Commun Nonlinear Sci Numer Simulat 15 (2010) 675689

Contents lists available at ScienceDirect

Commun Nonlinear Sci Numer Simulat

journal homepage: www.elsevier.com/locate/cnsns

Optimum signal and image recovery by the method of alternating

projections in fractional Fourier domains
Ahmet Serbes, Lutye Durak *
Yildiz Technical University, Department of Electronics and Communications Engineering, Yildiz, Besiktas, 34349 Istanbul, Turkey

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents a signal and image recovery scheme by the method of alternating pro-
Received 17 April 2009 jections onto convex sets in optimum fractional Fourier domains. It is shown that the frac-
Received in revised form 5 May 2009 tional Fourier domain order with minimum bandwidth is the optimum fractional Fourier
Accepted 5 May 2009
domain for the method employing alternating projections in signal recovery problems. Fol-
Available online 12 May 2009
lowing the estimation of optimum fractional Fourier transform orders, incomplete signal is
projected onto different convex sets consecutively to restore the missing part. Using a pri-
PACS:
ori information in optimum fractional Fourier domains, superior results are obtained com-
02.40.Ft
02.30.Nw
pared to the conventional Fourier domain restoration. The algorithm is tested on 1-D linear
43.60.Wy frequency modulated signals, real biological data and 2-D signals presenting chirp-type
87.63.lm characteristics. Better results are obtained in the matched fractional Fourier domain, com-
87.57.nt pared to not only the conventional Fourier domain restoration, but also other fractional
87.57.nf Fourier domains.
2009 Elsevier B.V. All rights reserved.
Keywords:
Method of alternating projections
Fractional Fourier transform
Signal recovery
Image recovery
Fractional Fourier transform order
estimation

1. Introduction

In many signal processing and optical applications, the need for signal recovery arises when some part of a signal is miss-
ing or incomplete. An example of these applications is the obscuration of EEG signals with functional magnetic resonance
imaging (fMRI) in the presence of burst noise [1]. In case of 2-D signals, estimating the background of an image behind
an obstacle is a problem of signal recovery in image processing [2,3]. Recovery of lost blocks in block-based image and video
coded signals [4,5] is another recent problem, since the encoded bit stream might be damaged during transmission on unre-
liable networks, especially in real-time video transmission. Packet loss in some packet switching networks including asyn-
chronous transfer mode (ATM) is common in most networks [6] that cause some part of the image or video to be lost because
of different reasons such as congestion and buffer overow. Alternating projections onto convex sets have been studied in
the elds of interpolation, extrapolation, and restoration in the conventional Fourier domain [7,8]. Moreover, image recovery
[25], signal synthesis [9], and other areas [10,11] including medical imaging and tomography [12] have also been
investigated.

* Corresponding author. Tel.: +90 212 383 24 98; fax: +90 212 383 24 86.
E-mail addresses: aserbes@yildiz.edu.tr (A. Serbes), lutye@ieee.org (L. Durak).

1007-5704/$ - see front matter 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.cnsns.2009.05.013
676 A. Serbes, L. Durak / Commun Nonlinear Sci Numer Simulat 15 (2010) 675689

In this work, we develop an algorithm for signal and image recovery problems of one and 2-D signals employing the
method of alternating projections (MAP) in optimum fractional Fourier domains. Conveying signal representations to all ob-
lique domains between time and frequency on the timefrequency plane, fractional Fourier transform (FrFT) provides a com-
prehensive analysis especially for linear frequency modulated (LFM) signals. LFM signals are very often employed in
applications such as radar, sonar, and seismic signal processing. Nature itself uses frequency modulated signals for some
of its radar processing applications in bats, dolphins, whales, etc. We apply the method of alternating projections (MAP)
in the presence of mono and multi-component LFM signals in optimum fractional Fourier domains. The optimum fractional
Fourier domain is estimated by novel methods given in this paper using the celebrated genetic algorithms (GA). We have
tested the proposed system by a synthetic chirp signal, a real bat echolocation data, and synthetic images and showed that
the proposed system is an optimum solution among the MAP-based reconstruction schemes.
The MAP reconstruction algorithm is a special case of the iterative alternating projections onto convex sets (POCS) meth-
od. The POCS algorithm computes the intersection points of convex sets by projecting onto these sets alternately. Let M con-
vex sets be dened by C k ; k 1; 2; . . . ; M; and they possess the prior information on the characteristics of the signal. The
prior information has to be a type of constraint, i.e., non-negativity constraint, band-limitidness constraint, constant area
constraint, bounded energy constraint or any combination of these. The solution is assured to be an element of the intersec-
T
tion of the convex sets M m1 C m each dened by a constraint [10]. Dening P m as the projection operator onto the corre-
sponding convex set C m , the projection of a signal x onto C m is the single step of the iteration and can be expressed as
Pm x. The iterations are carried out among other sets where xn1 PM ; . . . ; P 2 P 1 xn is the nth iteration executed till the con-
vergence is obtained. There are faster methods improving the convergence rate in [10,13] and a survey that gives an over-
view to convex optimization methods in [14].
FrFT has found many applications in the last decade mostly in the eld of signal processing, optics, and image processing
[1517]. The ath-order FrFT operator is the ath power of the conventional Fourier transform (FT) operator. The zeroth order
FrFT of a signal is the signal itself and the rst order transform of a signal is the ordinary FT of the signal. FrFT can also be
thought as an interpolation between the identity operator and the FT operator. One of the most signicant properties of the
FrFT is the rotation property. Taking the ath-order FrFT of a signal rotates the signal in the timefrequency plane by an angle
of a ap=2 [17] in the clockwise direction. The rotation in the timefrequency domain may lead to a key result for the stated
problem that the bandwidth of the signal may decrease signicantly in an FrFT domain. For example, an LFM signal can be
transformed into an impulse signal in the matched FrFT order, which is shown to be the optimum transform order for our
problem.
FrFT domain applications of POCS-based algorithms have been studied for signal enhancement [18], extrapolation [19],
interpolation [20] and restoration [21]. Cetin et. al. [18] studied the enhancement of resolution of prototypical wave-elds
using the information of M different FrFT-domain information of the signal. Sharma et. al. [19] have shown that the perfor-
mance of MAP-based systems for signal extrapolation depends on the FrFT order. However, it was not mentioned how to
estimate the optimal FrFT order where the performance is maximum. Guven et al. [21] discuss the restoration of signals
using M different FrFT-domain information for prototypical signals and also mentioning several constraints.
In this work, missing parts of different signals are recovered in the optimum FrFT order by projecting the signals onto
different closed convex sets. We show that the optimum FrFT order has to be estimated in advance for the best performance.
We have developed GA-based schemes to nd the optimum FrFT order. For this purpose, a time-bandwidth ratio (TBR) and
essential-bandwidth related measures are suggested as the tness functions of the GA scheme. Satisfactory results are ob-
tained for both the FrFT order estimation and for the signal recovery problem. Simulation results conrm that maximum
performance is obtained at the optimum FrFT order for MAP-based signal recovery problem.
The organization of the paper is as follows. The denition of the FrFT, some of its properties, and optimum FrFT order
estimates are given in Section 2. Section 3 introduces to the MAP in both conventional Fourier domain and FrFT domains.
Evolutionary algorithms for the optimization of FrFT order estimation are presented in Section 4. In Section 5 simulation re-
sults for different scenarios are given. The paper concludes in Section 6.

2. The fractional Fourier transform order estimation

The ath-order FrFT of a continuous signal f u in L2 R is represented by fa u and can be expressed for 0 < jaj < 4,
p Z 1    0
fa u 1  | cot a f u0 exp |p cot au2  2 csc auu0 cot au02 du 1
1

where a ap=2 [17].The transformation operator is dened as Fa . When the FrFT order a 2, the signal becomes
f2 u f u which is equivalent to taking the FT of the signal twice and when a 4; f4 u f0 u f u is the signal itself.
The kernel of the FrFT for a 0 or a 4 is du  u0 where d: represents the Dirac-delta function. For the FrFT order a 1,
the kernel turns out to be the ordinary FT kernel, and when a 3 the kernel is the inverse FT kernel. Two consecutive FrFTs
with orders, a and b is equivalent to a single FrFT with an order a b, whereas the inverse of the ath-order FrFT is the FrFT
with order a. Computational complexity of FrFT of a signal with length N is of order Nlog2 N [22].
The Wigner distribution (WD) of the FrFT of a signal at order a 2a=p is the rotated form of the WD of the original signal
with an angle a in the clock-wise direction. This property can be thought as rotating the timefrequency axis counter clock-
 A. Serbes, L. Durak / Commun Nonlinear Sci Numer Simulat 15 (2010) 675689 677

wise. As in Fig. 1, the FrFT of a signal rotates the timefrequency axis of the signal clockwise and the new timefrequency
coordinate system makes an angle of a ap=2 with the former coordinate system. Consequently, the magnitude-squared
FrFT of a signal is the integral projection of a signals WD onto the rotated axis by the angle a.
Consider an LFM signal xt A exp|bt 2 2pft with a chirp rate of b. At the FrFT-order of b 1, the LFM signal is
transformed to a Dirac-delta distributed signal and the inverse FT of the transformed signal, which corresponds to the
bth-order FrFT is a pure sinusoidal. The order of transformation b 1 is the optimum transformation order for the present
problem. The signal is in its most compact form at this optimum transform domain and it is the most suitable form for MAP,
as the MAP takes advantage of the uncertainty principle. If any signal xu is time-limited, it can not be band-limited and vice
versa: If the signal is band-limited, it can not be time-limited. If the frequency of a signal varies with time, there exists at
least one optimum FrFT order at which the bandwidth is minimum. To nd the optimum transform degree, two different
suggestions are proposed. The algorithm to nd the minimum bandwidth is chosen to be GA, which is a global heuristic
search technique providing robust FrFT-order estimation. Genetic algorithms is introduced and simulation results are given
in Section 4.

2.1. FrFT order estimation by using maximum fractional time-bandwidth ratio

The timefrequency support of a signal xt is proportional to its time-width T x and its frequency-domain bandwidth Bx ,
where T x and Bx are dened as

hR i2
t  gt 2 jxtj2 dt
Tx 2
kxk
 2
2
R 2
f  gf jXf j df
Bx 3
kxk

where Xf is the ordinary Fourier transform of xt; k  k is the norm operator, and the mean values gt and gf are
R
tjxtj2 dt
gt ; 4
kxk
R
f jXf j2 df
gf : 5
kxk
The aim is to nd the optimum FrFT order which minimizes the bandwidth in the transform domain. Fig. 2a presents a
signal with timefrequency support of T x  Bx . As shown in Fig. 2b, the optimum FrFT order minimizes the bandwidth of
the transformed signal, hence maximizes the time-support of the signal as dictated by the uncertainty principle. The or-
der a 1 that minimizes the fractional width is the optimum transformation order. Here, we introduce a cost function
to nd the optimum FrFT order which maximizes the transform domain bandwidth: Fractional time-bandwidth ratio
(FrTBR)
T x;a
FrTBRa 6
Bx;a

Fig. 1. Applying FrFT to a signal rotates the timefrequency axis counter clock-wise by an angle of a for that signal.
678 A. Serbes, L. Durak / Commun Nonlinear Sci Numer Simulat 15 (2010) 675689

Fig. 2. A signal has a time frequency support as shown in (a). The rotation property of the FrFT alters the timefrequency support. The optimum order of
FrFT minimizes the transformed signals bandwidth, while maximizing its time-width.

where T x;a and Bx;a are the widths at the fractional Fourier domains of order a and a 1, respectively, which can be explicitly
expressed as
 2
2
R
u  gu;a jxa uj2 du
T x;a ; 7
kxk
 2
2
R
u  gu;a1 jxa1 uj2 du
Bx;a ; 8
kxk
where gu;a and gu;a1 are the mean values at the fractional Fourier domains of orders a and a 1. The order that maximizes
TBR is the optimum transform order, which minimizes the width at fractional Fourier domain of order a 1 while maximiz-
ing the width in the ath-order fractional Fourier domain. Since Bx;a equals to T x;a1 , FrTBR can be expressed as
 2
2
R
u  gu;a jxa tj2 du
FrTBRa  2
2 : 9
R
u  gu;a1 jxa1 tj2 du

When discrete-time signals are of interest, discrete fractional time-bandwidth ratio comes as a substitute for the FrTBR. The
derivation of discrete fractional time-bandwidth ratio (DFrTBR) is straightforward
 2
2
P
k  gk;a k jxa kj2
DFrTBRa
2 ; 10
P 2
k  gu;a1 k jxa1 kj2

where gk;a k and gk;a1 k are

P
kjxa kj2
gk;a ; 11
kxk
P
kjxa1 kj2
gk;a1 : 12
kxk
FrTBR is applicable if signals of interest have a single component in the observation frame.

2.2. FrFT order estimation by using minimum essential-bandwidth

As our observations are limited, all practical signals have to be limited in time. Uncertainty principle imposes signals to be
either limited in time or limited in frequency. However, some time-limited signals carry almost all their power in certain
frequencies, which is called as the essential-bandwidth. The essential-bandwidth differs from the well-known conventional
bandwidth denition, i.e., essential bandwidth of sum of two signals with bandwidths X1 and X2 at different center frequen-
cies is at most X1 X2 . On the other hand, according to the conventional bandwidth denitions this summation may be
much bigger than the essential-bandwidth. This is illustrated at Fig. 3.
Let us consider a signal that is composed of multi-chirp components, such as a bats or a bottlenose dolphins echolocation
signal. These signals have chirps with almost same chirp-rates, but different time and frequency shifts [23]. Signals of this
 A. Serbes, L. Durak / Commun Nonlinear Sci Numer Simulat 15 (2010) 675689 679

Fig. 3. The essential-bandwidth is X1 X2 . However, the bandwidth may be far greater than the essential-bandwidth.

type have similar optimum transform-domain magnitude characteristics as the signal shown in Fig. 3. Minimizing the FrTBR
may not work well for this type of signals, since rotation in timefrequency axis will not minimize the width at the optimum
transformation order. Actually, for these types of signals, rotation in time frequency axis will minimize only the essential-
bandwidth at the optimum transformation order. In this section we propose a new method to nd the optimum transform
order that minimizes the essential-bandwidth. Let xu be a signal and xa u be its ath-order FrFT. The optimum transform
order minimizes the function
Z  
Ia g jxa uj2 du; 13

where gu is a hard-limiting function


1 u P uthr
gu ; 14
0 u 6 uthr
with a threshold uthr 2 R . It is convenient to choose uthr proportional to the power of the signal or amplitude at the trans-
form domain. We choose uthr proportional to the power of the signal as,
kxk
uthr ; m  1; 15
m
where m is chosen as 70 for all simulations. The choice of m effects the convergence rates signicantly, but we do not overly
concern ourselves about choosing the optimum m values. The essential-bandwidth of signals in the transform domain is ex-
pressed as
X 2 fujg jxa uj > uthr g: 16
where X is the essential-bandwidth, which is minimal at the optimum FrFT order.

3. Method of alternating projections in signal recovery

MAP is a special application of the well-known alternating projections onto convex sets (POCS) algorithm, which is used
for convex optimization problems. Papoulis [7] and Gercberg [24] were the rst to use alternating projections for restoration
of signals unintentionally. Youla [25] recognized that PapoulisGerchbergs method is a special application of MAP. The MAP
is used for interpolating or extrapolating band-limited signals in the conventional Fourier transform domain [2,3], which
takes advantage of the well-known fact that time-limited signals cannot be band-limited and vice-versa. As the observations
are limited in time, real-world signals can be considered as time-limited and the bandwidth cannot be limited. However,
these inconsistent constraints meet in some minimum-norm solution.
MAP has been applied to signal or image recovery problems [3], when some portion of the data is incomplete, missing or
misreceived. Suppose that a portion of a band-limited signal xu is missing between u2 < u < u1 . The MAP-based restoration
algorithm in the conventional Fourier transform domain is as follows.
680 A. Serbes, L. Durak / Commun Nonlinear Sci Numer Simulat 15 (2010) 675689

1. Using prior information about the bandwidth X of the signal, low-pass lter the signal xu with cut-off frequency X.
2. Replace the existing correct portion of the signal.
3. Return to step 1 and continue till the convergence is obtained.

There are two convex sets in this case,

C 1 fxujXx 0; jxj R Xg 17
 
 P1 xu x2 < u < x1
C 2 xu 18
xu otherwise:
where P 1 denotes the projection of the signal onto the convex set C 1 . Additional convex sets, i.e., bounded-energy or non-
negativity constraints, may be added in accordance with the problem.
When some part of a signal f u is missing, the signal can be considered as a multiplication of xu ruf u where f u is
the original signal that we are trying to recover and ru is a rectangular function, which is 0 in an interval u1 ; u2  and 1 else-
where. Even if f u is essentially-bandlimited, xu is obviously not bandlimited, since bandlimited functions have no discon-
tinuities. There are two constraints; the bandlimiting constraint, which is equal to zero-discontinuity constraint, and the
constraint that the known part of the signal remains unchanged. These set of constraints meet in a point of minimum-norm
solution [9].
To summarize, in the conventional Fourier transform domain case, when some portion of a signal is non-existent, and a
pre-knowledge about the signals bandwidth is available, the signal is rst ltered using the apriori information about the
bandwidth. When the signal is ltered, the missing portion starts to take some values, which are relevant to the original sig-
nals values. Afterwards, the known portion of the signal is replaced, which in turn causes discontinuities at the meeting
points of the known portion and the unknown portion. If the bandwidth of original signal is minimal, e.g., when the signal
is a sine wave, the discontinuity will be minimal even after the rst iteration, since discontinuities require large bandwidths.
However, the Fourier transform domain MAP may not be an optimum solution for non-stationary signals, such as mono- or
multi-component LFM signals sweeping large frequency band.
In this work, we restore missing parts of a chirp signal, a real bat echolocation data and 2-D images. We extend the idea of
FrFT domain MAP into 1-D and 2-D signal recovery problem, with estimating optimum-FrFT order selection process using
GA. The algorithm is developed by introducing the optimum FrFT-order selection process

1. Estimate the optimum FrFT order a and nd the minimum FrFT domain bandwidth X,
2. Filter the signal xa u using transform-domain bandwidth information X,
3. Replace the existing correct portion of the signal,
4. Return to step 2 and continue till the convergence is achieved.

The FrFT domain bandwidth is estimated in advance using (16). In Step 2, FrFT domain ltering is applied at the matched
FrFT order which minimizes the bandwidth. The matched FrFT order can also be determined for mono-component LFM-type
signals by alternative methods, such as in [26,27].
There are two convex sets, respectively,
C 1 fxujxa x 0; x R Xg 19
 
 P1 xu u2 < u < u1
C 2 xu 20
xu otherwise:
where P1 denotes the projection of the signal onto C 1 . Even though the bandlimited signals cannot be limited in time, inter-
section of the two convex sets C 1 2 C 2 is an empty set, but the inconsistent constraints lead the convex sets to a minimum
error solution in the mean square sense.
The n-th step of the iteration can be expressed as xn1 u P 2 P1 xn u, for positive integer values of n. The error is assured
to be minimum after limn!1 xn u achieves a limit cycle [10,11]. The MAP procedure can be extended to 2-D directly by
employing 2-D FrFT as in [28].
Band-pass ltering in the FrFT domain [29] for the rst convex set can be expressed as
   p Z X0
u  X0
Fa xa urect 1 | cot a exp|pcot au2  2 csc auu0 cot au02 xa udu0 21
2X X

where rect: is the rectangle function. In the following section GA is briey introduced and used to nd the optimum FrFT
order by using the FrTBR or essential-bandwidth as cost functions.

4. Optimization in FrFT order estimation

Objective functions may have different characteristics, i.e., they may neither be smooth, nor be non-linear functions.
Other than that, the objective function may be hard to differentiate or it may be impossible to differentiate with known-tech-
 A. Serbes, L. Durak / Commun Nonlinear Sci Numer Simulat 15 (2010) 675689 681

niques in order to employ gradient-based optimization algorithms. GA employs a global stochastic optimization search
method to nd the global optima with linear or non-linear, differentiable or non-differentiable, constrained or uncon-
strained, equality or inequality types of objective functions [30].
GA is a powerful heuristic optimization tool that mimics the natural selection and adaptation process [31,32], which
has been applied to many areas including face detection [33], electromagnetics [34], classication of cancers [35], and
even molecular biology [36]. The standard procedure of GA method of stochastic optimization is to maximize a function
Fx for x 2 fH  Rn g, where H is the parameter space. The populations are randomly initialized by n-dimensional binary
strings with N chromosomes. After computing the tness for each individual (chromosome) in the current population, the
selection operator populates the t chromosomes only as stated by the natural selection. A single, multi-point, or a scat-
tered crossover procedure may be followed to constitute the children as shown in Fig. 4. The multi-point crossover gen-
erates random crossover points and exchanges the corresponding bits between parents to produce the children. The
scattering crossover uses a random binary vector as a mask to select the genes from the parents, i.e., if the rst bit of
the mask is a 1, the gene is selected from the rst parent and if the rst gene is a 0 the gene is selected from the second,
or vice versa. The mutation operator is employed just after crossover operation with a pre-determined mutation probabil-
ity to ensure that the search operation will continue even if a local optimum is converged. This operation is simply a ip of
a bit in the bit-string of the chromosome. The elitism inserts highly t chromosomes into the next-generation gene pool
[37,38]. The iterations are repeated until the maximum number of generations is reached or the alternation in the tness
function is below some pre-determined value.
The steps of the employed algorithm is as follows.

1. Randomly generate an initial population X0 x1 ; x2 ; x3 ; . . . ; xN of coded binary strings of length N for each xi ,
2. Compute the tness Fxi for each chromosome in the current population,
3. Employing the selection, populate the t chromosomes only,
4. Generate new chromosomes by mating the current parents,
5. Apply mutation to the children with pre-assigned mutation probability,
6. As stated by the elitist strategy, select a portion of elitist parent chromosomes to take its place in the next generation,
7. Return to the step 2 until the goal is met.

We have used a double vector (64 genes) for a population size of 20 chromosomes with a rank-based [38,39] tness scal-
ing function. The stochastic uniform sampling scheme [30] has been chosen for selection operation. A deeper insight into the
selection functions can be found in [40]. The crossover is made by using a scattered multi-point crossover function. The
mutation is performed with Gaussian mutation function, such that a random number is generated by a normally distributed
function with a shrinking variance. The shrink parameter decreases the variance after each iterate. Two most successful indi-
viduals in the current generation are guaranteed to survive to the subsequent generation. The tolerance of the tness func-
tion is chosen as 106 . The search space a 2 0; 2 and the parameter space H 2 f0; 2; 0; 2; ; . . . ; 0; 2g are used for all
problems.

Fig. 4. Examples of (a) three-point crossover, in which horizontal lines represent the crossover points (b) scattered crossover.
682 A. Serbes, L. Durak / Commun Nonlinear Sci Numer Simulat 15 (2010) 675689

 Optimizing FrFT order by maximizing FrTBR

Our proposed algorithm of nding the optimum FrFT order which maximizes FrTBR using the GA is constructed as fol-
lows. The model of maximum FrTBR is observed in (9) for the optimum FrFT order. According to the proposed model, the
FrFT-order a minimizing the FrTBR is;
 2
2
R
u  gu;a jxa uj2 du
^ arg max
a
2 ; 22
a R 2
u  gu;a1 jxa1 uj2 du

where a^ is the optimal FrFT order estimate. GA aims to nd the optimum solution through successive generations, in which
selection, crossover, and mutation operators are carried out in single step of the iteration.

 Optimizing FrFT order by minimizing the essential-bandwidth

We propose another method to nd the optimum FrFT order by minimizing the essential-bandwidth. We minimize (13)
with respect to a
Z  
^ arg min
a g jxa uj2 du: 23
a

We have estimated the essential-bandwidth of signals as

b 2 fujg jxa^ uj > uthr g;
X 24
where Xb is the essential-bandwidth estimate of the signal and uthr is dened by (15). This way, the optimum FrFT order and
the essential-bandwidth of the signal are estimated together.

 Performance analysis of the optimum FrFT order selection

We present simulation studies of the proposed methods for nding the optimum FrFT order for various scenarios and dis-
cuss the cons and pros. For the rst scenario, a chirp signal is generated,
h i
x1 u Gu exp | b0 u  u0 2 2pbu  u0 25

with a chirp rate of b0 1=3, time shift u0 0, and frequency shift b 0, where Gu is the Gaussian envelope of the signal.
The proposed GA models are operated for this chirp signal.
For a second scenario, two chirp signals with equal amplitudes, time and frequency shifts, but different chirp rates, such
as b 1=3 and 4/9 are added together to form a multi-component chirp signal. The expected optimum FrFT order for the
summation of the two signals is 1 a1 a2 =2, where a1 and a2 are the corresponding chirp rates,
   
1 4
x2 u Gu exp | u2 Gu exp | u2 : 26
3 9
The third scenario is a multi-component chirp signal with components of equal amplitudes and chirp rates, but different
time and frequency shifts. In this case, two different signals are constituted where each one is a summation of two different
chirp components with the same chirp rate of 1/3. The rst signal x3 u has different time-shifts for each chirp component, as
0 and 3. The second signal x4 u has frequency shifts of 0 and 0.5 for each of its components. The optimum FrFT order is re-
lated to the chirp rate, so the expected optimum FrFT order estimate is 1 1=3 for both of these signals,
  h i
1
x3 u Gu exp | u2 Gu exp | 1=3u  32 ; 27
3

 

1 1
x4 u Gu exp | u  32 Gu exp | u2 pu : 28
3 3
The proposed algorithms are used to test these scenarios. The results are summarized in Table 1. The maximum FrTBR or the
minimum essential-bandwidth methods are both suitable for computing the optimum FrFT order for signals with the same
time and frequency shifts. For signals with different time and/or frequency shifts, e.g., a bat, whale or a dolphins echoloca-
tion signal, apparently the minimum essential bandwidth provides the optimum FrFT order.

5. Simulations and results

The experiments are held for MAP in FrFT domains for four different scenarios. The problem is a special application of
linear degradation: recovery of missing parts of non-stationary signals. In the rst scenario, a chirp signal with some missing
 A. Serbes, L. Durak / Commun Nonlinear Sci Numer Simulat 15 (2010) 675689 683

Table 1
Performance comparison of proposed algorithms for the scenarios.

Maximum TBR Minimum EB Exp*

Avg. Std % Err Avg. Std. % Err. a
x1 u 1.335 7.3E-4 0.15 1.333 3.7E-3 0.00 1.333
x2 u 1.396 8.2E-4 0.50 1.392 9.5E-4 0.22 1.389
x3 u 1.398 4.3E-4 4.88 1.333 3.8E-3 0.00 1.333
x4 u 1.448 5.9E-4 8.62 1.332 3.4E-3 0.07 1.333
*
Expected Optimum FrFT Order.

parts is considered. After estimating the optimum transform domain, the signal is optimally reconstructed using the opti-
mum fractional Fourier domain reconstruction scheme.
The second scenario is to recover a real bat echolocation signal, which is composed of multi-chirp components. We have
cropped the bat signal at an arbitrary length and recovered the missing cropped part. Optimum transformation order is esti-
mated just before the reconstruction procedure.
The third scenario is a recovery of a directional 2-D chirp image, in which some random part is cut off. We have employed
a 2-D Hilbert transform before nding the optimum transformation order, which is explained in detail.
The fourth scenario is similar to the third scenario, except that the 2-D image is now non-directional. The Hilbert trans-
form is applied to the image, optimum fractional Fourier transformation order is estimated and the reconstruction procedure
is followed.
The performance of the proposed algorithm is tested by mean square error (MSE) sense and compared with the conven-
tional Fourier domain reconstruction scheme. The mean square error in 1-D is dened as
R
jxu  ^xuj2 du
MSE1D ; u 2 u1 ; u2 ; 29
jjxujj

where u1 and u2 are the boundaries of the missing part of the signal, xu and ^ xu are the desired and obtained signals,
respectively. Extension of the MSE formulation to a 2-D signal xu; v can be easily done by replacing the boundaries with
u 2 u1 ; u2  and v 2 v 1 ; v 2 , where u1 ; u2 are the boundaries in the u-direction and v 1 ; v 2 are the boundaries in the v-
direction,
R
jxu; v  ^xu; v j2 dudv
MSE2D ; u 2 u1 ; u2 ; v 2 v 1 ; v 2 ; 30
jjxu; v jj

5.1. Recovery of a chirp signal

The rst scenario of simulations is held by a synthetic Gaussian amplitude-modulated analytical chirp signal given by
(25). The signal is composed of 2048 samples and the parameters are b 1=3; b 0 and u0 0. We crop out approximately
10% of the signal, which can be seen in the Fig. 5a. The MSE decreases as the iteration counts increases as shown in Fig. 5b
and reaches a limit cycle after 200 iterations at a negligible low error rate in the optimum FrFT order. Fig. 5b also shows that
the conventional FT domain method seems to be unsuccessful for the LFM signal recovery problem even after 400 iterations.
The reason for that the proposed method outperforms the conventional Fourier domain reconstruction is that LFM signals
are transformed into a dirac-delta distributed signal in its optimum transform domain, which is the minimum transform-
domain-width. However in the conventional Fourier domain, LFM signals have large bandwidth and lower performance.
The error and the desired signal are shown in Fig. 5c, after 150 iterations. The obtained signal and the original signal may
seem to mostly overlapping, since the error is small. Compared to the other FrFT domains including the conventional Fourier
domain, the error rate is slightly smaller for the optimum FrFT order, which is shown in Fig. 5d. The simulations show that
the matched order FrFT domain restoration is superior to other FrFT domains including the classical Fourier domain by
means of the MSE and the convergence rate. In this scenario it is shown that one optimum way of restoring LFM signals
is the matched-order FrFT domain restoration.

5.2. Recovery of the bat echolocation signal

In this section, we recover a digitized 2.8 ms echolocation of a long large brown bat (Eptesicus Fuscus) (available at:
http://www.dsp.rice.edu/software/TFA/RGK/BAT/batsig.bin.Z), in which 6.5% of it is cropped out in advance. The echoloca-
tion signal, which is shown in Fig. 6a is composed of multi-chirp components.
The optimum FrFT angle for the bat echolocation signal is the angle in which the essential-bandwidth of the signal com-
ponents is minimal. Fig. 6b shows the WD of the analytic bat echolocation signal after Hilbert transform and the optimum
transformation angle which is perpendicular to the angle of the chirps with respect to the time axis.
684 A. Serbes, L. Durak / Commun Nonlinear Sci Numer Simulat 15 (2010) 675689

Fig. 5. (a) A Gaussian amplitude-modulated LFM signal with cropped part of length 200 and a total length of 2048 samples, (b) MSE with respect to the
iteration number for the cropped LFM signal in the optimum FrFT order vs. the conventional Fourier domain restoration. (c) The recovered signal, the
original signal and the error signal after 150 iterations in the optimum FrFT order and (d) MSE vs. FrFT order after 200 iterations. The error rate is much
smaller at the optimum FrFT order.

The optimum FrFT order is estimated using the proposed minimum essential bandwidth method as about 0:391p rad
(70.4 degrees), which corresponds to the FrFT order of about 0.782. When the optimal FrFT is applied to the signal, the axis
is rotated counter-clockwise and minimum essential-bandwidth is obtained.
The randomly cropped echolocation signal and the restored signal after 100 iterations are shown in Fig. 6c and d, respec-
tively. The recovered signal bears all the properties of the original echolocation signal. In particular, we present the original
signal, the matched FrFT domain reconstructed signal and the conventional Fourier domain restored signal together for only
the cropped region after 200 iterations for each domain in Fig. 6e. The original signal and the optimum FrFT domain restored
signal seem to overlap, since the error is minor. Fig. 6f shows the error signals for both the conventional Fourier domain res-
toration and the optimum FrFT domain restoration after 200 iterations. Fig. 6e and f zooms only to the cropped region, since
the error takes place only in the cropped part.
The mean square error decreases gradually to a limit for both the Fourier domain restoration and the optimum FrFT do-
main restoration as shown on Fig. 6g. The convergence rate and the MSE of the optimum FrFT domain restoration is superior
to that of the Fourier domain restoration. In Fig. 6h, the mean square error with respect to the MSE is plotted. As it is clearly
seen, the best results are obtained at the optimum FrFT order.

5.3. Recovery of a directional 2-D chirp image

Some part of a background image may be covered by an obstacle or some part of an image data may be missing due to
burst noise. For this case, the original signal may be recovered using the FrFT-MAP by employing 2-D FrFT.
To achieve an analytical image, a 2-D Hilbert transform has to be employed in advance. There is no known exact solution
for a 2-D Hilbert transform, which transforms images into pure analytic images. However, most of them transform images
into pseudo-analytic images [41]. In order to obtain a pseudo-analytic image, we have used a Fourier transform based
algorithm,
IH xx ; xy Ixx ; xy |Ih xx ; xy 31
where, | is the imaginary unit, Ixx ; xy is the Fourier transform of the original image, IH xx ; xy is the Fourier transform of
the Hilbert-transformed image, and Ih xx ; xy is the additive part, which converts the image to an analytic-image. The in-
 A. Serbes, L. Durak / Commun Nonlinear Sci Numer Simulat 15 (2010) 675689 685

Fig. 6. (a) The bat echolocation signal, (b) its WD and the optimum FrFT degree. (c) The cropped signal and (d) the recovered signal after 100 iterations. (e)
The original signal, the optimum FrFT domain and the FT domain reconstructed signals after 200 iterations, and (f) comparison of the error signals for the
signals obtained in (e) are zoomed in for cropped part only. (g) MSE vs. iteration number for the optimum FrFT domain and conventional FT domain. (h) MSE
vs. FrFT order after 200 iterations.
686 A. Serbes, L. Durak / Commun Nonlinear Sci Numer Simulat 15 (2010) 675689

verse Fourier transforms of the Ih xx ; xy has to be a real image, so that the real part of the image is not changed. Since
IH xx ; xy is the analytic image of the original image, it can be expressed as,

Ixx ; xy ; 0 < xx ; xy < p=2:
IH xx ; xy 32
0; otherwise:
The Ih xx ; xy can easily be found by rearranging the terms in the above equation,
Ih xx ; xy |Ixx ; xy  |IH xx ; xy : 33
Thus, the analytic additive part of the image can be written as,
 
Ih x; y |Re Ixx ; xy  |IH xx ; xy 34

Fig. 7. (a) The original image. (b) The cropped image. (c) Recovered image after the rst iteration at the optimum FrFT estimate. (d) Recovered image after
ve iterations at the optimum FrFT estimate. (e) MSE vs. iteration. (f) MSE vs. FrFT order.
 A. Serbes, L. Durak / Commun Nonlinear Sci Numer Simulat 15 (2010) 675689 687

where Ih x; y; Ix; y, and IH x; y are the additive part of the analytic image, the original image, and the Hilbert transformed
image, respectively. As for all methods, we have also obtained pseudo-analytic images. All the images that are used for sim-
ulations are rst Hilbert transformed in order to estimate the optimum FrFT order in advance.
A 2-D chirp signal can be expressed as
h i
xu; v gu; v exp | au u  u0 2 2bu u  u0 2bv v  v 0 av v  v 0 2 35

where gu; v stands for the 2-D Gaussian envelope, au and av are the chirp rates, u0 and v 0 are time shifts, and bu and bv are
frequency shifts. In this scenario, a 2-D synthetic chirp image is generated. This image is a directional image, having a chirp
signal in only 1-D (Fig. 7a) The image is rst Hilbert transformed and the optimum FrFT order is estimated. The image is of

Fig. 8. (a) The original image. (b) The cropped image. (c) Recovered image after the rst iteration at the optimum FrFT estimate. (d) Recovered image after
ve iterations at the optimum FrFT estimate. (e) MSE vs. iteration. (f) MSE vs. FrFT order.
688 A. Serbes, L. Durak / Commun Nonlinear Sci Numer Simulat 15 (2010) 675689

128  128 and randomly chosen 25  25 part of it is cropped (Fig. 7b). The signal is recovered at the optimum FrFT order and
satisfactory results are obtained even after the rst iteration (Fig. 7c). Fig. 7d shows the recovered signal after ve iterations.
According to Fig. 7e, the convergence rate and the MSE of the optimum FrFT domain restoration is much better that the FT
domain restoration and the error rates are nearly zero after a few iterations. The error rate is about 0.02 for the FrFT domain
restoration, while this rate is about 0.1 for the FT domain restoration.
Fig. 7f shows the error rates after 50 iterations for different FrFT orders including the optimum FrFT order and the ordin-
ary FT, which corresponds to rst order FrFT. The undisguised optimum FrFT order 1.2 can be easily seen from this gure.

5.4. Recovery of a non-directional 2-D chirp image

A non-directional 2-D 128  128 chirp signal is produced as in Fig. 8a and some 25  25 part of it is cropped as shown in
Fig. 8b. The image is Hilbert transformed in advance by the method given above and the signal is recovered at the optimum
FrFT order and satisfactory results are obtained even after the rst iteration (Fig. 8c). Fig. 8d shows the recovered signal after
ve iterations. The error rates remain in a level that the human eye cannot differentiate the original image and the recon-
structed image.
According to Fig. 8e, the MSE of the optimum FrFT domain restoration is superior to that of the FT domain restoration and
the error rates are nearly zero after a few iterations. The error rate is much more larger for the FT domain reconstruction and
it is clear that the proposed method outperforms the classical FT domain restoration.
Similar to the previous scenario Fig. 8f shows the error rates after 50 iterations for different FrFT orders including the opti-
mum FrFT order and the ordinary FT, which corresponds to rst order FrFT. As it can be observed from the gure, the opti-
mum FrFT order a 1:2 is the order, in which the error rate is minimum. The error rate at this order is about zero.

6. Conclusions

We have restored missing parts of different signals in their optimum FrFT order and showed that the proposed recovery
scheme is optimum for time-varying signals and images. Compared to the conventional Fourier domain restoration scheme,
the proposed algorithm offers lower MSE rates and fast convergence rates. In order to achieve the best performance rates, a
novel optimum FrFT estimation scheme based on an evolutionary algorithm is provided.
The proposed algorithm is an optimum solution for missing, lost or misreceived LFM-type signals used in radar, sonar, or
seismic data. This method can be used in optics and image processing as well.
The application of the method for interpolation and extrapolation purposes is straightforward. The method can also be
directly applied to convert non-uniformly sampled signals to uniformly sampled signals.

Acknowledgements

The authors are supported by the Scientic and Technological Research Council of Turkey, TUBITAK under the grant of
Project No. 105E078. The authors thank Curtis Condon, Ken White, and Al Feng of the Beckman Institute of the University
of Illinois for the bat data and for permission to use it in this paper.

References

[1] Salek-Haddadi A, Friston KJ, Lemieux L, Fish DR. Studying spontaneous EEG activity with fMRI. Brain Res Rev 2003;43:11033.
[2] Youla DC, Webb H. Image restoration by the method of convex projections: Part 1 - Theory. IEEE Trans Med Imaging 1982;MI-1(2):8194.
[3] Youla DC, Webb H. Image restoration by the method of convex projections: Part 2 - Applications and numerical results. IEEE Trans Med Imaging
1982;MI-1(2):95101.
[4] Park J, Park DC, Marks II RJ, El-Sharkawi ES. Recovery of Image Blocks Using the Method of Alternating Projections. IEEE Trans Image Process
2005;14(4):46175.
[5] Sun H, Kwok W. Concealment of damaged block transform coded images using projections onto convex sets. IEEE Trans Image Process
1995;4(4):4707.
[6] Verbiest W, Pinnoo L, Voeten B. The impact of ATM concept on video coding. IEEE J Select Areas Commun 1988;7:73951.
[7] Papoulis A. A new algorithm in spectral analysis and band limited extrapolation. IEEE Trans Circ Syst 1975;CAS-22(9):73542.
[8] Jorge P, Ferreira SG. Interpolation and the discrete PapoulisGerchberg algorithm. IEEE Trans Signal Process 2003;42(10):2596606.
[9] Goldburg M, Marks II RJ. Signal synthesis in the presence of inconsistent set of constraints. IEEE Trans Circ Syst 1985;CAS-32(7):64763.
[10] Combettes PL. The foundations of set theoretic estimation. Proc IEEE 1993;81(2):182208.
[11] Marks II R. Alternating projections onto convex sets. In: Jansson PA, editor. Deconvolution of images and spectra. New York: Academic Press; 1984. p.
476501.
[12] Sezan MI, Stark H. Tomographic image reconstruction from incomplete view data by convex projections and direct Fourier method. IEEE Trans Med
Imaging MI-3 1984:918.
[13] Combettes PL. The convex feasibility problem in image recovery. In: Hawkes P, editor. Advances in imaging and electron physics, vol. 95. New
York: Academic Press; 1996. p. 155270.
[14] Bauschke HB, Borwein JM. On projection algorithms for solving convex feasibility problems. SIAM Rev 1996;38(3):367426.
[15] Erden MF, Kutay MA, Ozaktas HM. Repeated ltering in consecutive fractional Fourier domains and its applications to signal restoration. IEEE Trans
Signal Process 1999;47(5):145862.
[16] Mendlovic D, Ozaktas HM. Fractional Fourier transforms and their optical implementation: I. J Opt Soc Am A 1993;10:187581.
[17] Ozaktas HM, Zalevski Z, Kutay MA. The fractional Fourier transform with applications in optics and signal processing. Wiley and Sons; 2001.
[18] Cetin AE, Ozaktas H, Ozaktas HM. Resolution enhancement of low resolution waveelds with POCS algorithm. Electron Lett 2003;39(25):180810.
 A. Serbes, L. Durak / Commun Nonlinear Sci Numer Simulat 15 (2010) 675689 689

[19] Sharma KK, Joshi SD. Extrapolation of signals using the method of alternating projections in fractional Fourier domains. Signal Image Video Process
2008;2(2):17782.
[20] Sharma KK, Joshi SD. Papoulis-like generalized sampling expansions in fractional Fourier domains and their application to superresolution. Opt
Commun 2007;278:529.
[21] Guven HE, Ozaktas HM, Cetin AE, Barshan B. Signal recovery from partial fractional Fourier domain information and its applications. IET Signal Process
2008;4(2):1525.
[22] Ozaktas HM, Arikan O, Kutay MA. Digital computation of the fractional Fourier transform. IEEE Trans Signal Process 1996;44(9):214150.
[23] Thomas AJ, Moss CF, Vater M. Echolocation in bats and dolphins. London: The University of Chicago Press; 2004.
[24] Gerchberg RW. Super-resolution through error energy reduction. Opt Acta 1974;21:70920.
[25] Youla DC. Generalized image restoration by method of alternating orthogonal projections. IEEE Trans Circ Syst 1978;CAS-25(9):694702.
[26] Durak L, Arikan O. Short-time Fourier transform: two fundamental properties and an optimal implementation. IEEE Trans Signal Process 2003;51-
5:123142.
[27] Akay O, Boudreaux-Bartels FG. Fractional convolution and correlation via operator methods and an application to detection of linear FM Signals. IEEE
Trans Signal Process 2001;49-5:97993.
[28] Pei SC, Yeh MH. Two dimensional fractional Fourier transform. Signal Process 1998;67:1059.
[29] Zayed A. A convolution and product theorem for the fractional Fourier transform. IEEE Signal Process Lett 1998;5(4):1013.
[30] Goldberg D. Genetic algorithms in optimization search and machine learning. Massasuchettes: Addison-Wesley; 1989.
[31] Tang KS, Man KF, Kwong S, He QH. Genetic algorithms and their applications. IEEE Signal Process Mag 1996;13(6):2237.
[32] Goldberg D, Kendall G. Genetic algorithms. In: Burke EK, Kendall G, editors. Search methodologies: introductory tutorials in optimization and decision
support techniques. New York: Springer; 2005. p. 97126.
[33] Jun-Su J, Jong-Hwan K. Fast and robust face detection using evolutionary pursuit. IEEE Trans Evol Comput 2008;12(5):54261.
[34] Haupt RL, Werner DH. Genetic algorithms in electromagnetics. Pennsilvania: IEEE Press; 2007.
[35] Dolled-Filhart M, Ryden L, Cregger M, Jirstrom K, Harigopal M, Camp RL, et al. Classication of breast cancer using genetic algorithms and tissue
microarrays. Clin Cancer Res 2006;12:645968.
[36] Pond SLK, Posada D, Gravenor MB, Woelk CH, Frost SDW. Automated phylogenetic detection of recombination using a genetic algorithm. Mol Biol Evol
2006;23(10):1891901.
[37] Thierens D, Goldberg DE. Elitist recombination: an integrated selection recombination GA. Proc IEEE World Congress Comput Intell 1994;1:50812.
[38] Thierens D. Selection schemes, elitist recombination, and selection intensity. In: Proceedings of 7th international conference on genetic algorithms, vol.
7; 1997. p. 152159.
[39] Baker JE. Reducing bias and inefciency in the selection algorithm. In: Proceedings of international conference on genetic algorithms, vol. 2; 1987. p.
1421.
[40] Blickle T, Thiele L. A comparison of selection schemes used in genetic algorithms, TIK-Report, Computer Engineering and Communication Networks
Lab., Swiss Federal Institute of Technology, Zurich, vol. 11; 1995.
[41] Bose N, Prabhu K. Two-dimensional discrete Hilbert transform and computational complexity aspects in its implementation. IEEE Trans Acoust Speech
Signal Process 1979;27(4):35660.