IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 58, no.

5, May 2011 1037

Synthetic Aperture Focusing of Ultrasonic

Data From Multilayered Media Using an
Omega-K Algorithm
Martin H. Skjelvareid, Student Member, IEEE, Tomas Olofsson, Yngve Birkelund, Member, IEEE,
and Yngvar Larsen, Member, IEEE

Abstract—The synthetic aperture focusing technique main processing. In 1978, Stolt [4] introduced what is
(SAFT) is used to create focused images from ultrasound now called the frequency-wavenumber, or ω-k, algorithm,
scans. SAFT has traditionally been applied only for imaging
in a single medium, but the recently introduced phase shift
performing all processing in the temporal frequency and
migration (PSM) algorithm has expanded the use of SAFT to spatial frequency domains. This approach proved to be
multilayer structures. In this article we present a similar fo- significantly faster than the other methods available at
cusing algorithm called multi-layer omega-k (MULOK), which the time, and it has since become widely used in many
combines PSM and the ω-k algorithm to perform multilayer related fields. Its main disadvantage is that it requires the
imaging more efficiently. The asymptotic complexity is shown
to be lower for MULOK than for PSM, and this is confirmed
wave velocity of the propagating medium to be constant.
by comparing execution times for implementations of both al- In the same year, Gazdag [5] introduced the phase shift
gorithms. To facilitate the complexity analysis, a detailed de- migration (PSM) algorithm, which also operates in the
scription of algorithm implementation is included, which also frequency-wavenumber domain. Although PSM is not as
serves as a guide for readers interested in practical implemen- fast as the ω-k algorithm, it allows the wave velocity to
tation. Using data from an experiment with a multilayered
structure, we show that there is essentially no difference in
vary with depth.
image quality between the two algorithms. Both the ω-k and the PSM algorithm are based on ex-
trapolating the backscattered wavefield from the plane in
which it is recorded down into the region to be imaged.
I. Introduction In Fourier analysis of wave fields, this is also known as
angular spectrum propagation [6], [7].

S ynthetic aperture processing is used in radar, so- Building on previous work within sonar and radar imag-
nar, seismic, and ultrasound imaging. The technique ing, synthetic aperture focusing was introduced to the field
is based on emitting a wave into a region of interest, re- of NDT ultrasonics in the 1970s, and came to be known
cording the backscattered echoes, and repeating this for as the synthetic aperture focusing technique (SAFT) [8].
several positions. Recorded data are subsequently com- Although the time-domain delay-and-sum method was the
bined to create a large synthetic aperture, yielding a high- starting point, frequency domain algorithms were soon ad-
resolution image of the reflectivity in the region. opted by the ultrasonic community, yielding low execution
Synthetic aperture imaging is in many ways similar to times at the cost of increased memory usage [9]–[13]. In
imaging using arrays of multiple transmitters and/or re- recent years, there has also been a growing interest in
ceivers. Although there are several advantages to using adapting such algorithms for arrays [14], [15]. Research-
such arrays, they do not offer any increase in resolution ers have mainly focused on imaging in a single, constant-
compared with monostatic synthetic aperture imaging [1], velocity medium, but some time-domain methods for
[2]. Because monostatic imaging systems also have low multilayer structures have been developed [16], [17], and
complexity, cost, and spatial requirements, they remain recently Olofsson [18] introduced the use of the frequency-
relevant in modern applications. domain PSM algorithm for processing multilayer data. In
Within the field of seismic imaging, synthetic aperture this article, we combine the PSM and ω-k algorithms to
techniques are known as migration techniques. Computer- create a more efficient multilayer algorithm, and extend
based processing of seismic data started in the 1970s [3], the experimental setup to include three-layered media.
and the initial methods were limited to time-space do- There are several applications for ultrasound imaging
of multilayer structures. One very important case is that
of immersion scans, in which an object is immersed in
Manuscript received November 18, 2010; accepted February 22, 2011.
The authors thank Breivoll Inspection Technologies and the Research water and several pulse-echo measurements are performed
Council of Norway for funding this work. by scanning the transducer in the water layer above the
M. H. Skjelvareid and Y. Birkelund are with the University of Trom- object. The water and the object constitutes a two-layer
sø, Department of Physics and Technology, Tromsø, Norway (e-mail:
martin.skjelvareid@breivoll.no). structure, and to properly focus the backscattered echoes
T. Olofsson is with Uppsala University, Department of Engineering from within the object, the focusing algorithm has to take
Sciences, Uppsala, Sweden. the wave velocities of both media into account. This is also
Y. Larsen is with Norut, the Northern Research Institute, Department
of Earth Observation, Tromsø, Norway. the case for other multi-layered structures, for example,
Digital Object Identifier 10.1109/TUFFC.2011.1904 steel pipes lined with a corrosion-resistant layer.

0885–3010/$25.00 © 2011 IEEE

1038 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 58, no. 5, May 2011

The main advantages of immersion scans, as opposed to

contact scans, are that the water layer acts as a good and
uniform couplant for the acoustic waves, and that there
is no friction to cause transducer wear. The transducer
used in immersion scans is usually geometrically focused
to provide a good lateral resolution in the transducer focal
zone. Synthetic aperture techniques represent an alterna-
tive that can be used to obtain a high lateral resolution
that is independent of depth.
As previously mentioned, the ω-k algorithm has proven
to be very efficient for single-layer processing, and the
PSM algorithm is capable of imaging structures where the
wave velocity varies with depth. In 1989, Kim [19] in-
troduced a method for seismic imaging which combines
the advantages of both algorithms. He assumed that the
geological structure of the earth can be approximated as
a finite number of layers with constant wave velocity and
used the PSM algorithm to extrapolate the wavefield down
to the interfaces between the layers. He then used the ω-k
algorithm to effectively image the interior of each layer. In
this paper, we demonstrate that the same approach can be Fig. 1. (a) Scanned pulse-echo measurements and (b) the exploding re-
used for ultrasonic imaging of multilayer structures, and flector model.
we will refer to this as the multi-layer omega-k algorithm
(MULOK). The algorithm is compared with the PSM al- nario in which we assume that the scatterers themselves
gorithm in terms of both computational complexity and are sources of acoustic energy.
image quality. Fig. 1 illustrates a B-scan of a two-layer geometry, where
To simplify the treatment of the algorithms, a two- a transducer is scanned along the x-axis and a pulse-echo
dimensional geometry is considered here, but an extension measurement is performed for each x position. A scatterer
to three dimensions is straightforward. A real ultrasound is present in the second layer, and because the two layers
application is usually three-dimensional in nature, but if have different wave velocities, both the transmitted and
the region of interest is homogenous along one direction, a the reflected wave are refracted at the interface between
two-dimensional approximation can be made without any the layers.
loss of information. Fig. 1 shows how the exploding reflector model is ap-
To compare the asymptotical complexity of the PSM plied to the same geometry. The scatterer is assumed to
and MULOK algorithms, a detailed description of the al- spontaneously radiate a wave which travels toward an ar-
gorithm implementations has been included. As an added ray of transducer positions, undergoing the same refrac-
benefit, the description also serves as a guide for practical tion as for the pulse-echo case. In order for the time delay
implementation. of the pulses to remain the same under the exploding re-
The remainder of this article is organized as follows: flector model as for the actual pulse-echo measurement,
In Section II, the theories of the PSM and MULOK algo- an effective wave velocity of half the actual velocity must
rithms are presented, and in Section III, the asymptotic be assumed,
complexity of each algorithm is analyzed and compared
with actual execution times. An experiment illustrating cl
ĉ l = , (1)
the imaging performance of both algorithms is presented 2
in Section IV, before we state our conclusions in Section
V. Details of algorithm implementation and complexity where l is the layer index. The model is general and can be
analysis have been placed in the Appendix, to maintain applied for an arbitrary number of layers.
the flow of the article for the general reader. In most ultrasonic measurements, there are multiple
reflectors, and the wave field measured by the transducer
is then approximated as a superposition of waves from
II. Theory several exploding reflectors. This assumption holds true
as long as multiple reflections between scatterers can be
A. The Exploding Reflector Model neglected.

Derivation of migration algorithms for the monostatic B. Extrapolation of Wave Fields in the Fourier Domain
case are often based on the exploding reflector model [20],
[21], which simplifies the inverse imaging problem. It re- For migration algorithms, the purpose of wave field ex-
duces the two-way pulse-echo scenario to a one-way sce- trapolation is to calculate the wave field at an arbitrary
skjelvareid et al.: focusing of data from multilayered media using an ω-k algorithm 1039

depth from measurements performed at a given depth, +∞

here denoted Z. In the following derivation we will assume I (x, Z + ∆z ) = ∫∫ P(w, k x, Z )e ik ∆ze ik x dk x dw.
z x
(7)
that all scatterers are located in the half-space z > Z. −∞
Let p(t, x, z) denote the wave field generated by a set
of exploding reflectors, and assume that only longitudinal Note that inserting t = 0 into (5) reduces the inverse
waves are present in the medium. The shear wave com- transform with respect to ω to a simple integral over ω.
ponents of the wave field can generally be ignored if the Eq. (7) can be used iteratively to create an image line by
medium is not a solid, or if the waves have a near-normal line, by applying it for all depths Z + Δz to be imaged. In
incidence angle at interfaces to solids, as is often the case seismic processing, this is called phase shift migration
in immersion scans. (PSM), referring to the phase shift given by the term
Assume now that the wave field is recorded along the e ik z Dz.
line z = Z, so that p(t, x, Z) is known. Note that because
all scatterers are assumed to be located in the half-space E. Imaging Through Stolt Transform
z > Z, all recorded waves are traveling in the negative
z direction. The Fourier transform of the recorded wave Eq. (7) is very similar to an inverse Fourier transform
field is given by of P(ω, kx, Z), but it has an e ik z Dz kernel rather than an
+∞
1 e−iωt kernel. It can be recast into a proper inverse Fourier
P(ω, k x, Z ) =
4π 2 ∫∫ p(t, x, Z )e −i(k x −ωt)dx dt, (2)
x
transform by a change of variables from ω to kz. Integrals
−∞ in the form of Fourier transforms are of particular interest,
as they can be calculated using the computationally effi-
where 1/(4π 2) is a normalization constant. It can be shown cient fast Fourier transform (FFT) and its inverse coun-
[7] that the wave field can be extrapolated to any depth terpart (IFFT).
Z + Δz by multiplication with a complex exponential, We obtain an expression for ω by using the relation
given in (4), and assuming, as in Section II-B, that ω and
P(w, k x, Z + ∆z ) = P(w, k x, Z ) ⋅ e ik z (w,k x )⋅∆z, (3)
kz have opposite signs, so that
where kz(ω, kx) is given by w(k z, k x ) = − sgn(k z ) ⋅ cˆ k x2 + k z2. (8)
2
w
k z(w, k x ) = −sgn(w) ⋅ − k x2. (4) By substituting (8) into (7), we get
cˆ2
+∞
The sign function ensures that the kz value represents a
wave traveling in the negative z direction, and the effective I (x, Z + ∆z ) = ∫∫ P(k z, k x, Z )e ik ∆ze ik x dk x dk z, (9)
z x

wave velocity ĉ = c/2 is used because the exploding reflec- −∞

tor model is assumed. An inverse Fourier transform is
used to obtain the extrapolated wave field in time-space where
coordinates,
P(k z, k x, Z ) = A(k z, k x ) × P(w(k x, k z ), k x, Z ), (10)
+∞

p(t, x, Z + ∆z ) = ∫∫ P(w, k x, Z )e ik ∆ze i(k x −wt)dk x dw. (5)

z x
and
−∞
∂w(k z, k x ) cˆ
A(k z, k x ) = = . (11)
C. Imaging Condition ∂k z 1+
k x2
k z2

Given an expression for the wave field p(t, x, z), an im-

aging condition is needed to obtain an image I(x,z) of the We see here that, relative to the original wave field
exploding reflectors. For the exploding reflector model, P(ω, kx, Z), the substitution of variables leads to a multi-
the imaging condition is to set t = 0 [21], so that plication with an amplitude factor A(kz, kx) and a shift in
ω given by ω (kz, kx).
I (x, z ) = p(t = 0, x, z ). (6)
F. Adaptation to Multilayer Case
The wave field emitted from a scatterer is maximally con-
centrated in space at the origin of the time axis, and thus Assume now that we have several layers with poten-
the imaging condition in (6) is chosen to produce a maxi- tially different wave velocities, as shown in Fig. 2. Let the
mally focused image. layers be numbered with l = 1, 2, …, L, and let dl and cl
denote the thickness and wave velocity of layer l, respec-
D. Imaging Using Phase Shift Migration tively. The top of the uppermost layer is denoted Z1, and
the interfaces between the layers are denoted Zl, so that
l −1
Applying the imaging condition of (6) to (5), we get the top of layer l is given by Zl = Z1 + ∑m =1d m.
1040 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 58, no. 5, May 2011

l
P(w, k x, Z l ) ∝ P(w, k x, Z 1) ⋅ e i ∑m =1 k zmd m . (15)

Eqs. (13) and (15) constitute the basis for PSM imag-
ing of several layers. The imaging procedure for layer l can
be summarized as follows:

1) Calculate the wave field at the top of the layer inter-

face, P(ω, kx, Zl), using (15).
2) For each depth Zl + Δz to be imaged within the
layer:
a) Shift the wave field downwards with Δz by mul-
tiplying with a phase factor, according to (13), to
obtain P(ω, kx, Zl + Δz).
Fig. 2. Illustration of horizontally layered geometry.
b) Create an image line I(x, Zl + Δz) by integrating
with regard to ω and inverse transforming with
regard to kx, according to (7).
We know from (3) that in a medium with constant
sound speed, the wave field P(ω, kx, Z) can be calculated Instead of creating an image of each layer line by line,
by multiplying a reference wave field P(ω, kx, Z) with a the Stolt transform introduced in Section II-E can be
phase factor e ik z Dz. In the multilayer case, the field cannot adapted to the multilayer case to enable imaging of an en-
be extrapolated through several layers directly, because kz tire layer through a single inverse Fourier transform. This
is a function of the medium velocity cˆ,l is the approach we call MULOK. The complete image for
layer l is then given by
w2
k zl = −sgn(w) ⋅ − k x2, (12) +∞
cˆl2
I l(x, ∆z ) = ∫∫ P(k z, k x, Z l )e −ik ∆ze ik x dk x dk z,
z x
∆z < d l,
where the layer dependence is indicated by index l. How- −∞
(16)
ever, extrapolation within each layer is still possible,
where
P(w, k x, Z l + ∆z ) = P(w, k x, Z l ) ⋅ e ik zl ∆z, ∆z < d l, (13)
P(k z, k x, Z l ) = Al(k z, k x ) × Pl(wl(k x, k z ), k x, Z l ), (17)
where P(ω, kx, Zl) denotes the field at interface Zl, and
cˆl
P(ω, kx, Zl + Δz) denotes the field at depth z = Zl + Δz. Al(k z, k x ) = , and (18)
k x2
If the wave field is to be extrapolated to more than one 1+
k zl2
layer, the transmission of waves through the layer interfac-
es has to be considered. The transmission factor between wl(k z, k x ) = − sgn(k zl ) ⋅ cˆl k x2 + k zl2 . (19)
different media is generally a complex function, dependent
on both incident angle and the acoustic impedances of the The imaging procedure to create an image of layer l can
media [22]. However, for most commonly available ultra- be summarized in the following way:
sound transducers, the directivity of the transducer limits
the emitted and received wave fields to a relatively small 1) Calculate the wave field at the top of the layer,
angle interval. We will therefore assume that the trans- P(ω, kx, Zl), using (15).
mission factors are approximately independent of incident 2) Use (17)–(19) to perform the variable transforma-
angle, so that the wave fields directly above and directly tion from ω to kzl.
below an interface are proportional; 3) Inverse transform to obtain the image within the
layer, Il(x, Δz), using (16).
P(w, k x, Z l−) ∝ P(w, k x, Z l+), (14)

where the plus and minus signs are used to indicate the G. Comments on Theoretical Resolution
upper and lower side of the interface, respectively. Because
we are mainly interested in relative amplitudes within The lateral resolution of a synthetic aperture image is
each layer, the amplitude scaling effect imposed by the dependent on the bandwidth of the kx spectrum [23], and
interfaces is considered here to be unimportant to the im- this bandwidth is limited by the effective length L of the
aging problem. transducer. A common rule of thumb for the single-layer
Assuming proportionality across interfaces, the wave case is that this makes the maximum lateral resolution
field at an arbitrary interface Zl can, within a scaling fac- approximately L/2 [13]. Here we will argue that this limit
tor, be calculated from the wave field measured at Z1, is also relevant for the multilayer case.
skjelvareid et al.: focusing of data from multilayered media using an ω-k algorithm 1041

According to Snell’s law, the kx wavenumber of a wave TABLE I. Complexity for the Individual Steps of the PSM
incident on a layer interface must remain the same after Algorithm.
transmission into the next medium. Thus, the refraction Operation Complexity
of the wave does not in itself alter the horizontal wave- Initial Fourier transform O(MN log MN)
number, but the transmission factors between media are Phase shift to Zl O(LMN)
generally dependent on incident angle, making the inter- Phase shift to Zl + Δzl O(MN2)
Summation over ω O(MN2)
face a filter for the kx spectrum. The bandwidth of the kx Inverse transform (kx) O(MN log MN)
spectrum is only maintained as long as the transmission
factors are approximately uniform within the divergence
angle of the transducer, but in practice, this requirement TABLE II. Complexity of the Individual Steps of the MULOK
is fulfilled for many transducer designs and material com- Algorithm.
binations of interest. For example, for an immersion scan Operation Complexity
of copper using a 2.25-MHz, 10-mm-diameter transducer, Initial Fourier transform O(MN log MN)
the echo transmission factor varies by only approximately Phase shift to Zl O(LMN)
6% within the transducer beam [18]. As long as the kx Interpolation from ω to kz O(LMN log N)
spectrum bandwidth can be assumed to be the same for Amplitude scaling O(LMN)
the single-layer and multilayer case, the theoretical lateral Inverse Fourier transform O(LMN ∙ log MN)
resolution of L/2 is also the same.
TABLE III. Asymptotic Complexities PSM and MULOK,
Regarding N, M, and L Separately.
III. Algorithmic Complexity
Algorithm N M L
PSM O(N2) O(M log M) O(L)
A. Asymptotic Complexity
MULOK O(N log N) O(M log M) O(L)

The effectiveness of an algorithm is often quantified by

analyzing how the number of operations grows as the size
of the input data tends toward infinity, and this asymp- ber of time samples, N. The simulations were performed
totic complexity is denoted using big-O notation. In the on a dual-core 2-GHz laptop with 2 GB of RAM, running
case of the PSM and MULOK algorithms, the size of the a 64-bit Linux version of Matlab R2009b.
input data are given by the number of time samples, N, The resulting execution times are plotted in Fig. 3.
the number of measurement positions, M, and the number Note that both the x- and y-axes of the plots are loga-
of layers, L. Note that the sampling frequency is assumed rithmic. Figs. 3(a) and 3(b) show the execution times as
to be constant, so that the number of frequency samples functions of N when the number of layers L is 2 and 5, re-
within the transducer bandwidth is proportional to the spectively. To compare the asymptotic complexities with
number of time samples. The asymptotic complexities the execution times of the simulation, lines corresponding
for each individual step of the algorithms are analyzed to N 2 and N log N have been added to the plot, normal-
in part E of the Appendix, and are listed in Tables I and ized to intersect with the execution time for the highest
II. The overall complexity of each algorithm is given by value of N.
the algorithm step with the highest-order complexity. To We find from Figs. 3(a) and 3(b) that PSM has a lower
analyze this, we consider the complexities with regard to execution time than MULOK for small values of N, but
N, M, and L separately, assuming that the two remaining that MULOK is much faster for larger values of N. For
variables are kept constant. The highest-order complexi- example, for L = 5 and N = 16 384, MULOK is approxi-
ties are summarized in Table III. MULOK is seen to have mately 15 times faster than PSM, and for L = 2 and N =
a lower complexity than PSM with regard to N, because 16 384, MULOK is approximately 30 times faster. Com-
N log N < N 2, whereas the complexity with regard to M paring the results for L = 2 and L = 5, we see that the
and L is the same for both algorithms. execution time of the PSM algorithm is approximately the
same for both cases, whereas the time for the MULOK
B. Empirical Evaluation of Execution Times algorithm is significantly higher for L = 5. Thus, using a
greater number of layers shifts the crossover point between
To do a realistic comparison of the two algorithms, sev- the algorithms to a higher N value. The reason that MU-
eral simulated processings were performed in Matlab (The LOK is more heavily influenced by L than PSM can be
MathWorks, Natick, MA). For convenience, test data sets found by comparing the complexities for each algorithmic
were made using matrices with random numbers rather step, listed in Tables I and II. PSM has only one step
than actual ultrasonic scans. This should not affect the whose complexity is proportional to L, whereas almost
execution times of the algorithms, as they are only depen- all steps in MULOK have complexities proportional to L.
dent on the size of the data sets, and not on their content. The execution times are seen to correspond well to the
The number of measurement positions, M, was set to 128, asymptotic complexities indicated by the dashed lines,
and the processing times were measured for different num- particularly for larger values of N. This indicates that the
1042 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 58, no. 5, May 2011

Fig. 3. Execution time simulation results: (a) L = 2, M = 128; (6) L = 5, M = 128.

formed using a 2.25-MHz transducer with 6 mm diameter,

moved in steps of 1 mm, with a sampling frequency of
12.5 MHz. N = 1040 time samples were recorded at each
of the M = 111 measurement positions.
The upper block was 31-mm-thick poly(methyl meth-
acrylate) (PMMA). The lower block was 50-mm-thick,
aluminum. Each of the blocks had four side-drilled holes,
which were all 1.6 mm in diameter and 30 mm deep. The
vertical spacing between the holes was 6  mm in the PMMA
block and 10 mm in the aluminum block; the horizontal
spacing was 20 mm for both blocks. The blocks were also
shifted horizontally approximately 10 mm, so that the up-
per holes would not create a dominating shadow for the
lower holes. The steel pins were 0.3 mm in diameter, with
a vertical spacing of 5 mm and a horizontal spacing of
20 mm. The scan line was centered over the holes and pins
in the y direction.
To compensate for limited dynamic range in the ultra-
sonic acquisition system, a time-dependent damping/am-
plification was applied during the scan. The water-PMMA
interface echo was damped −10 dB, and the echoes from
Fig. 4. Experimental setup. the aluminum layer were amplified by +20 dB.
The Hilbert transform was used to create an analytic
signal for each time series in the raw data, and the B-scan
general trends seen here can be assumed to be valid in envelope was estimated by taking the absolute value of
general, even if the given execution times presented here the analytic signal. The resulting raw data image is shown
are valid only for one specific implementation of the al- in Figs. 5(a) and 5(d) plotted with a 25 and a 50 dB dy-
gorithms. namic range, respectively. Two different dynamic ranges
were used to emphasize the difference between high-am-
plitude and low-amplitude effects. The front echo from
IV. Experiment the water-PMMA interface is seen as a horizontal line at
approximately 100 μs, and echoes from the PMMA-alu-
In Section III, it was shown that the asymptotical com- minum and the aluminum-water interfaces are visible at
plexity of MULOK is lower than that of PSM with regard approximately 123 and 137 μs, respectively.
to N. To validate that MULOK is also capable of produc- The longitudinal wave velocities for water, PMMA, and
ing the same image quality as PSM, an experiment with a aluminum are approximately 1480, 2730, and 6320 m/s,
three-layer structure was conducted. and the difference in velocity can be seen clearly in the B-
Two test blocks with side-drilled holes were placed on scan, because the apparent thicknesses of the layers on the
top of each other and immersed in a water tank, and an time axis are far from their actual thicknesses. The echoes
array of four steel pins was placed over the two blocks, as from the steel pins and the side-drilled holes are seen as
shown in Fig. 4. A B-scan of the arrangement was per- four reflections in each of the layers, and the width of
skjelvareid et al.: focusing of data from multilayered media using an ω-k algorithm 1043

Fig. 5. Comparison of raw data and images focused by PSM and MULOK, displayed on a decibel intensity scale. The dynamic range of the top
row is 25 dB: (a) raw data, (b) PSM image, (c) MULOK image; the dynamic range of the bottom row is 50 dB: (d) raw data, (e) PSM image, (f)
MULOK image.

the reflections increases with depth because of the diver- The raw data was processed with both the MULOK and
gence of the emitted transducer pulse. There are also some the PSM algorithms, and the resulting images are shown in
weaker reflections cluttering the image in both the PMMA Figs. 5(b), 5(c), 5(d) and 5(f), plotted with absolute value
and aluminum layers. These are caused by multiple reflec- on a decibel intensity scale. Like the raw data, the images
tions of the scattterers. are shown with both 25 and 50 dB dynamic ranges.
1044 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 58, no. 5, May 2011

Fig. 6. Close-ups of the responses from the rightmost scatterer in each layer, interpolated to a higher resolution. Layer 1, water: (a) raw data, (b)
PSM image, and (c) MULOK image. Layer 2, PMMA: (d) raw data, (e) PSM image, and (f) MULOK image. Layer 3, aluminum: (g) raw data, (h)
PSM image, and (i) MULOK image. The amplitude difference between each contour is 6 dB.

It is evident from the images in Fig. 5 that the images images has been adjusted so that it covers the same depth
produced by PSM and MULOK are visually very simi- interval as the raw B-scan image. With this compensa-
lar. The differences between PSM and MULOK will be tion, we can see that the point scatterer response after
discussed later. The reflections from the scatterers have focusing is quite similar for each layer. Although there is
been focused, resulting in an improved lateral horizontal a much higher level of background noise in the aluminum
resolution that is approximately the same for all scatter- layer, the high-amplitude part of the response still has ap-
ers, independent of depth or layer. Multiple reflections proximately the same extent as for the other layers. Note
have been partially focused or defocused, depending on also that the only visible differences between the PSM and
how close they are in time to their original scatterers. For MULOK images are within the background noise.
example, the reflection seen at approximately t = 130 μs, After close inspection of the point scatterer responses
x = 30 mm in the raw data appears to be caused by a in Fig. 6, the −6-dB widths of the raw data images were
scatterer in the aluminum layer, but it is actually a mul- found to be 6.7, 13.1, and 21.2 mm, for the water, PMMA,
tiple reflection of the leftmost scatterer in the water layer. and aluminum layers, respectively. The corresponding −6-
In the focused images, the reflection has been defocused dB widths are 3.1, 3.1, and 3.8 mm for both the PSM and
into a curve, because it did not originate in the aluminum MULOK images, indicating that there is no practical dif-
layer. ference in lateral resolution between the two algorithms.
The improvement in lateral resolution offered by the The −6-dB widths also correspond well to the theoretical
PSM and MULOK algorithms is highlighted in Fig. 6, resolution of approximately half the transducer diameter
using close-up images of the rightmost scatterer in each (3 mm), as discussed in Section II-G. The aluminum layer
layer. The extent of the z-axis in the PSM and MULOK has a slightly broader response compared with the other
skjelvareid et al.: focusing of data from multilayered media using an ω-k algorithm 1045

two layers, and this is probably mainly because of the Appendix

higher level of background noise. A. Matrix Representation of Discrete Data

In Section II, the theory was outlined for the case of

V. Discussion and Conclusions continuous signals, and it was also assumed that the wave-
field in the measurement plane is known for all x and t. In
practice, the wave field must be sampled discretely both
We have presented a new algorithm for processing of in time and space, for a finite time period and over a finite
ultrasonic data from multilayer structures, called MU- section of the x-axis.
LOK, and compared it with the PSM algorithm. Both We assume that at each measurement position, a pulse
algorithms represent an extension of the SAFT concept to is emitted, and Nt time samples are recorded, correspond-
the case in which the imaged geometry consists of layers ing to time instants t 1, t 2, , t N t . The measurement is per-
with differing wave velocities. formed at M different x-positions, x1, x2, …, xM. Time
In theory, the algorithms can focus through an arbi- samples are equally spaced with ΔT = 1/fs, where fs is the
trary number of layers. However, because the transmission sampling frequency, and the x positions are equally spaced
factor between layers is generally less than one, both the with ΔX. Assuming that the measurement is done at
transmitted pulse and the backscattered echo will be sig- depth Z1, the discrete data set can be organized in a ma-
nificantly damped at each interface. Thus, in practice, the trix Ptx[Z1], with element pij corresponding to time instant
signal-to-noise ratio is likely to limit the number of layers ti of the pulse-echo measurement at position xj:
that can be imaged. In addition, multiple reflections from
within the first layers may interfere with echoes from lay- x1 x2 … xM
ers further down. t1  p 11 p 12 … p 1M 
 
It is assumed for both the PSM and the MULOK algo- Ptx[Z 1] = t 2  p 21 p 22   . (20)
     
rithms that the layers are all horizontal. PSM can also be 
used to focus data from geometries in which the interfaces tNt  p N t1 … … p N tM 

between layers are planar and non-horizontal, but this re-
quires an additional processing step in which the wave Note that some zero-padding of Ptx in the x-direction
field is rotated [24]. The same step can be applied to use may be required to avoid spatial aliasing in the focused
MULOK on non-horizontal layers. image [26]. The discrete Fourier transform of Ptx[Z1] is
We have shown, both theoretically and by numeri- also an Nt × M matrix, denoted P ˆ wk [Z 1]. The elements of
x

cal simulation, that the MULOK algorithm has a lower P̂wk x [Z 1] correspond to ω in the range [−π fs, π fs], but only
asymptotic complexity than the PSM algorithm. How- elements that correspond to the transducer passband are
ever, the simulations also indicated that the effective- significantly different from zero. Because the frequency
ness of the MULOK algorithm decreases as the number spectra of real valued signals are symmetric, we can also
of layers increases, whereas the execution time of the limit the processing to positive ω values. Denoting the up-
PSM algorithm is more or less independent of number per and lower cutoff frequency for the transducer fmin and
of layers. fmax, we define Pwk x [Z 1] as the subset of P̂wk x [Z 1] correspond-
We have also demonstrated that the image quality and ing to ω ∈ 2π[fmin, fmax]:
lateral resolution is approximately the same for both algo-
rithms. Note, however, that if the interpolation step of the k x1 k x 2 … k xM
MULOK algorithm is not performed accurately enough, w1  P11 P12 … P1M 
 
the focused image will contain visible artifacts. Thus, the Pwk x [Z 1] = w 2  P21 P22   , (21)
     
accuracy of the interpolation should be adjusted according 
wN w  PN w1 … … PN wM 
to the image quality required. There are also variations 
of the ω-k algorithm which do an approximate, but effi-
cient, mapping from ω to kz without any interpolation, for where ω1 = 2p  N t(f min/f s)  , w N w = 2p  N t(f max/f s)  , and
example using the chirp z-transform [25]. Modifying MU- the step size is Δω = 2π fs/Nt. The relationship between
LOK to accommodate such methods is seen as a subject Nω and Nt is given by the ratio of transducer bandwidth
for future work. to sampling frequency:
Taking all factors into account, we see that the choice
N w/N t ≈ (f max − f min)/f s. (22)
between PSM and MULOK relies both on the geometry
to be imaged and the resources available for implementa- Also, assuming that the Fourier transform output is ar-
tion. If the number of input samples is relatively large, the ranged so that the zero wavenumber is centered, and that
number of layers is low, and the interpolation between ω M is even, the kx wavenumbers are given by
and kz can be executed efficiently and accurately, MULOK
can produce the same image quality as PSM in a much 2p
⋅ [−M /2, −M /2 + 1,…, 0,…, M /2 − 1.]
kx =
more efficient manner. If these requirements are not ful- ∆X ⋅ M
filled, PSM may be a better alternative. (23)
1046 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 58, no. 5, May 2011

B. Wave Field Extrapolation

We know from (12) and (13) that a wavefield at depth

Zl can be shifted to an arbitrary depth Zl + Δz within
layer l by multiplication with a complex exponential e ik zl Dz,
where kzl is given by

w2
k zl = −sgn(w) ⋅ − k x2. (24)
cˆl2

The frequency-wavenumber spectrum should be limited to

propagating waves, corresponding to real-valued kzl, and
this requirement is fulfilled as long as the square root ar-
Fig. 7. Interpolation points in the ω-kx domain.
gument of (24) is positive. All elements of Pwk x for which
(w 2/cˆl2) - k x2 < 0 should therefore be set to zero.
Let Kzl be the discrete matrix representation of kzl. It where the sign function is omitted because only positive
is an Nω × M matrix given by ω values are included in the processing. kzl denotes the
vector containing the discrete kz values, and it is squared
 k zl(w 1, k x1) … k zl(w 1, k xM ) 
  element by element. The interpolated values also have to
K zl =      . (25) be scaled according to (28), and the amplitude factors for
 
 k zl(w N w, k x1) … k zl(w N w, k xM )  kxj are given by

The extrapolation from depth Zl to depth Zl + Δz is cˆl

a l (j ) = . (31)
performed with an entry-wise multiplication: 1 + k xj2 /k zl2

Pwk x [Z l + ∆z ] = Pwk x [Z l ]  exp(iK zl ∆z ). (26)

Assuming that all kz values are negative, the kzl vector
  is given by
C. Stolt Transformation of Variables
k z,l = −∆k z,l ⋅ [0,1, , N k z − 1]T . (32)
We know from Section II-F that the transformation
from the (ω, kx) domain to the (kz, kx) domain is given by Δkz,l is the step size between each value and N k z is the
total number of values; these should be chosen to cover
P(k z, k x, Z l ) = Al(k z, k x ) × Pl(wl(k z, k x ), k x, Z l ), (27) the range of possible kz values and also avoid any aliasing.
This is fulfilled if
where
2p fs
∆k z,l ≤ ⋅ , (33)
cˆl cˆl N t
Al(k z, k x ) = , and (28)
k x2
1+ and
k zl2

wl(k z, k x ) = − sgn(k zl ) ⋅ cˆl k x2 + k zl2 . (29) 2p f max/ˆ

cl
N kz ≥ . (34)
∆k z,l
The discrete version of P(ω, kx, Zl), Pwk x [Z l ], is computed
for a finite, equally spaced set of ω values. Similarly, the D. Algorithm Description
discrete version of P(kz, kx, Zl), denoted Pk zk x [Z l ], should be
computed for an equally spaced set of kz. However, the Fig. 8 shows the flow of the PSM and the MULOK al-
mapping given by ωl(kz, kx) does not in general coincide gorithms, from the input ultrasonic data (denoted Ptx[Z1])
with the equally spaced values of ω in P(ω, kx, Zl), and to the focused image (denoted Izx). Ptx[Z1] is first Fourier
therefore a interpolation step is needed. The situation is transformed, and the elements corresponding to the trans-
illustrated in Fig. 7, where the original ω-kx grid is indi- ducer passband are extracted. Then, for each layer, the
cated as black dots. For each discrete value of kx, the wavefield is multiplied with the phase factor
exp ( i ∑m =1K zmd m ) to shift it from Z1 down to the top of
l −1
spectrum has to be interpolated to a new set of ω values,
indicated with gray crosses. layer l, given by Zl. For the first layer, the phase factor is
Using (29), we find that for a given kxj, the ω values to set equal to 1, resulting in zero phase shift. The shifted
be interpolated for are given by the vector wavefield is a common starting point for both algorithms,
and the subsequent steps for PSM and MULOK are shown
wlip(j) = cˆl ⋅ k xj2 + k 2zl , (30) on the left and the right side of the flowchart, respectively.
skjelvareid et al.: focusing of data from multilayered media using an ω-k algorithm 1047

the al(j) vector. Only part of the result corresponds to z

values within the layer, that is, Δz ∈ [0, dl]. For each layer,
this part is cut out and saved as the local image Izz[l].
When all the layers have been processed, the subimages
Izz[l] are stacked on top of each other to form the complete
image of all of the layers.

E. Asymptotic Complexity

Having defined both the PSM and the MULOK imple-

mentations, it is possible to study the computational com-
plexity of the two. Big-O notation is used to describe the
growth rate of operations as function of the size of the
input data. As described in part D, the raw data Ptx[Z1] is
an Nt × M matrix, corresponding to Nt time samples and
M measurement positions. After Fourier transforming the
data, an Nω × M submatrix of the result is extracted for
use in the subsequent processing, where Nω is proportion-
al to Nt. The number of z-lines for the PSM algorithm,
denoted Nz, and the number of kz values for the MULOK
algorithm, denoted N k z , are also proportional to Nt. For
the sake of asymptotic complexity analysis, we can ignore
all such proportionality constants, and set Nt = Nω = Nz
= N k z = N. The number of layers is denoted by L.
The initial Fourier transform, from (t, x) to (ω, kx), is a
two-dimensional transform with complexity O(MN log MN).
The following multiplication with a phase factor to calcu-
late Pwk x [Z l ] is an entry-wise multiplication that is per-
formed L − 1 times. The complexity of this operation is
thus O(NML).
For the PSM algorithm, the wave field is multiplied
with yet another phase factor. This multiplication is per-
formed N times, once for each image line, and thus the
complexity for all image lines is O(MN2). The summa-
tion over ω is also performed N times, resulting in a total
complexity of O(MN2). Finally, the last operation is the
inverse Fourier transform of an M-length vector, which is
O(MN log M). The complexities of the individual steps of
the PSM algorithm are summarized in Table I.
For the MULOK algorithm, the calculation of Pwk x [Z l ] is
followed by an interpolation step. The complexity of this
step depends on the type of interpolation utilized, but in
this work, the following method was used: Pwk x [Z l ] was first
interpolated to a denser rectangular grid by inverse Fou-
Fig. 8. Flowchart for Stolt and phase shift algorithms. rier transforming along the ω dimension, zero padding,
and Fourier transforming back again. This operation is
O(MN log N). The final interpolation was subsequently
The PSM algorithm is based on forming an image line performed by linear interpolation between points on this
i(Zl + Δz) for each depth to be imaged. The first opera- denser grid. This operation consists of a search to find the
tion is to compute Pwk x [Z l + ∆z l ] by multiplying with the two closest ω values and calculating a weighted sum of P
additional phase factor exp (iKzlΔzl). An image line is for these values; the corresponding complexity is
then formed by summing over all ω and inverse Fourier O(MN log N). Because the 2-D interpolation is performed
transforming the resulting vector. once for each layer, the overall complexity is then
For the Stolt imaging algorithm, the next step after O(LMN log N). Multiplication with the amplitude factors
calculating Pwk x [Z l ] is to interpolate from the original ω-kx al(j) is O(LMN), and the inverse Fourier transform to pro-
grid to the kx-kz grid. This is done by looping through all duce the image is O(LMN log MN). The complexities of
kx values, interpolating for each column of Pwk x [Z l ] with the the different steps of the MULOK algorithm are summa-
ω values given by the wlip(j) vector, and multiplying with rized in Table II.
1048 IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , vol. 58, no. 5, May 2011

Acknowledgments [23] P. T. Gough and D. W. Hawkins, “Unified framework for modern

synthetic aperture imaging algorithms,” Int. J. Imaging Syst. Tech-
nol., vol. 8, no. 4, pp. 343–358, 1998.
The authors thank Prof. T. Stepinski for inviting M. H. [24] T. Olofsson, M. H. Skjelvareid, and A. Barkefors, “Ultrasonic imag-
Skjelvareid to the Signals and Systems Group at Uppsala ing of immersed objects using migration techniques,” in European
Conf. Synthetic Aperture Radar 2010, 2010, pp. 442–445.
University, where most of this research was carried out. [25] R. Lanari, “A new method for the compensation of the SAR range
cell migration based on the chirp z-transform,” IEEE Trans. Geosci.
Rem. Sens., vol. 33, no. 5, pp. 1296–1299, 1995.
[26] K. Gu, J. Wang, and J. Li, “Migration based SAR imaging for
References ground penetrating radar systems,” in IEE Proc. Radar, Sonar
Navig., vol. 151, no. 5, Oct. 2004.
[1] R. Y. Chiao and L. J. Thomas, “Analytic evaluation of sampled
aperture ultrasonic imaging techniques for NDE,” IEEE Trans. Ul-
trason. Ferroelectr. Freq. Control, vol. 41, no. 4, pp. 484–493, 1994.
[2] M. Soumekh, “Array imaging with beam-steered data,” IEEE Trans. Martin H. Skjelvareid (S’09) was born 1982 in
Image Process., vol. 1, no. 3, pp. 379–390, Jul. 1992. Steigen, Norway. He received his M.Sc. degree in
[3] J. B. Bednar, “A brief history of seismic migration,” Geophysics, vol. electronics and acoustics from the Norwegian Uni-
70, no. 3, pp. 3MJ–20MJ, 2005. versity of Science and Technology in 2008, and is
[4] R. H. Stolt, “Migration by Fourier transform,” Geophysics, vol. 43, currently a Ph.D. degree candidate at the Univer-
no. 1, pp. 23–48, Feb. 1978. sity of Tromsø, Norway. He is also employed at
[5] J. Gazdag, “Wave equation migration with the phase-shift method,” Breivoll Inspection Technologies, where he is
Geophysics, vol. 43, no. 7, pp. 1342–1351, 1978. working on methods for inline pipeline inspection
[6] J. W. Goodman, Introduction to Fourier Optics, 2nd ed., New York, using ultrasound. His current research interests
NY: McGraw-Hill, 1996. include ultrasonic thickness gauging, ultrasonic
[7] E. G. Williams, Fourier Acoustics. New York, NY: Academic, 1999. imaging, and signal processing.
[8] S. Doctor, T. Hall, and L. Reid, “SAFT—The evolution of a signal
processing technology for ultrasonic testing,” NDT Int., vol. 19, no.
3, pp. 163–167, Jun. 1986.
[9] K. Langenberg, M. Berger, T. Kreutter, K. Mayer, and V. Schmitz,
Tomas Olofsson was born 1968 in Sandviken,
“Synthetic aperture focusing technique signal processing,” NDT Int.,
Sweden. He received his M.Sc. degree in engineer-
vol. 19, no. 3, pp. 177–189, Jun. 1986.
ing physics in 1994 and his Ph.D. degree in signal
[10] K. Mayer, R. Marklein, K. Langenberg, and T. Kreutter, “Three-
processing in 2000, both from Uppsala University.
dimensional imaging system based on Fourier transform synthetic
He is currently working as an associate professor
aperture focusing technique,” Ultrasonics, vol. 28, pp. 241–255, Jul.
in the Signals and Systems group, Uppsala Uni-
1990.
versity, Sweden. His research concerns inference
[11] L. J. Busse, “Three-dimensional imaging using a frequency-domain
problems, in particular inverse problems in ultra-
synthetic aperture focusing technique,” in IEEE Trans. Ultrason.
sonics.
Ferroelectr. Freq. Control, vol. 39, no. 2, pp. 174–179, Mar. 1992.
[12] Y.-F. Chang and C.-C. Chern, “Frequency-wavenumber migration
of ultrasonic data,” J. Nondestruct. Eval., vol. 19, no. 1, pp. 1–10,
2000.
[13] T. Stepinski, “An implementation of synthetic aperture focusing Yngve Birkelund (S’99–M’03) received the
technique in frequency domain,” IEEE Trans. Ultrason. Ferroelectr. M.Sc. degree in electrical engineering and the
Freq. Control, vol. 54, no. 7, pp. 1399–1408, Jul. 2007. Ph.D. degree in physics from the Department of
[14] N. Pörtzgen, D. Gisolf, and G. Blacquière, “Inverse wave field ex- Physics and Technology, at the University of
trapolation: A different NDI approach to imaging defects,” IEEE Tromsø, Norway, in 1999 and 2003, respectively.
Trans. Ultrason. Ferroelectr. Freq. Control, vol. 54, no. 1, pp. 118– Dr. Birkelund joined the Faculty of Engineering
127, Jan. 2007. and Economics, Tromsø, University College, in
[15] A. J. Hunter, B. W. Drinkwater, and P. D. Wilcox, “The wavenum- 2003 as an Associate Professor, and became the
ber algorithm for full-matrix imaging using an ultrasonic array,” head of the Control System Bachelor Program in
IEEE Trans. Ultrason. Ferroelectr. Freq. Control, vol. 55, no. 11, pp. 2004. In 2007, he returned to his current position
2450–2462, Nov. 2008. at the University of Tromsø, Tromsø, Norway, as
[16] M. A. Haun, D. L. Jones, and W. D. O’Brien, Jr., “Adaptive focus- an Associate Professor in the Department of Physics and Technology. His
ing through layered media using the geophysical “time migration” area of research include signal processing, statistics, non-linear dynam-
concept,” in 2002 IEEE Ultrasonics Symp., vol. 2, pp. 1635–1638. ics, systems identification, time- frequency analysis, microwaves te-
[17] M. H. Skjelvareid and Y. Birkelund, “Ultrasound imaging using chonology for medical applications and ultrasonics for non-destructive
multilayer synthetic aperture focusing,” in ASME 2010 Pressure testing.
Vessels and Piping Conf., vol. 5, pp. 379–387.
[18] T. Olofsson, “Phase shift migration for imaging layered objects and
objects immersed in water,” IEEE Trans. Ultrason. Ferroelectr. Freq.
Control, vol. 57, no. 11, pp. 2522–2530, Nov. 2010. Yngvar Larsen (S’00–M’04) received the M.Sc.
[19] Y. C. Kim, R. Gonzalez, and J. R. Berryhill, “Recursive wavenum- degree in applied physics and the Ph.D. degree in
ber-frequency migration,” Geophysics, vol. 54, no. 3, pp. 319–329, physics from the University of Tromsø, Tromsø,
Mar. 1989. Norway, in 1999 and 2003, respectively. During
[20] J. F. Claerbout, Imaging the Earth’s Interior. Cambridge, MA: the summer of 1999, he was a Research Assistant
Blackwell Scientific Publications, 1985. with the NATO SACLANT Undersea Research
[21] G. M. Margrave, Numerical Methods in Exploration Seismology— Centre, La Spezia, Italy. From July 2001 to June
With Algorithms in MATLAB. University of Calgary, 2003, [Online]. 2002, he was a Visiting Researcher with the Uni-
Available: http://www.crewes.org/ResearchLinks/FreeSoftware/ versity of Minnesota, Minneapolis, MN. He is cur-
NumMeth.pdf rently a Senior Researcher with Norut. His current
[22] L. Brekhovskikh and O. Godin, Acoustics of Layered Media I: Plane research interests include signal processing for
and Quasi-Plane Waves, Springer Series on Wave Phenomena. New SAR, particularly the applications of interfero-
York, NY: Springer-Verlag, 1990. metric SAR.