. 23(b), which
may be interpreted as the local complexity and is about 8 times larger on the right than
on the left.

Atomization and Sprays


Image Processing Techniques for Velocity 773

7. CONCLUSION

A shadow imaging arrangement using ultrafast illumination was set up to capture time-
resolved shadow images of injection events from a single-hole test nozzle producing
fuel sprays in ambient atmospheric conditions. Double-pulse source illumination and
the synchronized action of a double-image camera allowed the system to obtain time-
resolved image-pairs containing spatially resolved information suitable for correlation
analysis. An in-house code was used to process the images, removing background noise
and isolating spatially resolved structure to generate accurate fluid structure velocity
information.
A thresholding procedure was applied to the images to separate the droplet cloud
from the jet core. Because of the dissymmetry of the spray, the analyses were applied
separately to each side of the spray. The distinctive behavior of the dissymmetric injector
provided a qualitative check on the effectiveness of the analytical methods.
The interface complexity of the main jet was characterized by a procedure based on
the curvature, and the projected area of the dispersed phase was used to estimate a spatial
probability for the occurrence of droplets.
All the observations on the dissymmetric injector under study may be summarized
as follows: (i) the interface velocity of the core jet is larger on the left than on the right;
(ii) the dispersed phase velocity has the same trend; (iii) the droplet (or liquid element)
velocity is smaller on the left than on the right; (iv) the complexity growth of the core
jet is smaller on the left than on the right; (v) the curvature distribution of the core jet
covers a smaller range on the left than on the right. One can assume there are relations
between velocity, droplet generation, and complexity of the jet, which might be revealed
by varying injectors and injection parameters.

ACKNOWLEDGMENT

The authors acknowledge the support of the French Agence Nationale de la Recherche
(ANR) under reference ANR-13-TDMO-0003.

REFERENCES
Bachalo, W. D., Experimental methods in multiphase flows, Int. J. Multiphase Flow, vol. 20, pp.
261–295, 1994.
Bradski, G. and Kaehler, A., Learning OpenCV: Computer Vision with the OpenCV Library,
Sebastopol, CA: O’Reilly Media, 2008.
Chaves, H., Kirmse, C., and Obermeier, F., Velocity measurements of dense diesel fuel sprays in
dense air, Atomization Sprays, vol. 14, no. 6, pp. 589–609, 2004.
Cheng, H. D. and Chen, Y. H., Fuzzy partition of two-dimensional histogram and its application
to thresholding, Patt. Recog., vol. 32, no. 5, pp. 825–843, 1999.

Volume 25, Number 9, 2015


774 Lounnaci et al.

Dumouchel, C., On the experimental investigation on primary atomization of liquid streams, Exp.
Fluids, vol. 45, no. 3, pp. 371–422, 2008.
Fielding, J., Long, M. B., Fielding, G., and Komiyama, M., Systematic errors in optical-flow
velocimetry for turbulent flows and flames, Appl. Opt., vol. 40, no. 6, pp. 757–764, 2001.
Hespel, C., Blaisot, J. B., Gazon, M., and Godard, G., Laser correlation velocimetry performance
in diesel applications: Spatial selectivity and velocity sensitivity, Exp. Fluids, vol. 53, no. 1,
pp. 245–264, 2012.
Kapur, J. N., Sahoo, P. K., Wong, A. K. C., A new method for gray-level picture thresholding
using the entropy of the histogram, Computer Vision, Graphics Image Processing, vol. 29, no.
3, pp. 273–285, 1985.
Kohse-Höinghaus, K. and Jeffries, J. B., eds., Applied Combustion Diagnostics, New York: Tay-
lor & Francis, 2002.
Krüger, S. and Grünefeld, G., Stereoscopic flow-tagging velocimetry, Appl. Phys. B: Lasers Opt.,
vol. 69, no. 5, pp. 509–512, 1999.
Leick, P., Bittlinger, G., and Tropea, C., Velocity measurements in the near nozzle region of
common-rail Diesel sprays at elevated back-pressures, In Proc. of 19th International Confer-
ence on Liquid Atomization and Spray Systems, ILASS-Europe ’04, Nottingham, 2004.
Li, C. H. and Lee, C. K., Minimum cross entropy thresholding, Pattern Recognition, vol. 26, no.
4, pp. 617–625, 1993.
Masataka, A., Physics behind diesel sprays, In Proc. of 12th Triennial International Conference
on Liquid Atomization and Spray Systems, ICLASS, Heidelberg, Germany, Sept. 2–6, 2012.
Marks, T. K., Hershey, J. R., and Movellan, J. R., Tracking motion, deformation, and texture
using conditionally gaussian processes, Pattern Analysis Machine Intelligence, IEEE Trans.,
vol. 32, no. 2, pp. 348–363, 2010.
Mingqiang, Y., Kidiyo, K., and Joseph, R., A survey of shape feature extraction techniques,
Pattern Recognition Techniques, Technology and Applications, 2008.
Mokhtarian, F., Khim Ung, Y., and Wang, K., Automatic fitting of digitised contours at multiple
scales through the curvature scale space technique, Comput. Graphics, vol. 29, pp. 961–971,
2005.
Otsu, N., A threshold selection method from gray-level histograms, IEEE Trans. Syst. Man. Cy-
bernetics, vol. 9, no. 1, pp. 62–66, 1979.
Purwar, H., Idlahcen, S., Rozé, C., Sedarsky, D., and Blaisot, J. B., Collinear, two-color opti-
cal Kerr effect shutter for ultrafast time-resolved imaging, Opt. Express, vol. 22, no. 13, pp.
15778–15790, 2014.
Raffel, M., Willert, C. E., Wereley, S. T., and Kompenhans, J., Particle Image Velocimetry: A
Practical Guide, New York: Springer, 2nd ed., 2007.
Ridler, T. and Calvard, S., Picture thresholding using an iterative selection method, IEEE Trans.
Syst. Man Cybernetics, vol. SMC-8, no. 8, pp. 630–632, 1978.
Sedarsky, D., Paciaroni, M.E., Linne, M. A., Gord, J. R., and Meyer, T. R., Velocity imaging for
the liquid-gas interface in the near field of an atomizing spray: Proof of concept, Opt. Lett.,
vol. 31, no. 7, pp. 906–908, 2006.
Sedarsky, D., Idlahcen, S., Blaisot, J. B., and Rozé, C., Planar velocity analysis of diesel spray

Atomization and Sprays


Image Processing Techniques for Velocity 775

shadow images, International Symposium on Multiphase flow and Transport Phenomena,


Agadir, Morocco, April 22–25, 2012.
Sedarsky, D., Idlahcen, S., Rozé, C., and Blaisot, J. B., Velocity measurements in the near field
of a diesel fuel injector by ultrafast imagery, Exp. Fluids, vol. 54, no. 2, pp. 1–12, 2013.
Seibert, J. A., Boone, J. M., and Lindfors, K. K., Flat-field correction technique for digital detec-
tors, Proc. SPIE, vol. 3336, pp. 348–354, 1998.
Shavit, U. and Chigier, N., Fractal dimensions of liquid jet interface under breakup, Atomization
Sprays, vol. 5, no. 6, pp. 525–543, 1995.
Tokumaru, P. T. and Dimotakis, P. E., Image correlation velocimetry, Exp. Fluids, vol. 19, pp.
1–15, 1995.
Yon, J., Jet Diesel haute pression en champ proche et lointain: Etude par imagerie, Ph.D. Thesis,
Université de Rouen, 2003.

Volume 25, Number 9, 2015

You might also like