July 1978 / Vol. 3, No.

1 / OPTICS LETTERS 27
Reconstruction of an object from the modulus
of its Fourier transform
J. R. Fienup
Environmental Research Institute of Michigan, P.O. Box 8618, Ann Arbor, Michigan 48107
Received February 23, 1978
We present a digital method for solving the phase-retrieval problem of optical-coherence theory: the reconstruc-
tion of a general object from the modulus of its Fourier transform. This technique should be useful for obtaining
high-resolution imagery from interferometer data.
Ordinarily, atmospheric turbulence limits the resolu-
tion of an image obtained through a large telescope.
There have been developed a number of interferometric
techniques,
1
such as speckle interferometry
2
and am-
plitude interferometry,
3
for making diffraction-limited
measurements of the modulus of the Fourier transform
of a space object. Although the autocorrelation of an
object can be directly computed from the Fourier
modulus, the object itself generally cannot be computed
without knowledge of the Fourier transform phase,
except for the special cases of centro-symmetric objects
and objects with nearby isolated point sources. If it
were possible to reconstruct an object from the modulus
of its Fourier transform, then very-high-resolution
imagery could be obtained from interferometer data.
One possible method of reconstructing an object
distribution from the modulus of its Fourier transform
is to use the dispersion relation and the locations of the
complex zeros of the Fourier modulus.
4
Another pos-
sibility is an iterative Newton-Raphson approach.
5
Neither of these two methods has proved practical for
complicated two-dimensional imagery. We describe
here a reconstruction method that works for general
objects, even with noisy Fourier modulus data.
Let the object be f (x) and its Fourier transform be
F(u) = JF(u)l exp[ii(u)] = 5(Jf(x)}
modifications of a fundamental nature to arrive at the
more powerful input-output approach.
A block diagram of the error-reduction approach is
shown in Fig. 1(a). At the kth iteration, gk (x), an es-
timate of the object, is Fourier transformed; the Fourier
transform is made to conform to the known modulus;
and the result is inverse-Fourier transformed, giving the
image gk'(X). Then the iteration is completed by
forming a new estimate of the object that conforms to
the object-domain constraints:
(2) gk+l(X) = Og'(X) X , Y
I, x e Y
where the region -y includes all points at which g'(x)
violates the constraints. The principal constraint is
that the object be nonnegative. An additional con-
straint that may be enforced is that the diameter may
not exceed the known diameter of the object (which is
half the diameter of the autocorrelation). The itera-
tions can be started by using a sequence of random
numbers for g
1
(x) or for 01(x).
The mean-squared error at each iteration can be de-
fined in the Fourier domain by
S T [IGk(u)I -IF(u)I]
2
du
EF
2
(3
fI' f (x) exp(-i2 7ru * x)dx,
= f__
(1)
where the vector position x represents a two-dimen-
sional spatial coordinate and u a spatial frequency. For
sky brightness objects, f(x) is a real, nonnegative
function. The problem is to find an object that is con-
sistent with all the known constraints: that it be non-
negative and that the modulus of its Fourier transform
equal the measured modulus, IF(u)l.
The problem is solved by an iterative approach, which
is a modified version of the Gerchberg-Saxton algo-
rithm that has been used in electron microscopy
6
and
other applications.
7
We first modified the
Gerchberg-Saxton algorithm to fit this problem merely
by using a new set of object constraints, and refer to this
as the error-reduction approach. We made other
T J I F(u)I
2
du
where Gk (u) = 5I(g, (x)}, or in the object (image) domain
by
INPUT g
g EC O G = IGI 0;
SATISFY SATISFY
OBJECT | FOURIER
CONSTRAINTS | CONSTRAINTS
g' -fH U~IG'=IFIe
(a)
OUTPUT g-
(b)
Fig. 1. (a) Block diagram of the error-reduction approach;
(b) block diagram of the system for the input-output con-
cept.
0146-9592/78/0700-0027$0.50/0 1978, Optical Society of America
(3)
28 OPTICS LETTERS / Vol. 3, No. 1 / July 1978
I-
z I
z
Fig.2.
lutions.
x
Reconstruction results: test object and 15 other
S S [gk'(x)]
2
dx
E02 -
4' [gk'(X)]
2
dx
In a way similar to that for the previous applicatic
it can be shown that the mean-squared error can o
decrease at each iteration,
6
,8 giving rise to the na
error-reduction approach. In practice, when
error-reduction approach is used for the present
plication for large two-dimensional images, the me
squared error decreases rapidly for the first few itE
tions but then decreases extremely slowly for later
erations, requiring an impractically large numbei
iterations for convergence.
In an attempt to speed up the convergence, we
veloped the more powerful input-output approach.
The input-output approach differs from the ern
reduction approach only in the object-domain opE
tion, and the other three operations are the sai
Together, those three operations, as shown in Fig. 1
can be viewed as a nonlinear system having an in
g(x) and an outputg'(x). A characteristic of this k
tem is that any output of the system has a Fou,
transform with a modulus equal to IF(u)l. Cor
quently, if the output can be forced to conform to
object-domain constraints, then it is a solution to
problem. Instead of modifying the last output, as in
(2), one can modify the previous input to form the i
input. The principle then used is similar to tha
negative feedback: compensate the input for the
lation of the constraints by the output. Therefore
order to drive the output to be nonnegative, a log
choice for the next input would be
gk+l(x) = gk (x), x6Y
lgk(x) -fgk'(x), xey'Y
was found to work better than using any one method for
all iterations.
When the iterative approach was used on one-di-
mensional objects, it found multiple solutions for the
same Fourier modulus data, depending on the initial
input used to start the iterations. Figure 2 shows an
example of an object and, superimposed, 15 different
solutions (each agreeing with the Fourier modulus to
within EF
2
< 10-3). In this case there is a uniqueness
problem, but at least the various solutions correlate
fairly well with the original object. For some other
one-dimensional objects there is little or no similarity
so- among the various solutions; for still others the solution
is unique.
The results obtained with complicated two-dimen-
sional objects were dramatic. Figure 3(a) shows a
two-dimensional object used for experiments that re-
(4) sembles a sun (of diameter 52 pixels in a field of 128 X
128 pixels-only the central 80 X 80 are shown) with
solar flares and sunspots. The modulus of its Fourier
transform is shown in Figure 3(b). For one test of the
nls, iterative method, the initial input used was the square
nmy of random numbers shown in Figure 3(c). Figures 3(d)
tmhe through 3(f) show the reconstruction results after 20,
the 230, and 600 iterations, respectively, having rms error
an- EO = 0.117,0.042, and 0.0055, respectively. In a second
ran- test, the initial input used was the circle of random
rarit- numbers shown in Figure 3(g). Figures 3(h) and 3(i)
it show the reconstruction results of that test after 2 it-
of erations (E
0
= 0.111) and 215 iterations (E
0
= 0.019),
de-
9,10
ror-
era-
me.
(b),
put
iys-
rier
ise-
the
the
Eq.
new
t of
Vio-
e, in
ical
(a) (b) (c)
(f)
(d) (e)
(5)
where ,B is a constant.
There are a number of methods of choosing gk+ 1(x),
two examples of which are given in Eqs. (2) and (5), re-
spectively, based on different points of view and dif-
ferent trade-offs inherent in the input-output approach.
Several different methods were found to succeed. A
particularly successful method of choosing gk+1(x) is
given by the first line of Eq. (2) combined with the
second line of Eq. (5). In most cases, the method of
choosing gk+1(x) was changed after each few iterations.
This periodic changing of methods after a few iterations
(g) (hJ (i)
Fig. 3. (a) Test object; (b) modulus of its Fourier transform;
(c) initial estimate of the object (first test); (d)-(f) recon-
struction results-number of iterations: (d) 20, (e) 230, (f)
600; (g) initial estimate of the object (second test); (h)-(i) re-
construction results-number of iterations: (h) 2, (i) 215.
July 1978 / Vol. 3, No. 1 / OPTICS LETTERS 29
(a) (b)
Fig. 4. (a) Reconstruction result for 2.2% rms noise; (b) re-
construction result for 11% rms noise.
respectively. As a matter of economy, the iterations
were halted before the error was driven to zero.
Comparing Figs. 3(f) and 3(i) with Fig. 3(a), we see
that the reconstruction results differ very little from the
original object. Note that inverted solutions are al-
lowed, since I51f(-x)UI = JJYlf(x))I. These results are
significant not only because they demonstrate a method
of finding solutions, but also because they suggest that
the uniqueness of the result is not a serious problem for
complicated two-dimensional objects.
The iterative approach has the flexibility of allowing
various types of constraints or a priori knowledge (such
as low-spatial-frequency phase information) to be en-
forced on the solution, which is helpful in causing the
iterations to converge more rapidly and in reducing the
possible ambiguities of the solution. Starting with a
good initial input also helps-compare Figure 3(e) with
Figures 3(h) and 3(i).
To test the sensitivity of this approach to noise, uni-
formly distributed random noise was added to F(u), and
the modulus of the resulting noisy Fourier data was used
to reconstruct the object. In one test, the ratio of the
rms noise to the Fourier modulus was 18% at the highest
spatial frequencies [where, as seen from Fig. 3(b), the
signal is weakest] and 2.2% overall. The initial input
used was the same as in Fig. 3(g). Figure 4(a) shows the
reconstruction result after 210 iterations (E
0
= 0.020).
This amount of noise degraded the reconstruction only
slightly. In another test, the relative rms noise was 91%
at the highest spatial frequencies and 11% overall.
Figure 4(b) shows the result after 300 iterations (E
0
=
0.059). In this case the fine details of the original object
were lost; nevertheless, the gross features of the object
(the low-spatial-frequency information) were still re-
constructed.
We have demonstrated an iterative approach to re-
constructing a general object from the modulus of its
Fourier transform. This data-processing method
represents a solution to the phase problem of optical
coherence theory and would allow the imaging of space
objects through the turbulent atmosphere using inter-
ferometer data. It is relatively fast, making it practical
for use on large two-dimensional images. It is not
highly sensitive to noise. Experimental results suggest
that the uniqueness problem is severe for one-dimen-
sional objects but may not be severe for complicated
two-dimensional objects. The error-reduction ap-
proach, a modification of the Gerchberg-Saxton algo-
rithm, was only partially successful; the more powerful
input-output approach achieved much faster conver-
gence. These iterative approaches should be valuable
for a number of other problems in optics in which only
partial information is known in each of two domains.
This work was begun under the direction of J. W.
Goodman at Stanford University, where the one-di-
mensional results were obtained
10
; the two-dimensional
results were obtained at ERIM under an internal re-
search and development program. Portions of this
paper were delivered at the October 1977 Annual
Meeting of the Optical Society of America in Toronto,
Canada.
References
1. R. H. T. Bates, IEEE Trans. Comput. C-24, 449 (1975).
2. A. Labeyrie, Astron. Astrophys. 6, 85 (1970); D. Y. Gezari,
A. Labeyrie, and R. V. Stachnik, Astrophys. Lett. 173, Li
(1972).
3. D. G. Currie, S. L. Knapp, and K. M. Liewer, Astrophys.
J. 187, 131 (1974).
4. For a list of references, see Ref. 4 in D. Kohler and L.
Mandel, J. Opt. Soc. Am. 63, 134 (1973).
5. B. R. Frieden and D. G. Currie, J. Opt. Soc. Am. 66, 1111A
(1976).
6. R. W. Gerchberg and W. 0. Saxton, Optik 35, 237
(1972).
7. N. C. Gallagher and B. Liu, Appl. Opt. 12,2328 (1973); R.
W. Gerchberg, Opt. Acta 21, 709 (1974).
8. B. Liu and N. C. Gallagher, Appl. Opt. 13, 2470 (1974).
9. J. R. Fienup, J. Opt. Soc. Am. 64, 1395A (1974).
10. J. R. Fienup, Ph.D. thesis, Stanford University, May
1975 (University Microfilms No. 75-25523), Chap. 5.