Research Article Vol. 58, No.

34 / 1 December 2019 / Applied Optics G11

Non-approximated Rayleigh–Sommerfeld
diffraction integral: advantages and
disadvantages in the propagation of
complex wave fields
CARLOS BUITRAGO-DUQUE AND JORGE GARCIA-SUCERQUIA*
School of Physics, Universidad Nacional de Colombia—Sede Medellin, A.A: 3840, Medellin 050034, Colombia
*Corresponding author: [email protected]

Received 19 June 2019; revised 28 July 2019; accepted 29 July 2019; posted 29 July 2019 (Doc. ID 370601); published 4 September 2019

Advantages and disadvantages of the non-approximated numerical implementation of the Rayleigh–Sommerfeld

diffraction integral (RSD) are revisited. In this work, it is shown that as trade-off for its large computation load, the
non-approximated RSD removes any limitation on the propagation range and does not introduce any artifact in the
computed wave field. A non-approximated GPU implementation of the RSD is contrasted with the angular spectrum,
the Fresnel transform, and a fast Fourier transform implementation of the RSD. The forecasted phase shift introduced
in the propagated wave fields as light is diffracted on complementary apertures and utilized as a metric to quantify the
performance of the tested methods. An application to numerical reconstructions with arbitrary shape and size of digital
recorded holograms from digital lensless holographic microscopy is presented. © 2019 Optical Society of America
https://doi.org/10.1364/AO.58.000G11

1. INTRODUCTION sampling of the terms involved in the computation has to be guar-

The most striking current applications of optics rely on the anteed, which introduces limitations on the range of validity of the
numerical propagation of wave fields. Numerical reconstruction different methods of numerical propagation [15–18].
of digitally recorded holograms, coined digital holography [1,2], Among the multiple methods of numerical propagation of
digital holographic microscopy (DHM) [3,4], Fourier ptycho- wave fields, up to the best knowledge of the authors, there is
graphic imaging [5,6], compressive sensing [7], optical encryp- no reported use of the non-approximated, non-Fourier imple-
tion [8,9], computer-generated holograms [10], and numerical mented Rayleigh–Sommerfeld diffraction (RSD) integral. Most
compensation of aberrations in optical systems [11], can be of the publications that make use of the RSD, utilize its convo-
named as examples. For these types of applications, the numeri- lution implementation [15,19–21], which bounds the pixel size
cal propagation of complex wave fields takes place in isotropic at the input and output planes to be equal, and limits its appli-
and homogeneous media, hence a scalar theory fully describes cability range; a different approach utilizes an innovative sampling
the physical phenomenon. scheme to remove the said limitation in the computation of non-
The Helmholtz equation encompasses the scalar theory of paraxial scalar diffracted wave fields [22]. In this paper, the use of
wave fields propagation [12]. In ideal problems, the Helmholtz the non-approximated, non-Fourier implemented RSD for the
equation is analytically solved by using different assumptions that numerical calculation of propagated wave fields is revisited.
lead to the methods of angular spectrum (AS) [13], Rayleigh– Despite its large computational complexity, it is shown that its
Sommerfeld and Fresnel–Kirchhoff diffraction formulae [12], implementation under the GPGPU programing paradigm makes
to name some of them. For real applications, as those mentioned its use technologically feasible, providing a limitless propagation
range between the very aperture and infinity, along with the
in the above paragraph, it is necessary to resort to numerical im-
capability of using input and output planes with arbitrary and
plementations to compute the propagation of wave fields. These
independent shape, size, and orientation.
numerical propagation methods are made technologically attrac-
tive by casting them into the form of discrete Fourier transforms,
which can be computed via a fast Fourier transform algorithm 2. PROPAGATION OF WAVE FIELDS
[14]. Similarly, as for the case of the analytical solutions, the The propagation of vector wave fields is fully described by
numerical propagation methods are meant to provide a tool means of the wave equation [12]. When the electromagnetic
for propagating wave fields from the very aperture to infinity. wave propagates through an isotropic and homogeneous
However, due to the use of Fourier’s formalism, the correct medium, at low numerical apertures (NAs) with no coupling

1559-128X/19/340G11-08 Journal © 2019 Optical Society of America

 G12 Vol. 58, No. 34 / 1 December 2019 / Applied Optics Research Article

of the energy from different polarization states, and the tech- A different approach to solve the Helmholtz equation is to con-
nique of separation of variables is used to derive the time- sider expressing the wave field U x, y, z in terms of spherical
independent solution to the wave equation, the latter turns into wavefronts. The simplest approach to this method is the
the Helmholtz equation: Huygens principle [12], expressing the propagated wavefront

2  as the amplitude superposition of the spherical wavelets pro-
∂ ∂2 ∂2



k 2
U x, y, z  0, (1) duced by any point on the input wavefront. A more elegant
∂x 2 ∂y 2 ∂z 2 way to express the propagated wave fields in terms of spherical
where k  2π∕λ is the wavenumber, with λ the wavelength of wavefronts is the use of Green functions and the appropriated
the electromagnetic scalar wave field U x, y, z. This equation boundary conditions [12,13]; Fresnel–Kirchhoff and Rayleigh–
can be solved with different methods [13]; the AS and RSD Sommerfeld diffraction integrals are two solutions of the
formula are some of the possible solutions to describe the Helmholtz equation within this framework with equivalent
propagation of scalar wave fields. Although both approaches performance in terms of accuracy in the far field. The latter
can be utilized to describe the propagation of wave fields, only expresses the propagated wave field as
in very few cases it is possible to find a fully analytical solution; Z Z

1 ∞ ∞ 1
therefore, the finding of propagated wave fields mainly relies on U x, y, z  − U x 0 , y 0 , 0 ik

2π −∞ −∞ r
numerical results from a diffraction integral that is derived from
the Helmholtz equation [2,23]. expikr
× cos χdx 0 dy 0 , (6)
The approach to solve Helmholtz equation via the AS is sup- r
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ported on considering a decomposition of the scalar wave field with r  x − x 0 2
y − y 0 2
z 2 being the distance from
U x, y, z in terms of plane waves traveling with cosine direc- one point on the input plane x 0 , y 0 , 0 to the point on the
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
tors αx  λf x , αy  λf y , and αz  1 − λ2 f 2x
f 2y , where output plane where one computes U x, y, z; χ denotes the
angle between the outward normal to output plane containing
f x and f y are the corresponding spatial frequencies. Hence, by
x, y, z and the position vector directed from the point
understanding the Fourier transform as the weighting factor of x 0 , y 0 , 0 to the point x 1 , y 1 , z, such that cos χ  z∕r.
a set of plane waves that synthetize a given function, the like- On considering the explicit form of the distance r, Eq. (6) al-
lihood of expressing a propagated wave field within the Fourier lows the understanding of the free space propagation as a linear
transform framework can be indicated. Specifically, a propa- space-invariant operation [13,24]. Hence, the convolution
gated scalar wave field U x, y, z is the result of the propagation theorem can be used to compute the propagated wave field
of an input scalar wave field U x 0 , y 0 , 0 at the input plane via Fourier transforms [20,21]:
x, y, 0. The AS of the input wave field,
Z ∞Z ∞ U x, y, z
Af x , f y , 0  U x 0 , y 0 , 0  
 
1 −1 1 expikr 0  z
−∞ −∞ − F F fU x 0 , y 0 , 0gF ik
0 ,
2π r r0 r0
× exp−i2πf x x 0
f y y 0 dx 0 dy 0 , (2)
(7)
is weighted by the transfer function of the free space to reach pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
the output plane x, y, z, yielding the AS of the propagated with r 0  x 2
y 2
z 2 ; this approach can be named the
wave field: convolution RSD (CRSD). On comparing Eqs. (5) and (7)

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi the reader can realize the direct equivalence between AS and
Af x ,f y ,z  Af x ,f y ,0exp ikz 1 − λ2 f 2x
f 2y  : (3) CRSD, because the analytical expression of the last Fourier
transform of Eq. (7) is exactly the propagation kernel of the
Equation (3) indicates that the only waves that can be propa- AS in Eq. (5). This analytical equivalence, derived from the
gated to the far field are those whose spatial frequencies fulfill convolutional approach of both methods, is not evident in
the condition 1∕λ2 ≥ f 2x
f 2y ; waves with spatial frequencies the numerical performance of them. This difference in perfor-
outside of this circle with radius 1∕λ in the spectrum domain mance of the AS and CRSD is due to the fact that while in the
constitute the set of evanescent waves [13]. From Eq. (3) the former an analytical expression of the amplitude transfer func-
propagated wave field can be therefore computed through an tion is used, in the latter the amplitude transfer function is the
inverse Fourier transform, numerical Fourier transform of the impulse response. This
Z ∞Z ∞ numerical Fourier transform of the impulse response introduces
U x, y, z  Af x , f y , z exp−2πf x x
f y ydxdy, numerical artefacts, inherited, for instance, from data truncation,
−∞ −∞ which lead to an overall different numerical performance of the
(4) analytically equivalent AS and CRSD.
A widely used method to compute propagated wave fields is
which in terms of the Fourier transform operator F f•g can be the Fresnel–Fraunhofer (FF) paraxial approximation [13,23].
readily written as This approach considers a Taylor expansion of the distance r,
U x,y,z limiting its range of application to the paraxial region where the

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi distance z between the input and output planes fulfills the
π
F −1 F fU x 0 ,y 0 ,0gexp ikz 1 − λ2 f 2x
f 2y  : (5) condition z 3 ≫ 4λ x 0 − x2
y 0 − y2 2MAX . The propagated
wave field can be computed as
 Research Article Vol. 58, No. 34 / 1 December 2019 / Applied Optics G13


−i expikz ik 2 ◯M 2 × R P. P
2
The casting of the Fourier transform operator
U x, y, z  exp x
y 2 
λz 2z F f•g  m n • exp − i2πmΔx 0 rΔf x
nΔy 0 sΔf y  into a
Z ∞Z ∞
 P P
ik 2 discrete Fourier transform kernel DFTf•g  m n •
× U x 0 , y0 , 0 exp x
y 0 
2
exp − i2πmr∕M
ns∕N , reduces that computational com-
−∞ −∞ 2z 0

 plexity to ◯2 × M log M × R log R. This casting imposes
−ik Δx 0 Δf x  1∕M and Δy0 Δf y  1∕R such that the propa-
× exp x x
y0 y dx 0 dy 0 , (8)
z 0 gated wave field via the AS yields
where one can recognize the computation of the propagated U pΔx,qΔy,z
wave field as a scaled Fourier transform: 2 3

 DFTU mΔx 0 ,nΔy 0 ,0
x y −i expikz ik 2 6 !
U , ,z  exp x
y2  
1
DFT−1 6  2  2  7
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 7:
λz λz λz 2z MR 4 r s 5
  ×exp iz λ 1−λ
2π 2
M Δx 0
N Δy
ik 2 0

× F U x 0 , y0 , 0 exp x 0
y20  : (9)

2z (12)
The equations presented above to compute the propagated
The use of the formalism of the AS imposes that the size of the
wave fields must be numerically implemented to their practical
pixels at the output plane equals that of the input plane. This
use.
feature limits the possibility of producing a magnification of the
propagated wave field with respect to the input one. As for pro-
3. NUMERICAL CALCULATION OF WAVE viding the AS with magnification, variations to the description
FIELDS PROPAGATION above presented can be found in the literature, which include,
for instance, zero padding [28–30] or the use of a chirp z trans-
The numerical calculation of the wave fields propagation requires
form operation [22].
the complete discretization of the expression for its computation;
In particular interest for the present work is the discretiza-
input and output planes, either in spatial or spatial frequency co-
tion of the RSD Eq. (6); the direct discretization leads to
ordinates, the known input wave field U x 0 , y 0 , 0, and the


propagation kernels that can include the Fourier transform oper- 1 X M X N
1
ator, have to be discretized. The set of constraints on data sam- U pΔx, qΔy, z  − U mΔx 0 , nΔy0 , 0 ik

2π m1 n1 r
pling and propagation kernels are analyzed elsewhere [22,25–27].
The procedure is illustrated with the discretization of the expikr z
× Δx Δy , (13)
numerical computation of the propagated wave fields via the r r 0 0
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
AS, namely Eq. (5). The continuous spatial coordinates at with r  pΔx − mΔx 0 2
qΔy − rΔy 0 2
z 2 . This
the input plane are discretized on a rectangular grid with pixel equation shows the possibility of computing one point of
size Δx 0 × Δy 0 , such that x 0  mΔx 0 and y 0  nΔy 0 , m and n the propagated wave field by means of OM × N  complex-
being integer numbers; correspondingly for the output plane valued operations. It follows that for an output plane of
x  pΔx and y  qΔy, with p and q integer numbers. R × Q pixels, the total of the complex-valued operations is
Accordingly, the spatial frequencies are also discretized as f x  OM × N × R × Q; this number is too large for most practical
rΔf x and f y  sΔf y . By introducing these discretizations on uses, rendering it to non-practical elapsed times of computing
Eq. (5), the propagated wave field is computed as of the order of hours, which has discouraged its use in the
U pΔx, qΔy, z present way. Despite this negative feature, the expression of
the propagated wave field as a direct summation allows its com-
R X
X Q
putation over areas with size and shape chosen completely at
 ArΔf x , sΔf y , 0 expiμz
r1 s1
will, which can sensibly reduce the computation elapsed time.
As was already pointed out in the result of Eq. (7), the dis-
× exp i2πpΔxrΔf x
qΔysΔf y Δf x Δf y , (10) tance between the output and the input points of calculation
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi allows the computation of the propagated wave fields via a con-
with μ  2π∕λ 1 − λ2 rΔf x 2
sΔf y 2  being the trans-
volution operation [15,19,20]. Indeed, this approach to the
fer function, R and Q the number of pixels along each direction computation of the propagated wave field is quite similar to
of the output plane, and the AS of the input wave field the AS, with the difference that in this case the transfer function
ArΔf x , sΔf y , 0 is numerically computed via a Fourier transform. Consequently,
the complete calculation can be casted to discrete Fourier trans-
X
M X
N
forms to use FFTs and therefore the size of the input and output
 U mΔx 0 , nΔy 0 , 0
planes has to be equal:
m1 n1
2 3
× exp −i2πmΔx 0 rΔf x
nΔy 0 sΔf y Δx 0 Δy 0 , (11) 1 DFTU mΔx 0 ,nΔy 0 ,0
U pΔx,qΔy,z  − DFT−1 4 h
i 5:
2πM R z
where M and N are the number of pixels along each direction ×DFT ik
1r expikr r r
of the input plane. For the particular case of square input and
output planes, the computation complexity of Eq. (10) is (14)
 G14 Vol. 58, No. 34 / 1 December 2019 / Applied Optics Research Article

With respect to the AS approach, this convolution method M Δx 20

exhibits a larger computational complexity of ◯3 × Zc  : (17)
λ
M log M × R log R. Also, as is the case for the AS, different For the two-dimensional case a similar expression is obtained
methods have been proposed in the literature to allow the propa- for the y-direction. Despite the wide acceptance of the former
gated wave fields to have a different scale size than those of the limit, it has been shown that this figure must be finely tuned.
input wave fields [31,32].
The revisited approach, obtained by sampling a 2π-phase jump
A similar procedure of discretization can be applied to the
with 3 pixels, leaves a gap between the application limits for the
calculation of the propagated wave fields by means of the FF,
AS and the FF and CRSD [18]. The non-overlapping of the
Eq. (9). The details can be read elsewhere [15,23], with the
valid propagation distances supposes an additional difficulty
result being

 for numerical diffraction applications, as the accuracy of the
pΔx qΔy results cannot be trusted with either method inside this critical
U , ,z
λz λz region. Owing to the non-approximated non-convolution RSD



 not using computation through the Fourier formalism, it is en-
−i expikz ik pλz 2 qλz 2
 exp
visioned not to have the formerly said limits for the other
λz 2z M Δx 0 N Δy 0
propagation methods.
× DFTU mΔx 0 , nΔy 0 , 0 To test the range of validity of the presented methods in this
ik work, a metric for evaluating their limits of application has been
× exp mΔx 0 2
nΔy 0 2 Δx 0 Δy 0 : (15) set by using Babinet’s principle for complementary apertures.
2z
This principle states that the sum of the amplitudes of the dif-
Equation (15) has the particular feature that the pixel size at the fracted wave fields produced by an obstacle U 1 x 1 , y 1 , z and its
output plane is determined by the experiment conditions such complementary aperture U 2 x 1 , y 1 , z equals that recorded when
that Δx  λz∕M Δx 0 and Δy  λz∕N Δy 0 . The last figures the wave field propagates in the free space U 0 x 1 , y 1 , z [12]. If
mean that for a given propagation distance z, there is a fixed the complementary apertures are taken in such a way that their
size relation between the input and output planes; this limita- superposition represents a circular aperture subtending an even
tion has been alleviated by different approaches [33–35]. In number of Fresnel’s zones in the observation plane [12], their
terms of its computational complexity for square input and amplitude sum in the optical axis will yield U 1 0, 0, z

output planes it is reduced to ◯M log M × R log R, approx- U 2 0, 0, z  0. This last relation translates into an analytical
imately the half of that required for the AS method. prediction that reads U 1 x 1 , y 1 , z  −U 2 x 1 , y 1 , z; that is,
the corresponding diffracted wave fields will have identical mag-
4. RANGE OF VALIDITY FOR NUMERICAL nitude and a π-phase difference in the optical axis, regardless of
IMPLEMENTATIONS OF SCALAR DIFFRACTION the screen shape. This relation is summarized in Fig. 1 using a
In terms of the propagation distance, the analytical formulae for superposition circle with a radius of 1 mm and an intricate in-
computing propagated wave fields presented in Section 2 are ternal shape for the complementary apertures, as presented in the
valid from the very aperture placed at the input plane to the three left upper panels. The lower panels represent the wave field
infinity. However, as their numerical implementations are com- amplitude diffracted by each aperture illuminated by plane waves
puted via the formalism of the Fourier transform, the sampling with a wavelength of 405 nm and propagated to 100 mm using a
theorem for all the involved terms has to be fulfilled [15]. The numerical implementation of FF. The amplitude superposition
correct sampling introduces constrains that relate the illuminat- matches the propagation of the circular aperture and displays
ing wavelength, the size and number of pixels at the input the expected dark spot in the optical axis, zoomed-in in the
plane, and the propagation distance. In turn, the latter is no right-most panel in the upper row.
longer valid from the very aperture to the infinity; rather The fulfilling of the aforementioned predicted phase shift can
the range of validity of the propagation distance is restricted thus be taken as the metric for the range of validity of the pres-
for the different methods to guarantee the correct sampling. ently studied methods. To do so, the phase difference between
By examining the number of pixels required to sample a the numerically propagated wave fields of the complementary
π-phase jump at the largest spatial frequency in all the methods, apertures is measured in the optical axis while the propagation
one can find a critical distance that splits the range of valid use distance is changed. The radius of the circular aperture is con-
of each approach. This condition of the phase jump sampling is sequently modified for each distance to ensure that an even num-
expressed for the largest spatial frequency as ber of Fresnel’s zones are being subtended. A more detailed
2sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi3 description of this testing method can be read elsewhere [18].

2
2
λ 4 1 2 M ∕2 − 1 5, (16) This predicted phase difference is used to evaluate the range
≥ 1−λ 2
− 1−λ
2Z c 2Δx 0 M Δx 0 of validity of the AS, FF, RSD, and CRSD propagation meth-
ods. The results are summarized in Fig. 2, with the normalized
where M is the number of pixels along the x-direction and Δx 0 distance against the limit proposed in literature, as presented in
its pixel size. Considering the Taylor’s expansion of the square Eq. (17), and the phase difference normalized to the expected
root and some direct algebra, one arrives to the result that con- value (π).
stitutes an upper limit for the AS and a lower limit for FF and The highlighted regions in Fig. 2 emphasize the ranges
the CRSD. This result that provides the critical propagation where negligible disturbances respect to the forecasted phase
distance Z c is given by [16,17,36] shift are achieved with each method. Unlike the others,
 Research Article Vol. 58, No. 34 / 1 December 2019 / Applied Optics G15

Fig. 1. Babinet’s principle of complementary apertures using a circle of 1 mm as free space. The apertures in the left-most upper panels were
illuminated with plane waves of 405 nm and propagated to 100 mm using Fresnel–Fraunhofer diffraction, such that the circle subtended an even
number of Fresnel’s zones. The sum of the obstacle’s wave field U 1 and the aperture’s wave field U 2 appropriately matches the free space propagation
U 0 and present the expected dark spot in the optical axis shown in the right-most upper panel.

Fig. 2. Normalized phase difference of the propagated wave fields of the aperture and its complement as the propagation distance changes, with
Z c  M Δx 20 ∕λ. The highlighted distance ranges emphasize the methods with negligible disturbances. The RSD shows an even behavior all over the
propagation distance.

the RSD has no major deviations from the expected value at phase map via the use of the RSD shown in Fig. 3(e) confirms
any point over the complete propagation distance, from the that it can recover the information entirely and to an excellent
very aperture to the far field. This unique feature shows that, extent. No phase distortion is introduced by the RSD as it can
despite its long-elapsed time of computation, the RSD is the be further seen by comparing the insets of each panel, in which
only propagation method that can be used all over the propa- a surface plot of the region bounded by the yellow square
gation range with no major distortion of its accuracy. is shown.
To further explore the accuracy of the propagation methods Despite its accurate calculations and full-range applicability,
in phase retrieval, an optical pure phase wave field was numeri- the RSD has a computational complexity ON 4  for input and
cally modeled. It was propagated near to Z c , and backpropa- output planes of N × N . This becomes its main setback, as even
gated to the initial plane using each of the studied methods. a medium-sized image needs more processing time than most
Figure 3 shows the results; the modeled pure phase wave field practical applications may allow. It has been shown in this work
with 1024 × 1024 pixels is illustrated in Fig. 3(a). The utilized that the attempts to overcome this difficulty via its implemen-
propagation distance was z  61.2 mm with λ  405 nm and tation as a convolution CRSD present difficulties inhered from
Δx 0  5.2 μm. The results for the AS, FF, CRSD, and RSD the Fourier formalism. However, to keep the calculation com-
are shown in Figs. 3(b), 3(c), 3(d), and 3(e), respectively. All the pletely free of approximations, which renders full accuracy
evaluated methods but the RSD introduce distortions in the and range of propagations, the complexity order cannot be re-
recovered phase map. The AS highly distorts the map, ruining duced. Nonetheless, a GPGPU implementation of the RSD can
any possibility of phase measurement. FF introduces a sort of be used to increase the effective throughput and reduce the
phase tilt, and strong corruptions in the smaller details. The elapsed processing time. Taking advantage of the installed power
CRSD recovers the overall shape of the original map, but it in a regular computer hosting a Graphics Processing Unit
introduces a great number of local distortions. The retrieved GeForce GTX 580 by NVIDIA, the parallelization of the
 G16 Vol. 58, No. 34 / 1 December 2019 / Applied Optics Research Article

Fig. 3. Recovering of a phase object at a given distance. (a) Initial phase distribution. Recovered phase after propagating near to Z c  M Δx 20 ∕λ
and back using (b) angular spectrum, (c) Fresnel–Fraunhofer transform, (d) convolution Rayleigh–Sommerfeld, (e) Rayleigh–Sommerfeld diffrac-
tion integral. The insets in each panel plot the region bounded by the yellow square.

RSD can achieve a 170-time reduction of computation time the use of the RSD on the reconstruction of DLHM
compared with a regular CPU-based calculation on an Intel holograms, a self-assembled monolayer of micrometer-sized
Core i7-2600. With this GPU-based implementation of the polystyrene spheres was illuminated with a point-source of λ 
RSD, it turns into a full-range, full-accurate way of numerical 405 nm at a distance of 149 μm from it. On a square comple-
computing diffraction integrals within technologically attractive mentary metal–oxide–semiconductor sensor with side length of
timeframes, as can be shown in the application to numerical 12.3 mm located 15 mm away from the illumination source,
reconstruction of digitally recorded holograms from digital lens- the DLHM hologram was recorded. The reconstruction of the
less holographic microscopy. full-size 1024 × 1024 hologram in a CPU-based implementa-
tion of the RSD could take an overwhelming 6 days and a half
of processing time, which is far from being technologically fea-
5. APPLICATION TO NUMERICAL sible. However, introducing the GPU-based implementation
RECONSTRUCTION OF DIGITALLY RECORDED achieves an effective 170-fold reduction for the full-size
HOLOGRAMS reconstruction, diminishing the total computation time to
Digital holographic microscopy has powered science and tech- 45 min. This resulting reconstruction in intensity is displayed
nology with the capability of accessing the amplitude and phase in Fig. 4(a).
at microscopic scale. The simplest architecture available for A further improvement in the processing time is inherited
DHM is that supported by the microscopy principle proposed from the complete independence that the RSD allows between
by Gabor [37] implemented in the digital world. In digital lens- the size and number of points of the input and output planes. If
less holographic microscopy (DLHM) [38,39] the sample is one has a prior knowledge of the spatial distribution of the re-
illuminated by a spherical wavefront to record on a digital cam- constructed object, gained, for instance, from a faster but approxi-
era the amplitude superposition of the unperturbed portion of mated method, it is possible to limit the computation to a region
the spherical wave and the light scattered by the specimen. To of interest (ROI) which can be of arbitrary shape, orientation, and
ensure a high spatial resolution, the holograms must be re- even contain multiple propagation distances. Taking advantage of
corded with high NAs [4], which introduces an unavoidable this unique feature of the RSD, the total processing time would
technical difficulty for the backpropagation process that is be proportional to the amount of points contained within the
required to retrieve the specimen information; most of the ROI, with each point being calculated in just a couple of milli-
numerical propagation algorithms, as previously stated, cannot seconds. Applying this idea, an unevenly shaped ROI in the des-
modify at will the size of the propagated wave field limiting tination plane was selected in the area outlined in yellow in
their application to high-NA propagations. Fig. 4(a). With the input image being 1024 × 1024 pixels in size,
The use of the RSD, due to its lack of approximations, the individual computation of each data point in the destination
allows the reconstruction of DLHM holograms without further plane takes around 3.5 ms; thus, for the selected ROI containing
modifications to the algorithm or the input field. For testing 8.6 thousand pixels, the reconstruction can be obtained in 30 s,
 Research Article Vol. 58, No. 34 / 1 December 2019 / Applied Optics G17

Fig. 4. Reconstruction of a DLHM hologram of a self-assembled monolayer of micrometer-sized polystyrene spheres. (a) Full-size intensity
reconstruction of 1024 × 1024 hologram. The region outlined in yellow, originally bounded by 141 × 141 pixels, was independently reconstructed
into a 512 × 512 image in (b) intensity and (c) phase. The same full-size DLHM hologram has been reconstructed via the AS in panel (d) and by
CRSD in panel (e).

which is a technologically attractive time. Additionally, as the area is not available for the two latter methods, indicating an extra
output plane discretization is completely independent from the reason to use the RSD approach for the reconstruction of this type
input parameters in the RSD, it is possible to reconstruct this of hologram, when the computation time has no importance.
same ROI into a higher pixel count to improve the visualization
quality. Figures 4(b) and 4(c) show the reconstruction in intensity
and phase, respectively, in an extended version of the aforemen- 6. CONCLUSIONS
tioned ROI now containing 120 thousand pixels whose compu- The Rayleigh–Sommerfeld diffraction integral has been revis-
tation takes a total of approximately 4 min; as reference, the ited within the framework of the numerical calculation of
resulting image is of 512 × 512 pixels while the original ROI propagation of wave fields. The angular spectrum, direct non-
was bounded by a box of 141 × 141 pixels, showing another ex- approximated integration, convolution method, and Fresnel–
ample of the freedom of computation that the RSD provides for Fraunhofer diffraction formula have been presented as possible
the propagation of wave fields. To compare the performance of solutions of the scalar Helmholtz equation. The numerical im-
the RSD on the reconstruction of DLHM holograms with the plementation of the above-said methods has also been shown to
results provided by the AS and CRSD, Figs. 4(d) and 4(e) present discuss some of their particular features.
the corresponding full-size 1024 × 1024 intensity reconstruc- An emphasis has been focused on studying the range of valid
tions, respectively. Contrasting panels (a), (d), and (e), one can computation for each method. Based on the analysis of the cor-
see the superior performance of the RSD over the AS and rect sampling of the propagation kernel for each method, its
CRSD in terms of the quality of the reconstructed images. range of use has been recalled. The validation of the range
Furthermore, the possibility of choosing at will the reconstruction has been done by testing the Babinet´s forecasted phase shift
 G18 Vol. 58, No. 34 / 1 December 2019 / Applied Optics Research Article

introduced in the diffraction patterns of complementary 15. J. Li and P. Picart, “Calculating diffraction by fast Fourier transform,” in
Digital Holography (Wiley, 2012), pp. 77–114.
apertures. The results indicate that among the tested methods,
16. D. Mendlovic, Z. Zalevsky, and N. Konforti, “Computation considera-
the only one that reproduces the forecasted phase shift from tions and fast algorithms for calculating the diffraction integral,”
the input to the output plane with no error is the non- J. Mod. Opt. 44, 407–414 (1997).
approximated Rayleigh–Sommerfeld diffraction integral. 17. D. Mas, J. Garcia, C. Ferreira, L. M. Bernardo, and F. Marinho, “Fast
In addition to the exact calculation of the propagated wave algorithms for free-space diffraction patterns calculation,” Opt.
Commun. 164, 233–245 (1999).
fields, the non-approximated Rayleigh–Sommerfeld diffraction 18. R. Castañeda, W. Toro, and J. Garcia-Sucerquia, “Evaluation of the
integral enables the computation between input and output limits of application for numerical diffraction methods based on basic
planes with arbitrary size, shape, and orientation. The trade- optics concepts,” Optik 126, 5963–5970 (2015).
off for all the here-mentioned features is the large computational 19. J. A. C. Veerman, J. J. Rusch, and H. P. Urbach, “Calculation of the
complexity it exhibits. The use of the GPGPU programming Rayleigh-Sommerfeld diffraction integral by exact integration of the
fast oscillating factor,” J. Opt. Soc. Am. A 22, 636–646 (2005).
paradigm eases the elapsed time of computing, setting the option 20. V. Nascov and P. C. Logofătu, “Fast computation algorithm for the
of numerical calculation of propagated wave fields with the non- Rayleigh-Sommerfeld diffraction formula using a type of scaled con-
approximated Rayleigh–Sommerfeld diffraction integral within a volution,” Appl. Opt. 48, 4310–4319 (2009).
technologically attractive framework. 21. F. Shen and A. Wang, “Fast-Fourier-transform based numerical inte-
The non-approximated Rayleigh–Sommerfeld diffraction in- gration method for the Rayleigh-Sommerfeld diffraction formula,”
Appl. Opt. 45, 1102–1110 (2006).
tegral has been applied to numerical reconstruction of digitally 22. M. Hillenbrand, D. P. Kelly, and S. Sinzinger, “Numerical solution of
recorded holograms from digital lensless holographic microscopy. nonparaxial scalar diffraction integrals for focused fields,” J. Opt. Soc.
The reconstructed images have spatial resolution in the microm- Am. A 31, 1832–1841 (2014).
eter-size range at variable reconstruction fields of view, showing 23. O. K. Ersoy, Diffraction, Fourier Optics and Imaging (Wiley, 2006).
24. U. Schnars and W. Juptner, “Digital recording and numerical,” Inst.
some of the versatility provided by this method of computing
Phys. Publ. 13, 17 (2002).
propagated wave fields. 25. D. G. Voelz and M. C. Roggemann, “Digital simulation of scalar optical
diffraction: revisiting chirp function sampling criteria and conse-
Funding. Universidad Nacional de Colombia—Sede quences,” Appl. Opt. 48, 6132–6142 (2009).
Medellin. 26. D. P. Kelly, “Numerical calculation of the Fresnel transform,” J. Opt.
Soc. Am. A 31, 755–764 (2014).
27. L. Onural, “Sampling of the diffraction field,” Appl. Opt. 39, 5929–5935
REFERENCES (2000).
1. U. Schnars, “Direct phase determination in hologram interferometry 28. T. Shimobaba, K. Matsushima, T. Kakue, N. Masuda, and T. Ito,
with use of digitally recorded holograms,” J. Opt. Soc. Am. A 11, “Scaled angular spectrum method,” Opt. Lett. 37, 4128–4130 (2012).
2011–2015 (1994). 29. T. Shimobaba, Y. Nagahama, T. Kakue, N. Takada, N. Okada,
2. P. Picart and L. Juan-chang, Digital Holography (Wiley, 2012). Y. Endo, R. Hirayama, D. Hiyama, and T. Ito, “Calculation reduction
3. E. Cuche, F. Bevilacqua, and C. Depeursinge, “Digital holography method for color digital holography and computer-generated holo-
for quantitative phase-contrast imaging,” Opt. Lett. 24, 291–293 gram using color space conversion,” Opt. Eng. 53, 24108 (2014).
(1999). 30. T. Shimobaba, J. Weng, T. Sakurai, N. Okada, T. Nishitsuji, N.
4. M. K. Kim, Digital Holographic Microscopy. Principles, Techniques, Takada, A. Shiraki, N. Masuda, and T. Ito, “Computational wave op-
and Aplications (Springer, 2011). tics library for C++: CWO++ library,” Comput. Phys. Commun. 183,
5. K. Giewekemeyer, M. Beckers, T. Gorniak, M. Grunze, T. Salditt, and 1124–1138 (2012).
A. Rosenhahn, “Ptychographic coherent x-ray diffractive imaging in 31. M. Leclercq and P. Picart, “Method for chromatic error compensation
the water window,” Opt. Express 19, 1037–1050 (2011). in digital color holographic imaging,” Opt. Express 21, 26456–26467
6. G. Zheng, R. Horstmeyer, and C. Yang, “Wide-field, high-resolution (2013).
Fourier ptychographic microscopy,” Nat. Photonics 7, 739–745 32. D. Claus and J. M. Rodenburg, “Pixel size adjustment in coherent dif-
(2013). fractive imaging within the Rayleigh–Sommerfeld regime,” Appl. Opt.
7. P. Clemente, V. Durán, E. Tajahuerce, P. Andrés, V. Climent, and J. 54, 1936–1944 (2015).
Lancis, “Compressive holography with a single-pixel detector,” Opt. 33. J. F. Restrepo and J. Garcia-Sucerquia, “Magnified reconstruction of
Lett. 38, 2524–2527 (2013). digitally recorded holograms by Fresnel-Bluestein transform,” Appl.
8. E. Tajahuerce and B. Javidi, “Encrypting three-dimensional informa- Opt. 49, 6430–6435 (2010).
tion with digital holography,” Appl. Opt. 39, 6595–6601 (2000). 34. F. Zhang, I. Yamaguchi, and L. P. Yaroslavsky, “Algorithm for
9. T. Nomura and B. Javidi, “Securing information using a digital holo- reconstruction of digital holograms with adjustable magnification,”
graphic technique,” Opt. Comput. 4089, 619–624 (2000). Opt. Lett. 29, 1668–1670 (2004).
10. W. Dallas, “Computer-generated holograms,” in The Computer in 35. T. Shimobaba, T. Kakue, N. Okada, M. Oikawa, Y. Yamaguchi, and T.
Optical Research, B. R. Frieden, ed. (Springer-Verlag, 1980), Ito, “Aliasing-reduced Fresnel diffraction with scale and shift opera-
Vol. 41, pp. 291–366. tions,” J. Opt. 15, 075405 (2013).
11. K. W. Seo, Y. S. Choi, E. S. Seo, and S. J. Lee, “Aberration compen- 36. M. Sypek, C. Prokopowicz, and M. Gorecki, “Image multiplying and
sation for objective phase curvature in phase holographic micros- high-frequency oscillations effects in the Fresnel region light propaga-
copy,” Opt. Lett. 37, 4976–4978 (2012). tion simulation,” Opt. Eng. 42, 3158–3164 (2003).
12. M. Born and E. Wolf, Principles of Optics, 7th ed. (Cambridge 37. D. Gabor, “Microscopy by reconstructed wave-fronts,” Proc. R. Soc.
University, 2005). London A 197, 454–487 (1949).
13. J. W. Goodman, Introduction to Fourier Optics, 3rd ed. (Roberts and 38. W. Xu, M. H. Jericho, I. A. Meinertzhagen, and H. J. Kreuzer, “Digital
Company Publishers, 2005). in-line holography for biological applications,” Proc. Natl. Acad. Sci.
14. M. Frigo and S. G. Johnson, “FFTW: an adaptive software architecture U. S. A. 98, 11301–11305 (2001).
for the FFT,” in Proceedings of the 1998 IEEE International 39. J. Garcia-Sucerquia, W. Xu, S. K. Jericho, P. Klages, M. H. Jericho,
Conference On Acoustics, Speech and Signal Processing (1998), and H. J. Kreuzer, “Digital in-line holographic microscopy,” Appl. Opt.
Vol. 3, pp. 1381–1384. 45, 836–850 (2006).