Optical correlators with A-kBth-law

nonlinearity: optimal and suboptimal solutions

Leonid P. Yaroslavsky

A computer simulation of nonlinear correlators with 1-k2th-law nonlinearity has been implemented. The
nonlinearity is applied to the input-image power spectrum, with either a matched filter or a phase-only
filter representing the reference object. Optimal correlators 1with an exactly known power spectrum of
the background-image component2 and suboptimal correlators have been studied in order to establish
potential limits and achievable figures for the correlators discrimination capability in a target location in
a cluttered background. For the suboptimal correlators, different values of the nonlinearity index k
have been investigated for two methods of the images power-spectrum estimation and for different
degrees on the limitation of the nonlinearitys dynamic range. The results show that the nonlinear
correlators promise significant improvement in the correlators discrimination capability and provide
important information for evaluating the practical aspects of the correlators design.
Key words: Optical correlator, matched filter, phase-only filter, phase-only correlator, discrimination
capability, target location.

1. Introduction A nonlinear correlator with k 5 1 for the input-

Optical correlators with a nonlinearity applied to the
input-image spectrum or to the target-object spec-
trum have attracted considerable attention of re-
searchers looking for ways to improve the perfor-
mance of optical correlators in pattern recognition
and target location.14 An important subclass of
such correlators is optical correlators with 1-k2th-law
nonlinearity, which is described by the relationship
Out 5 1 0 In 0 222k,
where In and Out are input and output signals,
respectively, of the nonlinear transformation. A
schematic diagram of these correlators is shown in
Fig. 1.
When this research was performed, the author was with the
Biomedical Engineering and Instrumentation Program, National
Center for Research Resources, National Institutes of Health,
Building 13@3N17, 13 South Drive MSC 5766, Bethesda, Maryland
50892-5766. He is now with the Department of Interdisciplinary
Studies, Faculty of Engineering, Tel Aviv University, Tel Aviv
69978, Israel, and is on leave from the Institute of Information
Transmission Problems, Russian Academy of Sciences, Yermolovoy
Ulitsa 19, Moscow 101447, Russia.
Received 11 November 1994; revised manuscript received 18
January 1995.
0003-6935@95@203924-09$06.00@0.
r 1995 Optical Society of America.
image spectrum and with no nonlinearity applied to
the target-object spectrum was described in Ref. 5.
It was shown that such a correlator approximates the
correlator that provides the maximal ratio of the
signal value in the point of the output plane, where
the target object is located, to the standard deviation
of the signal in the background part of the output
plane. This correlator is described by its frequency
response as
112
Hopt 5 O*@ 0 B 0 2, 122
where O* is the complex conjugate of the reference-
object Fourier spectrum and 0 B 0 2 is the power spec-
trum of the background component of the input
image. We refer to it as the optimal correlator 1OC2.
Implementation of the OC promises the highest pos-
sible discrimination capability with relation to the
background objects in the image. However, the de-
sign of the OC requires exact knowledge of the power
spectrum of the background objects. This spectrum
is not known before the target is located. One can
attempt to estimate it from the observed image if the
size of the target is reasonably small with regard to
the entire image size. Computer and optical experi-
ments have shown that even the simplest direct use of
the observed image power spectrum as an estimation
of the background spectrum provides considerable
improvement in discrimination capability.68 How-

3924 APPLIED OPTICS @ Vol. 34, No. 20 @ 10 July 1995

 Table 1. Variety of Correlators Implemented in the Computer Model

Correlator Type e k n bl Lim

Conventional cor- 0 0 0 0 `
relator 1matched
filter2
Correlator with the 0 0 1@2 0 `
phase-only filter
Phase-only corre- 0 1@2 1@2 $0 `
lator
Suboptimal corr- 0 and 1 .0 0 .0 ,`
elators

the unknown position of the target object and over the

Fig. 1. Schematic diagram of the optical correlators with 1-k2th-
law nonlinearity.
ever, the design and applications of the OC require
answers to a number of practical questions, such as
c A number of methods for power-spectrum estima-
tion can be used. One of the simplest of these
methods in optical implementation is spectrum linear
smoothing 1LS2. How much can one expect to gain in
the correlators discrimination capability by using
this method and other similar methods, and how well
do such methods approximate the optimal correlator?
c How sensitive is the correlators discrimination
capability to such design factors as the accuracy of
realization of the nonlinearity and the limitation of its
dynamic range?
c How can one compare the discrimination capa-
bility of the OC and suboptimal correlators with
1-k2th-law nonlinearity with that of the phase-only-
filter correlator and the phase-only correlator, which
are known to have the highest light efficiency?911
signal sensors noise realizations, then the image
power spectrum averaged over these random factors
has to be used as the estimation. If the target
position is assumed to have uniform probability distri-
bution over the entire image area, modification of the
input image power spectrum 0 I 0 2 by adding to it the
target-object power spectrum 0 O 0 2 can be recom-
mended as an estimate of the background spectrum
0 B 0 2 averaged over possible target positions.5 This
method is called the ATS method. The correspond-
ing suboptimal filter is, in this case,
112
Hsubopt 5 O*@1 0 I 0 2 1 0 O 0 22. 132
Spectrum averaging over the sensors noise can be
substituted by a sort of smoothing of the input-image
power spectrum. The simplest implementation is
spectrum linear smoothing by a convolution. This
method is called the LS method. The corresponding
suboptimal filter is
122
Hsubopt 5 O*@CONV5h, 0 I 0 26, 142
where CONV1h, ?2 is a convolution operator with the
point-spread function h. In the particular case of

It is the purpose of this paper to provide experimen-

tal data obtained by computer simulation to answer
these questions. The results are obtained for real-
life images and seem to be representative enough to
use as guidelines in evaluating the feasibility of
optical implementation of nonlinear correlators with
improved discrimination capability. The results
complement the comparison data and the data on the
sensitivity of the OC to quantization of the input-
image power spectrum recently published in Ref. 12.
2. Description of the Simulation Experiments
The implementation of the OC requires obtaining an
estimation of the image-background component power
spectra. If the correlator is designed to have the Fig. 2. Image and its fragments used in the experiments as the
highest discrimination capability on the average over test images.

10 July 1995 @ Vol. 34, No. 20 @ APPLIED OPTICS 3925

optical correlators with a nonlinear light-sensitive
medium placed in the correlators Fourier plane, this
is the point-spread function of the mediums transpar-
ency function.
An important issue in the design of nonlinear
optical correlators is the nonlinear mediums dynamic
range. All optical light-sensitive materials have a
limited dynamic range. This limitation also has to
be taken into account in designing the computer
model.
Therefore the computer model has been designed to
implement the following signal-processing scheme:
c 5 IFTEI ? LDRCACONV3h1bl2, 1 0 I 02 1 e 0 O 0224B2kD
? 1 0 O 0222n ? O*F,
where c is the correlator output signal, I is the
input-image Fourier spectrum, O* is the complex
conjugate of the reference-object Fourier spectrum,
IFT is the inverse Fourier transform operator, 12 is
the elementwise product of vectors, CONV3h1bl2, 1.24 is
a bl-fold linear convolution operator with point-
3926 APPLIED OPTICS @ Vol. 34, No. 20 @ 10 July 1995
spread function h, and LDR is a pointwise nonlinear
operator, where
x, x , Lim
LDR1x2 5 5x 5 Lim, otherwise
. 162
The model parameters bl, e, k, n, and Lim define
different processing modes. The parameter bl de-
fines the degree of spectrum LS, or blur; when bl 5 0,
no signal smoothing is performed. The parameter e
takes one of two values, 0 or 1, and determines
whether the target objects power spectrum is used to
modify the observed image power spectrum 3Eq. 1324.
The parameter k defines nonlinear transformation
152 applied to the estimation of the observed input im-
ages power spectrum. The parameter n defines non-
linear transformation applied to the target objects
power spectrum. It was used to simulate the phase-
only filter and the phase-only correlator 1n 5 1@22.
The parameter Lim defines the degree of the dynamic
range limitation.
The model covers a broad variety of nonlinear
Fig. 3. Discrimination capability of nonlinear correlators with
kth-law nonlinearity and an exactly known image power spectrum
as a function of the index k 1k 5 1 corresponds to the OC2: 1a2
sorted sequence of SNRVs at output of the conventional correlator
1k 5 02 for the set of test images, 1b2 gain in SNRV values compared
with the conventional correlator as a function of k for the set of test
images, 1c2 gain in SNRM values compared with the conventional
correlator as a function of k for the set of test images.
correlators. Some specific cases are shown in Table
1. The term suboptimal correlators in this table
means the family of correlators that uses an estima-
tion of the image power spectrum rather than the
exact power spectrum, which has to be used in the
optimal correlator design.
In the experiments, a range of values for the
nonlinearity index k has been tested. For the expo-
nent n, only those cases in which n 5 0 1matched filter2
and n 5 1@2 1phase-only filter2 were tested. As
mentioned above, the convolution operator was imple-
mented as a bl-fold repetition 1bl 5 0, 1, . . . , 72 of the
elementary blur operator with a circular symmetric
frequency response:
H1 f 2 5 exp12f 2@0.12,
where f [ 30, 14 is spatial frequency normalized to the
correlators bandwidth.
In the experiments with limitation of the nonlinear
material dynamic range, the parameter Lim was
selected as a 11@4lim2th fraction of the signal maxi-
mum with lim 5 0, 1, . . . , 7. Thus the case in which
lim 5 0 corresponded to no dynamic range limitation,
and the case in which lim 5 7 corresponded to a
limitation at the level of 11@163842th of the signal
maximal value.
The experiments were conducted with 16 128 3 128
pixel fragments of a 512 3 512 pixel satellite photo-
graph of a rural area as test images 1Fig. 22 and a
small circular spot with a diameter of ,5 pixels as the
target object embedded within the test images. The
optical Fourier transform was approximated by the
172 two-dimensional discrete Fourier transform. Two
performance measures of the correlators discrimina-
tion capability were computed in each experiment:
the ratio of the object signal maximum to the

Fig. 4. Discrimination capability of the suboptimal correlators with 1-k2th-law nonlinearity as a function of the index k and the degree of
spectrum smoothing for a typical image from the test set: 1a2 SNRV for the LS method of spectrum estimation, 1b2 SNRV for the ATS
method of spectrum estimation, 1c2 SNRM for the LS method of spectrum estimation, 1d2 SNRM for the ATS method of spectrum
estimation.

10 July 1995 @ Vol. 34, No. 20 @ APPLIED OPTICS 3927

Fig. 5. Gain in the nonlinear correlators discrimination capability for the LS method for estimation of the background-image power
spectrum compared with that for the direct use of the observed image power spectrum: 1a2 gain in SNRV in relation to the conventional
correlator-matched filter, 1b2 same as 1a2 but for SNRM.

signal standard deviation over the background area
in the correlation plane, or the signal-to-noise ratio in
terms of the background-signal variance 1SNRV2, and
the ratio of the highest object signal maximum to the
maximum over the background area in the correlation
plane, or the signal-to-noise ratio in terms of the
background-signal maximum 1SNRM2. The former
connects the experimental results with the analysis
presented in Ref. 5, whereas the latter is a more
straightforward indicator of the correlators ability to
locate a target on a cluttered background. In order
to compare the discrimination capability of subopti-
mal correlators with the potential discrimination
capability offered by the OC, the latter has also been
implemented in the model. In this case, to allow
exact estimation of the image power spectra, the
target object was not embedded in the test images.
3. Experimental Results and Discussion
A. Optimal Correlator
The results for OCs designed individually for each
image from the set of test images are represented in
Figs. 31a231c2 along with those for the conventional
correlator 1matched filter2. Figure 31a2 shows that
the SNRV values at the output of the conventional
correlator are almost uniformly distributed in the
broad range from ,0.3 to 3 for the set of test images.
This distribution allows the suggestion that the se-
lected set of images is representative enough. Val-

Fig. 6. Comparison of the optimal correlator 1curve 02 with the correlators with the LS 1curve 12 and ATS 1curve 22 methods for spectrum
estimation with the optimized smoothing index bl: 1a2 gain in SNRV in relation to the conventional correlator-matched filter, 1b2 same as
1a2 but for SNRM.

3928 APPLIED OPTICS @ Vol. 34, No. 20 @ 10 July 1995

ues for the SNRV and the SNRM as functions of the
nonlinearity index k obtained for all the test images
are plotted in Figs. 31b2 and 31c2, respectively. The
plots clearly indicate that

112 The nonlinear correlators with the matched

filter and with 1-k2th nonlinearity in the Fourier plane
can potentially outperform the conventional correla-
tor considerably in terms of the ability to locate a
target on a cluttered background. The observed gain
is in the range from ,20 to more than 200 for SNRV
and from ,15 to more than 100 for SNRM.
122 Although the optimal value of k in terms of the
SNRV values is equal to 1, the optimal value of k in
terms of the SNRM always tends to be slightly higher;
on the average, it is approximately 10% higher. It is
remarkable that SNRM decays relatively slowly for
values of k higher than the optimal value. Observa- Fig. 7. One-dimensional cross sections of outputs of the conven-
tions show that for k . 1 the correlators discrimina- tional matched filter 1k 5 0, bl 5 02 and the nonlinear correlator
tion is basically due to a small number of frequency with LS estimation of the image power spectrum 1k 5 1, bl 5 42 for
components of the signals, that is, the optimal filter one of the test images with the target object at coordinate 65.
acts as a bandpass filter with a very narrow band-
width.
c The LS method outperforms the ATS method of
spectrum estimation slightly if the degree of image
B. Suboptimal Correlators with 1-k2th-Law Nonlinearity and spectrum smoothing is properly selected; the ATS
Empirical Estimation of the Image Power Spectrum method is much less sensitive to the degree of image
As mentioned above two estimation methods of esti- spectrum smoothing and provides considerable im-
mating the background image spectrum have been provement even without spectrum smoothing.
studied for correlators with 1-k2th-law nonlinearity: c The optimal value of the nonlinearity index k
the LS method 3e 5 0 in Eq. 1524, and the ATS method tends to be slightly lower than 1 for the correlators
3e 5 1 in Eq. 1524. Experimental SNRV and SNRM performance evaluation in terms of SNRV, whereas it
values for these methods are plotted in Figs. 41a241c2 is close to 1 for the performance evaluation in terms of
for a typical image from the test set. Averaged data SNRM; small deviations of k from its optimal value
for the LS method and two values of the smoothing are not very critical.
parameter bl are plotted in Figs. 51a2 and 51b2 to c A considerable degree of spectrum smoothing is
illustrate the gain in the correlators discrimination required for achieving better discrimination capabil-
capability obtained by spectrum smoothing 1bl 5 4 ity; the optimal degree of smoothing corresponds to
the resolution power of the nonlinear material, from
corresponds to the highest average gain over the set of
15 to 25 times lower than that required for the image.
test images2 compared with the direct use of the
observed image spectrum as an estimation of the
background-image power spectrum. Averaged data
for the entire set of images and for the optimal value
of the smoothing parameter bl 5 4 are plotted in Figs.
61a2 and 61b2. These results demonstrate the follow-
ing conclusions:

c Suboptimal nonlinear correlators provide signifi-

cant 1on the average, ,20-fold for SRNV and eight-
fold for SNRM2 improvement of the discrimination
capability compared with that of the conventional
matched filter 1a vivid comparison is presented in
Fig. 72.
c Smoothing the observed image power spectrum
as a method of power-spectrum estimation improves
the nonlinear correlators discrimination capability
considerably; the optimal value of the nonlinearity
index k tends to be from 0.7 to 0.8 when no smoothing
is applied to the observed image power spectrum,
whereas spectrum smoothing requires higher values Fig. 8. Losses in the discrimination capability of the OC as a
for k. function of the dynamic range limitation degree.

10 July 1995 @ Vol. 34, No. 20 @ APPLIED OPTICS 3929

Fig. 9. Gain in the nonlinear correlators discrimination capability for the LS method of spectrum estimation and limitation of the
nonlinear medias dynamic range: 1a2 gain in SNRV with relation to the conventional correlator-matched filter; 1b2 same as 1a2 but for
SNRM.

c A twofoldfivefold gap still remains between the
discrimination capability of the suboptimal correla-
tors and that potentially achievable for an exactly
known power spectrum of the background-image com-
ponent.
Limitation of the nonlinearitys dynamic range is
an important issue in designing nonlinear correlators.
For the OC, one would expect that the limitation of
the nonlinearitys dynamic range deteriorates the
correlators discrimination capability. The simula-
tion has shown that, although this deterioration does
occur, it is not too severe 1Fig. 82. For the subopti-
mal correlator with the LS method of spectrum
estimation, the experiments have shown that the
dynamic range limitation does not necessarily deterio-
rate the correlators discrimination capability. More-
over, a sort of trade-off is possible between the degree
of the dynamic range limitation and the nonlinearity
index k: the higher the limitation, the higher the
nonlinearity index required for better discrimination
capability. This is illustrated in Figs. 91a2 and 91b2,
which show that appropriate selection of the nonlin-
earity index k provides practically the same discrimi-
nation capability as that of the nonlinear correlator
with an unlimited dynamic range. It is remarkable
that the dynamic range limitation at the level of 1@4
to 1@16 of the smoothed-spectrum dynamic range

Fig. 10. Discrimination capability of the correlators with the phase-only filter 1pof 2 and 1-k2th-law nonlinearity as a function of the index k
and the degree of spectrum smoothing for a typical image from the test set: 1a2 SNRV for the LS method of spectrum estimation, 1b2 SNRM
for the LS method of spectrum estimation.

3930 APPLIED OPTICS @ Vol. 34, No. 20 @ 10 July 1995

may even improve, although slightly, the discrimina-
tion capability of the correlator without the dynamic
range limitation. This can be attributed to the fact
that power-spectrum estimation by the spectrum LS
method is not perfect. Correlators with a strongly
restricted dynamic range have higher light efficiency
than do correlators with no restrictions, because,
according to Eq. 162, the nonlinear material can be
made entirely transparent in the spectral plane wher-
ever the signal that modulates it exceeds the limita-
tion threshold. Therefore it might be advisable to
use the observed trade-off between the nonlinearity
index and the degree of dynamic range limitation to
improve the correlators light efficiency while preserv-
ing its discrimination capability.
C. Phase-Only Filter and Phase-Only Correlators
The use of phase-only filters instead of matched filters
in optical correlators is advocated by the high light
efficiency of phase-only filters. However, experi-
ments show that the discrimination capability of the
correlators with phase-only filters is significantly
lower than that of the optimal and suboptimal nonlin-
ear correlators. The corresponding graphs for SNRV
and SNRM for the nonlinear correlators with the
phase-only filter representing the reference object are
given in Figs. 101a2 and 101b2 for the same test image
as in Fig. 4. One can see from Fig. 10 that the
optimal nonlinearity index k in this case is approxi-
mately k 5 0.5, which corresponds to the phase-only
correlator. The graphs also show that spectrum
smoothing performed before the spectrums nonlin-
ear transformation improves the discrimination capa-
bility of the phase-only correlators, although their
discrimination capability remains lower than that of
the suboptimal nonlinear correlators with a matched
Fig. 11. Comparison of the discrimination capability of the OC
1exact opt. corr.2, suboptimal nonlinear optical correlators with the
LS method of power spectrum estimation 1nlin. opt. corr.2, phase-
only correlator 1POCorr2, phase-only filter 1POF corr2, and matched
filter 1MF corr2 for the set of test images in terms of SNRV.
filter representing the target object. A comprehen-
sive comparison of all the correlators described is
presented in Fig. 11. These plots demonstrate a
hierarchy of correlators with 1-k2th-law nonlinearity in
terms of their discrimination capability.
4. Conclusion
The discrimination capability of nonlinear correlators
with the 1-k2th-law nonlinearity applied to the input-
image power spectrum and with either a matched
filter or a phase-only filter representing the reference
object has been studied experimentally by the use of
computer simulation. The results show that
c One can significantly improve the correlators
discrimination capability by the appropriate selection
of the nonlinearity index k and the method of power-
spectrum estimation.
c Nonlinear correlators with improved discrimina-
tion capability can be implemented by the use of
low-resolution nonlinear materials with a reasonably
limited dynamic range.
c A trade-off exists between the nonlinearity in-
dex k and the degree of dynamic range limitation; a
higher degree of limitation requires a higher value
of k.
c Combination of input-image spectrum smooth-
ing, high nonlinearity index k, and a high degree of
the dynamic range limitation promises considerable
improvement in the nonlinear correlators light effi-
ciency while preserving their high discrimination
capability and deserves further research.
c The twofoldfivefold gap in the discrimination
capability between OCs and suboptimal correlators
motivates further research into improved methods of
image power-spectrum estimation.
References
1. T. Szoplik and K. Chalasinska-Macukow, Towards nonlinear
optical processing, in International Trends in Optics, J. W.
Goodman, ed. 1Academic, Boston, San Diego, New York, 19912,
pp. 451464.
2. B. V. K. V. Kumar and L. Hasselbrook, Performance measures
for correlation filters, Appl. Opt. 29, 29973006 119902.
3. B. Javidi, Generalization of the linear matched filter concept
to nonlinear matched filters, Appl. Opt. 29, 12151224 119902.
4. O. K. Ersoy and M. Zeng, Nonlinear matched filtering, J.
Opt. Soc. Am. A 6, 636648 119892.
5. L. P. Yaroslavsky, The theory of optimal methods for localiza-
tion of objects in pictures, in Progress in Optics, E. Wolf, ed.
1Elsevier, Amsterdam, 19932, Vol. XXXII, pp. 147201.
6. V. N. Dudinov, V. A. Krishtal and L. P. Yaroslavsky, Localiza-
tion of objects in images by the coherent optics techniques,
Geod. Cartogr. 1, 4246 119772.
7. L. P. Yaroslavsky, Digital Picture Processing. An Introduc-
tion. 1Springer-Verlag, Berlin, Heidelberg, 19852, Chap. 9.
8. J. Campos, F. Turon, L. P. Yaroslavsky, and M. J. Yzuel, Some
filters for reliable recognition and localization of objects by
optical correlators: a comparison, Int. J. Opt. Comput. 2,
341366 119932.
10 July 1995 @ Vol. 34, No. 20 @ APPLIED OPTICS 3931
 9. H. Bartelt, Unconventional correlators, in Optical Signal
Processing, J. L. Horner, ed. 1Academic, San Diego, 19872, pp.
97128.
10. D. L. Flannery and J. L. Horner, Fourier optical signal
processors, Proc. IEEE, 77, 15111527 119892.
11. K. Chalasinska-Macukow, Generalized matched spatial fil-
3932 APPLIED OPTICS @ Vol. 34, No. 20 @ 10 July 1995
ters with optimum light efficiency, in Optical Processing and
Computing, H. H. Arsenault, T. Szoplik, and B. Macukow, eds.
1Academic, Boston, 19892, pp. 3145.
12. V. Kober, L. Yaroslavsky, J. Campos, and M. Yzuel, Optimal
filter approximation by means of phase-only filter with quanti-
zation, Opt. Lett. 19, 978980 119942.