Compressive Phase Retrieval
Compressive Phase Retrieval
Compressive Phase Retrieval
by
Lei Tian
B.S., Tsinghua University (2008)
S.M., Massachusetts Institute of Technology (2010)
Submitted to the Department of Mechanical Engineering
in partial fulfillment of the requirements for the degree of
Doctor of Philosophy
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
June 2013
c Massachusetts Institute of Technology 2013. All rights reserved.
Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Department of Mechanical Engineering
May 18, 2013
Certified by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
George Barbastathis
Professor
Thesis Supervisor
Accepted by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
David E. Hardt
Chairman, Department Committee on Graduate Students
Abstract
Recovering a full description of a wave from limited intensity measurements remains
a central problem in optics. Optical waves oscillate too fast for detectors to measure
anything but timeaveraged intensities. This is unfortunate since the phase can reveal
important information about the object. When the light is partially coherent, a
complete description of the phase requires knowledge about the statistical correlations
for each pair of points in space. Recovery of the correlation function is a much more
challenging problem since the number of pairs grows much more rapidly than the
number of points.
In this thesis, quantitative phase imaging techniques that works for partially coherent illuminations are investigated. In order to recover the phase information with
few measurements, the sparsity in each underly problem and efficient inversion methods are explored under the framework of compressed sensing. In each phase retrieval
technique under study, diffraction during spatial propagation is exploited as an effective and convenient mechanism to uniformly distribute the information about the
unknown signal into the measurement space.
Holography is useful to record the scattered field from a sparse distribution of particles; the ability of localizing each particles using compressive reconstruction method
is studied. When a thin sample is illuminated with partially coherent waves, the transport of intensity phase retrieval method is shown to be effective to recover the optical
path length of the sample and remove the effect of the illumination. This technique
is particularly suitable for Xray phase imaging since it does not require a coherent
source or any optical components. Compressive tomographic reconstruction, which
makes full use of the priors that the sample consists of piecewise constant refractive
indices, are demonstrated to make up missing data. The third technique, known as
the phase space tomography (PST), addresses the correlation function recovery problem. Implementing the PST involves measuring many intensity images under spatial
propagation. Experimental demonstration of a compressive reconstruction method,
which finds the sparse solution by decomposing the correlation function into a few mutually uncorrelated coherent modes, is presented to produce accurate reconstruction
even when the measurement suffers from the missing cone problem in the Fourier
3
domain.
Thesis Supervisor: George Barbastathis
Title: Professor
Acknowledgments
First of all, I would like to express my gratitude to my advisor, Prof. George Barbastathis. He has been not only a great advisor but also a sincere friend throughout all
these years. The most important thing I learned from him for research is to always
look for answers from the basics. I am also grateful for his encouragement when I
lack of confidence, and the many great experiences outside of MIT I got thanks to
his support, such as Singapore, and many conferences.
I would like to thank my thesis committee members: Prof. Colin Sheppard, Prof.
Ramesh Raskar, and Prof. Peter So for taking time to offer valuable advice on my
work.
I would also like to thank the members at the 3D Optical Systems group for all
their help throughout my graduate school life. I joined the group with almost no
idea about optics, the whiteboard drawings and lab demos from Jose A. Dominguez
Caballero and Nick Loomis are the most memorable lessons during my first semester,
and I would like to thank for all the helps from them. Laura Waller introduced
me the TIE, and we have been collaborators and good friends ever since, which I
feel very grateful. I would like to thank Se Baeh Oh, and Zhengyun Zhang for the
insightful discussions on partial coherence and phase space and their contributions to
the phase space tomography project; Jonathan Petruccelli for all the helps he gave
me in various projects we have been working on together; Justin Lee for his helps on
compressed sensing and his hospitalities in many occasions; ChihHao Chang, Yuan
Luo, Baile Zhang, Dipanjan Bhattacharya, and Yongjin Sung for their helps in both
research and career advices; Qin Miao for sharing the Xray imaging data; Jason
Ku for his American culture lessons; Nikhil Vadhavkar, Adam Pan, and Kelli Xu
for their helps at the last stage before my defense; Seongkeun Cho for proofreading
the thesis. I also enjoyed and feel thankful for the opportunities to learn about
the nanooptics research from Nader Shaar, Tony Nichol, Satoshi Takahashi, Martin
Deterre, Chee Wee Tan, Se Young Yang, Hyungryul (Johnny) Choi, and Jeong-gil
Kim. Furthermore, I had my best memories with my Chinese friends in this group,
5
Yi Liu, ChihHung (Max) Hsieh, Yubo Duan, Yuanhao (Howard) Huang, Hanhong
Gao, YenSheng Lu, Xiaogang (Larry) Liu, Wensheng Chen, Zhi Chen and Hongyi
(Thomas) Xu, which I would never forget.
Lastly and specially, I would like to thank my family for their love and support.
Contents
1 Introduction
1.1
17
18
1.1.1
Interferometric/holographic techniques . . . . . . . . . . . . .
19
1.1.2
20
1.2
21
1.3
24
27
2.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
2.2
30
2.2.1
Forward model . . . . . . . . . . . . . . . . . . . . . . . . . .
30
2.2.2
Backpropagation method . . . . . . . . . . . . . . . . . . . .
33
2.2.3
34
2.3
Numerical simulations . . . . . . . . . . . . . . . . . . . . . . . . . .
35
2.4
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
40
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
3.2
42
3.2.1
Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
3.2.2
43
46
3.3.1
46
3.3
Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
3.3.2
3.4
51
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
53
55
4.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
55
4.2
56
4.3
Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
4.4
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
60
63
5.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
63
5.2
Forward model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
65
5.3
68
5.4
71
5.5
73
79
6.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
79
6.2
81
6.3
Numerical simulations . . . . . . . . . . . . . . . . . . . . . . . . . .
85
6.4
90
6.5
96
6.6
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
97
103
7.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.2
Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.3
7.4
119
123
10
List of Figures
1-1 (a) A coherent wave has a welldefined amplitude and phase as illustrated by the set of blue curves. (b) Partially coherent light contains a
statistical mixture of coherent fields; here, the sets of red, green, and
blue curves indicate three different coherent fields that are mixed in
space. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
18
23
24
31
2-2 The total error q using the compressive reconstruction method (solid
red) and BPM-based method (dashed blue) at different seeding densities measured by the geometrical occlusion factor Rg . . . . . . . . . .
11
36
2-3 Left: Sample holograms; middle: real part of a depth slice from the
BPMobtained focal stack; right: real part of a depth slice from the
focal stack estimated by the compressive reconstruction method. First
row: Rg = 0.011; second row: Rg = 0.096; third row: Rg = 0.35. . . .
37
39
52
3-2 Experimental measurements and results for (ac) the sample with uniform illumination, (df) the illumination modulation mask alone, and
(gk) both sample and modulation mask in place. (a,d,g) Infocus intensity measurements. (b,e,h) Intesnity difference between defocused
measurements. (f) The scalar phase reconstruction for the illumination
with only the field modulation mask in place. Sample thickness reconstructed from (c) Eq. (3.28) with uniform illumination, (i) Eq. (3.28)
with both sample and modulation mask in place, (j) Eq. (3.29), assuming a purephase sample, and (k) Eq. (3.27), assuming a sample
containing both amplitude and phase variations. . . . . . . . . . . . .
54
4-1 (a) The NLD regularizing function and (b) the magnitude of the flux
as a function of the normalized magnitude of the gradient of the phase. 58
4-2 (a) Original phase, (b) intensity derivative, (c) direct, (d) Tikhonov
regularized, (e) TVregularized, (f) Weickert function regularized and
(g) the hybrid function regularized solutions. (f) Phase crosssections
taken along the dashedline in (a). . . . . . . . . . . . . . . . . . . .
59
60
65
68
72
12
74
75
5-6 Reconstruction results for the real part of the refractive index. (a)
Fourier based TIE solver + FBP; (b) Fourier based TIE solver +
Fourier domain tomographic inversion; (c) Compressive reconstruction
by solving Eq. (5.13). The three crosssections are taken from the
three orthogonal planes going through the center of the sample. . . .
76
77
6-1 Experimental arrangement of PST for recovering the 2D mutual intensity J at the plane of incidence O. . . . . . . . . . . . . . . . . . . . .
82
6-2 Experimental arrangement of PST for recovering the 4D mutual intensity J at the plane of incidence O. A pair of cylindrical lenses oriented
perpendicularly are used to introduce astigmatism to the measurement.
Intensities are measured at planes with axial coordinate zo . . . . . . .
84
86
6-4 The first nine coherent modes of the mutual intensity in Fig. 6-3(a).
(a) Theoretical modes, and (b) LRMR estimates. . . . . . . . . . . .
87
6-5 Eigenvalues of the mutual intensity in Fig. 6-3(a). (a) Theoretical values, (b) FBP estimates, (c) LRMR estimates, and (d) absolute errors
in the LRMR estimates versus mode index.
. . . . . . . . . . . . . .
87
6-6 Oversampling rate versus relative MSE of LRMR estimates. The input
field is a GSMS with parameters I = 36 and c = 18. The noisy data
is generated with different SNR from (a) an additive random Gaussian
noise model, and (b) a Poisson noise model. . . . . . . . . . . . . . .
13
89
6-7 Experimental arrangement for 2D PST. A 1D rectangular slit is illuminated by a partially coherent light. The coherence of the illumination
is controlled by the size of the slit placed at the front focal plane of
the condenser. Free space diffraction patterns after the object slit are
recorded at multiple axial locations. . . . . . . . . . . . . . . . . . . .
90
91
6-9 (a) Real and (b) imaginary parts of the radial slices in Ambiguity
space from Fourier transforming the vectors of intensities measured at
corresponding propagation distances. . . . . . . . . . . . . . . . . . .
91
6-10 Real part of the reconstructed mutual intensity from (a) FBP; (b)
LRMR method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
92
93
94
6-13 (a) LRMR estimated coherent modes of the mutual intensities in Fig. 610(b), and (b) coherent modes of the mutual intensities in Fig. 6-12(b),
calculated via use of the van CittertZernike theorem, and assumption
of incoherent illumination. . . . . . . . . . . . . . . . . . . . . . . . .
95
97
6-15 Missing data problem in the 4D PST. The white region indicates the
locations of the measured data in the Ambiguity space, and the black
region the inaccessible points due to the limitation of the experimental
arrangement. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
99
6-16 Real part of the mutual intensity from (a) LRMR, (b) FBP reconstructions, and (c) the van CittertZernike theorem estimation. . . . . . . 100
6-17 Eigenvalues estimated from (a) LRMR, (b) FBP reconstructions, and
(c) the van CittertZernike theorem. . . . . . . . . . . . . . . . . . . 101
6-18 The first 9 coherent modes estimated from (a) LRMR, and (b) the van
CittertZernike theorem. . . . . . . . . . . . . . . . . . . . . . . . . . 102
7-3 Sampling of the SWDF using an array of three lenslets. (a) Oneto
one mapping from the SWDF to the detector coordinate according to
u = (xo lw)/(f ) as the angular spread of the SWDF is narrower
than the numerical aperture of a lenslet. (b) Multiple points in the
SWDF domain contribute to detector pixels in the crosstalk region
as the angular spread of the incident field is wider than the numerical
aperture of a lenslet, which produces the 0th order crosstalk. . . . . . 110
15
7-4 Left: highly incoherent; middle: highly coherent; and right: partially
coherent case. (a) Total output intensity is composed of (b) SWDF
term and (c) total contribution from crosstalk terms. The total cross
talk is composed of (d) 0th order crosstalk and (e) total of higher
order crosstalk. All the intensities are normalized to the maximum
value in the total output. The horizontal axis is the spatial coordinate
normalized by the width of a lenslet. . . . . . . . . . . . . . . . . . . 113
7-5 Comparison of WDF (solid red line), SWDF (dashed blue lines) and
measured intensity (dotted green lines) for (a) highly incoherent (c =
0.01w), (b) highly coherent (c = 20w), and (c) partially coherent
(c = 0.1w) incident light. . . . . . . . . . . . . . . . . . . . . . . . . 114
7-6 Error metric Rerror in solid blue curve, crosstalk power fraction Rcrosstalk
in dashed green curve, and signal broadening metric Rconv in red dotted curve as functions of the normalized coherence length of incident
light c /w. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
16
Chapter 1
Introduction
Optical waves have amplitude and phase, but light oscillates too fast for electronic
detectors to measure anything more than timeaveraged intensities (amplitude). This
is unfortunate since the phase of an optical wave can reveal important information
about the object. Transparent objects, such as most of unstained biological cells and
tissues, do not change the intensity of light passing through them, but introduce phase
delays due to variations of thicknesses and/or refractive indices. Knowledge about
these phase delays allows physical properties of the object, otherwise invisible, to
be inferred.
When light is partially coherent, a complete description of the phase of the optical wave requires a more detailed treatment. For instance, light from a lightbulb (a
partially coherent source) measured at any single point in space will fluctuate randomly on a time scale of femtoseconds. To gain quantitative information from these
fluctuations, we can use statistical tools to specify the correlations for each pair of
points in an optical wave [70, 49]. The challenge is that the number of pairs grows
much more rapidly than the number of points (e.g. 103 points have 106 possible
combinations of pairs). Another useful and intuitive description of partial coherence is via the superposition of mutually uncorrelated coherent modes [70, 49]: fully
coherent light contains only a single coherent mode; while partially coherent light
consists of a mixture of coherent modes at any single point in space, as illustrated
in Fig. 1-1. Partially coherent illumination is widely used in modern imaging and
17
(a)
(b)
Figure 1-1: (a) A coherent wave has a welldefined amplitude and phase as illustrated
by the set of blue curves. (b) Partially coherent light contains a statistical mixture
of coherent fields; here, the sets of red, green, and blue curves indicate three different
coherent fields that are mixed in space.
manufacturing systems, such as microscopes, telescopes, Xray imaging systems, and
photolithography systems, comprehensive characterization of the wave fields may
create opportunities for innovation in these areas by exploiting the extra degrees of
freedom in partially coherent fields.
1.1
Phase contrast was first introduced in the early 1900s [132, 131]. It solves the problem
of directly visualizing phase from a single intensity measurement, for which Frits
Zernike won the Nobel prize in 1953. However, the greyscale value obtained from a
phase contrast image has a nonlinear relationship with the underlying phase, yielding
only a qualitative description of the object.
In order to quantitatively recover the phase information, computational methods
can be applied to the measured intensity data to account for the measurement process.
This idea in principle is applicable to any phase contrast imaging system; in practice,
however, a physical measurement that maps the underlying phase to the observed
intensity greatly influence the quality of the inversion result. For example, a nonlinear
mapping may not be easily inverted, or some spatial frequency components may
be lost during the measurement. Here, two types of quantitative phase imaging
techniques are investigated: (1) interferometric techniques recover the phase difference
18
between the unknown field and a known reference field by extracting information from
the interference term in a interferogram; (2) propagation based techniques rely on
measurements of diffraction during spatial propagation and computational inversion
of the propagation operator.
1.1.1
Interferometric/holographic techniques
There are many experimental arrangements for interferometry [117, 95], but the main
idea is that an unknown field f (x, y) = Af (x, y) exp[i(x, y)], where Af is amplitude
and is phase, is interfered with a known reference wave r(x, y) = Ar (x, y) exp[i0 (x, y)]
and the measured intensity is
I(x, y) = |Af (x, y) exp(i(x, y)) + Ar (x, y) exp(i0 (x, y))|2
= If (x, y) + Ir (x, y) + 2Af (x, y)Ar (x, y) cos[(x, y) 0 (x, y)], (1.1)
where the first two terms are the intensities of the two fields and the cosine term is the
interference term. Depending on the experimental arrangements, is either directly
related to the phase of the underlying object or the wavefront of the propagated field
from the object plane. In the first case, may be recovered by computationally extracting the interference term from I; the phase unwrapping algorithm is often also
needed when the phase difference is larger than 2 in the postprocessing [117]. For
the second case, the recorded intensity is also known as a hologram [95]. The interference term encodes the information about the defocus distance. A computational
method that accounts for propagation and other reconstruction artifacts in order to
find the infocus objects is often the concern. A detailed discussion on holographic
imaging is in Chapter 2.
When the underlying field is partially coherent, interferometric techniques can
also be applied. The measured intensity takes a similar form as the coherent case
[Eq. (1.1)], except that the contrast of the interference term is modified by the corre19
lation (x, y) between the two fields, as given by the following expression,
I(x, y) = If (x, y) + Ir (x, y) + 2
q
If (x, y)Ir (x, y)(x, y).
(1.2)
Note that often in practical the two fields are originated from different parts of the
same field, and in this case is the degree of coherence of this field [70, 49]. Equation 1.2 is the principle of interferometricbased coherence state measurement techniques [110, 59, 41, 71].
1.1.2
2 I
= (I ),
z
(1.3)
where denotes the gradient in the lateral dimension. The energy flux (the Poynting vector) F along the transverse direction is related to the gradient of the phase
by F = I . The phase is solved given the knowledge about the longitudinal
derivative I/z, which can be estimated by finite difference method based on the
intensity measurements in two or more planes [105, 120]. Extensions to the original
20
TIE and computational methods to invert the equation will be discussed in detail in
Chapters 35.
For partially coherent fields, a similar technique known as phase space tomography
(PST) allows the reconstruction of the correlation function at a transverse plane
from a stack of many propagated intensity measurements [91]. Experiments and
reconstruction methods for the PST will be discussed in Chapter 6.
Another method to estimate the energy flux vector is to use an array of lenslets,
such as in a ShackHartmann sensor [90]. The direction of the vector of the incoming
wave averaged over the underlying lenslets aperture is estimated by the location of
the focal spot after the wave propagates to the focal plane of the lens. It turns out
that the same setup can be used to estimate the coherence state of partially coherent
light. Its properties and limitations will be discussed in detail in Chapter 7.
1.2
object
image-forming optics
image
(a)
object
optics
measurement computation
image
(b)
Figure 1-2: Architectures of (a) conventional imaging and (b) computational imaging.
x(2)
x(2)
y = Ax
xl1
xl2
x(1)
(a)
y = Ax
x(1)
(b)
Figure 1-3: Geometry of the (a) `2 and (b) `1 recovery for an unknown vector x =
[x(1), x(2)]. A single measurement Ax = y can be graphically represented by the red
line in both figures. Since the cost function for kxk`2 is a circle, while for kxk`1 is
a diamond, the solution for either problem is found by the intersection between the
cost function and the line.
(1.4)
where n is the length of the unknown object vector, S is the sparsity and measures the
compressibility of a signal, is the incoherence parameter, and C is a small positive
constant. The optimal solution x of a linear system y = Ax with measurement y is
found by solving the following `1 minimization problem
min kxk`1
where kxk`1 =
subject to Ax = y,
(1.5)
vector of length two in Fig. 1-3. The solution of `1 minimization will lie on one of
the axes with very high probability, the sparsest solution for this problem is obtained
by this method since only one coefficient is nonzero in this case. The application of
CS is demonstrated by a simple example in Fig. 1-4. The original unknown signal
consists of 10 spikes [in (a)], which means that the signal is sparse in its natural
representation. The measurement is carried out by taking samples from the discrete
23
2
0
50
100
6
0
(a)
Fourier transform
Samples
50
100
1
2
0
(b)
50
100
(c)
Figure 1-4: Numerical example of compressed sensing. (a) A sparse realvalued signal
containing 10 spikes; (b) The Fourier transform (the blue curve) of the signal in (a) is
sampled at 40 random locations (as marked by the red circles); (c) The reconstruction
by `1 minimization.
Fourier transform of the signal. 40 samples are taken at random locations [as marked
by the red circles in (b)] in the simulation. Taking Fourier transform is an incoherent
measurement since a signal contains a single spike in the spatial domain results in
an evenly spreadout signal in its Fourier domain. The `1 minimization accurately
recovers the unknown signal, whose result is shown in (c).
A fundamental problem addressed by this thesis is how to best recover phase
from a minimal set of measurements. Sparse models and compressive measurement
and reconstruction techniques are investigated for different phase retrieval problems.
Diffraction during spatial propagation provides a convenient and effective way of
evenly mixing the phase information of the object. Efficient compressive inversion
methods are designed to recover the phase using much fewer measurements.
1.3
Chapter 2 studies holography under the CS framework. When the complex field
scattered from particlelike objects are recorded as an interference pattern, the information about the objects are mixed on the hologram during propagation. The
compressive method and its limitation for reconstructing the 3D information about
the particles will be discussed.
24
The TIE is originally derived for a coherent field; the difficulty in extending the
TIE from coherent to partially coherent fields arises from the fact that the latter
does not have a welldefined phase, as a partially coherent field experiences random
fluctuations over time. Chapter 3 presents a new form of the TIE and the experimental
demonstration of recovering the optical path length (OPL) of a thin sample under
partially coherent illumination.
The phase reconstruction from the TIE often suffers from lowfrequency artifacts due to the physics. Chapter 4 describes an iterative reconstruction method,
called nonlinear diffusion (NLD) regularization, to mitigate these artifacts under the
assumption that the sample is piecewise constant.
Chapter 5 presents the TIE tomography in the Xray regime using a tabletop
Xray microfocus source. A compressive reconstruction method to alleviate the TIE
lowfrequency artifacts and tomography highfrequency streaking noise from limited
sampling angles is demonstrated.
The propagation based correlation function recovery technique, the phase space
tomography (PST), for partially coherent fields is described in Chapter 6. Unphysical
artifacts often arise in the recovered correlation function using traditional reconstruction method because of inaccessible data in the experiment. A compressive reconstruction method that looks for a sparse solution in the coherent mode superposition
representation is demonstrated experimentally for both 1D and 2D fields.
Chapter 7 investigates the use of a lenslet array to measure the coherence state
of partially coherent light. Properties and limits of this techniques are studied by
numerical examples. Design parameters and constraints are suggested.
Finally, Chapter 8 states conclusions and future work.
25
26
Chapter 2
Compressive holography applied to
particulate flows
2.1
Introduction
The fundamental problem addressed by holography is that of recording, and later reconstructing, both the amplitude and the phase of a coherent optical wave based on
interferometric principle. A hologram is formed by the interference between the scattered field from a coherently illuminated object and a reference wave. Conventional
holography is recorded on a photographic film or plate. In the last decades, Digital
holography (DH) becomes increasingly popular due to advances in megapixel electronic sensors, e.g. CCD and CMOS, with high spatial resolution and high dynamic
range. Compared to traditional holography, DH records holograms on an electronic
detector array, which does not have the cumbersome requirement of film positioning, handling, and processing. In addition, with direct access to the hologram in
the digital form, various signal processing techniques can be applied to extract useful information from the hologram during reconstruction. DH has proven to be a
powerful computational imaging technique in a number of applications. For example,
quantitative phase information can be obtained by DH microscopy using the offaxis
geometry [73]. The alternative inline DH geometry, in which both the illumination
and the reference waves are served by the same beam, is also widely used due to its
27
simple optical geometry and better stability. However, since the ghost terms, e.g.
twin image and halo, cannot be easily removed with a single image, inline DH has
shown to be more suitable for 3D localization applications. For example, it has been
used for localizing particulate objects in flow cytometry [31, 96], holographic particle
imaging velocimetry (PIV) [38, 85, 61], marine biological imaging [39, 75], and two
phase flows [113, 111]. In these applications, the object under interrogation typically
consists of many pointlike scatterers, such as particles, blood cells, and air bubbles,
sparsely distributed in a 3D volume. The benefit of using the inline DH is the ability to capture the shape and 3D position information of the scatterers within the 3D
volume in a single shot. Since the scatterers are typically some distance away from
the detector, the hologram records some defocused information about each scatterer.
The task of digital reconstruction is to find the infocus information.
Traditional reconstruction methods typically consist of two separate steps. First,
a focal stack is generated by the backpropagation method (BPM) [94], in which
the hologram is convolved with a series of depth dependent freespace propagation
point spread functions. An image in this focal stack often contains both infocus
features and diffraction fringes from outoffocus objects. The goal in the next step
is to isolate infocus objects by using image segmentation techniques. A common
approach is to apply an experimentally determined threshold to a focus metric, such
as edge sharpness [75], minimum amplitude/intensity [42, 113] and correlation coefficient [130], to each plane in the focal stack. The advantage of this method is that
the algorithm can be implemented very efficiently. In practice, however, the values
computed from the focus metrics are often sensitive to noise, resulting spurious objects in the segmentation results. Furthermore, the localization accuracy from this
method is limited by the Nyquist sampling rate induced by the finite pixel size of the
detector [62, 45].
An alternative holographic reconstruction method is to treat the hologram as the
output from a linear system with the unknown object as the input signal. Direct
inversion will not produce satisfactory results because the equation describing the
system is underdetermined and it contains infinite numbers of possible solutions.
28
For the application of imaging particulate flows, an important question is to understand how the quality of the reconstruction is affected by the seeding density of
particles. In this chapter, the performance of the compressive and BPMbased methods are compared quantitatively by simulating holograms resulting from scatterings
of air bubbles immersed in water with increasing seeding density. The simulation
results show that the compressive reconstruction method provides better results over
a wide range of seeding densities. The theoretical limit of the seeding density is analyzed in the compressed sensing framework and found in good agreement with the
results from the simulation.
29
2.2
2.2.1
Forward model
The inline hologram is a record of the interference between a plane reference wave
Er of amplitude a and the scattered field E from a 3D object. A schematic diagram of
the experimental geometry for the inline holography is shown in Fig. 2-1. Assuming
that the wavefront of the illumination beam is not significantly disturbed due to the
presence of the object and a = 1 without loss of generality, the intensity g recorded
on the detector located at the z = 0 plane is
g(x, y) = |1 + E(x, y, 0)|2
= 1 + |E(x, y, 0)|2 + E (x, y, 0) + E(x, y, 0),
(2.1)
y
Er (x0 , y 0 , z 0 )f (x0 , y 0 , z 0 )h(x x0 , y y 0 , 0 z 0 )dx0 dy 0 dz 0 ,
2
(2.2)
where f is the object function, and h is the depth dependent point spread function
for freespace propagation. Under the paraxial approximation,
exp[i2(z z 0 )/]
h(x x , y y , z z ) =
exp
(z z 0 )
0
30
i
0 2
0 2
[(x x ) + (y y ) ] .
(z z 0 )
(2.3)
y
Scattered wave (E)
Hologram ( g )
3D object ( f )
Together with expression for the plane reference wave Er (x0 , y 0 , z 0 ) = exp(i2z 0 /),
the total scattered field is
1
y
i
0 0 0
0 2
0 2
f (x , y , z ) 0 exp 0 [(x x ) + (y y ) ] dx0 dy 0 dz 0 .
E(x, y, 0) =
2
z
z
i
i y hx
0 0 0
0
0
0
0
=
f (x , y , z ) exp{i2(ux + vy )}dx dy
Next, let us consider the discretization model for Eq. (2.4). Assume the object
is discretized into Nx Ny Nz voxels with lateral spacing , and axial spacing
z . Without loss of generality, assume the number of samples in both lateral dimensions are the same and Nx = Ny = N . Equation (2.4) is rewritten as the following
31
discretized form
En1 n2
"
!
#
pn1 +qn2
pm1 +qm2
iz X X X X X
2
2
2
fm1 m2 l ei2 N
eilz (p +q )u ei2 N
,
=
N 2 l
p
q
m m
1
(2.5)
where the matrix element for a physical variable is denoted by the same letter used
in the continuous model with a subscript indicating the index, and the value is determined by fm1 m2 l = f (m1 , m2 , lz ) and En1 n2 = E(n1 , n2 ). The spatial
frequency sampling pitch is u = 1/(N ), with p and q indicating the indices along
the two directions.
It is also useful to define a 2D matrix f (l) denoting the lth 2D slice from the
(l)
3D object matrix, whose element is defined by fm1 m2 = fm1 m2 l . Consider the term
enclosed by the square bracket in Eq. (2.5) by rewriting it as the following form
E(l) = H (l) f (l) ,
(2.6)
where E(l) is a vector denoting the field contributed from the lth object slice, f (l) is a
(l)
(l)
vector form of f (l) with its entries defined as f(N 1)m1 +m2 = fm1 m2 , and
1 (l)
H (l) = F2D
Q F2D
(2.7)
is the freespace propagation operator for the lth object slice, which applies the 2D
Fourier transform F2D of f (l) , multiplies a depth dependent quadratic phase function
1
Q(l) = exp{ilz (p2 + q 2 )2u }, and then applies the inverse Fourier transform (F2D
).
The total scattered field is the sum of all the E(l) , and takes the following form
E =
E(l) = Hf
H (1) H (2)
32
f (1)
f (2)
(Nz )
H
.
..
f (Nz )
(2.8)
where E is a vector form of E with its entries defined as E(N 1)m1 +m2 = Em1 m2 , and
H is the propagation operator for all the object slices.
The final form for the discretization model of Eq. (2.1) is
g = 1 + |Hf |2 + H f + Hf ,
(2.9)
where g is a vector containing samples from g, whose entries are defined as g(N 1)n1 +n2 = gn1 n2 .
2.2.2
Backpropagation method
In the BPM, a focal stack is generated by convolving the hologram with a series
of depth dependent point spread functions for freespace propagation.In its discrete
form, the backpropagation operator is simply H , where
conjugate transpose of a matrix. The focal stack s generated by the BPM from the
hologram g is given by
s = H g
= H 1 + H |Hf |2 + H (Hf ) + H Hf .
(2.10)
Since H (m) H (n) = H (m+n) , it can be easily verified that the last term in Eq. (2.10) is
(Nz 1)
(1)
(1)
IN
H
H
f
(1)
(N
2)
(2)
z
H
f
IN
H
H Hf =
..
..
..
..
..
.
.
.
.
.
(Nz 1)
(Nz 2)
(Nz )
IN
f
H
H
f (1)
H (1) f (2) + H (2) f (3) + + H (Nz 1) f (Nz )
(2)
f
H (1) f (1) + H (1) f (3) + + H (Nz 2) f (Nz )
+
=
..
..
.
.
f (Nz )
H (Nz 1) f (1) + H (Nz 2) f (2) + + H (1) f (Nz 1)
(2.11)
where the object vector f is recovered as shown in the first term along with the outof
focus term in the second vector. In addition, the focal stacks are further corrupted by
33
the halo [the second term in Eq. (2.10)] and twinimage [the third term in Eq. (2.10)]
terms.
2.2.3
y =
2Hr 2Hi
fr
fi
+ |Hf |2
Ax + e,
(2.12)
imization problem,
= arg min kxkTV such that y = Ax,
x
(2.13)
where kxkTV is defined as a function of the magnitude of the gradient of the object
function according to [92]
kxkTV =
XXX
l
m1
|x(l)
m1 m2 |,
(2.14)
m2
where denotes the finite difference operator in the lateral dimension. In practice,
the following regularization problem is often solved instead to account for imperfect
measurement [20, 10]
minimize
1
ky Axk2 + kxkTV ,
2
(2.15)
2.3
Numerical simulations
10
compressive
Rg=0.011
BPM
Rg=0.35
10
q
0
10
Rg=0.086
10
10
Rg
10
10
Figure 2-2: The total error q using the compressive reconstruction method (solid red)
and BPM-based method (dashed blue) at different seeding densities measured by the
geometrical occlusion factor Rg .
the reference wave is an undisturbed plane wave, the hologram is calculated as the
interference between the total scattered field and the plane reference wave. Holograms
containing 8, 16, 32, 64, 128, 256, 512, 1024 bubbles are calculated. For each data
set containing the same number of bubbles, 20 holograms generated from different
distributions of bubbles are simulated.
We qualify the seeding density with a dimensionless variable called the geometrical
occlusion factor Rg , defined as
Rg =
Np
r2
total crosssection area of all bubbles
,
area of the hologram
(N )2
(2.16)
where Np is the number of bubbles, r is the mean radius, and Rg measures the
fractional area taken up by the parallel projection of all bubbles on the hologram
plane.
During the reconstruction using either method, the estimated object function is
expressed as a 256 256 30 matrix corresponding to discretization of the actual
36
Hologram
BPM
Compressive
Rg=0.011
(Np=16)
Rg=0.086
(Np=128)
Rg=0.35
(Np=512)
Figure 2-3: Left: Sample holograms; middle: real part of a depth slice from the
BPMobtained focal stack; right: real part of a depth slice from the focal stack
estimated by the compressive reconstruction method. First row: Rg = 0.011; second
row: Rg = 0.096; third row: Rg = 0.35.
37
volume with lateral spacing = 2m and axial spacing z = 250m. The BPM
estimated object is obtained by applying a threshold on the intensity focal stack
matrix. Since the diameter of a bubble is much smaller than z , each bubble is
represented as a single disk located at some depth slice in the true object matrix. Note
that within the region of a disk where a bubble is located, the reconstructed values
from the two methods are different because they have different physical meanings.
It is not straightforward to relate the value to some physical quantity, such as the
refractive index or the scattering potential, because the Mie scattering is not a linear
function of either of these quantities [12]. When comparing the reconstruction results
produced by either method with the ground truth, the value within each disk is
ignored as long as the disk is correctly detected. The quality of the reconstruction
is measured by the number of falsely detected voxels weighted by the value in the
original intensity focal stack. The total reconstruction error q is defined as
|value of false voxel|2
;
2
all voxels |voxel value of the true object matrix|
P
q=P
(2.17)
The total error q of each reconstruction method is plotted as function of Rg in Fig. 22. The error of the BPMbased method grows approximately linearly as Rg increases
when both are plotted in the logscale. The results from the compressive reconstruction method remain accurate until Rg > 0.086 (Np =128). Beyond that value of Rg ,
the error grows rapidly as the seeding density increases. Reconstruction slices from
sample holograms with Rg = 0.011, Rg = 0.086, and Rg = 0.35 are shown in Fig. 23. A sample hologram at each Rg is shown in the first column. It is seen that the
quality of the BPM reconstruction degrades as the number of bubbles increases, as
shown in the middle column; it is impossible to find a single threshold value to isolate
the infocus bubbles when the degradation is severe. The compressive reconstruction
method successfully finds the infocus bubbles and removes most of the unwanted
artifacts in the first two cases, as shown in the third column.
It should be noted that the signaltonoise ratio (SNR) of a measurement by an
inline hologram decreases as the density of bubbles increases. For an ideal hologram
38
10
SNR
SNR
10
10
Rg=0.011
Rg=0.086
Rg=0.35
10
10
10
10
Rg section ratio
Geometric cross
10
Figure 2-4: The SNR of a hologram decreases as the density of bubbles increases.
with no random measurement noise, the noise is entirely contributed from the halo
term, so the SNR is
SNR =
(2.18)
As Rg increases, the SNR deceases almost linearly when both are plotted in the log
scale), as shown in Fig. 2-4. We expect the qualities of the hologram at high seeding
densities are even worse since the effects of multiple scatterings and the deviation
from an ideal planar reference wave become severe.
Next, we study the maximum Rg allowed for accurate reconstruction under the
compressed sensing framework. Let s denote the sparsity of the problem, i.e. the
number of nonzero coefficients of the unknown expressed in the sparsity basis. Since
the TV basis is used in our problem, s can be estimated by the total number of edge
voxels for all the bubbles:
s Np 2
r/,
(2.19)
where the average number of edge voxels for a bubble is estimated by the number of
voxels on the perimeter of a disk of radius r. To accurately reconstruct an ssparse
39
(2.20)
where C is a small positive constant and is another constant called the coherence
parameter, which measures the correlation between the measurement matrix and the
sparsity basis. In our problem, M = N 2 , and N0 = 2N 2 Nz . As a result, the number
of bubbles Np that can be successfully reconstructed should satisfy
Np
N 2
,
2C
r log(2N 2 Nz )
(2.21)
r
.
2C log(2N 2 Nz )
(2.22)
After substituting the values for all the variables used in the simulation and assuming
C = 1 [28] in Eq. (2.22), the maximum value of Rg is 0.168.
2.4
Discussion
Our simulation result and theoretical analysis based on compressed sensing both
indicate that the maximum geometrical occlusion factor Rg allowed for accurate reconstruction is around 0.086. Equation (2.21) shows that the maximum number
of particles that can be accurately reconstructed increases as the number of pixels (spacebandwidth product) of the hologram increases and the size of a particle
decreases; the number limit is not very sensitive to the discretization in the axial
direction since Nz only affects the results in its logscale. A nonlinear forward model
that properly treats the halo term as part of the useful signal may help to improve
the reconstruction.
40
Chapter 3
Transport of intensity imaging with
partially coherent illumination
3.1
Introduction
While interferometric methods, such as holography, are useful for phase measurement
for a coherent field, their performance is often limited due to the extreme sensitivity to
mechanical stability and environmental conditions and the requirement for a coherent
(laser) source and reference beam. In cases where analogous optical elements are
difficult to manufacture, such as Xray [80], electron [58], and neutron imaging [2],
the propagationbased (noninterferometric) techniques, which relies on measurements
of diffraction patterns of the unknown field upon propagation, are more suitable.
The transport of intensity imaging is a noniterative propagationbased technique,
which allows the recovery of the unknown phase by solving a linear equation, i.e.
the transport of intensity equation (TIE) by measuring intensities in closely spaced
planes [109, 105, 120]. It is easy to implement since the measurements can be taken
by simply displacing the detector or the sample along the optical axis.
The TIE is originally derived for a coherent field; the difficulty in extending the
TIE from coherent to partially coherent fields arises from the fact that the latter
does not have a welldefined phase, as a partially coherent field experiences random
fluctuations over time. Although it is possible to decompose a partially coherent
41
field into a sum of fully coherent fields through a process known as coherent mode
decomposition [127], it is shown that a weighted average of the phases of all the modes
does not match well with the TIEobtained phase [135]. An alternative interpretation
of the TIEobtained phase is by a scalar phase whose gradient is proportional to the
timeaveraged transverse component of the Poynting vector [84]. In this chapter,
we clarify the TIEobtained phase in terms of the optical path length (OPL) of a
thin sample and show that the OPL can be accurately recovered even with partially
coherent illumination from TIE measurement, the effects of the illumination may
be removed provided its transverse flux vector is first characterized by another TIE
measurement. The remainder of this chapter is structured as follows: the TIE with
coherent illumination will be reviewed in Section 3.2; the partially coherent TIE (PC
TIE) for recovering the OPL of a thin sample and the experimental verification of
this equation will be presented in Section 3.3.
3.2
3.2.1
Following the original work of Teague [109], TIE for a monochromatic coherent field
can be derived directly from the paraxial wave equation. Consider a monochromatic
coherent paraxial field of the form
U (x, z) =
p
Ic (x, z) exp[i(x, z)],
(3.1)
where Ic is the intensity, is the phase, x denotes a 2D vector specifying the position
in a plane transverse to the optical axis along z, and the subscript c denotes the
properties of coherent light. The propagation of the field U (x, z) satisfies the following
paraxial wave equation
U (x; z)
i 2
=
U (x; z),
z
2k
42
(3.2)
where is the 2D gradient operator in the transverse plane. To derive the TIE,
consider the propagation of intensity by differentiating Ic with respect to z, and since
Ic = U U ,
Ic (x; z)
i
=
[U (x; z) U (x; z) U (x; z) U (x; z)].
z
2k
(3.3)
i
[U (x; z) U (x; z) U (x; z) U (x; z)].
2k
(3.4)
1
Ic (x; z) (x; z).
k
(3.5)
Finally, the TIE, which relates the phase to variations in the propagation of intensity,
is given by
1
Ic (x; z)
= [Ic (x; z) (x; z)].
z
k
3.2.2
(3.6)
The TIE states that the phase distribution on a plane of constant z can be recovered
by solving a partial differential equation (PDE) with the knowledge of the intensity
derivative along the optical axis. However, since the intensity derivative cannot be
measured directly, finite difference methods are used to approximate the derivative
from two or more defocused images in practice [120]. For example, the central difference scheme can be implemented by capturing two defocused images at planes
located symmetrically before and after the focal plane, (the focal plane is assumed to
be located at z = 0 without loss of generality), to estimate the derivative with third
43
order accuracy:
Ic (x, z) Ic (x, z)
Ic (x; 0)
=
+O
z
2z
3 Ic (x; 0)
z 3
,
(3.7)
zu
zu
U x+
;0 U x
; 0 exp(i2x u) d2 x,
2
2
(3.8)
where u is the spatial frequency variable. Assume that the field U (x; 0) at the focal
plane varies slowly with respect to changes in u, in which case the field can be
approximated by the first order Taylor expansion term as
U
1
zu
; 0 = U (x; 0) + zu U (x; 0).
x
2
2
(3.9)
Substituting Eq. (3.9) into Eq. (3.8) and keeping only the first order terms with
respect to z yields
Ic (x; z) = Ic (x; 0)
z
[Ic (x; 0) (x; 0)].
k
(3.10)
(3.11)
The condition in (3.11) implies that the high spatial frequency components of the field
can only be captured with a relatively small defocus distance so that the condition
still holds; whereas low spatial frequency components tolerate a much larger defocus
distance. When z is large, the high spatial frequency components of the field produce diffraction fringes, which violate the linearity assumption inherent in the TIE.
44
The loss of high spatial frequency information results in a lowpass filtering effect
on both the measured data and the recovered phase. A small z avoids loss of high
frequency information, however, may result in measurement with low signaltonoise
ratio in practice, as the differences between the defocused images may be too small
as compared to the noise level. An optimal z is object and system dependent, and
needs to be found experimentally.
In many practical situations, the concern is to recover the optical properties of a
sample rather than the phase of the incident field. Consider a thin sample characterized by the following transmission function
g(x) =
T (x) exp[i(x)],
(3.12)
where T (x) is the transmittance of the sample, which characterizes the total attenuation property of the sample; is the phase of the sample, which is equal to the
product of the OPL though the sample with the wavenumber of the illumination,
(x) = kOPL. If the sample consists of homogenous material of known refractive
index n, the physical thickness L of the sample is related to the OPL by L = OPL/n.
For a sample consisting of multiple materials with different refractive indices, tomographic TIE measurement is required to fully recover the 3D index distribution;
experimental results of TIE tomography will be discussed in Chapter 5. Here, the
thin sample approximation will be used throughout this chapter.
When a sample is illuminated by a plane wave with constant intensity I0 , the
phase of the field at the plane immediately after the sample is entirely due to the
samples phase. The transverse flux vector at this plane is fully characterized by the
absorption and phase of the sample:
Fc (x; z) =
1
I0 T (x) (x; z).
k
(3.13)
1 Ic (x; z)
= 2 (x; z).
I0
z
(3.14)
Equation 3.14 is a Poisson equation, stating that the change of intensity at small
defocus is proportional to the Laplacian of the phase of the sample. When the sample
contains both amplitude and phase variations, the full solution of TIE requires to solve
two Poisson equations, as described below. First, we introduce an auxiliary function
(x; z) as the solution of the following equation,
(x; z) = I0 T (x) (x; z).
(3.15)
Ic (x; z)
,
z
(3.16)
which is a Poisson equation which can be solved for (x; z). Next, the second Poisson
equation relating the phase can be obtained by rearranging Eq. (3.15) to get
3.3
3.3.1
(x; z)
= 2 (x; z).
I0 T (x)
(3.17)
A partially coherent field does not have a well defined phase since the field at any
point in the space experiences statistical fluctuations over time. To fully characterize
a partially coherent field, secondorder correlations between pairs of points need to
be used. The correlation function for a stationary quasimonochromatic spatially
partially coherent field on a plane of constant z is the mutual intensity
J(x1 , x2 ; z) = hU (x1 ; z)U (x2 ; z)i ,
46
(3.18)
(3.19)
where the subscript p denotes the properties of partially coherent light. Note that J
is a complexvalued correlation function and defined over the 4D space spanned by
the pair of 2D position vectors, {x1,2 }. Therefore, although one can still measure a
focal stack of intensities and apply the TIE in order to reconstruct a 2D function in
this plane, the recovered phase clearly does not provide the phase of J.
Since a partially coherent field can be represented in terms of an ensemble of
coherent fields, each of which satisfies the paraxial wave equation, the propagation
equation of intensity can be obtained by differentiating Ip with respect to z, with the
substitution of Eqs. (3.2) into (3.19), yielding the PCTIE:
Ip (x; z)
= Fp (x; z),
z
(3.20)
where Fp (x; z) is the transverse flux vector of partially coherent field, and Fp (x; z) is
defined as [70]
i
hU (x; z) U (x; z) U (x; z) U (x; z)i
2k
i
x0
x0
.
= x0 J x + , x ; z
2k
2
2
x0 =0
Fp (x; z) =
(3.21)
Equation 3.20 is the analogous energy conservation equation for partially coherent
field under paraxial propagation, which states that the transverse flux can be characterized given the knowledge of intensity derivatives along the propagation direction.
Since Fp (x; z) is a vector over a transverse plane, it can be uniquely decomposed
into a curlfree component and a divergencefree component via the Helmholtz decomposition [84]:
Fp (x; z) = s + A(x).
47
(3.22)
Substitution of Eq. (3.22) into Eq. (3.20) results in a Poissons equation for s :
Ip (x; z)
= 2 s ,
z
(3.23)
which can be solved for specified boundary conditions, yielding Fp uniquely up to the
addition of A(x). Note that Fp retrieved by this procedure has a similar form
as the coherent flux in Eq. (3.5), in which the gradient of the scalar phase function
s measures the rate of the transverse energy flow as the field propagates. The
coherent case can be treated as a special case of this result where the scalar phase
function equals to the phase of the complex field up to a multiplicative constant.
While the complex field representation for a coherent light can be fully determined
with the knowledge of Fc and Ic ; it is not sufficient to characterize the mutual intensity
with only the knowledge of Fp and Ip for a partially coherent light expect for trivial
cases.
In many practical applications, it is only important to characterize the optical
properties of a sample rather than the full coherence state of the incident field. It will
be shown in the following that it is still possible to recover the optical thickness of a
sample by modifying the PCTIE to account for the partially coherent illumination.
A thin sample, described by Eq. (3.12), is illuminated by a partially coherent field
with mutual intensity Ji (x1 , x2 ; z) and intensity Ii (x; z), where the subscript i denotes
the properties of the illumination. At the plane immediately after the sample, the
mutual intensity is given by
Jt (x1 , x2 ) = Ji (x1 , x2 ; z)g(x1 )g (x2 ) = hUi (x1 )g(x1 )Ui (x2 )g (x2 )i ,
(3.24)
where the subscript t denotes the properties of the field immediately after the sample,
and the intensity at the same plane is
It (x; z) = Ii (x; z)T (x).
(3.25)
The propagation of the intensity It (x; z) follows the PCTIE in Eq. (3.20); the trans48
verse energy flux vector Ft can be obtained by substituting Ui (x; z)g(x) for U (x; z)
in Eq. (3.21), yielding
1
Ft (x; z) = T (x)Fi (x; z) + It (x; z) (x).
k
(3.26)
Equation 3.26 shows that two terms contribute to the total flux Ft : the first term
characterizes the flux of the illumination modified by the absorption of the sample;
the second term characterizes the flux due to the presence of the sample, which takes a
similar form as the one for the coherent case in Eq. (3.13). Substituting the expression
of Ft (x; z) into Eq. (3.20) yields the PCTIE for the field immediately after the thin
sample:
It (x; z)
1
+ [T (x)Fi (x; z)] = [It (x; z) (x)].
z
k
(3.27)
Equation 3.27 shows that the phase of a sample can still be recovered by conducting
TIE measurements even if the illumination is partially coherent. The only difference
from the coherent TIE is that the term [T (x)Fi (x; z)] needs to be properly
accounted for. To do that, we assume the illumination flux is divergence free so it
is fully described by a scalar phase by Fi = s . s at the sample plane can be
calibrated by first taking intensity focal stacks without the sample in place and then
recovered by solving Eq. (3.23). Next, the samples phase can be retrieved by taking
the second set of TIE measurements with the sample in place and solving Eq. (3.27).
Note that when Fi (x; z) = 0, Eq. (3.27) simplifies to
It (x; z)
1
= [It (x; z) (x)],
z
k
(3.28)
which has the exact same form as the coherent TIE. This implies that the optical
thickness of a sample can be retrieved by taking a single set of TIE measurement as
long as the illumination does not have transverse flux at the sample plane. This is
approximately the case when the sample is illuminated by Kohler illumination in a
standard brightfield microscope [105].
Another special case is when the sample is purephase, where Eq. (3.27) reduces
49
(3.29)
where Eq. (3.20) is used to express the lefthandside as the difference between the
two intensity derivative measurements taken with and without the sample in place.
The result of Eq. (3.27) assumes that intensity derivative can be measured accurately. In reality, one is limited to estimate the derivative from finite difference
measurements. The validity of Eq. (3.27), taking the finite displacement distance
z) of
along z into account, can be examined as follows. The Fourier transform I(u;
the intensity taken at the plane defocused by z from the focal plane (assumed at
z = 0) can be written as [56]
z) =
I(u;
s
zu
zu
zu
zu
Ji x +
,x
T x+
T x
2
2
2
2
zu
zu
x
exp i x +
2
2
2
exp(i2x u) d x,
(3.30)
where u is the spatial frequency variable. Assume that all functions of zu vary
slowly with respect to changes in u and
|zu|2 1,
(3.31)
Eq. (3.30) can be simplified by taking up to the first order terms with respect to
zu, yielding
z) =
I(u;
Z
[T (x)Ii (x; 0) + i2zT (x)u Fi (x; 0) + izIt (x; 0)u (x)]
exp(i2x u) d2 x.
(3.32)
50
z
[It (x; 0) (x)]. (3.33)
k
3.3.2
The experimental arrangement used to test the PCTIE is illustrated in Fig. 3-1.
=620nm and bandwidth 20nm is placed before
An LED with central wavelength
a pinhole with a diameter d = 500m; a diffuser is placed in between to create
a uniform angular distribution of light. We use a Kohler illumination system, by
placing a condenser lens at one focal length (f =50mm) away from the pinhole, to
create a uniform intensity distribution after the lens. If we assume that the lens is
perfect and neglect the effect of finite apertures, the field after the lens consists of
plane waves propagating at different directions. If we further assume the primary
source is incoherent with uniform intensity I0 , which is a good approximation in our
case, then the mutual intensity at any arbitrary plane z after the lens can be obtained
according to the vanCittert Zernike theorem as
I0 d
Ji (x1 , x2 , z) =
J1
k|x2 x1 |
kd|x2 x1 |
2f
,
(3.34)
4f system
(unit magnification)
Diffuser
Condenser
LED
f = 50mm
pinhole
500m
Field
modulation
mask
Unknown
phase
object
Detector
Figure 3-1: Imaging system used for the partially coherent TIE
of applying the coherent TIE, without accounting for the illumination, is shown in
Fig. 3-2(i), which clearly mixes illumination flux with samples phase. The samples
thickness reconstructed from Eq. (3.29), assuming a purephase sample, is shown
in Fig. 3-2(j), while the result of applying Eq. (3.27) is shown in Fig. 3-2(k). The
difference between Figs. 3-2(j) and (k) is minimal, due to the fact that the infocus
intensity of the sample exhibits minimal variation, i.e. it is nearly purephase.
3.4
Conclusion
The PCTIE thats allows recovery of the OPL of a thin sample and accounts for
the effect of partially coherent illumination is demonstrated. For illuminations with
transverse symmetries, this equation reduces to the same form as the coherent TIE.
Otherwise, two sets of TIE measurements are required in order to account for the
nonuniform illumination flux.
53
0.9
50
0.8
0
0.7
50
0.5
0.6
thickness (nm)
0.5
(a)
(b)
(c)
0.9
100
0.4
0.5
0.2
0.8
0
0.7
0.2
0.5
0.6
0.4
0.5
(d)
0.5
0.6
(e)
(f)
50
0.9
0.6
0.8
0.4
0.2
0.7
0
0.6
0.2
0.5
0.4
50
thickness (nm)
0.8
0.6
0.4
0.8
(g)
(h)
(i)
50
(j)
100
50
thickness (nm)
50
thickness (nm)
50
100
100
(k)
Figure 3-2: Experimental measurements and results for (ac) the sample with uniform
illumination, (df) the illumination modulation mask alone, and (gk) both sample
and modulation mask in place. (a,d,g) Infocus intensity measurements. (b,e,h)
Intesnity difference between defocused measurements. (f) The scalar phase reconstruction for the illumination with only the field modulation mask in place. Sample
thickness reconstructed from (c) Eq. (3.28) with uniform illumination, (i) Eq. (3.28)
with both sample and modulation mask in place, (j) Eq. (3.29), assuming a pure
phase sample, and (k) Eq. (3.27), assuming a sample containing both amplitude and
phase variations.
54
Chapter 4
Transport of intensity phase
imaging with nonlinear diffusion
regularization
4.1
Introduction
The transport of intensity equation (TIE) follows from the wave equation under paraxial propagation and in the simplest case relates the first derivative of intensity along
the optical axis to the Laplacian of the phase in the transverse direction. In practice, recovery of phase from the TIE involves estimating the first intensity derivative
from finite differences taken between two or more intensity images at different propagation distances [120, 121, 112] followed by numerical inversion of the Laplacian
operator. Phase reconstructed from noisy images using the TIE often contains unwanted lowfrequency artifacts [83], which existing denoising methods often do not
remove effectively [58, 83]. On the other hand, proper adaptation of prior knowledge
of the object can significantly improve recovery, as has been shown in numerous other
inverse problems. In this chapter, we consider the prior, often encountered in practice, of the objects consisting of regions of piecewise constant phase and demonstrate
how to incorporate this prior into the TIE solution as nonlinear diffusion (NLD)
55
regularization.
4.2
Consider a pure phase object oriented perpendicular to the optical axis, z, whose optical thickness is specified by (x, y). If this sample is illuminated by a normally incident plane wave of intensity I0 , the intensity immediately after the object, I(x, y, z),
satisfies a simplified form of the TIE
k I(x, y, z)
= 2 (x, y),
I0
z
(4.1)
where k = 2/ is the wave number. This has the form of Poissons equation in 2D. If
I/z is known along with suitable boundary conditions for the phase, the phase can
be uniquely computed from Eq. (4.1). For a more general phase and amplitude object,
the full solution of the TIE can be obtained by solving two Poissons equations [109],
the reconstructed phase has been shown to be unique up to phase singularities [55].
To understand lowfrequency artifacts in the TIE solution, it is instructive to
consider the TIE in the Fourier domain. Consider a single Fourier component of
the phase, which has spatial variation proportional to sin(u0 x). From Eq. (4.1) it
is clear that the change in intensity upon propagation, I/z, is proportional to
u20 (a more rigorous analysis for a brightfield microscope may be found in [3, 97]).
Therefore, low spatial frequency components yield a much weaker signal than those
with high spatial frequency. The inversion of the TIE to recover phase involves
multiplication in the Fourier domain by the transfer function, H = 1/(u2 + v 2 ),
where u and v are the spatial frequency variables along x and y, respectively. This has
the effect of significantly amplifying low frequency noise in the phase reconstruction.
Traditional linear denoising utilizes a Tikhonov regularized transfer function, HTR =
h
i
2
(u2 + v 2 ) / (u2 + v 2 ) + , where is a regularization parameter. HTR behaves like
a highpass filtered version of H and controls the lower cutoff frequency.
An alternative nonlinear method is to write the TIE in terms of a linear model
56
x
1
kf Ak2 +
(||) dxdy.
2
(4.2)
0 (||)
.
||
(4.3)
1.67s term from the Weikert function, resulting instead in the hybrid NLD flux
FH (s) = 1 exp(3.86/s12 ).
(4.4)
This hybrid NLD regularization function and the magnitude of the flux are shown in
Fig. 4-1(a) and (b), respectively. Below the critical gradient, FH behaves like the TV
function; and above the critical gradient, it behaves as the Weickert function.
3.0
TV
Weickert
Hybrid
2.5
1.5
2.0
F 1.0
1.5
1.0
0.5
0.5
0
1.0
s
(a)
2.0
3.0
1.0
s
(b)
2.0
3.0
Figure 4-1: (a) The NLD regularizing function and (b) the magnitude of the flux as
a function of the normalized magnitude of the gradient of the phase.
Pseudocode to implement the optimization in Eq. (4.2) based on the iterative
shrinkage/thresholding(IST) algorithm [10] is as follows:
(0)
1. Initialize b = A f ;
2. For i = 1, 2, . . .,
(i)
(2.a) apply gradient descent with a step size for the data fidelity term: a =
(i1)
(i1)
b
+ A (f Ab );
(i)
(i)
(i)
(i)
(i)
(2.b) apply NLD with time t such that t/ = : b = a +t 0 (|a |)a /|a | ;
(i1)
3. Stop and yield the final estimate = b
when the change in the minimizing
functional in Eq. (4.2) between consecutive steps falls below some threshold.
For large t, we implement the NLD with n timemarching steps of step size dt
(t = ndt) using the additive operator splitting (AOS) scheme [124]. Additionally,
before computing the flux at each step, we smoothen the gradient by convolution
58
1
0.8
0.6
0.4
0.2
0
0.2
(a)
0.01
(b)
(d)
(c)
hybrid
direct
original
Phase (rads)
0.8
0.6
0.4
0.2
0
0.2
0
(e)
(f)
(g)
550
Position (m)
1100
(h)
Figure 4-2: (a) Original phase, (b) intensity derivative, (c) direct, (d) Tikhonov
regularized, (e) TVregularized, (f) Weickert function regularized and (g) the hybrid
function regularized solutions. (f) Phase crosssections taken along the dashedline
in (a).
4.3
Simulations
We demonstrate our method by a numerical example: a pure phase object with the
phase distribution given in Fig. 4-2(a) is illuminated by a normally incident plane
wave. The intensity derivative in Fig. 4-2(b) was generated by simulating over
and underfocused images, each defocused from the infocus plane by a normalized
distance znorm = 8 (znorm = z NA2 /, where z is the physical distance and NA=0.14
is the numerical aperture). Each defocused image was corrupted by Poisson noise
such that the signaltonoise ratio(SNR) in this intensity derivative data is 0dB. Low
frequency artifacts are clearly seen in the directly inverted result in Fig. 4-2(c). The
effect of Tikhonov regularization with the regularizer = 1012 (chosen to minimize
the average perpixel rootmeansquare (RMS) error) is shown in Fig. 4-2(d). We
implemented the NLD regularization with the TV, Weickert and hybrid functions
and the results are shown in Fig. 4-2(e), (f), and (g), respectively. The TV function
59
x 10 4
Direct
Hybrid
Weickert
TV
2
SNR (dB)
4.4
Discussion
We have demonstrated that the use of the piecewise constant priors in the TIE by NLD
regularization to account for lowfrequency artifacts while preserving sharp edges and
60
introduced a new function combining the desired properties of TV and Weickert NLD.
It should be noted that, aside from the TV case, the NLD regularization explored here
is not a provably convex problem; therefore, the recovered phase is not necessarily
the global minimizer of Eq. (4.2). This problem also occurs in edge-preserving NLD
since the regularization functions, that preserve edges and smooth lowfrequency
artifacts are not convex. However, in NLD it is generally observed that discretization
due to finite differences (or Gaussian smoothing of the gradient prior to performing
the NLD) greatly stabilizes the results [125]. Based on our observations, this appears
to be the case for the technique presented here as well. For phase maps that are not
piecewise constant, but have strong sparsity priors in other domains, e.g. the phase
is piecewise continuous or consist of a small set of Zernike modes, similar techniques
may prove useful for removing lowfrequency artifacts.
61
62
Chapter 5
Compressive Xray phase
tomography based on transport of
intensity
5.1
Introduction
Traditional tomography with hard Xrays recovers the attenuation of an object. Attenuation does not always provide good contrast when imaging objects made of materials with low electron density, e.g. soft tissues. In these cases, richer information is
often contained in the phase, i.e. the optical thickness of the sample [76, 89, 129, 35],
since the Xray phase shift is almost 103 times larger than the Xray attenuation
for lowZ elements [107], such as hydrogen, carbon, nitrogen and oxygen. Xray
interferometry based on MachZehnder geometry using Bragg crystals [76] provides
Xray phase information; however, it suffers from extreme sensitivity to mechanical
stability and alignment, the coherent synchrotron source is necessary for good fringe
contrast. The diffraction enhanced Xray phase imaging technique measures the refraction angle after an Xray beam passing through a sample using crystal analyzers.
Enhancement of the Xray phase contrast has been demonstrated using this technique [29], however, the experimental setup is also very complicated and relies on a
63
Sample (n)
x (rotational axis)
Area detector
x
Light source
x0
zs
y0
y
y
ys
z=-z0
z=0
z=d
5.2
Forward model
A schematic diagram of the imaging geometry for Xray TIE tomography is shown
in Fig. 5-1. A quasimonochromatic source with mean wavelength is located at
the plane of z = z0 with coordinates x0 = (x0 , y0 ). Here, we consider a planar
incoherent source with intensity distribution I0 , which is a good approximation for a
tabletop Xray source. The mutual intensity of the incident field Ji at the plane of
z = 0 with coordinates x = (x, y), under the paraxial approximation, is [70]:
1
Ji (x1 , x2 ) = 2 2 exp
z0
ik
x1 x2
2
2
|x1 | |x2 |
H0
2z0
z0
(5.1)
where the exponential term measures the wavefront curvature in the illumination; the
term H0 is source dependent:
H0
x1 x2
z0
=
x1 x2
I0 (x0 ) exp i2
x0
z0
d2 x0 ,
(5.2)
which relates the mutual intensity of the partially coherent illumination to the Fourier
transform of the intensity distribution of the primary incoherent source.
65
(5.3)
The intensity after light passing through the object is recorded by an area detector
located at the plane of z = d with coordinates x0 = (x0 , y 0 ). Under the paraxial
approximation, the intensity I on the detector plane is given by the following propagation integrals:
1
I(x ) = 2 2
d
0
ZZZZ
2
ik
0
2
0
2
J (x1 , x2 ) exp
|x x1 | |x x2 |
d x1 d2 x2 .
2d
0
(5.4)
Substitute Eqs. (5.15.3) into Eq. (5.4), and then take the Fourier transform on both
side of Eq. (5.4), yielding
x s
0u
0u
u
1
u
d
d
I
= 2 2 H0
T x
T x+
Ms
z0
Mi
2
2
d0 u
d0 u
exp i x
x+
exp(i2x u) d2 x
2
2
66
(5.5)
where denotes the Fourier transform, u = (u, v) is the spatial frequency variable,
x = (x1 + x2 )/2, Ms = (z0 + d)/z0 and Mi = (z0 + d)/d are the geometrical magnification factors for the sample and the source, respectively, and d0 = z0 d/(z0 + d) is
effective propagation distance. Assuming that |d0 u| is small compared to the variations in the transmission and phase functions, Eq. (5.5) can be expanded to the first
order in d0 u, yielding
u
I
Ms
Z
1
u
0
2
= 2 2 H0
T(u) [id T (x)u xy (x)] exp(i2x u) d x .
z0
Mi
(5.6)
By taking the inverse Fourier transform on both sides, Eq. (5.6) can be expressed as
M2
k
g(x) 0 [Ms2 I(Ms x) Id (x)] = 2 i 2 I0 (Mi x) xy [T (x)xy (x)],
d
z0
(5.7)
where denotes the convolution operator. The infocus intensity measurement at the
plane immediately after the sample is Id (x) =
Mi2
I (Mi x)
2 z02 0
modified TIE that includes the filtering effect due to the finite size of an incoherent
source. Since the total intensity can be treated as an incoherent superposition of
intensities generated from every points on the source, each of which illuminates the
sample with a slightly different displacement with respect to the center, the overall
effect on the final intensity is the convolution with intensity distribution of the source.
When the sample is weakly attenuating such that T 1, Id is entirely determined
by the intensity profile of the source. If we further ignore the filtering effect due to
the finite source size, Eq. (5.7) is simplified to
k
g(x) 0 [Is (Ms x) Ii (Ms x)] = Ii (Ms x)2xy (x),
d
(5.8)
where Ii is the background image taken without the sample in place. Equation 5.7
is a form of the TIE that uses two images taken with and without the sample in
place and no detector motion in order to recover the phase projection of a weakly
attenuating sample. The measurement process is illustrated in Fig. 5-2.
In a tomographic measurement, the sample is rotated about the x axis and the
67
no object
Ii
z
object
z
Figure 5-2: Experimental process for a pure phase sample. A background image Ii is
first taken without the object in place; images I with the sample in place are taken
next at the same plane.
phase projection for a given rotation angle can be expressed as:
(x; ) =
(5.9)
The set of measurements g obtained from many angles of rotation can be arranged
into a vector g. The forward model that relates the unknown n to the data takes the
simple form
g = P R n An.
(5.10)
where the TIE operator P is performed on each phase projection and the forward
operator A = P R takes the projection of n by R at each angle.
5.3
(5.11)
where the total variation (TV) function knkTV is our sparsity basis, defined as
knkTV =
Xq
(x n)2 + (y n)2 + (z n)2 ,
(5.12)
and x , y , and z are the finite difference operators in the three spatial dimensions. The solution of the optimization in Eq. (5.11) is found by solving the following
regularization problem instead to account for imperfect measurement [20, 10]
minimize
1
kg Anr k2 + knr kTV ,
2
(5.13)
where the regularization parameter controls the relative weight of the TV constraint
as compared to the data fidelity term kg A nr k2 . We adapt the twostep iterative
shrinkage/thresholding algorithm (TwIST) [10] to solve this minimization problem.
When implementing the algorithm, nr is discretized into a N N N matrix
with voxel length . The operator P is performed in the Fourier domain:
P I0 xy T xy
1
1
= Fxy
Hi Hx Fxy T Fxy
Hx Fxy
1
1
+Fxy
Hi Hy Fxy T Fxy
Hy Fxy
(5.14)
1
1
where the gradient operator xy is replaced by (Fxy
Hx Fxy , Fxy
Hy Fxy ) and the en-
tries of Hx and Hy at (p, q) are given by Hx,pq = i2p/(N ) and Hy,pq = i2q/(N ),
1
respectively; Fxy and Fxy
denote the discrete Fourier and inverse Fourier trans-
form, respectively, with respect to the (x, y) variables; the transfer function matrix due to the incoherent source is considered by Hi , whose entry is determined by
Hi,pq = H0 (p/(Mi N ), q/(Mi N ))/2 z02 ; and denotes entrywise multiplication. For a weakly attenuating sample illuminated by an ideal point source, the TIE
69
(5.15)
where H is the transfer function matrix with entries defined as Hpq = 4 2 (p2 +
q 2 )/(N )2 .
The operator R can be implemented in either the spatial domain [50] or the Fourier
domain [43]. Generally, it is observed that the performance of the Fourier domain
method is more robust to discretization and noise, which has been demonstrated in
both Xray CT [82] and MRI [68]. We adopt the Fourier domain method to write R
as
R = Fy1 SFyz,x
(5.16)
where Fyz,x denotes the discrete 2D Fourier transform taken for the cross section at a
1
1
1
constant x, S takes radial slices in the Fourier domain, and Fy1 = [Fy,
, Fy,
, , Fy,
],
1
2
N
1
where Fy,
denotes the 1D inverse Fourier transform of a radial slice along the angle
m
m . Since the radial slices need to be taken from points equally spaced along the
radial direction, while the result of Fyz,x is on a Cartesian grid, S requires interpolation and then resampling (gridding) [43]. We implement S together with Fyz,x by
adapting the non-uniform FFT (NUFFT) algorithm [43].
(5.17)
where FNU is the NUFFT operator. For a weakly attenuating sample illuminated by
an ideal point source, A is further simplified as
1
A = Fxy
H Fx FNU .
70
(5.18)
5.4
The experimental setup for the Xray TIE tomography is shown in Fig. 5-3. A
microfocus source (Hamamatsu, L812103) with a circular focal spot size of 5m in
diameter, located at z0 = 0.765, was operated at 20kVp to produce a diverging Xray
beam with central wavelength = 0.062nm. For a beetle sample, intensity images
were obtained with an Xray detector consisting of a CsI:Tl scintillator (RMD Inc.)
coupled to an EMCCD (Andor, ixon, 512512 pixels, 16m pixel size) by an imaging
system with 6 demagnification; the effective pixel size of an image directly seen at
the scintillator is 96m. The scintillator was placed at d = 1.711m. We assume the
beetle sample is weakly attenuating. During the tomographic measurement, a single
image was taken at every 5 degrees with 6.7 seconds exposure time. The intensity of
the incident beam Ii was calibrated by taking a single background image without the
sample in place.
The coherence width xc of the illumination at the sample plane is estimated by
calculating the width of the main lobe of the Fourier transform according to Eq. (5.2),
yielding xc = 19m. Since xc is much smaller than the finest feature size of the sample,
the lowpass filtering effect due to the finite source was ignored for the inversion. We
have also verified that blurring due to the finite spectral bandwidth of the source
is orders of magnitude smaller than xc , so our model based on monochromaticity
assumption applies in this experiment.
During the reconstruction, we first compute g for each angle, four examples of
which are shown in Fig. 5-4. Reconstruction results using three different methods
are compared in Fig. 5-6. In (a), the Fourier domain TIE solver with Tikhonov
regularization chosen to provide optimal results is used to compute phase projections
at each angle (the reconstructed phase projections for the measurements in Fig. 5-4
are shown in Fig. 5-5), and then the FBP method with a RamLak filter is applied for
the tomographic inversion. In (b), the same phase projections are inverted using the
Fourier domain tomographic reconstruction method. The results from both methods
suffer from severe streaking artifacts due to missing samples between slices in the
71
Rad-icon
CMOS camera
48m pixel size
20002000 pixels
object on
rotation
stage
z0
z0+d
Hamamatsu
micro-focus source
5m source
72
Fourier domain. Lowfrequency artifacts (blurring) around edges are also observable
but are less severe as compared to a single phase projection reconstruction. This
is likely due to denser sampling around the origin resulting from the intersection
of the Fourier slices. Both artifacts can be greatly suppressed using compressive
reconstruction with TV minimization, Eq. (5.11), whose results are shown in (c),
since TV minimization favors large structures with sharp edges. A 3D rendering of
the refractive index reconstruction by the compressive method is shown in Fig. 5-7.
5.5
A forward model that relates the refractive index distribution of the underlying sample to the intensity measurements of the straightthrough Xray beam has been developed. A compressive reconstruction method that simultaneously suppresses both
high and lowfrequency reconstruction artifacts is proposed and demonstrated experimentally for a weakly attenuating object in a microfocus Xray system.
The current forward model shows that the intensity distribution of the primary
incoherent source acts like a filter in the Fourier domain. The fringe contrast of the
measured intensity data is controlled by the source size. When the blurring due to
lowpass filtering from a mediumsized source is not severe, deconvolution techniques
might help to improve the phase reconstruction. From an opposite perspective, the
intensity distribution of the primary source provides extra degrees of freedom for
designing the phase imaging system.
Compressed sensing techniques work best when the measurement is incoherent,
i.e. the sparse information in the unknown is evenly spread out in the measurement
[19]. This is achieved in our model by the projection operator R. The measurement could be made more incoherent through the use of source coding or coded
apertures [69].
73
x (mm)
x (mm)
8
10
10
12
12
14
14
2
8
10
(a)
y (mm)
12
14
22
44
66
xx (mm)
(mm)
x (mm)
8
10
(b)
y (mm)
12
14
0.03
0.03
2mm
0.02
0.02
0.01
0.01
00
0.01
0.01
88
0.02
0.02
10
10
10
12
12
12
0.04
0.04
14
14
14
0.05
0.05
0.03
0.03
0.06
0.06
2
8
(c)
y (mm)
10
12
14
22
44
66
88
(d)
10
10
(mm)
yy(mm)
12
12
14
14
74
(a)
(b)
200
2mm
0
200
400
600
800
(c)
(d)
Figure 5-5: Phase projection reconstructions of the beetle sample at (a) = 0 , (b)
= 30 , (c) = 60 , (d) = 90 .
75
x 10
x 10
6
6
22
00
2
2
4
4
(a)
6
x1010
6
x1010
8
8
44
22
00
2
2
4
(b)
7
x 10
8x 106
8
6
6
22
00
2
2
(c)
Figure 5-6: Reconstruction results for the real part of the refractive index. (a) Fourier
based TIE solver + FBP; (b) Fourier based TIE solver + Fourier domain tomographic
inversion; (c) Compressive reconstruction by solving Eq. (5.13). The three cross
sections are taken from the three orthogonal planes going through the center of the
sample.
76
77
78
Chapter 6
Compressive Phase Space
Tomography
6.1
Introduction
In the previous three chapters, the TIE method is shown to be able to recover the
phase based on measurements of propagated intensities. When the light is partially
coherent, the TIE technique cannot recover the full statistical correlation information
about the field. In this chapter, a technique, called phase space tomography (PST),
which also relies on taking an axial stack of intensity images while the beam propagates, will be described. By using a different method to process the data, PST allows
the recovery of the full correlation function of the underlying partially coherent field.
Correlation functions provide complete characterization of wave fields in several
branches of physics, e.g. the mutual intensity of stationary quasimonochromatic
partially coherent light [70], and the density matrix of conservative quantum systems
(i.e., those with a timeindependent Hamiltonian) [11]. Classical mutual intensity
expresses the joint statistics between two points on a wavefront, and it is traditionally measured using interferometry: two sheared versions of a field are overlapped
in a Young, MachZehnder, or rotational shear [59, 71] arrangement, and twopoint
ensemble statistics are estimated as time averages by a slow detector under the assumption of ergodicity [70, 49].
79
As an alternative to interferometry, PST is an elegant method to measure correlation functions. In classical optics, PST involves measuring the intensity under spatial
propagation [79, 91, 114] or time evolution [8]. In quantum mechanics, analogous
techniques apply [118, 99, 64, 63]. However, the large dimensionality of the unknown
state makes tomography difficult. In order to recover the correlation matrix corresponding to just n points in space, a standard implementation would require at least
n2 data points.
Compressed sensing [19, 20, 40] exploits sparsity priors to recover missing data
with high confidence from a few measurements derived from a linear operator. Here,
sparsity means that the unknown vector contains only a small number of nonzero
entries in some specified basis. Lowrank matrix recovery (LRMR) [25, 27] is a
generalization of compressed sensing from vectors to matrices: one attempts to reconstruct a highfidelity and lowrank description of the unknown matrix from very
few linear measurements.
In this chapter, the experimental measurement and verification of the correlation
function of a classical partially coherent field using LRMR will be presented. It is
worth noting that LRMR came about in the context of compressive quantum state
tomography (QST) [53], which utilizes different physics to attain the same end goal of
reconstructing the quantum state. In PST, one performs tomographic projection measurements, rotating the Wigner space between successive projections by evolving the
wave function [79, 91]. This is directly analogous to the classical optical experiment
we are presenting here, where we perform intensity measurements (i.e., tomographic
projections in Wigner space) and utilize propagation along the optical axis to rotate
the Wigner space between projections. The difference lies in the fact that in QST
the state is recovered via successive applications of the Pauli dimensionalityreducing
operator, and there is no need to evolve the state. Nevertheless, both approaches lead
to the same Hermitian LRMR problem, as long as the assumption of a quasipure
unknown state is satisfied. In [52], it was shown that estimation of a lowrank matrix
of dimension n and rank r requires only O(rn ln n) to O(rn ln2 n) data points.
The lowrank assumption for classical partially coherent light anticipates a source
80
6.2
(6.1)
i 2
x1 x2
2
J(x1 , x2 ) exp (x1 x2 ) exp i2
xo dx1 dx2 .
z
z
(6.2)
(6.3)
where P denotes the propagation operator that combines both the quadratic phase
81
xo
Figure 6-1: Experimental arrangement of PST for recovering the 2D mutual intensity
J at the plane of incidence O.
and Fourier transform operations in Eq. (6.2), tr() computes the trace, and xo denotes the lateral coordinate at the observation plane. By changing variables x =
(x1 + x2 ) /2, x0 = x1 x2 and Fourier transforming the mutual intensity with respect
to x we obtain the Ambiguity Function (AF) [14, 15, 115]
0
A (u , x ) =
x0
x0
J x + ,x
exp (i2u0 x) dx.
2
2
(6.4)
(6.5)
where I is the Fourier transform of the vector of measured intensities with respect
to xo . Thus, radial slices of the AF may be obtained from Fourier transforming the
vectors of intensities measured at corresponding propagation distances, and from the
AF the mutual intensity of a 1D field can be recovered by an additional inverse Fourier
transform, subject to sufficient sampling.
The extension of Eq. (6.5) for a 2D partially coherent field [described by the 4D
82
(6.6)
2
x1 x22 y12 y22
J(x1 , y1 , x2 , y2 ) exp i
+
zxa
zya
x1 x2
y1 y2
exp i2
xo +
yo ,
zxb
zyb
(6.7)
where
1/z1 + 1/(zo z1 ) 1/f1
z1 (zo z1 )
, zxb = zo
,
1/(zo z1 ) 1/f1
f1
1/z2 + 1/(zo z2 ) 1/f2
z2 (zo z2 )
= z2
, zyb = zo
,
1/(zo z2 ) 1/f2
f2
zxa = z1
zya
(6.8)
and f1 and f2 are the focal length of the cylindrical lenses oriented along x and y,
respectively. The 2D Fourier transform of the measured intensity is related to the AF
by
zxb
zyb
(6.9)
which shows that the effective propagation distances in x and y are zxa and zya ,
respectively. By changing the distances (z1 , z2 , and zo ) and/or the focal lengths (f1
and f2 ), zxa and zya can be varied independently.
83
xo
yo
z1
zo
z2
Figure 6-2: Experimental arrangement of PST for recovering the 4D mutual intensity
J at the plane of incidence O. A pair of cylindrical lenses oriented perpendicularly
are used to introduce astigmatism to the measurement. Intensities are measured at
planes with axial coordinate zo .
In both cases, a linear model for compressive PST can formulated by first arranging
the measured intensity data in the Ambiguity space. The mutual intensity is defined
as the sparse unknown to solve for. To relate the unknowns (mutual intensity) to
the measurements (AF), the centerdifference coordinatetransform is first applied,
expressed as a linear transformation T upon the mutual intensity J, followed by
Fourier transform F, and adding measurement noise e as
A = F T J + e.
(6.10)
rank(J)
A = F T J,
X
i 0, and
i = 1.
(6.11)
k J k
A = F T J,
X
i 0, and
i = 1,
(6.12)
P
where the nuclear norm is the sum of the singular values i = |i |, k J k = i i .
This problem is convex and a number of numerical solvers can be applied to solve it.
In our implementation, we used the singular value thresholding (SVT) method [18].
The output estimate after each iteration of SVT typically has a sub-normalized total
P
energy, i.e.
i i < 1; we compensated for this by renormalizing at the end of each
iteration [53].
6.3
Numerical simulations
(6.13)
0.02
x2
0.015
0.01
0.005
x1
(a)
(b)
0.025
0.02
0.02
0.015
x2
x2
0.015
0.01
0.01
0.005
0.005
0
x1
x1
(c)
(d)
Figure 6-3: (a) Input mutual intensity of a GSMS with paramters I = 17 and
c = 13, (b) data point locations in the Ambiguity space, mutual intensities estimated
by (c) FBP and (d) LRMR methods.
and
I(x) = exp
x2
2I2
,
(x1 x2 )2
,
(x1 x2 ) = exp
2c2
(6.14)
where I determines the spatial extent of the source, and c is proportional to the
coherence length and determines the number of coherent modes in the input source.
The eigenvalues of GSMS are never zero (analytical solution given in [103]). We
defined the number of modes (rank of the source) r as the first r modes containing
the 99% of the total energy.
One example is shown in Fig. 6-3(a). The parameters in this example are I = 17
86
0.2
0.2
0.2
0.2
50
100
150
0.2
50
100
150
0.2
0.2
0.2
0.2
0.2
50
100
150
0.2
0.2
50
100
150
0.2
50
100
0.2
0.2
0.2
0.2
0.2
150
50
100
150
0.2
0.2
50
100
150
0.2
50
100
150
0.2
0.2
0.2
0.2
0.2
50
100
150
0.2
50
100
150
0.2
0.2
0.2
0.2
0.2
0.2
100
150
50
100
150
50
100
150
50
100
150
(a)
100
150
50
100
150
50
100
150
0.2
0.2
50
50
0.2
50
100
150
(b)
0.7
0.5
0.6
0.4
0.5
0.3
Eigenvalue
Eigenvalue
Figure 6-4: The first nine coherent modes of the mutual intensity in Fig. 6-3(a). (a)
Theoretical modes, and (b) LRMR estimates.
0.4
0.3
0.2
0.1
0.2
0.1
0.1
0
0
10
15
Modes
20
25
0.2
0
30
10
(a)
15
Modes
20
25
30
20
25
30
(b)
3
0.7
x 10
0.6
4
Absolute Error
Eigenvalue
0.5
0.4
0.3
0.2
1
0.1
0
0
10
15
Modes
20
25
0
0
30
(c)
10
15
Modes
(d)
Figure 6-5: Eigenvalues of the mutual intensity in Fig. 6-3(a). (a) Theoretical values, (b) FBP estimates, (c) LRMR estimates, and (d) absolute errors in the LRMR
estimates versus mode index.
where data from both the near field and the far field are missing due to the finite
range of camera scanning motion allowed in the actual experiment. The missing cone
around the u0 -axis is due to missing data from near field, while the data missing from
far field results in the missing cone around the x0 -axis. Both cones have an apex angle
of 20 degrees.
For comparison, the data are first processed using the traditional filtered-backprojection
(FBP) method [60]. Applying the Fourier-slice theorem to Eq. (6.5) implies that the
1D Fourier transform of a radial slice in the Ambiguity space (an intensity measurement) is related to a projection in the AFs 2D Fourier space (the Wigner space [5, 7]).
The Wigner distribution function (WDF) is related to the mutual intensity by
Z
W (x, u) =
x0
x0
J x + ,x
2
2
(6.15)
To implement the FBP method, each intensity projection is first filtered by a Ram
Lak kernel apodized by a Hamming window; the estimated WDF is obtained by
back-projecting all the filtered intensities, and then an inverse Fourier transform is
applied to produce the mutual intensity. Figure 6-3(c) shows the reconstructed mutual intensity following this procedure. Three types of artifacts can be seen in this
reconstruction. First, the reconstructed mutual intensity has lower values along the
diagonal of the matrix due to the missing cones. However, this is unphysical because
a correlation function should always have maximum value at zero separation. Second, the estimated degree of coherence is lower than the original field. The third
artifact is the high frequency noise around the diagonal of the matrix, which is due
to undersampling between the radial slices. All these artifacts have been greatly
suppressed or completely removed by LRMR, whose reconstruction result is shown
in Fig. 6-3(d). The disappearance of the correlation peak along the diagonal (i.e.,
the intensity) when we use FBP for the reconstruction can be best explained with
the help of Figure 6-3(b). Going from the Ambiguity space to the mutual intensity
space involves Fourier transforming along horizontal lines, parallel to the u0 axis. The
diagonal in particular corresponds to the line x0 = 0. It can be easily seen that, due
88
0.07
0.07
SNR = Inf
SNR = 30dB
SNR = 20dB
SNR = 10dB
0.06
0.05
Relative MSE
Relative MSE
0.05
0.04
0.03
0.04
0.03
0.02
0.02
0.01
0.01
0
0
SNR = 30dB
SNR = 20dB
SNR = 10dB
0.06
0
0
6 7 8 9 10 11 12 13 14 15
# of samples / r
6 7 8 9 10 11 12 13 14 15
# of samples / r
Figure 6-6: Oversampling rate versus relative MSE of LRMR estimates. The input
field is a GSMS with parameters I = 36 and c = 18. The noisy data is generated
with different SNR from (a) an additive random Gaussian noise model, and (b) a
Poisson noise model.
to the missing cone, pretty much all the data are missing from that line, except near
the origin; thus resulting in a lowpass filtering effect. The fact that the compressive
reconstruction method manages to restore the physically correct values of the correlation along the diagonal corroborates that the missing cone is successfully retrieved
in our LRMR reconstruction. The FBP reconstruction may also be compared quantitatively to the compressive reconstruction in terms of the global degree of coherence
P 2
parameter
= P |i i |i [102, 6], which was found as 0.150 and 0.617, respectively; the
i
Figure 6-7: Experimental arrangement for 2D PST. A 1D rectangular slit is illuminated by a partially coherent light. The coherence of the illumination is controlled
by the size of the slit placed at the front focal plane of the condenser. Free space
diffraction patterns after the object slit are recorded at multiple axial locations.
parameters I = 36 and c = 18 (rank r = 9). We generate noisy data with different
signal-to-noise ratio (SNR) from both an additive random Gaussian noise model and a
Poisson noise model. However, we emphasize that the reconstruction algorithm does
not make use of the noise statistics. For each SNR level, we repeat the simulation 100
times with different random noise terms, and then record the average relative meansquare-error (MSE) from the LRMR reconstruction. The ratio between the number
of samples taken from the intensity measurements and the rank r of the input mutual
intensity matrix determines the oversampling rate [21]. This rate is plotted versus
relative MSE for different SNR cases in Fig. 6-6. For good performance, the required
oversampling rate is at least 56 (the theoretical oversampling rate is on the order of
ln(256) = 5.5 according to [52]). Furthermore, the LRMR method is robust to noise
in the sense that the reconstruction degrades gracefully as the SNR decreases.
6.4
The experimental arrangement for 2D PST is illustrated in Fig. 6-7. The illumination
is generated by an LED with 620nm central wavelength and 20nm bandwidth. To
generate partially coherent illumination, a single slit of width 355.6m (0.01400 ) is
placed immediately after the LED and one focal length (75 mm) to the left of a
cylindrical lens. One focal length to the right of the lens, we place the second single
90
1500
0.03
1000
0.025
500
x (m)
0.02
0
0.015
500
0.01
1000
0.005
1500
19
26
34
41
50
58
67
77
87
99 111 125 141 160 181 207 239 280 334 409
z (mm)
1500
1500
1000
0.8
1000
500
0.6
500
0.4
500
0.2
1000
1500
0.02
x (m)
x (m)
0
u (m1)
0.01
500
1000
0
0.01
0.05
1500
0.02
0.02
(a)
0.05
0.01
0
u (m1)
0.01
0.02
(b)
Figure 6-9: (a) Real and (b) imaginary parts of the radial slices in Ambiguity space
from Fourier transforming the vectors of intensities measured at corresponding propagation distances.
slit of width 457.2m (0.01800 ), which is used as a onedimensional (1D) object.
The goal is to retrieve the mutual intensity immediately to the right of the object
from a sequence of intensity measurements at varying zdistances downstream from
the object, as described in the theory. We measured the intensities at 20 zdistances,
ranging from 18.2mm to 467.2mm, to the right of the object. The data are given
91
1500
1500
0.025
0.03
1000
1000
0.02
0.02
500
x2 (m)
x2 (m)
500
0.01
500
0.015
0
0.01
500
0.005
0
1000
1500
1000
0.01
1000
0
x1 (m)
1000
(a)
1500
0.005
1000
0
x1 (m)
1000
(b)
Figure 6-10: Real part of the reconstructed mutual intensity from (a) FBP; (b) LRMR
method.
0.4
0.35
0.3
0.3
Eigenvalue
Eigenvalue
0.25
0.2
0.1
0.2
0.15
0.1
0
0.05
0.1
0
10
15
Modes
20
25
0
0
30
(a)
10
15
Modes
20
25
30
(b)
Figure 6-11: Eigenvalues estimated by (a) FBP, and (b) LRMR method.
1500
1000
0.03
500
x2 (m)
0.035
Intensity
0.025
0.02
0.025
0.02
0.015
0.01
500
0.015
0.01
0.005
1000
0.005
0
1500
1500
1000
500
0
x(m)
500
1000
1500
1000
0
x1 (m)
(b)
0.35
0.07
0.3
0.06
0.25
0.05
Absolute Error
Eigenvalue
(a)
0.2
0.15
0.1
0.05
0
0
1000
0.04
0.03
0.02
0.01
10
15
Modes
20
25
0
0
30
(c)
10
15
Modes
20
25
30
(d)
Figure 6-12: (a) Intensity measured immediately to the right of the illumination slit;
(b) real part of van CittertZernike theorem estimated mutual intensity immediately
to the right of the object slit; (c) eigenvalues of the mutual intensity in (b); (d)
absolute error between the eigenvalues in Fig. 6-11(b) and 6-12(c) versus mode index.
94
0.2
0.2
0.2
0.2
1000
0
x (m)
1000
0.2
1000
0
x (m)
1000
0.2
0.2
0.2
1000
0
x1 (m)
1000
0.2
1000
0
x1 (m)
1000
0.2
0.2
0.2
0.2
0.2
1000
0
x1 (m)
0
x (m)
1000
1000
0
x1 (m)
1000
1000
0
x1 (m)
1000
1000
0
x (m)
1000
1000
0
x1 (m)
1000
1000
0
x1 (m)
1000
0.2
0.2
1000
1000
0.2
1000
0
x1 (m)
1000
0.2
(a)
0.2
0.2
0.2
0.2
0.2
1000
0
x (m)
1000
0.2
1000
0
x (m)
1000
0.2
0.2
0.2
0.2
0.2
0.2
1000
0
x1 (m)
1000
1000
0
x1 (m)
1000
0.2
0.2
0.2
0.2
0.2
1000
0
x1 (m)
1000
0.2
1000
0
x1 (m)
1000
(b)
Figure 6-13: (a) LRMR estimated coherent modes of the mutual intensities in Fig. 610(b), and (b) coherent modes of the mutual intensities in Fig. 6-12(b), calculated via
use of the van CittertZernike theorem, and assumption of incoherent illumination.
95
is still not perfectly compensated by the compressive approach, and because of other
experimental imperfections.
6.5
The experimental arrangement for the 4D PST is shown in Fig. 6-14. The illumination is generated by an LED with 620nm central wavelength and 20nm bandwidth.
A diffuser is used to generate a quasiuniform intensity distribution. A rectangular aperture (100m 300m) is placed immediately after the diffuser and one focal
length (100 mm) to the left of a spherical lens to control the coherence of the illumination. The goal is to retrieve the mutual intensity immediately to the right of the
circular iris object (640m). A pair of cylindrical lenses are used to intentionally
introduce astigmatism to the system; Lens 1 is oriented along x and Lens 2 is along
y, both of which have focal length 100mm. Lens 2 and the camera are placed on separate linear motion stages to allow them moving independently along the optical axis
z. For each lens figuration, an axial stack containing 23 intensity images is recorded
by displacing the camera between the back focal plane of Lens 1 (BFP1) and the
image plane of lens 2 (IP2). A total of 12 axial stacks are taken by displacing Lens 2
from the location immediately to the right of Lens 1 to BFP1. Each image contains
512512 pixels with pixel size 2.2m.
To process the data, each image is downsampled to 3232 due to memory limitation for the computation. The Fourier transforms of the downsampled intensity
images are arranged into the Ambiguity space according to Eq. (6.9). Due to the scanning range limitation, only a portion of the Ambiguity space is accessible, as shown by
the white region in Fig. 6-15; the missing data in the black region need to be compensated by the postprocessing. The LRMR and FBP reconstructed mutual intensities
are compared with the theoretical prediction based on the van CittertZernike theorem in Fig. 6-16. Similar to the 2D case, LRMR reconstruction alleviates most of
the artifacts in the FBP reconstruction. The reconstructed eigenvalues are shown in
Fig. 6-17. The LRMR reconstructed eigenvalues are very similar to the theoretical
96
300m
640m
100m
Coherence
aperture
Object
Cylindrical
lens 2
Cylindrical (along y) BFP
2
lens 1
Camera
(along x)
IP2
LED
Diffuser
Collimating
lens
BFP1
IP1
linear stage 1
linear stage 2
6.6
Discussion
In conclusion, we experimentally demonstrated compressive reconstruction of the mutual intensity function of a classical partially coherent source using phase space tomography. By exploiting the physically justifiable assumption of a quasipure source,
both measurement and postprocessing dimensionality are greatly reduced. We used
the van CittertZernike theorem to estimate the true mutual intensity function as a
way to crossvalidate the compressive reconstruction, and found indeed good agree97
ment.
In this classical experiment, we have the benefit that direct observation of the
object is available; thus, we were able to carry out quantitative analysis of the accuracy of the compressive estimate. In the quantum analogue of measuring a complete
quantum state, direct observation would have of course not been possible, but the
accuracy attained through the compressive estimate should be comparable, provided
the low entropy assumption holds [53].
Here we followed a much simplified version of the approach described in [26] which
showed that the complex operators describing the measurements should be uniformly
distributed in the ndimensional unit sphere, whereas we simply utilized free space
propagation. The phase masks described in [26] to implement optimal sampling are
outside the scope of the present work.
98
Inaccessible
Inaccessible
Inaccessible
v
Inaccessible
u
Figure 6-15: Missing data problem in the 4D PST. The white region indicates the
locations of the measured data in the Ambiguity space, and the black region the
inaccessible points due to the limitation of the experimental arrangement.
99
y2
y2
y1
x2
y1
x2
x1
x1
(a)
(b)
y2
y1
x2
x1
(c)
Figure 6-16: Real part of the mutual intensity from (a) LRMR, (b) FBP reconstructions, and (c) the van CittertZernike theorem estimation.
100
0.04
0.35
0.3
0.03
Eigenvalue
Eigenvalue
0.25
0.2
0.15
0.02
0.01
0.1
0.05
0
0
10
15
Modes
20
25
0.01
0
30
(a)
20
40
Modes
60
80
100
(b)
0.35
0.3
Eigenvalue
0.25
0.2
0.15
0.1
0.05
0
0
10
15
Modes
20
25
30
(c)
Figure 6-17: Eigenvalues estimated from (a) LRMR, (b) FBP reconstructions, and
(c) the van CittertZernike theorem.
101
(a)
(b)
Figure 6-18: The first 9 coherent modes estimated from (a) LRMR, and (b) the van
CittertZernike theorem.
102
Chapter 7
Wigner function measurement
using a lenslet array
7.1
Introduction
(7.1)
The intensity at any point is given by the integral of B over all directions
Z
I(r; z) =
B(r, p; z)d p =
(7.2)
Use of the radiance predates the wave theory of light, and it was initially described
by assigning non-negative values to all trajectories coming from source points. Such
descriptions are insufficient to model wave effects since these trajectories contain
103
(7.3)
1
S[Wi , Wp ](r, u),
f
104
(7.4)
ux
ux
...
probe
x0
x0
WDF
SWDF
(a)
detection cell
(b)
Figure 7-1: Illustration of a lens measurement of the SWDF in 1D. (a) A probe
scans across the WDF forms an SWDF measurement according to the convolution in
Eq. (7.5). (b) Points at x0 and along ux in the SWDF domain are mapped to detector
pixels at x0 + f ux at the lens Fourier plane according to Eq. (7.4).
where S[W, P ] denotes the smoothed WDF, defined as a convolution between two
WDFs:
Z
S[W, P ](r, u) =
(7.5)
and removes the need to scan. The advantage of using lenslet arrays is that the
SWDF may be measured in a single snapshot. However, if light passing through
different lenslets overlaps at the detector, the unique mapping implied by Eq. (7.4)
no longer holds. In this chapter, we conduct a more rigorous analysis, considering the
mapping between pixels on a detector and point samples of the SWDF in the presence
of crosstalk. For simplicity, we consider only scalar fields in one spatial dimension.
We show that the intensity at a detector pixel in general contains contributions from
multiple points of the SWDF due to crosstalk. In Section 2, we develop a rigorous
mathematical theory and show that both fully incoherent and fully coherent cases
have large amounts of crosstalk; it is caused by the large angular spread in the
former and high number of interference terms in the latter. In Section 3, we illustrate
tradeoffs between coherence and fidelity using a numerical example, showing that
there exists an optimal Goldilocks regime for array pitch, given the the coherence
width of the input light, such that crosstalk is reduced to a minimum without the
need for additional barriers to block light between lenslets. It is in this optimal regime
that each detector pixel corresponds to a single point in the SWDF domain, allowing
lenslet array systems to measure the SWDF with high accuracy.
7.2
Theory
xo
2w
w
Ji(x1,x2)
w
f
T (x) =
N
X
l=N
rect
x lw
w
i
2
exp (x lw) ,
f
(7.6)
J1 x + , x
= Ji x + , x
T x+
T x
,
2
2
2
2
2
2
(7.7)
where x and x0 are the center and difference coordinates, respectively, Ji is the mutual
intensity of the illumination immediately before the lenslet array; the subscript i
indicates that its associated function describes properties of the incident field at the
input plane, and we will use this notation through the rest of the chapter.
As a stepping stone to the full relationship between the incident field and the
observed intensity behind the lenslet array, we will first consider a simpler system
107
wherein we scan through the lenslets. That is, instead of letting light pass simultaneously through all the lenslets while recording the intensity image, we only let light
pass through one lenslet at a time, cycling through all the lenslets while still recording
a single image. This removes the effect of crosslenslet interference, whose derivation
we will consider later.
According to Eq. (7.4), each measurement samples the SWDF over spatial frequency with position fixed at the lenslets center, x = lw. The aperture of each
lenslet is a rect function of width w, and thus the weighting WDF is given by
Wr (x, u) =
x
sin[2u(w 2|x|)]
rect
u
w
(7.8)
1
S[Wi , Wr ]
f
lw
lw, xof
(7.9)
It is clear from this equation that the SWDF is sampled spatially at intervals of w,
the spacing of the lenslet centers. The sampling rate along the spatial frequency axis
in the SWDF is determined by both the detector pixel size and the linear mapping
u = (xo lw)/f between detector coordinate xo and spatial frequency coordinate
u. The mapping can be explained by the fact that (xo lw)/f equals to the angle
between the ray reaching the detector pixel at xo and the optical axis of the lth lenslet
under a small angle approximation. Note that if the angular spread of the SWDF is
large enough, each detector cell will include contributions to intensity not only from
the SWDF associated with its lenslet, but also from neighboring lenslets. This can
be prevented by increasing the size of the lenslets or by decreasing the angular spread
of the incident field by placing either a main lens with finite numerical aperture in
front of the array [78] or physical barriers between lenslets [33]. If we assume that
each detector cell measures only light from its associated lenslet, then we would have
108
(7.10)
N
1 X
xo lw
xo lw
=
S[Wi , Wr ] lw,
1 rect
.
f l=N
f
w
(7.11)
This term describes the additional light from other lenslets if the previous assumption
were to not hold. In this situation multiple points in the SWDF can contribute to
the same point on the detector.
To demonstrate the sampling described by Eqs. (7.97.11), an array containing
three lenslets (centered at w, 0, w) is shown in Fig. 7-3. According to Eq. (7.9),
three lines sampled at spatial coordinates w, 0, w parallel to the uaxis from the
SWDF are mapped to the detector plane (marked by different colors in Fig. 7-3).
To ensure onetoone mapping, the maximum spatial frequency um of the lth line
sample cannot exceed w/2f , as shown in case (a); otherwise, points at (lw, um ) and
[(l+1)w, um w/(f )] from the SWDF domain will be measured by the same detector
pixel at xo = lw + f um , as shown in case (b).
So far, we have only considered the incoherent superposition of light from all of the
lenslets, whereas light passing through all of the lenslets simultaneously should create
additional interference terms. Since light from lenslets separated by a distance greater
than the incident fields coherence width will not create appreciable interference when
mixed, it is useful to enumerate these crosstalk terms with an index n proportional
to the lenslet separation. All possible pairs of lenslets with indices l0 and l00 such that
109
xo
xo
w
-w
-w
-w
0
-w
SWDF
SWDF
mapping: u=(xo-lw)/f
(a)
(b)
Figure 7-3: Sampling of the SWDF using an array of three lenslets. (a) Onetoone
mapping from the SWDF to the detector coordinate according to u = (xo lw)/(f )
as the angular spread of the SWDF is narrower than the numerical aperture of a
lenslet. (b) Multiple points in the SWDF domain contribute to detector pixels in
the crosstalk region as the angular spread of the incident field is wider than the
numerical aperture of a lenslet, which produces the 0th order crosstalk.
(n)
|l0 l00 | = n > 0 contribute to the nth order crosstalk term Ic , given by
2
Ic(n) (xo ) =
f
N n
2
X x
Wi (x, u)Wr
l=N + n
2
xo lw
x lw, u
f
x xo
+ u nw dxdu.
cos 2
f
(7.12)
Note that when n is odd, l takes a value halfway between two integers, and thus Wr
is centered at the edge between the (l 1/2)th and (l + 1/2)th lenslets; when n is
even, l takes every integer value, thus Wr is centered at the lth lenslet. We expect
the n = 1 term to be significant even in highly incoherent fields, since some points
near the boundary between two neighboring lenslets are expected to be within the
coherence width of the field.
The total output intensity, considering all of the discussed effects, can be written
as the sum of three components
2N
X
n=1
110
Ic(n) (xo ).
(7.13)
A detailed derivation of this result, obtained by performing Fresnel propagation integrals on Eq. (7.7), is given in Appendix A. Equation (7.13) demonstrates that if all
orders of the crosstalk could be made small, then the measured intensity would be
an accurate representation of the SWDF. In order for the crosstalk to be negligible,
(0)
both the angular spread of the SWDF should be small [for Ic (xo )] and the coherence
(n)
width should be less than the width of a single lenslet [for Ic (xo )]. In order to optimally measure the SWDF, the angular and coherence widths of the SWDF should
be balanced so that as much of each lenslets detector cell is utilized as possible while
minimizing crosstalk. It should also be noted that even with minimal cross-talk, the
measurement yields only the SWDF; recovery of the mutual intensity (or WDF) of
the field still requires deconvolution of the SWDF with the aperture WDF.
7.3
Numerical Example
We study the effect of coherence width on the quality of the resulting measurement by
studying the following example. Let us consider a spatially homogeneous Gaussian
correlated Schellmodel incident field, which can be described by the following mutual
intensity,
(x2 x1 )2
Ji (x1 , x2 ) = exp
.
2c2
(7.14)
(7.15)
where u = 1/(2c ) quantifies the spatial frequency bandwidth of the WDF and
is proportional to the angular spread of the field. The SWDF resulting from the
convolution between the WDF of the input field and that of a rectangular aperture
is
w2
S[Wi , Wr ](x, u) =
2u
(u u0 )2
exp
2u2
111
sin(wu0 )
wu0
2
du0 .
(7.16)
N
X
xo lw
w2
1 rect
=
w
2u f l=N
2
Z
[(xo lw)/f u0 ]2
sin(wu0 )
exp
du0 .
2u2
wu0
(7.18)
The nth order crosstalk term by carrying out the integration in Eq. (7.12) is
2w2
Ic(n) (xo ) =
2u f
N n
2
X Z
l=N + n
2
[(xo lw)/f u0 ]2
exp
2u2
cos(2nwu0 )
(7.19)
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0.5
0.1
0.01
0.4
0.08
0.008
0.3
0.06
0.006
0.2
0.04
0.004
0.1
0.02
0.002
0
2
0.5
0.01
0.01
0.4
0.008
0.008
0.3
0.006
0.006
0.2
0.004
0.004
0.1
0.002
0.002
0.1
0.1
0.01
0.08
0.08
0.008
0.06
0.06
0.006
0.04
0.04
0.004
0.02
0.02
0.002
xo /w
xo /w
xo /w
Figure 7-4: Left: highly incoherent; middle: highly coherent; and right: partially
coherent case. (a) Total output intensity is composed of (b) SWDF term and (c)
total contribution from crosstalk terms. The total crosstalk is composed of (d)
0th order crosstalk and (e) total of higher order crosstalk. All the intensities are
normalized to the maximum value in the total output. The horizontal axis is the
spatial coordinate normalized by the width of a lenslet.
113
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.2
0.2
0.4
0.2
0
0.5
WDF
SWDF
Measured intensity
0
0
0.5 0.5
xo /w
highly incoherent (=0.01w)
0
xo /w
highly coherent (=20w)
0
0.5 0.5
0
xo /w
partially coherent (=0.1w)
0.5
Figure 7-5: Comparison of WDF (solid red line), SWDF (dashed blue lines) and
measured intensity (dotted green lines) for (a) highly incoherent (c = 0.01w), (b)
highly coherent (c = 20w), and (c) partially coherent (c = 0.1w) incident light.
coherent case, shown in the middle column (c = 20w). Here, most of the crosstalk
comes from higher order terms. The results for a partially coherent field (c = 0.1w)
is shown in the right column; crosstalk contributes minimally to the final intensity,
although both 0th order and higher order terms are present.
The effect of the convolution with the aperture on the accuracy of WDF estimation
is studied as follows. In Fig. 7-5 we compare the measured intensity to the actual
SWDF and WDF for each of the three different fields. Since the fields are spatially
homogeneous, we only look at slices along u of the WDF (solid red lines) and SWDF
(dashed blue lines), mapped to the xo coordinate. The intensity distribution behind
the central lenslet is shown using dotted green lines. In both the highly incoherent
and partially coherent cases, the SWDF and WDF are very similar, since the WDF
of the aperture is much smaller than any variations in the incident WDFs. In the
highly coherent case, the incident WDF is narrower in u than the aperture WDF,
and therefore the SWDF is significantly broadened by the convolution. In order to
recover the WDF from the measured intensity, deconvolution is necessary [30].
We define a total error metric Rerror by measuring the difference between the
output intensity and the original WDF as
Rerror =
(7.20)
Note that the sampled WDF refers to the values of the WDF over a single line in
xu space corresponding to the line of the SWDF mapped to the output intensity. To
quantify the crosstalk corruption in the output, define the crosstalk power fraction
Rcrosstalk as
Rcrosstalk =
(7.21)
The signal broadening caused by the convolution in the SWDF is also considered by
qualifying the difference between the SWDF and the WDF by the signal broadening
metric Rconv defined as
Rconv =
(7.22)
Here, both the SWDF and WDF are sampled over the same line.
All these variables are plotted as functions of the coherence of incident light (measured by the ratio of c to w) in Fig. 7-6. As seen in the dashed green curve, the contribution from crosstalk increases quickly as the field becomes less coherent. When
the field becomes more coherent, the contribution from crosstalk also increases until
it saturates to the point in which the field is coherent within the whole array. There
exists a partially coherent regime where the SWDF can be measured with minimal
crosstalk corruption. Depending on accuracy requirements, this regime may provide
acceptable measurements. For example, if less than 1% of crosstalk can be tolerated,
then the coherence width should be such that 0.02w < c < w. On the other hand,
signal broadening increases as the field becomes more coherent, making the SWDF a
less accurate estimate of the WDF in these situations. The total error metric, which
considers artifacts from both crosstalk and signal broadening, has a similar shape
to the crosstalk curve. The measurement deviates from the original WDF except
in a partially coherent region. If error needs to be at most 1%, then we would need
0.02w < c < 0.4w.
115
10
R error
R crosstalk
R conv
10
10
10 2
10
10
10
c/w
10
10
Figure 7-6: Error metric Rerror in solid blue curve, crosstalk power fraction Rcrosstalk
in dashed green curve, and signal broadening metric Rconv in red dotted curve as
functions of the normalized coherence length of incident light c /w.
7.4
Concluding Remarks
Although the numerical example was chosen explicitly to consider the effect of coherence width on the measurement of the SWDF using a lenslet array, this simple model
can also provide useful insights for a much broader class of fields whose intensity
varies slowly across the field, with features much wider than the coherence width. As
a rule of thumb, higher order (coherent) cross-talk can be reduced by ensuring that
the lenslet apertures are at least one coherence width in size. This makes intuitive
sense, since an aperture larger than the coherence width will not cause the incident
beam to diffract significantly, and any light that is diffracted from the aperture will
not interfere with that from neighboring lenslets. Both 0th and higher order crosstalk
can be reduced by ensuring the incident illuminations angular spread is such that
each lenslet primarily illuminates only the pixels lying within its detector cell, such
that there is a nearly onetoone mapping from SWDF space to each detector pixel.
It should also be noted that we have derived these results under the paraxial
approximation and that both the 0th and higher order crosstalk can include contributions for which the light propagates highly nonparaxially from one lenslet to its
neighbors. In these cases, we expect that a similar analysis can be performed using
116
nonparaxial versions of the Wigner function [128, 32], although this is outside the
scope of our current work.
As was discussed while analyzing the example, there are cases where the SWDF
is not an accurate estimate of the WDF. Performing deconvolution to recover the
WDF may benefit from techniques such as coded apertures [65, 47] and compressed
sensing [112, 134].
117
118
Chapter 8
Conclusion and future work
This thesis has explored various imaging methods and computational algorithms for
quantitative phase retrieval. A fundamental problem that the author hope to address
is the recovery of a full description of wave fields (either fully coherent or partially
coherent) with a small number of measurements. Compressed sensing provides an
intriguing recipe that allows the recovery of an unknown signal with the number of
measurements scaled by the information rate (sparsity) of the signal. A central
idea that has been explored throughout is that free space propagation provides easily
accessible phase information. For instance, by recording the propagated wavefront
through holography, 3D information can be efficiently encoded on a 2D plane; the
transport of intensity relies on defocus to measure the curvature of the phase; the
coherence state of light can be recovered by phase space tomography using a stack of
intensity measurements during propagation.
The main contributions of the thesis include:
Beyond this work, there are many more unexplored problems in the application of
novel sensing method to quantitative phase recovery problem. Venue to extend the
current work includes:
In holography, it is well known that the depth localization capability is much
worse than the one in the lateral dimension due to the missing spatial frequency
samples in the longitudinal direction. Compressive reconstruction might help
to improve the result. A rigorous study on the limit of axial localization improvement by using compressive reconstruction methods may be interesting for
future study.
The effect of the partially coherent illumination on the TIE is treated as artifacts
which needs to be compensated in the current study. However, active control
over illumination could also create opportunities on obtaining phase information
with better contrast or making the inverse problem better conditioned. Coded
source and/or combination of measurements with different source figurations
are interesting directions for future work.
120
Phase space tomography takes projection measurements of the Wigner distribution function while rotating the phase space. It may be possible to tailor the
sampling trajectories in the Wigner space beyond the tomographic approach.
To do that, optical elements, such as volume holograms, coded apertures or
nonlinear materials, may be introduced in the path between the unknown wave
and detector. Most light sources are not fully coherent; since partially coherent
waves contain many more degrees of freedom than coherent waves, they are
more difficult to deal with, yet they offer more opportunity for manipulation of
light in ways that coherent light cannot. This will lead to cases where information is encoded in or retrieved from the wave, with applications in optical
information science, lithography, X-ray sciences, medical imaging, and biological microscopy. It is also interesting to note that an analogous problem exists
in quantum mechanics for the recovery of quantum correlations (the density
matrix). The understanding of this problem might have further implications to
quantum information theory.
121
122
Appendix A
Proof of Eq. (7.13)
Appendix A: Proof of Eq. (7.13)
Assume the mutual intensity of the incident field is
Ji (x1 , x2 ) = hUi (x1 )Ui (x2 )i
(A.1)
x0
x0
Ji x + , x
exp(i2ux0 )dx0 ,
2
2
(A.2)
where
x=
x1 + x2 0
, x = x1 x2 ,
2
(A.3)
are the center and difference coordinates; u is the local spatial frequency variable.
The mutual intensity immediately to the right of the lenslet array is [70]
J1 (x1 , x2 ) = Ji (x1 , x2 )T (x1 )T (x2 ),
123
(A.4)
T (x) =
N
X
rect
l=N
x lw
w
i
2
exp (x lw) .
f
(A.5)
The intensity at the output plane xo at one focal length to the right of the lenslet
array is related to J1 by double Fresnel integrals under the paraxial approximation,
1 x
i
2
2
I(xo ) =
J1 (x1 , x2 ) exp
(xo x1 ) (xo x2 )
dx1 dx2 .
f
f
(A.6)
Substitutions of Eqs. (A.5) and (A.4) into Eq. (A.6) shows that the calculation of
I(xo ) requires in general carrying out double summations with respect to different
lenslet indices l1 and l2 . After some simplification, the output intensity is rewritten
as
N
N
1 X X x
x1 l1 w
x2 l2 w
I(xo ) =
Ji (x1 , x2 )rect
rect
f l =N l =N
w
w
1
2
2
2
2
2 w
exp i
(l1 l2 )
+ (x1 x2 )xo l1 wx1 + l2 wx2 dx1 dx2 .
f
2
(A.7)
The double summations are simply due to the fact that the intensity at the region
directly behind a particular lenslet results from both the field passing through the
local lenslet and crosstalk from the field passing through neighboring lenslets. We
proceed with the derivation by the change of variables in Eq. (A.3) and
m = l1 + l2 , n = l1 l2 .
(A.8)
The consequence of the change of variables in Eq. (A.8) can be studied as follows.
(A.9)
(A.10)
and
m = 2l, where l = N + |q|, N + |q| + 1, ..., N |q| 1, N |q|.
(A.11)
(A.12)
l1 = l + q, l2 = l q,
(A.13)
and
with
m = 2l, where l = N + |q|, N + |q| + 1, ..., N |q| 1, N |q|.
(A.14)
Notice that a term of fixed n contributes a nonzero value to I(xo ) only if the two rect
functions overlap. This implies that the separation x0 between the pair of correlating
points on the incident field can only take certain values, as determined by the following
125
inequalities
|x lw| < w/4,
(A.16)
(A.17)
Eq. (A.17) implies that x0 is bounded to a region of width 2w 4|x lw| centered
at nw. Also recall that the magnitude of mutual intensity is significantly larger than
zero at large separation distance x0 only if the field is highly coherent. This implies
that more terms in the summation over n need to be considered if the field is more
coherent. To simplify Eq. (A.15), we relate I(xo ) to the WDF of the incident field
and the WDF of a rectangular aperture of width w
Z
Wr (x, u) =
rect
x + x0 /2
w
rect
x x0 /2
w
exp(i2u x0 )dx0 ,
(A.18)
2N
1
X
n=2N +1
odd
2N
X
n=2N
even
|n|
X2
l=N +
N
|n|
X2
l=N +
|n|
2
|n|
2
xo lw
x xo
Wi (x, u)Wr x lw, u
+ u nw dxdu
exp i2
f
f
xo lw
x xo
Wi (x, u)Wr x lw, u
+ u nw dxdu.
exp i2
f
f
(A.19)
126
l=N + 2
n
N
2N
x xo
2 X X2 x
xo lw
cos 2
+ u nw dxdu.
+
Wi (x, u)Wr x lw, u
f n=2
f
f
n
even
l=N + 2
127
128
Bibliography
[1] E. H. Adelson and J. Y. A. Wang. Single lens stereo with a plenoptic camera.
IEEE Trans. Pattern Anal. Mach. Intell., 14(2):99106, 1992.
[2] B. Allman, P. McMahon, K. Nugent, D. Paganin, D. Jacobson, M. Arif, and
S. Werner. Imaging: phase radiography with neutrons. Nature, 408:158159,
2000.
[3] E. D. Barone-Nugent, A. Barty, and K. A. Nugent. Quantitative phaseamplitude microscopy I: optical microscopy. Journal of Microscopy, 206(3):194
203, 2002.
[4] H.O. Bartelt, K.-H. Brenner, and A.W. Lohmann. The Wigner distribution
function and its optical production. Optics Communications, 32(1):32 38,
1980.
[5] M. J. Bastiaans. The Wigner distribution function applied to optical signals
and systems. Optics Communications, 25:2630, April 1978.
[6] Martin J. Bastiaans. New class of uncertainty relations for partially coherent
light. J. Opt. Soc. Am. A, 1(7):711715, Jul 1984.
[7] Martin J. Bastiaans. Application of the Wigner distribution function to partially
coherent light. J. Opt. Soc. Am. A, 3(8):12271238, Aug 1986.
[8] M. Beck, M. G. Raymer, I. A. Walmsley, and V. Wong. Chronocyclic tomography for measuring the amplitude and phase structure of optical pulses. Opt.
Lett., 18(23):20412043, Dec 1993.
[9] M. Bertero and P. Boccacci. Introduction to inverse problems in imaging. Taylor
& Francis, 1998.
[10] J. M. Bioucas-Dias and M. A. T. Figueiredo. A New TwIST: Two-Step Iterative
Shrinkage/Thresholding algorithms for image restoration. IEEE Transactions
on Image Processing, 16:29923004, December 2007.
[11] K. Blum. Density matrix theory and applications. Plenum Press, 1981.
[12] M. Born and E. Wolf. Principles of Optics: Electromagnetic Theory of Propagation, Interference and Diffraction of Light. Cambridge University Press, 7
edition, October 1999.
129
[13] David J. Brady, Kerkil Choi, Daniel L. Marks, Ryoichi Horisaki, and Sehoon
Lim. Compressive holography. Opt. Express, 17(15):1304013049, Jul 2009.
[14] K.-H. Brenner, A.W. Lohmann, and J. Ojeda-Casta
neda. The ambiguity function as a polar display of the OTF. Opt. Commun., 44(5):323 326, 1983.
[15] K.-H. Brenner and J. Ojeda-Casta
neda. Ambiguity function and Wigner distribution function applied to partially coherent imagery. Opt. Acta., 31:213223,
February 1984.
[16] Andrei V. Bronnikov. Theory of quantitative phase-contrast computed tomography. J. Opt. Soc. Am. A, 19(3):472480, Mar 2002.
[17] Anna Burvall, Ulf Lundstrom, Per A. C. Takman, Daniel H. Larsson, and
Hans M. Hertz. Phase retrieval in x-ray phase-contrast imaging suitable for
tomography. Opt. Express, 19(11):1035910376, May 2011.
[18] J.-F. Cai, E. J. Cand`es, and Z. Shen. A singular value thresholding algorithm
for matrix completion. ArXiv: 0810.3286, October 2008.
[19] E. Cand`es, J. Romberg, and T. Tao. Robust uncertainty principles: exact signal
reconstruction from highly incomplete frequency information. IEEE Trans.
Inform. Theory, 52(2):489509, 2006.
[20] E. Cand`es, J. Romberg, and T. Tao. Stable signal recovery from incomplete and
inaccurate measurements. Comm. Pure Appl. Math., 59(8):12071223, August
2006.
[21] E. J. Cand`es, Y. Eldar, T. Strohmer, and V. Voroninski. Phase retrieval via
matrix completion. ArXiv: 1109.0573, September 2011.
[22] E. J. Cand`es and Y. Plan. Matrix completion with noise. ArXiv: 0903.3131,
March 2009.
[23] E.J. Candes and T. Tao. Near-optimal signal recovery from random projections:
Universal encoding strategies? IEEE Transactions on Information Theory,
52(12):54065425, Dec. 2006.
[24] Emmanuel J. Cand`es and Yaniv Plan. A probabilistic and ripless theory of
compressed sensing. IEEE Transactions on Information Theory, 57(11):7235
7254, 2011.
[25] Emmanuel J. Cand`es and Benjamin Recht. Exact matrix completion via convex
optimization. Found. Comput. Math., 9(6):717772, 2009.
[26] Emmanuel J. Cand`es, Thomas Strohmer, and Vladislav Voroninski. Phaselift:
exact and stable signal recovery from magnitude measurements via convex programming. ArXiv: 1109.4499v1, 09 2011.
130
[27] Emmanuel J. Cand`es and Terence Tao. The power of convex relaxation: nearoptimal matrix completion. IEEE Trans. Inform. Theory, 56:20532080, May
2010.
[28] Emmanuel J Cand`es and Michael B Wakin. An introduction to compressive
sampling. IEEE Signal Processing Magazine, 25(2):2130, 2008.
[29] D. Chapman, W. Thomlinson, RE Johnston, D. Washburn, E. Pisano, N. Gm
ur,
Z. Zhong, R. Menk, F. Arfelli, and D. Sayers. Diffraction enhanced x-ray imaging. Physics in medicine and biology, 42(11):2015, 1999.
[30] Henry N. Chapman. Phase-retrieval X-ray microscopy by Wignerdistribution
deconvolution. Ultramicroscopy, 66(3-4):153 172, 1996.
[31] F. C. Cheong, B. Sun, R. Dreyfus, J. Amato-Grill, K. Xiao, L. Dixon, and
D. G. Grier. Flow visualization and flow cytometry with holographic video
microscopy. Opt. Express, 17(15):1307113079, 2009.
[32] S. Cho, JC Petruccelli, and MA Alonso. Wigner functions for paraxial and
nonparaxial fields. Journal of Modern Optics, 56(17):18431852, 2009.
[33] Heejin Choi, Sung-Wook Min, Sungyong Jung, Jae-Hyeung Park, and Byoungho
Lee. Multiple-viewing-zone integral imaging using a dynamic barrier array for
three-dimensional displays. Opt. Express, 11(8):927932, Apr 2003.
[34] Christy Fernandez Cull, David A. Wikner, Joseph N. Mait, Michael Mattheiss,
and David J. Brady. Millimeter-wave compressive holography. Appl. Opt.,
49(19):E67E82, Jul 2010.
[35] TJ Davis, D Gao, TE Gureyev, AW Stevenson, and SW Wilkins. Phasecontrast imaging of weakly absorbing materials using hard x-rays. Nature,
373(6515):595598, 1995.
[36] Loc Denis, Dirk Lorenz, Eric Thiebaut, Corinne Fournier, and Dennis
Trede. Inline hologram reconstruction with sparsity constraints. Opt. Lett.,
34(22):34753477, Nov 2009.
[37] LS Dolin. Beam description of weakly-inhomogeneous wave fields. Izv. Vyssh.
Uchebn. Zaved. Radiofiz, 7:559563, 1964.
[38] J. A. Dominguez-Caballero and G. Barbastathis. Stability of inversion in digital
holographic particle imaging: Theory and experimental validation. In Frontiers
in Optics, page FThV4. Optical Society of America, 2008.
[39] Jose A. Dominguez-Caballero, Nick Loomis, Weichang Li, Qiao Hu, Jerome
Milgram, George Barbastathis, and Cabell Davis. Advances in plankton imaging
using digital holography. In Computational Optical Sensing and Imaging, page
DMB5. Optical Society of America, 2007.
131
[40] David L. Donoho. Compressed sensing. IEEE Trans. Inform. Theory, 52:1289
1306, 2006.
[41] Thomas Dresel, Gerd Hausler, and Holger Venzke. Three-dimensional sensing
of rough surfaces by coherence radar. Applied Optics, 31(7):919925, 1992.
[42] F. Dubois, C. Schockaert, N. Callens, and C. Yourassowsky. Focus plane detection criteria in digital holography microscopy by amplitude analysis. Opt.
Express, 14(13):58955908, June 2006.
[43] J.A. Fessler and B.P. Sutton. Nonuniform fast fourier transforms using min-max
interpolation. IEEE Transactions on Signal Processing, 51(2):560574, 2003.
[44] J. R. Fienup. Phase retrieval algorithms:
21(15):27582769, 1982.
a comparison.
Appl. Opt.,
[45] Corinne Fournier, Loc Denis, and Thierry Fournel. On the single point resolution of on-axis digital holography. J. Opt. Soc. Am. A, 27(8):18561862, Aug
2010.
[46] A. T. Friberg. On the existence of a radiance function for finite planar sources
of arbitrary states of coherence. J. Opt. Soc. Am., 69(1):192198, 1979.
[47] Michael E. Gehm, Scott T. McCain, Nikos P. Pitsianis, David J. Brady, Prasant
Potuluri, and Michael E. Sullivan. Static two-dimensional aperture coding for
multimodal, multiplex spectroscopy. Appl. Opt., 45(13):29652974, May 2006.
[48] RW Gerchberg. A practical algorithm for the determination of phase from image
and diffraction plane pictures. Optik, 35:237, 1972.
[49] J. W. Goodman. Statistical Optics. Wiley-Interscience, July 2000.
[50] R. Gordon, R. Bender, and G.T. Herman. Algebraic reconstruction techniques
(ART) for three-dimensional electron microscopy and x-ray photography. Journal of theoretical Biology, 29(3):471, 1970.
[51] A. Groso, R. Abela, and M. Stampanoni. Implementation of a fast method for
high resolution phase contrast tomography. Opt. Express, 14(18):81038110,
Sep 2006.
[52] D. Gross. Recovering low-rank matrices from few coefficients in any basis. IEEE
Trans. Inf. Theory, 57(3):1548 1566, march 2011.
[53] David Gross, Yi-Kai Liu, Steven T. Flammia, Stephen Becker, and Jens Eisert. Quantum state tomography via compressed sensing. Phys. Rev. Lett.,
105(15):150401, Oct 2010.
[54] Jean Pierre Guigay, Max Langer, Renaud Boistel, and Peter Cloetens. Mixed
transfer function and transport of intensity approach for phase retrieval in the
fresnel region. Opt. Lett., 32(12):16171619, Jun 2007.
132
[82] J OSullivan. A fast sinc function gridding algorithm for fourier inversion in
computer tomography. IEEE Transactions on Medical Imaging, 4(4):200207,
1985.
[83] D. Paganin, A. Barty, P. J. McMahon, and K. A. Nugent. Quantitative
phase-amplitude microscopy. III. The effects of noise. Journal of Microscopy,
214(1):5161, 2004.
[84] D. Paganin and K. A. Nugent. Noninterferometric phase imaging with partially
coherent light. Physical Review Letters, 80(12):25862589, 1998.
[85] G. Pan and H. Meng. Digital holography of particle fields: reconstruction by
use of complex amplitude. Appl. Opt., 42(5):827833, February 2003.
[86] Jae-Hyeung Park, Keehoon Hong, and Byoungho Lee. Recent progress in threedimensional information processing based on integral imaging. Appl. Opt.,
48(34):H77H94, 2009.
[87] Daniele Pelliccia, Andrei Y. Nikulin, Herbert O. Moser, and Keith A. Nugent.
Experimental characterization of the coherence properties of hard x-ray sources.
Opt. Express, 19(9):80738078, Apr 2011.
[88] P. Perona and J. Malik. Scale-space and edge detection using anisotropic diffusion. Pattern Analysis and Machine Intelligence, IEEE Transactions on,
12(7):629 639, July 1990.
[89] Franz Pfeiffer, Timm Weitkamp, Oliver Bunk, and Christian David. Phase retrieval and differential phase-contrast imaging with low-brilliance x-ray sources.
Nature Physics, 2(4):258261, 2006.
[90] B. C. Platt and R. Shack. History and principles of Shack-Hartmann wavefront
sensing. Journal of Refractive Surgery, 17, 2001.
[91] M. G. Raymer, M. Beck, and D. McAlister. Complex wave-field reconstruction
using phase-space tomography. Phys. Rev. Lett., 72(8):11371140, Feb 1994.
[92] Leonid I. Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation
based noise removal algorithms. Physica D: Nonlinear Phenomena, 60(14):259
268, 1992.
[93] Otmar Scherzer and Joachim Weickert. Relations between regularization and
diffusion filtering. Journal of Mathematical Imaging and Vision, 12(1):4363,
2000.
[94] U. Schnars and W. P. Juptner. Direct recording of holograms by a CCD target
and numerical reconstruction. Appl. Opt., 33:179181, January 1994.
[95] Ulf Schnars and Werner Jueptner. Digital holography: digital hologram recording, numerical reconstruction, and related techniques. Springer, 2004.
135
[96] Sungkyu Seo, Ting-Wei Su, Derek K. Tseng, Anthony Erlinger, and Aydogan
Ozcan. Lensfree holographic imaging for on-chip cytometry and diagnostics.
Lab Chip, 9:777787, 2009.
[97] Colin J. R. Sheppard. Defocused transfer function for a partially coherent
microscope and application to phase retrieval. J. Opt. Soc. Am. A, 21(5):828
831, May 2004.
[98] Lawrence Sirovich and Michael Kirby. Low-dimensional procedure for the characterization of human faces. JOSA A, 4(3):519524, 1987.
[99] D. T. Smithey, M. Beck, M. G. Raymer, and A. Faridani. Measurement of
the Wigner distribution and the density matrix of a light mode using optical
homodyne tomography: application to squeezed states and the vacuum. Phys.
Rev. Lett., 70(9):12441247, Mar 1993.
[100] Saowapak Sotthivirat and Jeffrey A. Fessler. Penalized-likelihood image reconstruction for digital holography. J. Opt. Soc. Am. A, 21(5):737750, May
2004.
[123] A. Wax and J. E. Thomas. Optical heterodyne imaging and Wigner phase space
distributions. Opt. Lett., 21(18):14271429, 1996.
[124] J. Weickert, B.M.T.H. Romeny, and M.A. Viergever. Efficient and reliable
schemes for nonlinear diffusion filtering. Image Processing, IEEE Transactions
on, 7(3):398 410, mar 1998.
[125] Joachim Weickert. A review of nonlinear diffusion filtering. In Scale-Space
Theory in Computer Vision, volume 1252 of Lecture Notes in Computer Science,
pages 128. Springer Berlin / Heidelberg, 1997.
[126] E. Wigner. On the quantum correction for thermodynamic equilibrium. Physical
Review, 40(5):07490759, June 1932.
[127] Emil Wolf. New theory of partial coherence in the space-frequency domain.
Part I: spectra and cross spectra of steady-state sources. J. Opt. Soc. Am.,
72(3):343351, Mar 1982.
[128] Kurt Bernardo Wolf, Miguel Angel Alonso, and Gregory W. Forbes. Wigner
functions for helmholtz wave fields. J. Opt. Soc. Am. A, 16(10):24762487, Oct
1999.
[129] Xizeng Wu and Hong Liu. Clinical implementation of x-ray phase-contrast
imaging: Theoretical foundations and design considerations. Medical physics,
30:2169, 2003.
[130] Y. Yang, B. Kang, and Y. Choo. Application of the correlation coefficient
method for determination of the focal plane to digital particle holography. Appl.
Opt., 47(6):817824, 2008.
[131] Frits Zernike. Phase contrast, a new method for the microscopic observation of
transparent objects. Physica, 9:686698, 1942.
[132] Frits Zernike. How I discovered phase contrast. Science, 121(3141):345, 1955.
[133] Z. Zhang and M. Levoy. Wigner distributions and how they relate to the
light field. In IEEE International Conference on Computational Photography
(ICCP), pages 110. IEEE, 2009.
[134] Zhengyun Zhang, Zhi Chen, Shakil Rehman, and George Barbastathis. Factored form descent: a practical algorithm for coherence retrieval. Opt. Express,
21(5):57595780, Mar 2013.
[135] Adam M. Zysk, Robert W. Schoonover, P. Scott Carney, and Mark A. Anastasio. Transport of intensity and spectrum for partially coherent fields. Opt.
Lett., 35(13):22392241, Jul 2010.
138